ARMISelLowering.cpp revision 0412957764de44f63657360a22ce5445a8c87628
1//===-- ARMISelLowering.cpp - ARM DAG Lowering Implementation -------------===// 2// 3// The LLVM Compiler Infrastructure 4// 5// This file is distributed under the University of Illinois Open Source 6// License. See LICENSE.TXT for details. 7// 8//===----------------------------------------------------------------------===// 9// 10// This file defines the interfaces that ARM uses to lower LLVM code into a 11// selection DAG. 12// 13//===----------------------------------------------------------------------===// 14 15#include "ARM.h" 16#include "ARMAddressingModes.h" 17#include "ARMConstantPoolValue.h" 18#include "ARMISelLowering.h" 19#include "ARMMachineFunctionInfo.h" 20#include "ARMRegisterInfo.h" 21#include "ARMSubtarget.h" 22#include "ARMTargetMachine.h" 23#include "llvm/CallingConv.h" 24#include "llvm/Constants.h" 25#include "llvm/Function.h" 26#include "llvm/Instruction.h" 27#include "llvm/Intrinsics.h" 28#include "llvm/GlobalValue.h" 29#include "llvm/CodeGen/CallingConvLower.h" 30#include "llvm/CodeGen/MachineBasicBlock.h" 31#include "llvm/CodeGen/MachineFrameInfo.h" 32#include "llvm/CodeGen/MachineFunction.h" 33#include "llvm/CodeGen/MachineInstrBuilder.h" 34#include "llvm/CodeGen/MachineRegisterInfo.h" 35#include "llvm/CodeGen/PseudoSourceValue.h" 36#include "llvm/CodeGen/SelectionDAG.h" 37#include "llvm/Target/TargetOptions.h" 38#include "llvm/ADT/VectorExtras.h" 39#include "llvm/Support/MathExtras.h" 40using namespace llvm; 41 42static bool CC_ARM_APCS_Custom_f64(unsigned &ValNo, MVT &ValVT, MVT &LocVT, 43 CCValAssign::LocInfo &LocInfo, 44 ISD::ArgFlagsTy &ArgFlags, 45 CCState &State); 46static bool CC_ARM_AAPCS_Custom_f64(unsigned &ValNo, MVT &ValVT, MVT &LocVT, 47 CCValAssign::LocInfo &LocInfo, 48 ISD::ArgFlagsTy &ArgFlags, 49 CCState &State); 50static bool RetCC_ARM_APCS_Custom_f64(unsigned &ValNo, MVT &ValVT, MVT &LocVT, 51 CCValAssign::LocInfo &LocInfo, 52 ISD::ArgFlagsTy &ArgFlags, 53 CCState &State); 54static bool RetCC_ARM_AAPCS_Custom_f64(unsigned &ValNo, MVT &ValVT, MVT &LocVT, 55 CCValAssign::LocInfo &LocInfo, 56 ISD::ArgFlagsTy &ArgFlags, 57 CCState &State); 58 59void ARMTargetLowering::addTypeForNEON(MVT VT, MVT PromotedLdStVT, 60 MVT PromotedBitwiseVT) { 61 if (VT != PromotedLdStVT) { 62 setOperationAction(ISD::LOAD, VT, Promote); 63 AddPromotedToType (ISD::LOAD, VT, PromotedLdStVT); 64 65 setOperationAction(ISD::STORE, VT, Promote); 66 AddPromotedToType (ISD::STORE, VT, PromotedLdStVT); 67 } 68 69 MVT ElemTy = VT.getVectorElementType(); 70 if (ElemTy != MVT::i64 && ElemTy != MVT::f64) 71 setOperationAction(ISD::VSETCC, VT, Custom); 72 if (ElemTy == MVT::i8 || ElemTy == MVT::i16) 73 setOperationAction(ISD::EXTRACT_VECTOR_ELT, VT, Custom); 74 setOperationAction(ISD::BUILD_VECTOR, VT, Custom); 75 setOperationAction(ISD::VECTOR_SHUFFLE, VT, Custom); 76 setOperationAction(ISD::SCALAR_TO_VECTOR, VT, Custom); 77 setOperationAction(ISD::CONCAT_VECTORS, VT, Custom); 78 if (VT.isInteger()) { 79 setOperationAction(ISD::SHL, VT, Custom); 80 setOperationAction(ISD::SRA, VT, Custom); 81 setOperationAction(ISD::SRL, VT, Custom); 82 } 83 84 // Promote all bit-wise operations. 85 if (VT.isInteger() && VT != PromotedBitwiseVT) { 86 setOperationAction(ISD::AND, VT, Promote); 87 AddPromotedToType (ISD::AND, VT, PromotedBitwiseVT); 88 setOperationAction(ISD::OR, VT, Promote); 89 AddPromotedToType (ISD::OR, VT, PromotedBitwiseVT); 90 setOperationAction(ISD::XOR, VT, Promote); 91 AddPromotedToType (ISD::XOR, VT, PromotedBitwiseVT); 92 } 93} 94 95void ARMTargetLowering::addDRTypeForNEON(MVT VT) { 96 addRegisterClass(VT, ARM::DPRRegisterClass); 97 addTypeForNEON(VT, MVT::f64, MVT::v2i32); 98} 99 100void ARMTargetLowering::addQRTypeForNEON(MVT VT) { 101 addRegisterClass(VT, ARM::QPRRegisterClass); 102 addTypeForNEON(VT, MVT::v2f64, MVT::v4i32); 103} 104 105ARMTargetLowering::ARMTargetLowering(TargetMachine &TM) 106 : TargetLowering(TM), ARMPCLabelIndex(0) { 107 Subtarget = &TM.getSubtarget<ARMSubtarget>(); 108 109 if (Subtarget->isTargetDarwin()) { 110 // Uses VFP for Thumb libfuncs if available. 111 if (Subtarget->isThumb() && Subtarget->hasVFP2()) { 112 // Single-precision floating-point arithmetic. 113 setLibcallName(RTLIB::ADD_F32, "__addsf3vfp"); 114 setLibcallName(RTLIB::SUB_F32, "__subsf3vfp"); 115 setLibcallName(RTLIB::MUL_F32, "__mulsf3vfp"); 116 setLibcallName(RTLIB::DIV_F32, "__divsf3vfp"); 117 118 // Double-precision floating-point arithmetic. 119 setLibcallName(RTLIB::ADD_F64, "__adddf3vfp"); 120 setLibcallName(RTLIB::SUB_F64, "__subdf3vfp"); 121 setLibcallName(RTLIB::MUL_F64, "__muldf3vfp"); 122 setLibcallName(RTLIB::DIV_F64, "__divdf3vfp"); 123 124 // Single-precision comparisons. 125 setLibcallName(RTLIB::OEQ_F32, "__eqsf2vfp"); 126 setLibcallName(RTLIB::UNE_F32, "__nesf2vfp"); 127 setLibcallName(RTLIB::OLT_F32, "__ltsf2vfp"); 128 setLibcallName(RTLIB::OLE_F32, "__lesf2vfp"); 129 setLibcallName(RTLIB::OGE_F32, "__gesf2vfp"); 130 setLibcallName(RTLIB::OGT_F32, "__gtsf2vfp"); 131 setLibcallName(RTLIB::UO_F32, "__unordsf2vfp"); 132 setLibcallName(RTLIB::O_F32, "__unordsf2vfp"); 133 134 setCmpLibcallCC(RTLIB::OEQ_F32, ISD::SETNE); 135 setCmpLibcallCC(RTLIB::UNE_F32, ISD::SETNE); 136 setCmpLibcallCC(RTLIB::OLT_F32, ISD::SETNE); 137 setCmpLibcallCC(RTLIB::OLE_F32, ISD::SETNE); 138 setCmpLibcallCC(RTLIB::OGE_F32, ISD::SETNE); 139 setCmpLibcallCC(RTLIB::OGT_F32, ISD::SETNE); 140 setCmpLibcallCC(RTLIB::UO_F32, ISD::SETNE); 141 setCmpLibcallCC(RTLIB::O_F32, ISD::SETEQ); 142 143 // Double-precision comparisons. 144 setLibcallName(RTLIB::OEQ_F64, "__eqdf2vfp"); 145 setLibcallName(RTLIB::UNE_F64, "__nedf2vfp"); 146 setLibcallName(RTLIB::OLT_F64, "__ltdf2vfp"); 147 setLibcallName(RTLIB::OLE_F64, "__ledf2vfp"); 148 setLibcallName(RTLIB::OGE_F64, "__gedf2vfp"); 149 setLibcallName(RTLIB::OGT_F64, "__gtdf2vfp"); 150 setLibcallName(RTLIB::UO_F64, "__unorddf2vfp"); 151 setLibcallName(RTLIB::O_F64, "__unorddf2vfp"); 152 153 setCmpLibcallCC(RTLIB::OEQ_F64, ISD::SETNE); 154 setCmpLibcallCC(RTLIB::UNE_F64, ISD::SETNE); 155 setCmpLibcallCC(RTLIB::OLT_F64, ISD::SETNE); 156 setCmpLibcallCC(RTLIB::OLE_F64, ISD::SETNE); 157 setCmpLibcallCC(RTLIB::OGE_F64, ISD::SETNE); 158 setCmpLibcallCC(RTLIB::OGT_F64, ISD::SETNE); 159 setCmpLibcallCC(RTLIB::UO_F64, ISD::SETNE); 160 setCmpLibcallCC(RTLIB::O_F64, ISD::SETEQ); 161 162 // Floating-point to integer conversions. 163 // i64 conversions are done via library routines even when generating VFP 164 // instructions, so use the same ones. 165 setLibcallName(RTLIB::FPTOSINT_F64_I32, "__fixdfsivfp"); 166 setLibcallName(RTLIB::FPTOUINT_F64_I32, "__fixunsdfsivfp"); 167 setLibcallName(RTLIB::FPTOSINT_F32_I32, "__fixsfsivfp"); 168 setLibcallName(RTLIB::FPTOUINT_F32_I32, "__fixunssfsivfp"); 169 170 // Conversions between floating types. 171 setLibcallName(RTLIB::FPROUND_F64_F32, "__truncdfsf2vfp"); 172 setLibcallName(RTLIB::FPEXT_F32_F64, "__extendsfdf2vfp"); 173 174 // Integer to floating-point conversions. 175 // i64 conversions are done via library routines even when generating VFP 176 // instructions, so use the same ones. 177 // FIXME: There appears to be some naming inconsistency in ARM libgcc: 178 // e.g., __floatunsidf vs. __floatunssidfvfp. 179 setLibcallName(RTLIB::SINTTOFP_I32_F64, "__floatsidfvfp"); 180 setLibcallName(RTLIB::UINTTOFP_I32_F64, "__floatunssidfvfp"); 181 setLibcallName(RTLIB::SINTTOFP_I32_F32, "__floatsisfvfp"); 182 setLibcallName(RTLIB::UINTTOFP_I32_F32, "__floatunssisfvfp"); 183 } 184 } 185 186 // These libcalls are not available in 32-bit. 187 setLibcallName(RTLIB::SHL_I128, 0); 188 setLibcallName(RTLIB::SRL_I128, 0); 189 setLibcallName(RTLIB::SRA_I128, 0); 190 191 if (Subtarget->isThumb()) 192 addRegisterClass(MVT::i32, ARM::tGPRRegisterClass); 193 else 194 addRegisterClass(MVT::i32, ARM::GPRRegisterClass); 195 if (!UseSoftFloat && Subtarget->hasVFP2() && !Subtarget->isThumb()) { 196 addRegisterClass(MVT::f32, ARM::SPRRegisterClass); 197 addRegisterClass(MVT::f64, ARM::DPRRegisterClass); 198 199 setTruncStoreAction(MVT::f64, MVT::f32, Expand); 200 } 201 202 if (Subtarget->hasNEON()) { 203 addDRTypeForNEON(MVT::v2f32); 204 addDRTypeForNEON(MVT::v8i8); 205 addDRTypeForNEON(MVT::v4i16); 206 addDRTypeForNEON(MVT::v2i32); 207 addDRTypeForNEON(MVT::v1i64); 208 209 addQRTypeForNEON(MVT::v4f32); 210 addQRTypeForNEON(MVT::v2f64); 211 addQRTypeForNEON(MVT::v16i8); 212 addQRTypeForNEON(MVT::v8i16); 213 addQRTypeForNEON(MVT::v4i32); 214 addQRTypeForNEON(MVT::v2i64); 215 216 setTargetDAGCombine(ISD::INTRINSIC_WO_CHAIN); 217 setTargetDAGCombine(ISD::SHL); 218 setTargetDAGCombine(ISD::SRL); 219 setTargetDAGCombine(ISD::SRA); 220 setTargetDAGCombine(ISD::SIGN_EXTEND); 221 setTargetDAGCombine(ISD::ZERO_EXTEND); 222 setTargetDAGCombine(ISD::ANY_EXTEND); 223 } 224 225 computeRegisterProperties(); 226 227 // ARM does not have f32 extending load. 228 setLoadExtAction(ISD::EXTLOAD, MVT::f32, Expand); 229 230 // ARM does not have i1 sign extending load. 231 setLoadExtAction(ISD::SEXTLOAD, MVT::i1, Promote); 232 233 // ARM supports all 4 flavors of integer indexed load / store. 234 for (unsigned im = (unsigned)ISD::PRE_INC; 235 im != (unsigned)ISD::LAST_INDEXED_MODE; ++im) { 236 setIndexedLoadAction(im, MVT::i1, Legal); 237 setIndexedLoadAction(im, MVT::i8, Legal); 238 setIndexedLoadAction(im, MVT::i16, Legal); 239 setIndexedLoadAction(im, MVT::i32, Legal); 240 setIndexedStoreAction(im, MVT::i1, Legal); 241 setIndexedStoreAction(im, MVT::i8, Legal); 242 setIndexedStoreAction(im, MVT::i16, Legal); 243 setIndexedStoreAction(im, MVT::i32, Legal); 244 } 245 246 // i64 operation support. 247 if (Subtarget->isThumb()) { 248 setOperationAction(ISD::MUL, MVT::i64, Expand); 249 setOperationAction(ISD::MULHU, MVT::i32, Expand); 250 setOperationAction(ISD::MULHS, MVT::i32, Expand); 251 setOperationAction(ISD::UMUL_LOHI, MVT::i32, Expand); 252 setOperationAction(ISD::SMUL_LOHI, MVT::i32, Expand); 253 } else { 254 setOperationAction(ISD::MUL, MVT::i64, Expand); 255 setOperationAction(ISD::MULHU, MVT::i32, Expand); 256 if (!Subtarget->hasV6Ops()) 257 setOperationAction(ISD::MULHS, MVT::i32, Expand); 258 } 259 setOperationAction(ISD::SHL_PARTS, MVT::i32, Expand); 260 setOperationAction(ISD::SRA_PARTS, MVT::i32, Expand); 261 setOperationAction(ISD::SRL_PARTS, MVT::i32, Expand); 262 setOperationAction(ISD::SRL, MVT::i64, Custom); 263 setOperationAction(ISD::SRA, MVT::i64, Custom); 264 265 // ARM does not have ROTL. 266 setOperationAction(ISD::ROTL, MVT::i32, Expand); 267 setOperationAction(ISD::CTTZ, MVT::i32, Expand); 268 setOperationAction(ISD::CTPOP, MVT::i32, Expand); 269 if (!Subtarget->hasV5TOps() || Subtarget->isThumb1Only()) 270 setOperationAction(ISD::CTLZ, MVT::i32, Expand); 271 272 // Only ARMv6 has BSWAP. 273 if (!Subtarget->hasV6Ops()) 274 setOperationAction(ISD::BSWAP, MVT::i32, Expand); 275 276 // These are expanded into libcalls. 277 setOperationAction(ISD::SDIV, MVT::i32, Expand); 278 setOperationAction(ISD::UDIV, MVT::i32, Expand); 279 setOperationAction(ISD::SREM, MVT::i32, Expand); 280 setOperationAction(ISD::UREM, MVT::i32, Expand); 281 setOperationAction(ISD::SDIVREM, MVT::i32, Expand); 282 setOperationAction(ISD::UDIVREM, MVT::i32, Expand); 283 284 // Support label based line numbers. 285 setOperationAction(ISD::DBG_STOPPOINT, MVT::Other, Expand); 286 setOperationAction(ISD::DEBUG_LOC, MVT::Other, Expand); 287 288 setOperationAction(ISD::RET, MVT::Other, Custom); 289 setOperationAction(ISD::GlobalAddress, MVT::i32, Custom); 290 setOperationAction(ISD::ConstantPool, MVT::i32, Custom); 291 setOperationAction(ISD::GLOBAL_OFFSET_TABLE, MVT::i32, Custom); 292 setOperationAction(ISD::GlobalTLSAddress, MVT::i32, Custom); 293 294 // Use the default implementation. 295 setOperationAction(ISD::VASTART, MVT::Other, Custom); 296 setOperationAction(ISD::VAARG, MVT::Other, Expand); 297 setOperationAction(ISD::VACOPY, MVT::Other, Expand); 298 setOperationAction(ISD::VAEND, MVT::Other, Expand); 299 setOperationAction(ISD::STACKSAVE, MVT::Other, Expand); 300 setOperationAction(ISD::STACKRESTORE, MVT::Other, Expand); 301 setOperationAction(ISD::DYNAMIC_STACKALLOC, MVT::i32, Expand); 302 setOperationAction(ISD::MEMBARRIER, MVT::Other, Expand); 303 304 if (!Subtarget->hasV6Ops()) { 305 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i16, Expand); 306 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i8, Expand); 307 } 308 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i1, Expand); 309 310 if (!UseSoftFloat && Subtarget->hasVFP2() && !Subtarget->isThumb()) 311 // Turn f64->i64 into FMRRD, i64 -> f64 to FMDRR iff target supports vfp2. 312 setOperationAction(ISD::BIT_CONVERT, MVT::i64, Custom); 313 314 // We want to custom lower some of our intrinsics. 315 setOperationAction(ISD::INTRINSIC_WO_CHAIN, MVT::Other, Custom); 316 317 setOperationAction(ISD::SETCC, MVT::i32, Expand); 318 setOperationAction(ISD::SETCC, MVT::f32, Expand); 319 setOperationAction(ISD::SETCC, MVT::f64, Expand); 320 setOperationAction(ISD::SELECT, MVT::i32, Expand); 321 setOperationAction(ISD::SELECT, MVT::f32, Expand); 322 setOperationAction(ISD::SELECT, MVT::f64, Expand); 323 setOperationAction(ISD::SELECT_CC, MVT::i32, Custom); 324 setOperationAction(ISD::SELECT_CC, MVT::f32, Custom); 325 setOperationAction(ISD::SELECT_CC, MVT::f64, Custom); 326 327 setOperationAction(ISD::BRCOND, MVT::Other, Expand); 328 setOperationAction(ISD::BR_CC, MVT::i32, Custom); 329 setOperationAction(ISD::BR_CC, MVT::f32, Custom); 330 setOperationAction(ISD::BR_CC, MVT::f64, Custom); 331 setOperationAction(ISD::BR_JT, MVT::Other, Custom); 332 333 // We don't support sin/cos/fmod/copysign/pow 334 setOperationAction(ISD::FSIN, MVT::f64, Expand); 335 setOperationAction(ISD::FSIN, MVT::f32, Expand); 336 setOperationAction(ISD::FCOS, MVT::f32, Expand); 337 setOperationAction(ISD::FCOS, MVT::f64, Expand); 338 setOperationAction(ISD::FREM, MVT::f64, Expand); 339 setOperationAction(ISD::FREM, MVT::f32, Expand); 340 if (!UseSoftFloat && Subtarget->hasVFP2() && !Subtarget->isThumb()) { 341 setOperationAction(ISD::FCOPYSIGN, MVT::f64, Custom); 342 setOperationAction(ISD::FCOPYSIGN, MVT::f32, Custom); 343 } 344 setOperationAction(ISD::FPOW, MVT::f64, Expand); 345 setOperationAction(ISD::FPOW, MVT::f32, Expand); 346 347 // int <-> fp are custom expanded into bit_convert + ARMISD ops. 348 if (!UseSoftFloat && Subtarget->hasVFP2() && !Subtarget->isThumb()) { 349 setOperationAction(ISD::SINT_TO_FP, MVT::i32, Custom); 350 setOperationAction(ISD::UINT_TO_FP, MVT::i32, Custom); 351 setOperationAction(ISD::FP_TO_UINT, MVT::i32, Custom); 352 setOperationAction(ISD::FP_TO_SINT, MVT::i32, Custom); 353 } 354 355 // We have target-specific dag combine patterns for the following nodes: 356 // ARMISD::FMRRD - No need to call setTargetDAGCombine 357 setTargetDAGCombine(ISD::ADD); 358 setTargetDAGCombine(ISD::SUB); 359 360 setStackPointerRegisterToSaveRestore(ARM::SP); 361 setSchedulingPreference(SchedulingForRegPressure); 362 setIfCvtBlockSizeLimit(Subtarget->isThumb() ? 0 : 10); 363 setIfCvtDupBlockSizeLimit(Subtarget->isThumb() ? 0 : 2); 364 365 if (!Subtarget->isThumb()) { 366 // Use branch latency information to determine if-conversion limits. 367 // FIXME: If-converter should use instruction latency of the branch being 368 // eliminated to compute the threshold. For ARMv6, the branch "latency" 369 // varies depending on whether it's dynamically or statically predicted 370 // and on whether the destination is in the prefetch buffer. 371 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo(); 372 const InstrItineraryData &InstrItins = Subtarget->getInstrItineraryData(); 373 unsigned Latency= InstrItins.getLatency(TII->get(ARM::Bcc).getSchedClass()); 374 if (Latency > 1) { 375 setIfCvtBlockSizeLimit(Latency-1); 376 if (Latency > 2) 377 setIfCvtDupBlockSizeLimit(Latency-2); 378 } else { 379 setIfCvtBlockSizeLimit(10); 380 setIfCvtDupBlockSizeLimit(2); 381 } 382 } 383 384 maxStoresPerMemcpy = 1; //// temporary - rewrite interface to use type 385 // Do not enable CodePlacementOpt for now: it currently runs after the 386 // ARMConstantIslandPass and messes up branch relaxation and placement 387 // of constant islands. 388 // benefitFromCodePlacementOpt = true; 389} 390 391const char *ARMTargetLowering::getTargetNodeName(unsigned Opcode) const { 392 switch (Opcode) { 393 default: return 0; 394 case ARMISD::Wrapper: return "ARMISD::Wrapper"; 395 case ARMISD::WrapperJT: return "ARMISD::WrapperJT"; 396 case ARMISD::CALL: return "ARMISD::CALL"; 397 case ARMISD::CALL_PRED: return "ARMISD::CALL_PRED"; 398 case ARMISD::CALL_NOLINK: return "ARMISD::CALL_NOLINK"; 399 case ARMISD::tCALL: return "ARMISD::tCALL"; 400 case ARMISD::BRCOND: return "ARMISD::BRCOND"; 401 case ARMISD::BR_JT: return "ARMISD::BR_JT"; 402 case ARMISD::RET_FLAG: return "ARMISD::RET_FLAG"; 403 case ARMISD::PIC_ADD: return "ARMISD::PIC_ADD"; 404 case ARMISD::CMP: return "ARMISD::CMP"; 405 case ARMISD::CMPZ: return "ARMISD::CMPZ"; 406 case ARMISD::CMPFP: return "ARMISD::CMPFP"; 407 case ARMISD::CMPFPw0: return "ARMISD::CMPFPw0"; 408 case ARMISD::FMSTAT: return "ARMISD::FMSTAT"; 409 case ARMISD::CMOV: return "ARMISD::CMOV"; 410 case ARMISD::CNEG: return "ARMISD::CNEG"; 411 412 case ARMISD::FTOSI: return "ARMISD::FTOSI"; 413 case ARMISD::FTOUI: return "ARMISD::FTOUI"; 414 case ARMISD::SITOF: return "ARMISD::SITOF"; 415 case ARMISD::UITOF: return "ARMISD::UITOF"; 416 417 case ARMISD::SRL_FLAG: return "ARMISD::SRL_FLAG"; 418 case ARMISD::SRA_FLAG: return "ARMISD::SRA_FLAG"; 419 case ARMISD::RRX: return "ARMISD::RRX"; 420 421 case ARMISD::FMRRD: return "ARMISD::FMRRD"; 422 case ARMISD::FMDRR: return "ARMISD::FMDRR"; 423 424 case ARMISD::THREAD_POINTER:return "ARMISD::THREAD_POINTER"; 425 426 case ARMISD::VCEQ: return "ARMISD::VCEQ"; 427 case ARMISD::VCGE: return "ARMISD::VCGE"; 428 case ARMISD::VCGEU: return "ARMISD::VCGEU"; 429 case ARMISD::VCGT: return "ARMISD::VCGT"; 430 case ARMISD::VCGTU: return "ARMISD::VCGTU"; 431 case ARMISD::VTST: return "ARMISD::VTST"; 432 433 case ARMISD::VSHL: return "ARMISD::VSHL"; 434 case ARMISD::VSHRs: return "ARMISD::VSHRs"; 435 case ARMISD::VSHRu: return "ARMISD::VSHRu"; 436 case ARMISD::VSHLLs: return "ARMISD::VSHLLs"; 437 case ARMISD::VSHLLu: return "ARMISD::VSHLLu"; 438 case ARMISD::VSHLLi: return "ARMISD::VSHLLi"; 439 case ARMISD::VSHRN: return "ARMISD::VSHRN"; 440 case ARMISD::VRSHRs: return "ARMISD::VRSHRs"; 441 case ARMISD::VRSHRu: return "ARMISD::VRSHRu"; 442 case ARMISD::VRSHRN: return "ARMISD::VRSHRN"; 443 case ARMISD::VQSHLs: return "ARMISD::VQSHLs"; 444 case ARMISD::VQSHLu: return "ARMISD::VQSHLu"; 445 case ARMISD::VQSHLsu: return "ARMISD::VQSHLsu"; 446 case ARMISD::VQSHRNs: return "ARMISD::VQSHRNs"; 447 case ARMISD::VQSHRNu: return "ARMISD::VQSHRNu"; 448 case ARMISD::VQSHRNsu: return "ARMISD::VQSHRNsu"; 449 case ARMISD::VQRSHRNs: return "ARMISD::VQRSHRNs"; 450 case ARMISD::VQRSHRNu: return "ARMISD::VQRSHRNu"; 451 case ARMISD::VQRSHRNsu: return "ARMISD::VQRSHRNsu"; 452 case ARMISD::VGETLANEu: return "ARMISD::VGETLANEu"; 453 case ARMISD::VGETLANEs: return "ARMISD::VGETLANEs"; 454 case ARMISD::VDUPLANEQ: return "ARMISD::VDUPLANEQ"; 455 } 456} 457 458/// getFunctionAlignment - Return the Log2 alignment of this function. 459unsigned ARMTargetLowering::getFunctionAlignment(const Function *F) const { 460 return getTargetMachine().getSubtarget<ARMSubtarget>().isThumb() ? 1 : 2; 461} 462 463//===----------------------------------------------------------------------===// 464// Lowering Code 465//===----------------------------------------------------------------------===// 466 467/// IntCCToARMCC - Convert a DAG integer condition code to an ARM CC 468static ARMCC::CondCodes IntCCToARMCC(ISD::CondCode CC) { 469 switch (CC) { 470 default: assert(0 && "Unknown condition code!"); 471 case ISD::SETNE: return ARMCC::NE; 472 case ISD::SETEQ: return ARMCC::EQ; 473 case ISD::SETGT: return ARMCC::GT; 474 case ISD::SETGE: return ARMCC::GE; 475 case ISD::SETLT: return ARMCC::LT; 476 case ISD::SETLE: return ARMCC::LE; 477 case ISD::SETUGT: return ARMCC::HI; 478 case ISD::SETUGE: return ARMCC::HS; 479 case ISD::SETULT: return ARMCC::LO; 480 case ISD::SETULE: return ARMCC::LS; 481 } 482} 483 484/// FPCCToARMCC - Convert a DAG fp condition code to an ARM CC. It 485/// returns true if the operands should be inverted to form the proper 486/// comparison. 487static bool FPCCToARMCC(ISD::CondCode CC, ARMCC::CondCodes &CondCode, 488 ARMCC::CondCodes &CondCode2) { 489 bool Invert = false; 490 CondCode2 = ARMCC::AL; 491 switch (CC) { 492 default: assert(0 && "Unknown FP condition!"); 493 case ISD::SETEQ: 494 case ISD::SETOEQ: CondCode = ARMCC::EQ; break; 495 case ISD::SETGT: 496 case ISD::SETOGT: CondCode = ARMCC::GT; break; 497 case ISD::SETGE: 498 case ISD::SETOGE: CondCode = ARMCC::GE; break; 499 case ISD::SETOLT: CondCode = ARMCC::MI; break; 500 case ISD::SETOLE: CondCode = ARMCC::GT; Invert = true; break; 501 case ISD::SETONE: CondCode = ARMCC::MI; CondCode2 = ARMCC::GT; break; 502 case ISD::SETO: CondCode = ARMCC::VC; break; 503 case ISD::SETUO: CondCode = ARMCC::VS; break; 504 case ISD::SETUEQ: CondCode = ARMCC::EQ; CondCode2 = ARMCC::VS; break; 505 case ISD::SETUGT: CondCode = ARMCC::HI; break; 506 case ISD::SETUGE: CondCode = ARMCC::PL; break; 507 case ISD::SETLT: 508 case ISD::SETULT: CondCode = ARMCC::LT; break; 509 case ISD::SETLE: 510 case ISD::SETULE: CondCode = ARMCC::LE; break; 511 case ISD::SETNE: 512 case ISD::SETUNE: CondCode = ARMCC::NE; break; 513 } 514 return Invert; 515} 516 517//===----------------------------------------------------------------------===// 518// Calling Convention Implementation 519// 520// The lower operations present on calling convention works on this order: 521// LowerCALL (virt regs --> phys regs, virt regs --> stack) 522// LowerFORMAL_ARGUMENTS (phys --> virt regs, stack --> virt regs) 523// LowerRET (virt regs --> phys regs) 524// LowerCALL (phys regs --> virt regs) 525// 526//===----------------------------------------------------------------------===// 527 528#include "ARMGenCallingConv.inc" 529 530// APCS f64 is in register pairs, possibly split to stack 531static bool f64AssignAPCS(unsigned &ValNo, MVT &ValVT, MVT &LocVT, 532 CCValAssign::LocInfo &LocInfo, 533 CCState &State, bool CanFail) { 534 static const unsigned RegList[] = { ARM::R0, ARM::R1, ARM::R2, ARM::R3 }; 535 536 // Try to get the first register. 537 if (unsigned Reg = State.AllocateReg(RegList, 4)) 538 State.addLoc(CCValAssign::getCustomReg(ValNo, ValVT, Reg, LocVT, LocInfo)); 539 else { 540 // For the 2nd half of a v2f64, do not fail. 541 if (CanFail) 542 return false; 543 544 // Put the whole thing on the stack. 545 State.addLoc(CCValAssign::getCustomMem(ValNo, ValVT, 546 State.AllocateStack(8, 4), 547 LocVT, LocInfo)); 548 return true; 549 } 550 551 // Try to get the second register. 552 if (unsigned Reg = State.AllocateReg(RegList, 4)) 553 State.addLoc(CCValAssign::getCustomReg(ValNo, ValVT, Reg, LocVT, LocInfo)); 554 else 555 State.addLoc(CCValAssign::getCustomMem(ValNo, ValVT, 556 State.AllocateStack(4, 4), 557 LocVT, LocInfo)); 558 return true; 559} 560 561static bool CC_ARM_APCS_Custom_f64(unsigned &ValNo, MVT &ValVT, MVT &LocVT, 562 CCValAssign::LocInfo &LocInfo, 563 ISD::ArgFlagsTy &ArgFlags, 564 CCState &State) { 565 if (!f64AssignAPCS(ValNo, ValVT, LocVT, LocInfo, State, true)) 566 return false; 567 if (LocVT == MVT::v2f64 && 568 !f64AssignAPCS(ValNo, ValVT, LocVT, LocInfo, State, false)) 569 return false; 570 return true; // we handled it 571} 572 573// AAPCS f64 is in aligned register pairs 574static bool f64AssignAAPCS(unsigned &ValNo, MVT &ValVT, MVT &LocVT, 575 CCValAssign::LocInfo &LocInfo, 576 CCState &State, bool CanFail) { 577 static const unsigned HiRegList[] = { ARM::R0, ARM::R2 }; 578 static const unsigned LoRegList[] = { ARM::R1, ARM::R3 }; 579 580 unsigned Reg = State.AllocateReg(HiRegList, LoRegList, 2); 581 if (Reg == 0) { 582 // For the 2nd half of a v2f64, do not just fail. 583 if (CanFail) 584 return false; 585 586 // Put the whole thing on the stack. 587 State.addLoc(CCValAssign::getCustomMem(ValNo, ValVT, 588 State.AllocateStack(8, 8), 589 LocVT, LocInfo)); 590 return true; 591 } 592 593 unsigned i; 594 for (i = 0; i < 2; ++i) 595 if (HiRegList[i] == Reg) 596 break; 597 598 State.addLoc(CCValAssign::getCustomReg(ValNo, ValVT, Reg, LocVT, LocInfo)); 599 State.addLoc(CCValAssign::getCustomReg(ValNo, ValVT, LoRegList[i], 600 LocVT, LocInfo)); 601 return true; 602} 603 604static bool CC_ARM_AAPCS_Custom_f64(unsigned &ValNo, MVT &ValVT, MVT &LocVT, 605 CCValAssign::LocInfo &LocInfo, 606 ISD::ArgFlagsTy &ArgFlags, 607 CCState &State) { 608 if (!f64AssignAAPCS(ValNo, ValVT, LocVT, LocInfo, State, true)) 609 return false; 610 if (LocVT == MVT::v2f64 && 611 !f64AssignAAPCS(ValNo, ValVT, LocVT, LocInfo, State, false)) 612 return false; 613 return true; // we handled it 614} 615 616static bool f64RetAssign(unsigned &ValNo, MVT &ValVT, MVT &LocVT, 617 CCValAssign::LocInfo &LocInfo, CCState &State) { 618 static const unsigned HiRegList[] = { ARM::R0, ARM::R2 }; 619 static const unsigned LoRegList[] = { ARM::R1, ARM::R3 }; 620 621 unsigned Reg = State.AllocateReg(HiRegList, LoRegList, 2); 622 if (Reg == 0) 623 return false; // we didn't handle it 624 625 unsigned i; 626 for (i = 0; i < 2; ++i) 627 if (HiRegList[i] == Reg) 628 break; 629 630 State.addLoc(CCValAssign::getCustomReg(ValNo, ValVT, Reg, LocVT, LocInfo)); 631 State.addLoc(CCValAssign::getCustomReg(ValNo, ValVT, LoRegList[i], 632 LocVT, LocInfo)); 633 return true; 634} 635 636static bool RetCC_ARM_APCS_Custom_f64(unsigned &ValNo, MVT &ValVT, MVT &LocVT, 637 CCValAssign::LocInfo &LocInfo, 638 ISD::ArgFlagsTy &ArgFlags, 639 CCState &State) { 640 if (!f64RetAssign(ValNo, ValVT, LocVT, LocInfo, State)) 641 return false; 642 if (LocVT == MVT::v2f64 && !f64RetAssign(ValNo, ValVT, LocVT, LocInfo, State)) 643 return false; 644 return true; // we handled it 645} 646 647static bool RetCC_ARM_AAPCS_Custom_f64(unsigned &ValNo, MVT &ValVT, MVT &LocVT, 648 CCValAssign::LocInfo &LocInfo, 649 ISD::ArgFlagsTy &ArgFlags, 650 CCState &State) { 651 return RetCC_ARM_APCS_Custom_f64(ValNo, ValVT, LocVT, LocInfo, ArgFlags, 652 State); 653} 654 655/// CCAssignFnForNode - Selects the correct CCAssignFn for a the 656/// given CallingConvention value. 657CCAssignFn *ARMTargetLowering::CCAssignFnForNode(unsigned CC, 658 bool Return) const { 659 switch (CC) { 660 default: 661 assert(0 && "Unsupported calling convention"); 662 case CallingConv::C: 663 case CallingConv::Fast: 664 // Use target triple & subtarget features to do actual dispatch. 665 if (Subtarget->isAAPCS_ABI()) { 666 if (Subtarget->hasVFP2() && 667 FloatABIType == FloatABI::Hard) 668 return (Return ? RetCC_ARM_AAPCS_VFP: CC_ARM_AAPCS_VFP); 669 else 670 return (Return ? RetCC_ARM_AAPCS: CC_ARM_AAPCS); 671 } else 672 return (Return ? RetCC_ARM_APCS: CC_ARM_APCS); 673 case CallingConv::ARM_AAPCS_VFP: 674 return (Return ? RetCC_ARM_AAPCS_VFP: CC_ARM_AAPCS_VFP); 675 case CallingConv::ARM_AAPCS: 676 return (Return ? RetCC_ARM_AAPCS: CC_ARM_AAPCS); 677 case CallingConv::ARM_APCS: 678 return (Return ? RetCC_ARM_APCS: CC_ARM_APCS); 679 } 680} 681 682/// LowerCallResult - Lower the result values of an ISD::CALL into the 683/// appropriate copies out of appropriate physical registers. This assumes that 684/// Chain/InFlag are the input chain/flag to use, and that TheCall is the call 685/// being lowered. The returns a SDNode with the same number of values as the 686/// ISD::CALL. 687SDNode *ARMTargetLowering:: 688LowerCallResult(SDValue Chain, SDValue InFlag, CallSDNode *TheCall, 689 unsigned CallingConv, SelectionDAG &DAG) { 690 691 DebugLoc dl = TheCall->getDebugLoc(); 692 // Assign locations to each value returned by this call. 693 SmallVector<CCValAssign, 16> RVLocs; 694 bool isVarArg = TheCall->isVarArg(); 695 CCState CCInfo(CallingConv, isVarArg, getTargetMachine(), RVLocs); 696 CCInfo.AnalyzeCallResult(TheCall, 697 CCAssignFnForNode(CallingConv, /* Return*/ true)); 698 699 SmallVector<SDValue, 8> ResultVals; 700 701 // Copy all of the result registers out of their specified physreg. 702 for (unsigned i = 0; i != RVLocs.size(); ++i) { 703 CCValAssign VA = RVLocs[i]; 704 705 SDValue Val; 706 if (VA.needsCustom()) { 707 // Handle f64 or half of a v2f64. 708 SDValue Lo = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(), MVT::i32, 709 InFlag); 710 Chain = Lo.getValue(1); 711 InFlag = Lo.getValue(2); 712 VA = RVLocs[++i]; // skip ahead to next loc 713 SDValue Hi = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(), MVT::i32, 714 InFlag); 715 Chain = Hi.getValue(1); 716 InFlag = Hi.getValue(2); 717 Val = DAG.getNode(ARMISD::FMDRR, dl, MVT::f64, Lo, Hi); 718 719 if (VA.getLocVT() == MVT::v2f64) { 720 SDValue Vec = DAG.getNode(ISD::UNDEF, dl, MVT::v2f64); 721 Vec = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v2f64, Vec, Val, 722 DAG.getConstant(0, MVT::i32)); 723 724 VA = RVLocs[++i]; // skip ahead to next loc 725 Lo = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(), MVT::i32, InFlag); 726 Chain = Lo.getValue(1); 727 InFlag = Lo.getValue(2); 728 VA = RVLocs[++i]; // skip ahead to next loc 729 Hi = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(), MVT::i32, InFlag); 730 Chain = Hi.getValue(1); 731 InFlag = Hi.getValue(2); 732 Val = DAG.getNode(ARMISD::FMDRR, dl, MVT::f64, Lo, Hi); 733 Val = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v2f64, Vec, Val, 734 DAG.getConstant(1, MVT::i32)); 735 } 736 } else { 737 Val = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(), VA.getLocVT(), 738 InFlag); 739 Chain = Val.getValue(1); 740 InFlag = Val.getValue(2); 741 } 742 743 switch (VA.getLocInfo()) { 744 default: assert(0 && "Unknown loc info!"); 745 case CCValAssign::Full: break; 746 case CCValAssign::BCvt: 747 Val = DAG.getNode(ISD::BIT_CONVERT, dl, VA.getValVT(), Val); 748 break; 749 } 750 751 ResultVals.push_back(Val); 752 } 753 754 // Merge everything together with a MERGE_VALUES node. 755 ResultVals.push_back(Chain); 756 return DAG.getNode(ISD::MERGE_VALUES, dl, TheCall->getVTList(), 757 &ResultVals[0], ResultVals.size()).getNode(); 758} 759 760/// CreateCopyOfByValArgument - Make a copy of an aggregate at address specified 761/// by "Src" to address "Dst" of size "Size". Alignment information is 762/// specified by the specific parameter attribute. The copy will be passed as 763/// a byval function parameter. 764/// Sometimes what we are copying is the end of a larger object, the part that 765/// does not fit in registers. 766static SDValue 767CreateCopyOfByValArgument(SDValue Src, SDValue Dst, SDValue Chain, 768 ISD::ArgFlagsTy Flags, SelectionDAG &DAG, 769 DebugLoc dl) { 770 SDValue SizeNode = DAG.getConstant(Flags.getByValSize(), MVT::i32); 771 return DAG.getMemcpy(Chain, dl, Dst, Src, SizeNode, Flags.getByValAlign(), 772 /*AlwaysInline=*/false, NULL, 0, NULL, 0); 773} 774 775/// LowerMemOpCallTo - Store the argument to the stack. 776SDValue 777ARMTargetLowering::LowerMemOpCallTo(CallSDNode *TheCall, SelectionDAG &DAG, 778 const SDValue &StackPtr, 779 const CCValAssign &VA, SDValue Chain, 780 SDValue Arg, ISD::ArgFlagsTy Flags) { 781 DebugLoc dl = TheCall->getDebugLoc(); 782 unsigned LocMemOffset = VA.getLocMemOffset(); 783 SDValue PtrOff = DAG.getIntPtrConstant(LocMemOffset); 784 PtrOff = DAG.getNode(ISD::ADD, dl, getPointerTy(), StackPtr, PtrOff); 785 if (Flags.isByVal()) { 786 return CreateCopyOfByValArgument(Arg, PtrOff, Chain, Flags, DAG, dl); 787 } 788 return DAG.getStore(Chain, dl, Arg, PtrOff, 789 PseudoSourceValue::getStack(), LocMemOffset); 790} 791 792void ARMTargetLowering::PassF64ArgInRegs(CallSDNode *TheCall, SelectionDAG &DAG, 793 SDValue Chain, SDValue &Arg, 794 RegsToPassVector &RegsToPass, 795 CCValAssign &VA, CCValAssign &NextVA, 796 SDValue &StackPtr, 797 SmallVector<SDValue, 8> &MemOpChains, 798 ISD::ArgFlagsTy Flags) { 799 DebugLoc dl = TheCall->getDebugLoc(); 800 801 SDValue fmrrd = DAG.getNode(ARMISD::FMRRD, dl, 802 DAG.getVTList(MVT::i32, MVT::i32), Arg); 803 RegsToPass.push_back(std::make_pair(VA.getLocReg(), fmrrd)); 804 805 if (NextVA.isRegLoc()) 806 RegsToPass.push_back(std::make_pair(NextVA.getLocReg(), fmrrd.getValue(1))); 807 else { 808 assert(NextVA.isMemLoc()); 809 if (StackPtr.getNode() == 0) 810 StackPtr = DAG.getCopyFromReg(Chain, dl, ARM::SP, getPointerTy()); 811 812 MemOpChains.push_back(LowerMemOpCallTo(TheCall, DAG, StackPtr, NextVA, 813 Chain, fmrrd.getValue(1), Flags)); 814 } 815} 816 817/// LowerCALL - Lowering a ISD::CALL node into a callseq_start <- 818/// ARMISD:CALL <- callseq_end chain. Also add input and output parameter 819/// nodes. 820SDValue ARMTargetLowering::LowerCALL(SDValue Op, SelectionDAG &DAG) { 821 CallSDNode *TheCall = cast<CallSDNode>(Op.getNode()); 822 MVT RetVT = TheCall->getRetValType(0); 823 SDValue Chain = TheCall->getChain(); 824 unsigned CC = TheCall->getCallingConv(); 825 bool isVarArg = TheCall->isVarArg(); 826 SDValue Callee = TheCall->getCallee(); 827 DebugLoc dl = TheCall->getDebugLoc(); 828 829 // Analyze operands of the call, assigning locations to each operand. 830 SmallVector<CCValAssign, 16> ArgLocs; 831 CCState CCInfo(CC, isVarArg, getTargetMachine(), ArgLocs); 832 CCInfo.AnalyzeCallOperands(TheCall, CCAssignFnForNode(CC, /* Return*/ false)); 833 834 // Get a count of how many bytes are to be pushed on the stack. 835 unsigned NumBytes = CCInfo.getNextStackOffset(); 836 837 // Adjust the stack pointer for the new arguments... 838 // These operations are automatically eliminated by the prolog/epilog pass 839 Chain = DAG.getCALLSEQ_START(Chain, DAG.getIntPtrConstant(NumBytes, true)); 840 841 SDValue StackPtr = DAG.getRegister(ARM::SP, MVT::i32); 842 843 RegsToPassVector RegsToPass; 844 SmallVector<SDValue, 8> MemOpChains; 845 846 // Walk the register/memloc assignments, inserting copies/loads. In the case 847 // of tail call optimization, arguments are handled later. 848 for (unsigned i = 0, realArgIdx = 0, e = ArgLocs.size(); 849 i != e; 850 ++i, ++realArgIdx) { 851 CCValAssign &VA = ArgLocs[i]; 852 SDValue Arg = TheCall->getArg(realArgIdx); 853 ISD::ArgFlagsTy Flags = TheCall->getArgFlags(realArgIdx); 854 855 // Promote the value if needed. 856 switch (VA.getLocInfo()) { 857 default: assert(0 && "Unknown loc info!"); 858 case CCValAssign::Full: break; 859 case CCValAssign::SExt: 860 Arg = DAG.getNode(ISD::SIGN_EXTEND, dl, VA.getLocVT(), Arg); 861 break; 862 case CCValAssign::ZExt: 863 Arg = DAG.getNode(ISD::ZERO_EXTEND, dl, VA.getLocVT(), Arg); 864 break; 865 case CCValAssign::AExt: 866 Arg = DAG.getNode(ISD::ANY_EXTEND, dl, VA.getLocVT(), Arg); 867 break; 868 case CCValAssign::BCvt: 869 Arg = DAG.getNode(ISD::BIT_CONVERT, dl, VA.getLocVT(), Arg); 870 break; 871 } 872 873 // f64 and v2f64 are passed in i32 pairs and must be split into pieces 874 if (VA.needsCustom()) { 875 if (VA.getLocVT() == MVT::v2f64) { 876 SDValue Op0 = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64, Arg, 877 DAG.getConstant(0, MVT::i32)); 878 SDValue Op1 = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64, Arg, 879 DAG.getConstant(1, MVT::i32)); 880 881 PassF64ArgInRegs(TheCall, DAG, Chain, Op0, RegsToPass, 882 VA, ArgLocs[++i], StackPtr, MemOpChains, Flags); 883 884 VA = ArgLocs[++i]; // skip ahead to next loc 885 if (VA.isRegLoc()) { 886 PassF64ArgInRegs(TheCall, DAG, Chain, Op1, RegsToPass, 887 VA, ArgLocs[++i], StackPtr, MemOpChains, Flags); 888 } else { 889 assert(VA.isMemLoc()); 890 if (StackPtr.getNode() == 0) 891 StackPtr = DAG.getCopyFromReg(Chain, dl, ARM::SP, getPointerTy()); 892 893 MemOpChains.push_back(LowerMemOpCallTo(TheCall, DAG, StackPtr, VA, 894 Chain, Op1, Flags)); 895 } 896 } else { 897 PassF64ArgInRegs(TheCall, DAG, Chain, Arg, RegsToPass, VA, ArgLocs[++i], 898 StackPtr, MemOpChains, Flags); 899 } 900 } else if (VA.isRegLoc()) { 901 RegsToPass.push_back(std::make_pair(VA.getLocReg(), Arg)); 902 } else { 903 assert(VA.isMemLoc()); 904 if (StackPtr.getNode() == 0) 905 StackPtr = DAG.getCopyFromReg(Chain, dl, ARM::SP, getPointerTy()); 906 907 MemOpChains.push_back(LowerMemOpCallTo(TheCall, DAG, StackPtr, VA, 908 Chain, Arg, Flags)); 909 } 910 } 911 912 if (!MemOpChains.empty()) 913 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other, 914 &MemOpChains[0], MemOpChains.size()); 915 916 // Build a sequence of copy-to-reg nodes chained together with token chain 917 // and flag operands which copy the outgoing args into the appropriate regs. 918 SDValue InFlag; 919 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i) { 920 Chain = DAG.getCopyToReg(Chain, dl, RegsToPass[i].first, 921 RegsToPass[i].second, InFlag); 922 InFlag = Chain.getValue(1); 923 } 924 925 // If the callee is a GlobalAddress/ExternalSymbol node (quite common, every 926 // direct call is) turn it into a TargetGlobalAddress/TargetExternalSymbol 927 // node so that legalize doesn't hack it. 928 bool isDirect = false; 929 bool isARMFunc = false; 930 bool isLocalARMFunc = false; 931 if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee)) { 932 GlobalValue *GV = G->getGlobal(); 933 isDirect = true; 934 bool isExt = (GV->isDeclaration() || GV->hasWeakLinkage() || 935 GV->hasLinkOnceLinkage()); 936 bool isStub = (isExt && Subtarget->isTargetDarwin()) && 937 getTargetMachine().getRelocationModel() != Reloc::Static; 938 isARMFunc = !Subtarget->isThumb() || isStub; 939 // ARM call to a local ARM function is predicable. 940 isLocalARMFunc = !Subtarget->isThumb() && !isExt; 941 // tBX takes a register source operand. 942 if (isARMFunc && Subtarget->isThumb() && !Subtarget->hasV5TOps()) { 943 ARMConstantPoolValue *CPV = new ARMConstantPoolValue(GV, ARMPCLabelIndex, 944 ARMCP::CPStub, 4); 945 SDValue CPAddr = DAG.getTargetConstantPool(CPV, getPointerTy(), 4); 946 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr); 947 Callee = DAG.getLoad(getPointerTy(), dl, 948 DAG.getEntryNode(), CPAddr, NULL, 0); 949 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex++, MVT::i32); 950 Callee = DAG.getNode(ARMISD::PIC_ADD, dl, 951 getPointerTy(), Callee, PICLabel); 952 } else 953 Callee = DAG.getTargetGlobalAddress(GV, getPointerTy()); 954 } else if (ExternalSymbolSDNode *S = dyn_cast<ExternalSymbolSDNode>(Callee)) { 955 isDirect = true; 956 bool isStub = Subtarget->isTargetDarwin() && 957 getTargetMachine().getRelocationModel() != Reloc::Static; 958 isARMFunc = !Subtarget->isThumb() || isStub; 959 // tBX takes a register source operand. 960 const char *Sym = S->getSymbol(); 961 if (isARMFunc && Subtarget->isThumb() && !Subtarget->hasV5TOps()) { 962 ARMConstantPoolValue *CPV = new ARMConstantPoolValue(Sym, ARMPCLabelIndex, 963 ARMCP::CPStub, 4); 964 SDValue CPAddr = DAG.getTargetConstantPool(CPV, getPointerTy(), 4); 965 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr); 966 Callee = DAG.getLoad(getPointerTy(), dl, 967 DAG.getEntryNode(), CPAddr, NULL, 0); 968 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex++, MVT::i32); 969 Callee = DAG.getNode(ARMISD::PIC_ADD, dl, 970 getPointerTy(), Callee, PICLabel); 971 } else 972 Callee = DAG.getTargetExternalSymbol(Sym, getPointerTy()); 973 } 974 975 // FIXME: handle tail calls differently. 976 unsigned CallOpc; 977 if (Subtarget->isThumb()) { 978 if (!Subtarget->hasV5TOps() && (!isDirect || isARMFunc)) 979 CallOpc = ARMISD::CALL_NOLINK; 980 else 981 CallOpc = isARMFunc ? ARMISD::CALL : ARMISD::tCALL; 982 } else { 983 CallOpc = (isDirect || Subtarget->hasV5TOps()) 984 ? (isLocalARMFunc ? ARMISD::CALL_PRED : ARMISD::CALL) 985 : ARMISD::CALL_NOLINK; 986 } 987 if (CallOpc == ARMISD::CALL_NOLINK && !Subtarget->isThumb()) { 988 // implicit def LR - LR mustn't be allocated as GRP:$dst of CALL_NOLINK 989 Chain = DAG.getCopyToReg(Chain, dl, ARM::LR, DAG.getUNDEF(MVT::i32),InFlag); 990 InFlag = Chain.getValue(1); 991 } 992 993 std::vector<SDValue> Ops; 994 Ops.push_back(Chain); 995 Ops.push_back(Callee); 996 997 // Add argument registers to the end of the list so that they are known live 998 // into the call. 999 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i) 1000 Ops.push_back(DAG.getRegister(RegsToPass[i].first, 1001 RegsToPass[i].second.getValueType())); 1002 1003 if (InFlag.getNode()) 1004 Ops.push_back(InFlag); 1005 // Returns a chain and a flag for retval copy to use. 1006 Chain = DAG.getNode(CallOpc, dl, DAG.getVTList(MVT::Other, MVT::Flag), 1007 &Ops[0], Ops.size()); 1008 InFlag = Chain.getValue(1); 1009 1010 Chain = DAG.getCALLSEQ_END(Chain, DAG.getIntPtrConstant(NumBytes, true), 1011 DAG.getIntPtrConstant(0, true), InFlag); 1012 if (RetVT != MVT::Other) 1013 InFlag = Chain.getValue(1); 1014 1015 // Handle result values, copying them out of physregs into vregs that we 1016 // return. 1017 return SDValue(LowerCallResult(Chain, InFlag, TheCall, CC, DAG), 1018 Op.getResNo()); 1019} 1020 1021SDValue ARMTargetLowering::LowerRET(SDValue Op, SelectionDAG &DAG) { 1022 // The chain is always operand #0 1023 SDValue Chain = Op.getOperand(0); 1024 DebugLoc dl = Op.getDebugLoc(); 1025 1026 // CCValAssign - represent the assignment of the return value to a location. 1027 SmallVector<CCValAssign, 16> RVLocs; 1028 unsigned CC = DAG.getMachineFunction().getFunction()->getCallingConv(); 1029 bool isVarArg = DAG.getMachineFunction().getFunction()->isVarArg(); 1030 1031 // CCState - Info about the registers and stack slots. 1032 CCState CCInfo(CC, isVarArg, getTargetMachine(), RVLocs); 1033 1034 // Analyze return values of ISD::RET. 1035 CCInfo.AnalyzeReturn(Op.getNode(), CCAssignFnForNode(CC, /* Return */ true)); 1036 1037 // If this is the first return lowered for this function, add 1038 // the regs to the liveout set for the function. 1039 if (DAG.getMachineFunction().getRegInfo().liveout_empty()) { 1040 for (unsigned i = 0; i != RVLocs.size(); ++i) 1041 if (RVLocs[i].isRegLoc()) 1042 DAG.getMachineFunction().getRegInfo().addLiveOut(RVLocs[i].getLocReg()); 1043 } 1044 1045 SDValue Flag; 1046 1047 // Copy the result values into the output registers. 1048 for (unsigned i = 0, realRVLocIdx = 0; 1049 i != RVLocs.size(); 1050 ++i, ++realRVLocIdx) { 1051 CCValAssign &VA = RVLocs[i]; 1052 assert(VA.isRegLoc() && "Can only return in registers!"); 1053 1054 // ISD::RET => ret chain, (regnum1,val1), ... 1055 // So i*2+1 index only the regnums 1056 SDValue Arg = Op.getOperand(realRVLocIdx*2+1); 1057 1058 switch (VA.getLocInfo()) { 1059 default: assert(0 && "Unknown loc info!"); 1060 case CCValAssign::Full: break; 1061 case CCValAssign::BCvt: 1062 Arg = DAG.getNode(ISD::BIT_CONVERT, dl, VA.getLocVT(), Arg); 1063 break; 1064 } 1065 1066 if (VA.needsCustom()) { 1067 if (VA.getLocVT() == MVT::v2f64) { 1068 // Extract the first half and return it in two registers. 1069 SDValue Half = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64, Arg, 1070 DAG.getConstant(0, MVT::i32)); 1071 SDValue HalfGPRs = DAG.getNode(ARMISD::FMRRD, dl, 1072 DAG.getVTList(MVT::i32, MVT::i32), Half); 1073 1074 Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(), HalfGPRs, Flag); 1075 Flag = Chain.getValue(1); 1076 VA = RVLocs[++i]; // skip ahead to next loc 1077 Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(), 1078 HalfGPRs.getValue(1), Flag); 1079 Flag = Chain.getValue(1); 1080 VA = RVLocs[++i]; // skip ahead to next loc 1081 1082 // Extract the 2nd half and fall through to handle it as an f64 value. 1083 Arg = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64, Arg, 1084 DAG.getConstant(1, MVT::i32)); 1085 } 1086 // Legalize ret f64 -> ret 2 x i32. We always have fmrrd if f64 is 1087 // available. 1088 SDValue fmrrd = DAG.getNode(ARMISD::FMRRD, dl, 1089 DAG.getVTList(MVT::i32, MVT::i32), &Arg, 1); 1090 Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(), fmrrd, Flag); 1091 Flag = Chain.getValue(1); 1092 VA = RVLocs[++i]; // skip ahead to next loc 1093 Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(), fmrrd.getValue(1), 1094 Flag); 1095 } else 1096 Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(), Arg, Flag); 1097 1098 // Guarantee that all emitted copies are 1099 // stuck together, avoiding something bad. 1100 Flag = Chain.getValue(1); 1101 } 1102 1103 SDValue result; 1104 if (Flag.getNode()) 1105 result = DAG.getNode(ARMISD::RET_FLAG, dl, MVT::Other, Chain, Flag); 1106 else // Return Void 1107 result = DAG.getNode(ARMISD::RET_FLAG, dl, MVT::Other, Chain); 1108 1109 return result; 1110} 1111 1112// ConstantPool, JumpTable, GlobalAddress, and ExternalSymbol are lowered as 1113// their target countpart wrapped in the ARMISD::Wrapper node. Suppose N is 1114// one of the above mentioned nodes. It has to be wrapped because otherwise 1115// Select(N) returns N. So the raw TargetGlobalAddress nodes, etc. can only 1116// be used to form addressing mode. These wrapped nodes will be selected 1117// into MOVi. 1118static SDValue LowerConstantPool(SDValue Op, SelectionDAG &DAG) { 1119 MVT PtrVT = Op.getValueType(); 1120 // FIXME there is no actual debug info here 1121 DebugLoc dl = Op.getDebugLoc(); 1122 ConstantPoolSDNode *CP = cast<ConstantPoolSDNode>(Op); 1123 SDValue Res; 1124 if (CP->isMachineConstantPoolEntry()) 1125 Res = DAG.getTargetConstantPool(CP->getMachineCPVal(), PtrVT, 1126 CP->getAlignment()); 1127 else 1128 Res = DAG.getTargetConstantPool(CP->getConstVal(), PtrVT, 1129 CP->getAlignment()); 1130 return DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, Res); 1131} 1132 1133// Lower ISD::GlobalTLSAddress using the "general dynamic" model 1134SDValue 1135ARMTargetLowering::LowerToTLSGeneralDynamicModel(GlobalAddressSDNode *GA, 1136 SelectionDAG &DAG) { 1137 DebugLoc dl = GA->getDebugLoc(); 1138 MVT PtrVT = getPointerTy(); 1139 unsigned char PCAdj = Subtarget->isThumb() ? 4 : 8; 1140 ARMConstantPoolValue *CPV = 1141 new ARMConstantPoolValue(GA->getGlobal(), ARMPCLabelIndex, ARMCP::CPValue, 1142 PCAdj, "tlsgd", true); 1143 SDValue Argument = DAG.getTargetConstantPool(CPV, PtrVT, 4); 1144 Argument = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, Argument); 1145 Argument = DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), Argument, NULL, 0); 1146 SDValue Chain = Argument.getValue(1); 1147 1148 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex++, MVT::i32); 1149 Argument = DAG.getNode(ARMISD::PIC_ADD, dl, PtrVT, Argument, PICLabel); 1150 1151 // call __tls_get_addr. 1152 ArgListTy Args; 1153 ArgListEntry Entry; 1154 Entry.Node = Argument; 1155 Entry.Ty = (const Type *) Type::Int32Ty; 1156 Args.push_back(Entry); 1157 // FIXME: is there useful debug info available here? 1158 std::pair<SDValue, SDValue> CallResult = 1159 LowerCallTo(Chain, (const Type *) Type::Int32Ty, false, false, false, false, 1160 CallingConv::C, false, 1161 DAG.getExternalSymbol("__tls_get_addr", PtrVT), Args, DAG, dl); 1162 return CallResult.first; 1163} 1164 1165// Lower ISD::GlobalTLSAddress using the "initial exec" or 1166// "local exec" model. 1167SDValue 1168ARMTargetLowering::LowerToTLSExecModels(GlobalAddressSDNode *GA, 1169 SelectionDAG &DAG) { 1170 GlobalValue *GV = GA->getGlobal(); 1171 DebugLoc dl = GA->getDebugLoc(); 1172 SDValue Offset; 1173 SDValue Chain = DAG.getEntryNode(); 1174 MVT PtrVT = getPointerTy(); 1175 // Get the Thread Pointer 1176 SDValue ThreadPointer = DAG.getNode(ARMISD::THREAD_POINTER, dl, PtrVT); 1177 1178 if (GV->isDeclaration()){ 1179 // initial exec model 1180 unsigned char PCAdj = Subtarget->isThumb() ? 4 : 8; 1181 ARMConstantPoolValue *CPV = 1182 new ARMConstantPoolValue(GA->getGlobal(), ARMPCLabelIndex, ARMCP::CPValue, 1183 PCAdj, "gottpoff", true); 1184 Offset = DAG.getTargetConstantPool(CPV, PtrVT, 4); 1185 Offset = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, Offset); 1186 Offset = DAG.getLoad(PtrVT, dl, Chain, Offset, NULL, 0); 1187 Chain = Offset.getValue(1); 1188 1189 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex++, MVT::i32); 1190 Offset = DAG.getNode(ARMISD::PIC_ADD, dl, PtrVT, Offset, PICLabel); 1191 1192 Offset = DAG.getLoad(PtrVT, dl, Chain, Offset, NULL, 0); 1193 } else { 1194 // local exec model 1195 ARMConstantPoolValue *CPV = 1196 new ARMConstantPoolValue(GV, ARMCP::CPValue, "tpoff"); 1197 Offset = DAG.getTargetConstantPool(CPV, PtrVT, 4); 1198 Offset = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, Offset); 1199 Offset = DAG.getLoad(PtrVT, dl, Chain, Offset, NULL, 0); 1200 } 1201 1202 // The address of the thread local variable is the add of the thread 1203 // pointer with the offset of the variable. 1204 return DAG.getNode(ISD::ADD, dl, PtrVT, ThreadPointer, Offset); 1205} 1206 1207SDValue 1208ARMTargetLowering::LowerGlobalTLSAddress(SDValue Op, SelectionDAG &DAG) { 1209 // TODO: implement the "local dynamic" model 1210 assert(Subtarget->isTargetELF() && 1211 "TLS not implemented for non-ELF targets"); 1212 GlobalAddressSDNode *GA = cast<GlobalAddressSDNode>(Op); 1213 // If the relocation model is PIC, use the "General Dynamic" TLS Model, 1214 // otherwise use the "Local Exec" TLS Model 1215 if (getTargetMachine().getRelocationModel() == Reloc::PIC_) 1216 return LowerToTLSGeneralDynamicModel(GA, DAG); 1217 else 1218 return LowerToTLSExecModels(GA, DAG); 1219} 1220 1221SDValue ARMTargetLowering::LowerGlobalAddressELF(SDValue Op, 1222 SelectionDAG &DAG) { 1223 MVT PtrVT = getPointerTy(); 1224 DebugLoc dl = Op.getDebugLoc(); 1225 GlobalValue *GV = cast<GlobalAddressSDNode>(Op)->getGlobal(); 1226 Reloc::Model RelocM = getTargetMachine().getRelocationModel(); 1227 if (RelocM == Reloc::PIC_) { 1228 bool UseGOTOFF = GV->hasLocalLinkage() || GV->hasHiddenVisibility(); 1229 ARMConstantPoolValue *CPV = 1230 new ARMConstantPoolValue(GV, ARMCP::CPValue, UseGOTOFF ? "GOTOFF":"GOT"); 1231 SDValue CPAddr = DAG.getTargetConstantPool(CPV, PtrVT, 4); 1232 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr); 1233 SDValue Result = DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), 1234 CPAddr, NULL, 0); 1235 SDValue Chain = Result.getValue(1); 1236 SDValue GOT = DAG.getGLOBAL_OFFSET_TABLE(PtrVT); 1237 Result = DAG.getNode(ISD::ADD, dl, PtrVT, Result, GOT); 1238 if (!UseGOTOFF) 1239 Result = DAG.getLoad(PtrVT, dl, Chain, Result, NULL, 0); 1240 return Result; 1241 } else { 1242 SDValue CPAddr = DAG.getTargetConstantPool(GV, PtrVT, 4); 1243 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr); 1244 return DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), CPAddr, NULL, 0); 1245 } 1246} 1247 1248/// GVIsIndirectSymbol - true if the GV will be accessed via an indirect symbol 1249/// even in non-static mode. 1250static bool GVIsIndirectSymbol(GlobalValue *GV, Reloc::Model RelocM) { 1251 // If symbol visibility is hidden, the extra load is not needed if 1252 // the symbol is definitely defined in the current translation unit. 1253 bool isDecl = GV->isDeclaration() && !GV->hasNotBeenReadFromBitcode(); 1254 if (GV->hasHiddenVisibility() && (!isDecl && !GV->hasCommonLinkage())) 1255 return false; 1256 return RelocM != Reloc::Static && (isDecl || GV->isWeakForLinker()); 1257} 1258 1259SDValue ARMTargetLowering::LowerGlobalAddressDarwin(SDValue Op, 1260 SelectionDAG &DAG) { 1261 MVT PtrVT = getPointerTy(); 1262 DebugLoc dl = Op.getDebugLoc(); 1263 GlobalValue *GV = cast<GlobalAddressSDNode>(Op)->getGlobal(); 1264 Reloc::Model RelocM = getTargetMachine().getRelocationModel(); 1265 bool IsIndirect = GVIsIndirectSymbol(GV, RelocM); 1266 SDValue CPAddr; 1267 if (RelocM == Reloc::Static) 1268 CPAddr = DAG.getTargetConstantPool(GV, PtrVT, 4); 1269 else { 1270 unsigned PCAdj = (RelocM != Reloc::PIC_) 1271 ? 0 : (Subtarget->isThumb() ? 4 : 8); 1272 ARMCP::ARMCPKind Kind = IsIndirect ? ARMCP::CPNonLazyPtr 1273 : ARMCP::CPValue; 1274 ARMConstantPoolValue *CPV = new ARMConstantPoolValue(GV, ARMPCLabelIndex, 1275 Kind, PCAdj); 1276 CPAddr = DAG.getTargetConstantPool(CPV, PtrVT, 4); 1277 } 1278 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr); 1279 1280 SDValue Result = DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), CPAddr, NULL, 0); 1281 SDValue Chain = Result.getValue(1); 1282 1283 if (RelocM == Reloc::PIC_) { 1284 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex++, MVT::i32); 1285 Result = DAG.getNode(ARMISD::PIC_ADD, dl, PtrVT, Result, PICLabel); 1286 } 1287 if (IsIndirect) 1288 Result = DAG.getLoad(PtrVT, dl, Chain, Result, NULL, 0); 1289 1290 return Result; 1291} 1292 1293SDValue ARMTargetLowering::LowerGLOBAL_OFFSET_TABLE(SDValue Op, 1294 SelectionDAG &DAG){ 1295 assert(Subtarget->isTargetELF() && 1296 "GLOBAL OFFSET TABLE not implemented for non-ELF targets"); 1297 MVT PtrVT = getPointerTy(); 1298 DebugLoc dl = Op.getDebugLoc(); 1299 unsigned PCAdj = Subtarget->isThumb() ? 4 : 8; 1300 ARMConstantPoolValue *CPV = new ARMConstantPoolValue("_GLOBAL_OFFSET_TABLE_", 1301 ARMPCLabelIndex, 1302 ARMCP::CPValue, PCAdj); 1303 SDValue CPAddr = DAG.getTargetConstantPool(CPV, PtrVT, 4); 1304 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr); 1305 SDValue Result = DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), CPAddr, NULL, 0); 1306 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex++, MVT::i32); 1307 return DAG.getNode(ARMISD::PIC_ADD, dl, PtrVT, Result, PICLabel); 1308} 1309 1310SDValue 1311ARMTargetLowering::LowerINTRINSIC_WO_CHAIN(SDValue Op, SelectionDAG &DAG) { 1312 MVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy(); 1313 unsigned IntNo = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue(); 1314 DebugLoc dl = Op.getDebugLoc(); 1315 switch (IntNo) { 1316 default: return SDValue(); // Don't custom lower most intrinsics. 1317 case Intrinsic::arm_thread_pointer: 1318 return DAG.getNode(ARMISD::THREAD_POINTER, dl, PtrVT); 1319 case Intrinsic::eh_sjlj_setjmp: 1320 SDValue Res = DAG.getNode(ARMISD::EH_SJLJ_SETJMP, dl, MVT::i32, 1321 Op.getOperand(1)); 1322 return Res; 1323 } 1324} 1325 1326static SDValue LowerVASTART(SDValue Op, SelectionDAG &DAG, 1327 unsigned VarArgsFrameIndex) { 1328 // vastart just stores the address of the VarArgsFrameIndex slot into the 1329 // memory location argument. 1330 DebugLoc dl = Op.getDebugLoc(); 1331 MVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy(); 1332 SDValue FR = DAG.getFrameIndex(VarArgsFrameIndex, PtrVT); 1333 const Value *SV = cast<SrcValueSDNode>(Op.getOperand(2))->getValue(); 1334 return DAG.getStore(Op.getOperand(0), dl, FR, Op.getOperand(1), SV, 0); 1335} 1336 1337SDValue 1338ARMTargetLowering::GetF64FormalArgument(CCValAssign &VA, CCValAssign &NextVA, 1339 SDValue &Root, SelectionDAG &DAG, 1340 DebugLoc dl) { 1341 MachineFunction &MF = DAG.getMachineFunction(); 1342 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>(); 1343 1344 TargetRegisterClass *RC; 1345 if (AFI->isThumbFunction()) 1346 RC = ARM::tGPRRegisterClass; 1347 else 1348 RC = ARM::GPRRegisterClass; 1349 1350 // Transform the arguments stored in physical registers into virtual ones. 1351 unsigned Reg = MF.addLiveIn(VA.getLocReg(), RC); 1352 SDValue ArgValue = DAG.getCopyFromReg(Root, dl, Reg, MVT::i32); 1353 1354 SDValue ArgValue2; 1355 if (NextVA.isMemLoc()) { 1356 unsigned ArgSize = NextVA.getLocVT().getSizeInBits()/8; 1357 MachineFrameInfo *MFI = MF.getFrameInfo(); 1358 int FI = MFI->CreateFixedObject(ArgSize, NextVA.getLocMemOffset()); 1359 1360 // Create load node to retrieve arguments from the stack. 1361 SDValue FIN = DAG.getFrameIndex(FI, getPointerTy()); 1362 ArgValue2 = DAG.getLoad(MVT::i32, dl, Root, FIN, NULL, 0); 1363 } else { 1364 Reg = MF.addLiveIn(NextVA.getLocReg(), RC); 1365 ArgValue2 = DAG.getCopyFromReg(Root, dl, Reg, MVT::i32); 1366 } 1367 1368 return DAG.getNode(ARMISD::FMDRR, dl, MVT::f64, ArgValue, ArgValue2); 1369} 1370 1371SDValue 1372ARMTargetLowering::LowerFORMAL_ARGUMENTS(SDValue Op, SelectionDAG &DAG) { 1373 MachineFunction &MF = DAG.getMachineFunction(); 1374 MachineFrameInfo *MFI = MF.getFrameInfo(); 1375 1376 SDValue Root = Op.getOperand(0); 1377 DebugLoc dl = Op.getDebugLoc(); 1378 bool isVarArg = cast<ConstantSDNode>(Op.getOperand(2))->getZExtValue() != 0; 1379 unsigned CC = MF.getFunction()->getCallingConv(); 1380 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>(); 1381 1382 // Assign locations to all of the incoming arguments. 1383 SmallVector<CCValAssign, 16> ArgLocs; 1384 CCState CCInfo(CC, isVarArg, getTargetMachine(), ArgLocs); 1385 CCInfo.AnalyzeFormalArguments(Op.getNode(), 1386 CCAssignFnForNode(CC, /* Return*/ false)); 1387 1388 SmallVector<SDValue, 16> ArgValues; 1389 1390 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) { 1391 CCValAssign &VA = ArgLocs[i]; 1392 1393 // Arguments stored in registers. 1394 if (VA.isRegLoc()) { 1395 MVT RegVT = VA.getLocVT(); 1396 1397 SDValue ArgValue; 1398 if (VA.needsCustom()) { 1399 // f64 and vector types are split up into multiple registers or 1400 // combinations of registers and stack slots. 1401 RegVT = MVT::i32; 1402 1403 if (VA.getLocVT() == MVT::v2f64) { 1404 SDValue ArgValue1 = GetF64FormalArgument(VA, ArgLocs[++i], 1405 Root, DAG, dl); 1406 VA = ArgLocs[++i]; // skip ahead to next loc 1407 SDValue ArgValue2 = GetF64FormalArgument(VA, ArgLocs[++i], 1408 Root, DAG, dl); 1409 ArgValue = DAG.getNode(ISD::UNDEF, dl, MVT::v2f64); 1410 ArgValue = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v2f64, 1411 ArgValue, ArgValue1, DAG.getIntPtrConstant(0)); 1412 ArgValue = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v2f64, 1413 ArgValue, ArgValue2, DAG.getIntPtrConstant(1)); 1414 } else 1415 ArgValue = GetF64FormalArgument(VA, ArgLocs[++i], Root, DAG, dl); 1416 1417 } else { 1418 TargetRegisterClass *RC; 1419 if (FloatABIType == FloatABI::Hard && RegVT == MVT::f32) 1420 RC = ARM::SPRRegisterClass; 1421 else if (FloatABIType == FloatABI::Hard && RegVT == MVT::f64) 1422 RC = ARM::DPRRegisterClass; 1423 else if (AFI->isThumbFunction()) 1424 RC = ARM::tGPRRegisterClass; 1425 else 1426 RC = ARM::GPRRegisterClass; 1427 1428 assert((RegVT == MVT::i32 || RegVT == MVT::f32 || 1429 (FloatABIType == FloatABI::Hard && RegVT == MVT::f64)) && 1430 "RegVT not supported by FORMAL_ARGUMENTS Lowering"); 1431 1432 // Transform the arguments in physical registers into virtual ones. 1433 unsigned Reg = MF.addLiveIn(VA.getLocReg(), RC); 1434 ArgValue = DAG.getCopyFromReg(Root, dl, Reg, RegVT); 1435 } 1436 1437 // If this is an 8 or 16-bit value, it is really passed promoted 1438 // to 32 bits. Insert an assert[sz]ext to capture this, then 1439 // truncate to the right size. 1440 switch (VA.getLocInfo()) { 1441 default: assert(0 && "Unknown loc info!"); 1442 case CCValAssign::Full: break; 1443 case CCValAssign::BCvt: 1444 ArgValue = DAG.getNode(ISD::BIT_CONVERT, dl, VA.getValVT(), ArgValue); 1445 break; 1446 case CCValAssign::SExt: 1447 ArgValue = DAG.getNode(ISD::AssertSext, dl, RegVT, ArgValue, 1448 DAG.getValueType(VA.getValVT())); 1449 ArgValue = DAG.getNode(ISD::TRUNCATE, dl, VA.getValVT(), ArgValue); 1450 break; 1451 case CCValAssign::ZExt: 1452 ArgValue = DAG.getNode(ISD::AssertZext, dl, RegVT, ArgValue, 1453 DAG.getValueType(VA.getValVT())); 1454 ArgValue = DAG.getNode(ISD::TRUNCATE, dl, VA.getValVT(), ArgValue); 1455 break; 1456 } 1457 1458 ArgValues.push_back(ArgValue); 1459 1460 } else { // VA.isRegLoc() 1461 1462 // sanity check 1463 assert(VA.isMemLoc()); 1464 assert(VA.getValVT() != MVT::i64 && "i64 should already be lowered"); 1465 1466 unsigned ArgSize = VA.getLocVT().getSizeInBits()/8; 1467 int FI = MFI->CreateFixedObject(ArgSize, VA.getLocMemOffset()); 1468 1469 // Create load nodes to retrieve arguments from the stack. 1470 SDValue FIN = DAG.getFrameIndex(FI, getPointerTy()); 1471 ArgValues.push_back(DAG.getLoad(VA.getValVT(), dl, Root, FIN, NULL, 0)); 1472 } 1473 } 1474 1475 // varargs 1476 if (isVarArg) { 1477 static const unsigned GPRArgRegs[] = { 1478 ARM::R0, ARM::R1, ARM::R2, ARM::R3 1479 }; 1480 1481 unsigned NumGPRs = CCInfo.getFirstUnallocated 1482 (GPRArgRegs, sizeof(GPRArgRegs) / sizeof(GPRArgRegs[0])); 1483 1484 unsigned Align = MF.getTarget().getFrameInfo()->getStackAlignment(); 1485 unsigned VARegSize = (4 - NumGPRs) * 4; 1486 unsigned VARegSaveSize = (VARegSize + Align - 1) & ~(Align - 1); 1487 unsigned ArgOffset = 0; 1488 if (VARegSaveSize) { 1489 // If this function is vararg, store any remaining integer argument regs 1490 // to their spots on the stack so that they may be loaded by deferencing 1491 // the result of va_next. 1492 AFI->setVarArgsRegSaveSize(VARegSaveSize); 1493 ArgOffset = CCInfo.getNextStackOffset(); 1494 VarArgsFrameIndex = MFI->CreateFixedObject(VARegSaveSize, ArgOffset + 1495 VARegSaveSize - VARegSize); 1496 SDValue FIN = DAG.getFrameIndex(VarArgsFrameIndex, getPointerTy()); 1497 1498 SmallVector<SDValue, 4> MemOps; 1499 for (; NumGPRs < 4; ++NumGPRs) { 1500 TargetRegisterClass *RC; 1501 if (AFI->isThumbFunction()) 1502 RC = ARM::tGPRRegisterClass; 1503 else 1504 RC = ARM::GPRRegisterClass; 1505 1506 unsigned VReg = MF.addLiveIn(GPRArgRegs[NumGPRs], RC); 1507 SDValue Val = DAG.getCopyFromReg(Root, dl, VReg, MVT::i32); 1508 SDValue Store = DAG.getStore(Val.getValue(1), dl, Val, FIN, NULL, 0); 1509 MemOps.push_back(Store); 1510 FIN = DAG.getNode(ISD::ADD, dl, getPointerTy(), FIN, 1511 DAG.getConstant(4, getPointerTy())); 1512 } 1513 if (!MemOps.empty()) 1514 Root = DAG.getNode(ISD::TokenFactor, dl, MVT::Other, 1515 &MemOps[0], MemOps.size()); 1516 } else 1517 // This will point to the next argument passed via stack. 1518 VarArgsFrameIndex = MFI->CreateFixedObject(4, ArgOffset); 1519 } 1520 1521 ArgValues.push_back(Root); 1522 1523 // Return the new list of results. 1524 return DAG.getNode(ISD::MERGE_VALUES, dl, Op.getNode()->getVTList(), 1525 &ArgValues[0], ArgValues.size()).getValue(Op.getResNo()); 1526} 1527 1528/// isFloatingPointZero - Return true if this is +0.0. 1529static bool isFloatingPointZero(SDValue Op) { 1530 if (ConstantFPSDNode *CFP = dyn_cast<ConstantFPSDNode>(Op)) 1531 return CFP->getValueAPF().isPosZero(); 1532 else if (ISD::isEXTLoad(Op.getNode()) || ISD::isNON_EXTLoad(Op.getNode())) { 1533 // Maybe this has already been legalized into the constant pool? 1534 if (Op.getOperand(1).getOpcode() == ARMISD::Wrapper) { 1535 SDValue WrapperOp = Op.getOperand(1).getOperand(0); 1536 if (ConstantPoolSDNode *CP = dyn_cast<ConstantPoolSDNode>(WrapperOp)) 1537 if (ConstantFP *CFP = dyn_cast<ConstantFP>(CP->getConstVal())) 1538 return CFP->getValueAPF().isPosZero(); 1539 } 1540 } 1541 return false; 1542} 1543 1544static bool isLegalCmpImmediate(unsigned C, bool isThumb) { 1545 return ( isThumb && (C & ~255U) == 0) || 1546 (!isThumb && ARM_AM::getSOImmVal(C) != -1); 1547} 1548 1549/// Returns appropriate ARM CMP (cmp) and corresponding condition code for 1550/// the given operands. 1551static SDValue getARMCmp(SDValue LHS, SDValue RHS, ISD::CondCode CC, 1552 SDValue &ARMCC, SelectionDAG &DAG, bool isThumb, 1553 DebugLoc dl) { 1554 if (ConstantSDNode *RHSC = dyn_cast<ConstantSDNode>(RHS.getNode())) { 1555 unsigned C = RHSC->getZExtValue(); 1556 if (!isLegalCmpImmediate(C, isThumb)) { 1557 // Constant does not fit, try adjusting it by one? 1558 switch (CC) { 1559 default: break; 1560 case ISD::SETLT: 1561 case ISD::SETGE: 1562 if (isLegalCmpImmediate(C-1, isThumb)) { 1563 CC = (CC == ISD::SETLT) ? ISD::SETLE : ISD::SETGT; 1564 RHS = DAG.getConstant(C-1, MVT::i32); 1565 } 1566 break; 1567 case ISD::SETULT: 1568 case ISD::SETUGE: 1569 if (C > 0 && isLegalCmpImmediate(C-1, isThumb)) { 1570 CC = (CC == ISD::SETULT) ? ISD::SETULE : ISD::SETUGT; 1571 RHS = DAG.getConstant(C-1, MVT::i32); 1572 } 1573 break; 1574 case ISD::SETLE: 1575 case ISD::SETGT: 1576 if (isLegalCmpImmediate(C+1, isThumb)) { 1577 CC = (CC == ISD::SETLE) ? ISD::SETLT : ISD::SETGE; 1578 RHS = DAG.getConstant(C+1, MVT::i32); 1579 } 1580 break; 1581 case ISD::SETULE: 1582 case ISD::SETUGT: 1583 if (C < 0xffffffff && isLegalCmpImmediate(C+1, isThumb)) { 1584 CC = (CC == ISD::SETULE) ? ISD::SETULT : ISD::SETUGE; 1585 RHS = DAG.getConstant(C+1, MVT::i32); 1586 } 1587 break; 1588 } 1589 } 1590 } 1591 1592 ARMCC::CondCodes CondCode = IntCCToARMCC(CC); 1593 ARMISD::NodeType CompareType; 1594 switch (CondCode) { 1595 default: 1596 CompareType = ARMISD::CMP; 1597 break; 1598 case ARMCC::EQ: 1599 case ARMCC::NE: 1600 // Uses only Z Flag 1601 CompareType = ARMISD::CMPZ; 1602 break; 1603 } 1604 ARMCC = DAG.getConstant(CondCode, MVT::i32); 1605 return DAG.getNode(CompareType, dl, MVT::Flag, LHS, RHS); 1606} 1607 1608/// Returns a appropriate VFP CMP (fcmp{s|d}+fmstat) for the given operands. 1609static SDValue getVFPCmp(SDValue LHS, SDValue RHS, SelectionDAG &DAG, 1610 DebugLoc dl) { 1611 SDValue Cmp; 1612 if (!isFloatingPointZero(RHS)) 1613 Cmp = DAG.getNode(ARMISD::CMPFP, dl, MVT::Flag, LHS, RHS); 1614 else 1615 Cmp = DAG.getNode(ARMISD::CMPFPw0, dl, MVT::Flag, LHS); 1616 return DAG.getNode(ARMISD::FMSTAT, dl, MVT::Flag, Cmp); 1617} 1618 1619static SDValue LowerSELECT_CC(SDValue Op, SelectionDAG &DAG, 1620 const ARMSubtarget *ST) { 1621 MVT VT = Op.getValueType(); 1622 SDValue LHS = Op.getOperand(0); 1623 SDValue RHS = Op.getOperand(1); 1624 ISD::CondCode CC = cast<CondCodeSDNode>(Op.getOperand(4))->get(); 1625 SDValue TrueVal = Op.getOperand(2); 1626 SDValue FalseVal = Op.getOperand(3); 1627 DebugLoc dl = Op.getDebugLoc(); 1628 1629 if (LHS.getValueType() == MVT::i32) { 1630 SDValue ARMCC; 1631 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32); 1632 SDValue Cmp = getARMCmp(LHS, RHS, CC, ARMCC, DAG, ST->isThumb(), dl); 1633 return DAG.getNode(ARMISD::CMOV, dl, VT, FalseVal, TrueVal, ARMCC, CCR,Cmp); 1634 } 1635 1636 ARMCC::CondCodes CondCode, CondCode2; 1637 if (FPCCToARMCC(CC, CondCode, CondCode2)) 1638 std::swap(TrueVal, FalseVal); 1639 1640 SDValue ARMCC = DAG.getConstant(CondCode, MVT::i32); 1641 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32); 1642 SDValue Cmp = getVFPCmp(LHS, RHS, DAG, dl); 1643 SDValue Result = DAG.getNode(ARMISD::CMOV, dl, VT, FalseVal, TrueVal, 1644 ARMCC, CCR, Cmp); 1645 if (CondCode2 != ARMCC::AL) { 1646 SDValue ARMCC2 = DAG.getConstant(CondCode2, MVT::i32); 1647 // FIXME: Needs another CMP because flag can have but one use. 1648 SDValue Cmp2 = getVFPCmp(LHS, RHS, DAG, dl); 1649 Result = DAG.getNode(ARMISD::CMOV, dl, VT, 1650 Result, TrueVal, ARMCC2, CCR, Cmp2); 1651 } 1652 return Result; 1653} 1654 1655static SDValue LowerBR_CC(SDValue Op, SelectionDAG &DAG, 1656 const ARMSubtarget *ST) { 1657 SDValue Chain = Op.getOperand(0); 1658 ISD::CondCode CC = cast<CondCodeSDNode>(Op.getOperand(1))->get(); 1659 SDValue LHS = Op.getOperand(2); 1660 SDValue RHS = Op.getOperand(3); 1661 SDValue Dest = Op.getOperand(4); 1662 DebugLoc dl = Op.getDebugLoc(); 1663 1664 if (LHS.getValueType() == MVT::i32) { 1665 SDValue ARMCC; 1666 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32); 1667 SDValue Cmp = getARMCmp(LHS, RHS, CC, ARMCC, DAG, ST->isThumb(), dl); 1668 return DAG.getNode(ARMISD::BRCOND, dl, MVT::Other, 1669 Chain, Dest, ARMCC, CCR,Cmp); 1670 } 1671 1672 assert(LHS.getValueType() == MVT::f32 || LHS.getValueType() == MVT::f64); 1673 ARMCC::CondCodes CondCode, CondCode2; 1674 if (FPCCToARMCC(CC, CondCode, CondCode2)) 1675 // Swap the LHS/RHS of the comparison if needed. 1676 std::swap(LHS, RHS); 1677 1678 SDValue Cmp = getVFPCmp(LHS, RHS, DAG, dl); 1679 SDValue ARMCC = DAG.getConstant(CondCode, MVT::i32); 1680 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32); 1681 SDVTList VTList = DAG.getVTList(MVT::Other, MVT::Flag); 1682 SDValue Ops[] = { Chain, Dest, ARMCC, CCR, Cmp }; 1683 SDValue Res = DAG.getNode(ARMISD::BRCOND, dl, VTList, Ops, 5); 1684 if (CondCode2 != ARMCC::AL) { 1685 ARMCC = DAG.getConstant(CondCode2, MVT::i32); 1686 SDValue Ops[] = { Res, Dest, ARMCC, CCR, Res.getValue(1) }; 1687 Res = DAG.getNode(ARMISD::BRCOND, dl, VTList, Ops, 5); 1688 } 1689 return Res; 1690} 1691 1692SDValue ARMTargetLowering::LowerBR_JT(SDValue Op, SelectionDAG &DAG) { 1693 SDValue Chain = Op.getOperand(0); 1694 SDValue Table = Op.getOperand(1); 1695 SDValue Index = Op.getOperand(2); 1696 DebugLoc dl = Op.getDebugLoc(); 1697 1698 MVT PTy = getPointerTy(); 1699 JumpTableSDNode *JT = cast<JumpTableSDNode>(Table); 1700 ARMFunctionInfo *AFI = DAG.getMachineFunction().getInfo<ARMFunctionInfo>(); 1701 SDValue UId = DAG.getConstant(AFI->createJumpTableUId(), PTy); 1702 SDValue JTI = DAG.getTargetJumpTable(JT->getIndex(), PTy); 1703 Table = DAG.getNode(ARMISD::WrapperJT, dl, MVT::i32, JTI, UId); 1704 Index = DAG.getNode(ISD::MUL, dl, PTy, Index, DAG.getConstant(4, PTy)); 1705 SDValue Addr = DAG.getNode(ISD::ADD, dl, PTy, Index, Table); 1706 bool isPIC = getTargetMachine().getRelocationModel() == Reloc::PIC_; 1707 Addr = DAG.getLoad(isPIC ? (MVT)MVT::i32 : PTy, dl, 1708 Chain, Addr, NULL, 0); 1709 Chain = Addr.getValue(1); 1710 if (isPIC) 1711 Addr = DAG.getNode(ISD::ADD, dl, PTy, Addr, Table); 1712 return DAG.getNode(ARMISD::BR_JT, dl, MVT::Other, Chain, Addr, JTI, UId); 1713} 1714 1715static SDValue LowerFP_TO_INT(SDValue Op, SelectionDAG &DAG) { 1716 DebugLoc dl = Op.getDebugLoc(); 1717 unsigned Opc = 1718 Op.getOpcode() == ISD::FP_TO_SINT ? ARMISD::FTOSI : ARMISD::FTOUI; 1719 Op = DAG.getNode(Opc, dl, MVT::f32, Op.getOperand(0)); 1720 return DAG.getNode(ISD::BIT_CONVERT, dl, MVT::i32, Op); 1721} 1722 1723static SDValue LowerINT_TO_FP(SDValue Op, SelectionDAG &DAG) { 1724 MVT VT = Op.getValueType(); 1725 DebugLoc dl = Op.getDebugLoc(); 1726 unsigned Opc = 1727 Op.getOpcode() == ISD::SINT_TO_FP ? ARMISD::SITOF : ARMISD::UITOF; 1728 1729 Op = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::f32, Op.getOperand(0)); 1730 return DAG.getNode(Opc, dl, VT, Op); 1731} 1732 1733static SDValue LowerFCOPYSIGN(SDValue Op, SelectionDAG &DAG) { 1734 // Implement fcopysign with a fabs and a conditional fneg. 1735 SDValue Tmp0 = Op.getOperand(0); 1736 SDValue Tmp1 = Op.getOperand(1); 1737 DebugLoc dl = Op.getDebugLoc(); 1738 MVT VT = Op.getValueType(); 1739 MVT SrcVT = Tmp1.getValueType(); 1740 SDValue AbsVal = DAG.getNode(ISD::FABS, dl, VT, Tmp0); 1741 SDValue Cmp = getVFPCmp(Tmp1, DAG.getConstantFP(0.0, SrcVT), DAG, dl); 1742 SDValue ARMCC = DAG.getConstant(ARMCC::LT, MVT::i32); 1743 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32); 1744 return DAG.getNode(ARMISD::CNEG, dl, VT, AbsVal, AbsVal, ARMCC, CCR, Cmp); 1745} 1746 1747SDValue ARMTargetLowering::LowerFRAMEADDR(SDValue Op, SelectionDAG &DAG) { 1748 MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo(); 1749 MFI->setFrameAddressIsTaken(true); 1750 MVT VT = Op.getValueType(); 1751 DebugLoc dl = Op.getDebugLoc(); // FIXME probably not meaningful 1752 unsigned Depth = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue(); 1753 unsigned FrameReg = (Subtarget->isThumb() || Subtarget->isTargetDarwin()) 1754 ? ARM::R7 : ARM::R11; 1755 SDValue FrameAddr = DAG.getCopyFromReg(DAG.getEntryNode(), dl, FrameReg, VT); 1756 while (Depth--) 1757 FrameAddr = DAG.getLoad(VT, dl, DAG.getEntryNode(), FrameAddr, NULL, 0); 1758 return FrameAddr; 1759} 1760 1761SDValue 1762ARMTargetLowering::EmitTargetCodeForMemcpy(SelectionDAG &DAG, DebugLoc dl, 1763 SDValue Chain, 1764 SDValue Dst, SDValue Src, 1765 SDValue Size, unsigned Align, 1766 bool AlwaysInline, 1767 const Value *DstSV, uint64_t DstSVOff, 1768 const Value *SrcSV, uint64_t SrcSVOff){ 1769 // Do repeated 4-byte loads and stores. To be improved. 1770 // This requires 4-byte alignment. 1771 if ((Align & 3) != 0) 1772 return SDValue(); 1773 // This requires the copy size to be a constant, preferrably 1774 // within a subtarget-specific limit. 1775 ConstantSDNode *ConstantSize = dyn_cast<ConstantSDNode>(Size); 1776 if (!ConstantSize) 1777 return SDValue(); 1778 uint64_t SizeVal = ConstantSize->getZExtValue(); 1779 if (!AlwaysInline && SizeVal > getSubtarget()->getMaxInlineSizeThreshold()) 1780 return SDValue(); 1781 1782 unsigned BytesLeft = SizeVal & 3; 1783 unsigned NumMemOps = SizeVal >> 2; 1784 unsigned EmittedNumMemOps = 0; 1785 MVT VT = MVT::i32; 1786 unsigned VTSize = 4; 1787 unsigned i = 0; 1788 const unsigned MAX_LOADS_IN_LDM = 6; 1789 SDValue TFOps[MAX_LOADS_IN_LDM]; 1790 SDValue Loads[MAX_LOADS_IN_LDM]; 1791 uint64_t SrcOff = 0, DstOff = 0; 1792 1793 // Emit up to MAX_LOADS_IN_LDM loads, then a TokenFactor barrier, then the 1794 // same number of stores. The loads and stores will get combined into 1795 // ldm/stm later on. 1796 while (EmittedNumMemOps < NumMemOps) { 1797 for (i = 0; 1798 i < MAX_LOADS_IN_LDM && EmittedNumMemOps + i < NumMemOps; ++i) { 1799 Loads[i] = DAG.getLoad(VT, dl, Chain, 1800 DAG.getNode(ISD::ADD, dl, MVT::i32, Src, 1801 DAG.getConstant(SrcOff, MVT::i32)), 1802 SrcSV, SrcSVOff + SrcOff); 1803 TFOps[i] = Loads[i].getValue(1); 1804 SrcOff += VTSize; 1805 } 1806 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other, &TFOps[0], i); 1807 1808 for (i = 0; 1809 i < MAX_LOADS_IN_LDM && EmittedNumMemOps + i < NumMemOps; ++i) { 1810 TFOps[i] = DAG.getStore(Chain, dl, Loads[i], 1811 DAG.getNode(ISD::ADD, dl, MVT::i32, Dst, 1812 DAG.getConstant(DstOff, MVT::i32)), 1813 DstSV, DstSVOff + DstOff); 1814 DstOff += VTSize; 1815 } 1816 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other, &TFOps[0], i); 1817 1818 EmittedNumMemOps += i; 1819 } 1820 1821 if (BytesLeft == 0) 1822 return Chain; 1823 1824 // Issue loads / stores for the trailing (1 - 3) bytes. 1825 unsigned BytesLeftSave = BytesLeft; 1826 i = 0; 1827 while (BytesLeft) { 1828 if (BytesLeft >= 2) { 1829 VT = MVT::i16; 1830 VTSize = 2; 1831 } else { 1832 VT = MVT::i8; 1833 VTSize = 1; 1834 } 1835 1836 Loads[i] = DAG.getLoad(VT, dl, Chain, 1837 DAG.getNode(ISD::ADD, dl, MVT::i32, Src, 1838 DAG.getConstant(SrcOff, MVT::i32)), 1839 SrcSV, SrcSVOff + SrcOff); 1840 TFOps[i] = Loads[i].getValue(1); 1841 ++i; 1842 SrcOff += VTSize; 1843 BytesLeft -= VTSize; 1844 } 1845 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other, &TFOps[0], i); 1846 1847 i = 0; 1848 BytesLeft = BytesLeftSave; 1849 while (BytesLeft) { 1850 if (BytesLeft >= 2) { 1851 VT = MVT::i16; 1852 VTSize = 2; 1853 } else { 1854 VT = MVT::i8; 1855 VTSize = 1; 1856 } 1857 1858 TFOps[i] = DAG.getStore(Chain, dl, Loads[i], 1859 DAG.getNode(ISD::ADD, dl, MVT::i32, Dst, 1860 DAG.getConstant(DstOff, MVT::i32)), 1861 DstSV, DstSVOff + DstOff); 1862 ++i; 1863 DstOff += VTSize; 1864 BytesLeft -= VTSize; 1865 } 1866 return DAG.getNode(ISD::TokenFactor, dl, MVT::Other, &TFOps[0], i); 1867} 1868 1869static SDValue ExpandBIT_CONVERT(SDNode *N, SelectionDAG &DAG) { 1870 SDValue Op = N->getOperand(0); 1871 DebugLoc dl = N->getDebugLoc(); 1872 if (N->getValueType(0) == MVT::f64) { 1873 // Turn i64->f64 into FMDRR. 1874 SDValue Lo = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, Op, 1875 DAG.getConstant(0, MVT::i32)); 1876 SDValue Hi = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, Op, 1877 DAG.getConstant(1, MVT::i32)); 1878 return DAG.getNode(ARMISD::FMDRR, dl, MVT::f64, Lo, Hi); 1879 } 1880 1881 // Turn f64->i64 into FMRRD. 1882 SDValue Cvt = DAG.getNode(ARMISD::FMRRD, dl, 1883 DAG.getVTList(MVT::i32, MVT::i32), &Op, 1); 1884 1885 // Merge the pieces into a single i64 value. 1886 return DAG.getNode(ISD::BUILD_PAIR, dl, MVT::i64, Cvt, Cvt.getValue(1)); 1887} 1888 1889/// getZeroVector - Returns a vector of specified type with all zero elements. 1890/// 1891static SDValue getZeroVector(MVT VT, SelectionDAG &DAG, DebugLoc dl) { 1892 assert(VT.isVector() && "Expected a vector type"); 1893 1894 // Zero vectors are used to represent vector negation and in those cases 1895 // will be implemented with the NEON VNEG instruction. However, VNEG does 1896 // not support i64 elements, so sometimes the zero vectors will need to be 1897 // explicitly constructed. For those cases, and potentially other uses in 1898 // the future, always build zero vectors as <4 x i32> or <2 x i32> bitcasted 1899 // to their dest type. This ensures they get CSE'd. 1900 SDValue Vec; 1901 SDValue Cst = DAG.getTargetConstant(0, MVT::i32); 1902 if (VT.getSizeInBits() == 64) 1903 Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v2i32, Cst, Cst); 1904 else 1905 Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v4i32, Cst, Cst, Cst, Cst); 1906 1907 return DAG.getNode(ISD::BIT_CONVERT, dl, VT, Vec); 1908} 1909 1910/// getOnesVector - Returns a vector of specified type with all bits set. 1911/// 1912static SDValue getOnesVector(MVT VT, SelectionDAG &DAG, DebugLoc dl) { 1913 assert(VT.isVector() && "Expected a vector type"); 1914 1915 // Always build ones vectors as <4 x i32> or <2 x i32> bitcasted to their dest 1916 // type. This ensures they get CSE'd. 1917 SDValue Vec; 1918 SDValue Cst = DAG.getTargetConstant(~0U, MVT::i32); 1919 if (VT.getSizeInBits() == 64) 1920 Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v2i32, Cst, Cst); 1921 else 1922 Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v4i32, Cst, Cst, Cst, Cst); 1923 1924 return DAG.getNode(ISD::BIT_CONVERT, dl, VT, Vec); 1925} 1926 1927static SDValue LowerShift(SDNode *N, SelectionDAG &DAG, 1928 const ARMSubtarget *ST) { 1929 MVT VT = N->getValueType(0); 1930 DebugLoc dl = N->getDebugLoc(); 1931 1932 // Lower vector shifts on NEON to use VSHL. 1933 if (VT.isVector()) { 1934 assert(ST->hasNEON() && "unexpected vector shift"); 1935 1936 // Left shifts translate directly to the vshiftu intrinsic. 1937 if (N->getOpcode() == ISD::SHL) 1938 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT, 1939 DAG.getConstant(Intrinsic::arm_neon_vshiftu, MVT::i32), 1940 N->getOperand(0), N->getOperand(1)); 1941 1942 assert((N->getOpcode() == ISD::SRA || 1943 N->getOpcode() == ISD::SRL) && "unexpected vector shift opcode"); 1944 1945 // NEON uses the same intrinsics for both left and right shifts. For 1946 // right shifts, the shift amounts are negative, so negate the vector of 1947 // shift amounts. 1948 MVT ShiftVT = N->getOperand(1).getValueType(); 1949 SDValue NegatedCount = DAG.getNode(ISD::SUB, dl, ShiftVT, 1950 getZeroVector(ShiftVT, DAG, dl), 1951 N->getOperand(1)); 1952 Intrinsic::ID vshiftInt = (N->getOpcode() == ISD::SRA ? 1953 Intrinsic::arm_neon_vshifts : 1954 Intrinsic::arm_neon_vshiftu); 1955 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT, 1956 DAG.getConstant(vshiftInt, MVT::i32), 1957 N->getOperand(0), NegatedCount); 1958 } 1959 1960 assert(VT == MVT::i64 && 1961 (N->getOpcode() == ISD::SRL || N->getOpcode() == ISD::SRA) && 1962 "Unknown shift to lower!"); 1963 1964 // We only lower SRA, SRL of 1 here, all others use generic lowering. 1965 if (!isa<ConstantSDNode>(N->getOperand(1)) || 1966 cast<ConstantSDNode>(N->getOperand(1))->getZExtValue() != 1) 1967 return SDValue(); 1968 1969 // If we are in thumb mode, we don't have RRX. 1970 if (ST->isThumb()) return SDValue(); 1971 1972 // Okay, we have a 64-bit SRA or SRL of 1. Lower this to an RRX expr. 1973 SDValue Lo = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, N->getOperand(0), 1974 DAG.getConstant(0, MVT::i32)); 1975 SDValue Hi = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, N->getOperand(0), 1976 DAG.getConstant(1, MVT::i32)); 1977 1978 // First, build a SRA_FLAG/SRL_FLAG op, which shifts the top part by one and 1979 // captures the result into a carry flag. 1980 unsigned Opc = N->getOpcode() == ISD::SRL ? ARMISD::SRL_FLAG:ARMISD::SRA_FLAG; 1981 Hi = DAG.getNode(Opc, dl, DAG.getVTList(MVT::i32, MVT::Flag), &Hi, 1); 1982 1983 // The low part is an ARMISD::RRX operand, which shifts the carry in. 1984 Lo = DAG.getNode(ARMISD::RRX, dl, MVT::i32, Lo, Hi.getValue(1)); 1985 1986 // Merge the pieces into a single i64 value. 1987 return DAG.getNode(ISD::BUILD_PAIR, dl, MVT::i64, Lo, Hi); 1988} 1989 1990static SDValue LowerVSETCC(SDValue Op, SelectionDAG &DAG) { 1991 SDValue TmpOp0, TmpOp1; 1992 bool Invert = false; 1993 bool Swap = false; 1994 unsigned Opc = 0; 1995 1996 SDValue Op0 = Op.getOperand(0); 1997 SDValue Op1 = Op.getOperand(1); 1998 SDValue CC = Op.getOperand(2); 1999 MVT VT = Op.getValueType(); 2000 ISD::CondCode SetCCOpcode = cast<CondCodeSDNode>(CC)->get(); 2001 DebugLoc dl = Op.getDebugLoc(); 2002 2003 if (Op.getOperand(1).getValueType().isFloatingPoint()) { 2004 switch (SetCCOpcode) { 2005 default: assert(0 && "Illegal FP comparison"); break; 2006 case ISD::SETUNE: 2007 case ISD::SETNE: Invert = true; // Fallthrough 2008 case ISD::SETOEQ: 2009 case ISD::SETEQ: Opc = ARMISD::VCEQ; break; 2010 case ISD::SETOLT: 2011 case ISD::SETLT: Swap = true; // Fallthrough 2012 case ISD::SETOGT: 2013 case ISD::SETGT: Opc = ARMISD::VCGT; break; 2014 case ISD::SETOLE: 2015 case ISD::SETLE: Swap = true; // Fallthrough 2016 case ISD::SETOGE: 2017 case ISD::SETGE: Opc = ARMISD::VCGE; break; 2018 case ISD::SETUGE: Swap = true; // Fallthrough 2019 case ISD::SETULE: Invert = true; Opc = ARMISD::VCGT; break; 2020 case ISD::SETUGT: Swap = true; // Fallthrough 2021 case ISD::SETULT: Invert = true; Opc = ARMISD::VCGE; break; 2022 case ISD::SETUEQ: Invert = true; // Fallthrough 2023 case ISD::SETONE: 2024 // Expand this to (OLT | OGT). 2025 TmpOp0 = Op0; 2026 TmpOp1 = Op1; 2027 Opc = ISD::OR; 2028 Op0 = DAG.getNode(ARMISD::VCGT, dl, VT, TmpOp1, TmpOp0); 2029 Op1 = DAG.getNode(ARMISD::VCGT, dl, VT, TmpOp0, TmpOp1); 2030 break; 2031 case ISD::SETUO: Invert = true; // Fallthrough 2032 case ISD::SETO: 2033 // Expand this to (OLT | OGE). 2034 TmpOp0 = Op0; 2035 TmpOp1 = Op1; 2036 Opc = ISD::OR; 2037 Op0 = DAG.getNode(ARMISD::VCGT, dl, VT, TmpOp1, TmpOp0); 2038 Op1 = DAG.getNode(ARMISD::VCGE, dl, VT, TmpOp0, TmpOp1); 2039 break; 2040 } 2041 } else { 2042 // Integer comparisons. 2043 switch (SetCCOpcode) { 2044 default: assert(0 && "Illegal integer comparison"); break; 2045 case ISD::SETNE: Invert = true; 2046 case ISD::SETEQ: Opc = ARMISD::VCEQ; break; 2047 case ISD::SETLT: Swap = true; 2048 case ISD::SETGT: Opc = ARMISD::VCGT; break; 2049 case ISD::SETLE: Swap = true; 2050 case ISD::SETGE: Opc = ARMISD::VCGE; break; 2051 case ISD::SETULT: Swap = true; 2052 case ISD::SETUGT: Opc = ARMISD::VCGTU; break; 2053 case ISD::SETULE: Swap = true; 2054 case ISD::SETUGE: Opc = ARMISD::VCGEU; break; 2055 } 2056 2057 // Detect VTST (Vector Test Bits) = vicmp ne (and (op0, op1), zero). 2058 if (Opc == ARMISD::VCEQ) { 2059 2060 SDValue AndOp; 2061 if (ISD::isBuildVectorAllZeros(Op1.getNode())) 2062 AndOp = Op0; 2063 else if (ISD::isBuildVectorAllZeros(Op0.getNode())) 2064 AndOp = Op1; 2065 2066 // Ignore bitconvert. 2067 if (AndOp.getNode() && AndOp.getOpcode() == ISD::BIT_CONVERT) 2068 AndOp = AndOp.getOperand(0); 2069 2070 if (AndOp.getNode() && AndOp.getOpcode() == ISD::AND) { 2071 Opc = ARMISD::VTST; 2072 Op0 = DAG.getNode(ISD::BIT_CONVERT, dl, VT, AndOp.getOperand(0)); 2073 Op1 = DAG.getNode(ISD::BIT_CONVERT, dl, VT, AndOp.getOperand(1)); 2074 Invert = !Invert; 2075 } 2076 } 2077 } 2078 2079 if (Swap) 2080 std::swap(Op0, Op1); 2081 2082 SDValue Result = DAG.getNode(Opc, dl, VT, Op0, Op1); 2083 2084 if (Invert) 2085 Result = DAG.getNOT(dl, Result, VT); 2086 2087 return Result; 2088} 2089 2090/// isVMOVSplat - Check if the specified splat value corresponds to an immediate 2091/// VMOV instruction, and if so, return the constant being splatted. 2092static SDValue isVMOVSplat(uint64_t SplatBits, uint64_t SplatUndef, 2093 unsigned SplatBitSize, SelectionDAG &DAG) { 2094 switch (SplatBitSize) { 2095 case 8: 2096 // Any 1-byte value is OK. 2097 assert((SplatBits & ~0xff) == 0 && "one byte splat value is too big"); 2098 return DAG.getTargetConstant(SplatBits, MVT::i8); 2099 2100 case 16: 2101 // NEON's 16-bit VMOV supports splat values where only one byte is nonzero. 2102 if ((SplatBits & ~0xff) == 0 || 2103 (SplatBits & ~0xff00) == 0) 2104 return DAG.getTargetConstant(SplatBits, MVT::i16); 2105 break; 2106 2107 case 32: 2108 // NEON's 32-bit VMOV supports splat values where: 2109 // * only one byte is nonzero, or 2110 // * the least significant byte is 0xff and the second byte is nonzero, or 2111 // * the least significant 2 bytes are 0xff and the third is nonzero. 2112 if ((SplatBits & ~0xff) == 0 || 2113 (SplatBits & ~0xff00) == 0 || 2114 (SplatBits & ~0xff0000) == 0 || 2115 (SplatBits & ~0xff000000) == 0) 2116 return DAG.getTargetConstant(SplatBits, MVT::i32); 2117 2118 if ((SplatBits & ~0xffff) == 0 && 2119 ((SplatBits | SplatUndef) & 0xff) == 0xff) 2120 return DAG.getTargetConstant(SplatBits | 0xff, MVT::i32); 2121 2122 if ((SplatBits & ~0xffffff) == 0 && 2123 ((SplatBits | SplatUndef) & 0xffff) == 0xffff) 2124 return DAG.getTargetConstant(SplatBits | 0xffff, MVT::i32); 2125 2126 // Note: there are a few 32-bit splat values (specifically: 00ffff00, 2127 // ff000000, ff0000ff, and ffff00ff) that are valid for VMOV.I64 but not 2128 // VMOV.I32. A (very) minor optimization would be to replicate the value 2129 // and fall through here to test for a valid 64-bit splat. But, then the 2130 // caller would also need to check and handle the change in size. 2131 break; 2132 2133 case 64: { 2134 // NEON has a 64-bit VMOV splat where each byte is either 0 or 0xff. 2135 uint64_t BitMask = 0xff; 2136 uint64_t Val = 0; 2137 for (int ByteNum = 0; ByteNum < 8; ++ByteNum) { 2138 if (((SplatBits | SplatUndef) & BitMask) == BitMask) 2139 Val |= BitMask; 2140 else if ((SplatBits & BitMask) != 0) 2141 return SDValue(); 2142 BitMask <<= 8; 2143 } 2144 return DAG.getTargetConstant(Val, MVT::i64); 2145 } 2146 2147 default: 2148 assert(0 && "unexpected size for isVMOVSplat"); 2149 break; 2150 } 2151 2152 return SDValue(); 2153} 2154 2155/// getVMOVImm - If this is a build_vector of constants which can be 2156/// formed by using a VMOV instruction of the specified element size, 2157/// return the constant being splatted. The ByteSize field indicates the 2158/// number of bytes of each element [1248]. 2159SDValue ARM::getVMOVImm(SDNode *N, unsigned ByteSize, SelectionDAG &DAG) { 2160 BuildVectorSDNode *BVN = dyn_cast<BuildVectorSDNode>(N); 2161 APInt SplatBits, SplatUndef; 2162 unsigned SplatBitSize; 2163 bool HasAnyUndefs; 2164 if (! BVN || ! BVN->isConstantSplat(SplatBits, SplatUndef, SplatBitSize, 2165 HasAnyUndefs, ByteSize * 8)) 2166 return SDValue(); 2167 2168 if (SplatBitSize > ByteSize * 8) 2169 return SDValue(); 2170 2171 return isVMOVSplat(SplatBits.getZExtValue(), SplatUndef.getZExtValue(), 2172 SplatBitSize, DAG); 2173} 2174 2175static SDValue BuildSplat(SDValue Val, MVT VT, SelectionDAG &DAG, DebugLoc dl) { 2176 // Canonicalize all-zeros and all-ones vectors. 2177 ConstantSDNode *ConstVal = dyn_cast<ConstantSDNode>(Val.getNode()); 2178 if (ConstVal->isNullValue()) 2179 return getZeroVector(VT, DAG, dl); 2180 if (ConstVal->isAllOnesValue()) 2181 return getOnesVector(VT, DAG, dl); 2182 2183 MVT CanonicalVT; 2184 if (VT.is64BitVector()) { 2185 switch (Val.getValueType().getSizeInBits()) { 2186 case 8: CanonicalVT = MVT::v8i8; break; 2187 case 16: CanonicalVT = MVT::v4i16; break; 2188 case 32: CanonicalVT = MVT::v2i32; break; 2189 case 64: CanonicalVT = MVT::v1i64; break; 2190 default: assert(0 && "unexpected splat element type"); break; 2191 } 2192 } else { 2193 assert(VT.is128BitVector() && "unknown splat vector size"); 2194 switch (Val.getValueType().getSizeInBits()) { 2195 case 8: CanonicalVT = MVT::v16i8; break; 2196 case 16: CanonicalVT = MVT::v8i16; break; 2197 case 32: CanonicalVT = MVT::v4i32; break; 2198 case 64: CanonicalVT = MVT::v2i64; break; 2199 default: assert(0 && "unexpected splat element type"); break; 2200 } 2201 } 2202 2203 // Build a canonical splat for this value. 2204 SmallVector<SDValue, 8> Ops; 2205 Ops.assign(CanonicalVT.getVectorNumElements(), Val); 2206 SDValue Res = DAG.getNode(ISD::BUILD_VECTOR, dl, CanonicalVT, &Ops[0], 2207 Ops.size()); 2208 return DAG.getNode(ISD::BIT_CONVERT, dl, VT, Res); 2209} 2210 2211// If this is a case we can't handle, return null and let the default 2212// expansion code take care of it. 2213static SDValue LowerBUILD_VECTOR(SDValue Op, SelectionDAG &DAG) { 2214 BuildVectorSDNode *BVN = dyn_cast<BuildVectorSDNode>(Op.getNode()); 2215 assert(BVN != 0 && "Expected a BuildVectorSDNode in LowerBUILD_VECTOR"); 2216 DebugLoc dl = Op.getDebugLoc(); 2217 2218 APInt SplatBits, SplatUndef; 2219 unsigned SplatBitSize; 2220 bool HasAnyUndefs; 2221 if (BVN->isConstantSplat(SplatBits, SplatUndef, SplatBitSize, HasAnyUndefs)) { 2222 SDValue Val = isVMOVSplat(SplatBits.getZExtValue(), 2223 SplatUndef.getZExtValue(), SplatBitSize, DAG); 2224 if (Val.getNode()) 2225 return BuildSplat(Val, Op.getValueType(), DAG, dl); 2226 } 2227 2228 return SDValue(); 2229} 2230 2231static SDValue LowerVECTOR_SHUFFLE(SDValue Op, SelectionDAG &DAG) { 2232 return Op; 2233} 2234 2235static SDValue LowerSCALAR_TO_VECTOR(SDValue Op, SelectionDAG &DAG) { 2236 return Op; 2237} 2238 2239static SDValue LowerEXTRACT_VECTOR_ELT(SDValue Op, SelectionDAG &DAG) { 2240 MVT VT = Op.getValueType(); 2241 DebugLoc dl = Op.getDebugLoc(); 2242 assert((VT == MVT::i8 || VT == MVT::i16) && 2243 "unexpected type for custom-lowering vector extract"); 2244 SDValue Vec = Op.getOperand(0); 2245 SDValue Lane = Op.getOperand(1); 2246 Op = DAG.getNode(ARMISD::VGETLANEu, dl, MVT::i32, Vec, Lane); 2247 Op = DAG.getNode(ISD::AssertZext, dl, MVT::i32, Op, DAG.getValueType(VT)); 2248 return DAG.getNode(ISD::TRUNCATE, dl, VT, Op); 2249} 2250 2251static SDValue LowerCONCAT_VECTORS(SDValue Op) { 2252 if (Op.getValueType().is128BitVector() && Op.getNumOperands() == 2) 2253 return Op; 2254 return SDValue(); 2255} 2256 2257SDValue ARMTargetLowering::LowerOperation(SDValue Op, SelectionDAG &DAG) { 2258 switch (Op.getOpcode()) { 2259 default: assert(0 && "Don't know how to custom lower this!"); abort(); 2260 case ISD::ConstantPool: return LowerConstantPool(Op, DAG); 2261 case ISD::GlobalAddress: 2262 return Subtarget->isTargetDarwin() ? LowerGlobalAddressDarwin(Op, DAG) : 2263 LowerGlobalAddressELF(Op, DAG); 2264 case ISD::GlobalTLSAddress: return LowerGlobalTLSAddress(Op, DAG); 2265 case ISD::CALL: return LowerCALL(Op, DAG); 2266 case ISD::RET: return LowerRET(Op, DAG); 2267 case ISD::SELECT_CC: return LowerSELECT_CC(Op, DAG, Subtarget); 2268 case ISD::BR_CC: return LowerBR_CC(Op, DAG, Subtarget); 2269 case ISD::BR_JT: return LowerBR_JT(Op, DAG); 2270 case ISD::VASTART: return LowerVASTART(Op, DAG, VarArgsFrameIndex); 2271 case ISD::SINT_TO_FP: 2272 case ISD::UINT_TO_FP: return LowerINT_TO_FP(Op, DAG); 2273 case ISD::FP_TO_SINT: 2274 case ISD::FP_TO_UINT: return LowerFP_TO_INT(Op, DAG); 2275 case ISD::FCOPYSIGN: return LowerFCOPYSIGN(Op, DAG); 2276 case ISD::FORMAL_ARGUMENTS: return LowerFORMAL_ARGUMENTS(Op, DAG); 2277 case ISD::RETURNADDR: break; 2278 case ISD::FRAMEADDR: return LowerFRAMEADDR(Op, DAG); 2279 case ISD::GLOBAL_OFFSET_TABLE: return LowerGLOBAL_OFFSET_TABLE(Op, DAG); 2280 case ISD::INTRINSIC_WO_CHAIN: return LowerINTRINSIC_WO_CHAIN(Op, DAG); 2281 case ISD::BIT_CONVERT: return ExpandBIT_CONVERT(Op.getNode(), DAG); 2282 case ISD::SHL: 2283 case ISD::SRL: 2284 case ISD::SRA: return LowerShift(Op.getNode(), DAG, Subtarget); 2285 case ISD::VSETCC: return LowerVSETCC(Op, DAG); 2286 case ISD::BUILD_VECTOR: return LowerBUILD_VECTOR(Op, DAG); 2287 case ISD::VECTOR_SHUFFLE: return LowerVECTOR_SHUFFLE(Op, DAG); 2288 case ISD::SCALAR_TO_VECTOR: return LowerSCALAR_TO_VECTOR(Op, DAG); 2289 case ISD::EXTRACT_VECTOR_ELT: return LowerEXTRACT_VECTOR_ELT(Op, DAG); 2290 case ISD::CONCAT_VECTORS: return LowerCONCAT_VECTORS(Op); 2291 } 2292 return SDValue(); 2293} 2294 2295/// ReplaceNodeResults - Replace the results of node with an illegal result 2296/// type with new values built out of custom code. 2297void ARMTargetLowering::ReplaceNodeResults(SDNode *N, 2298 SmallVectorImpl<SDValue>&Results, 2299 SelectionDAG &DAG) { 2300 switch (N->getOpcode()) { 2301 default: 2302 assert(0 && "Don't know how to custom expand this!"); 2303 return; 2304 case ISD::BIT_CONVERT: 2305 Results.push_back(ExpandBIT_CONVERT(N, DAG)); 2306 return; 2307 case ISD::SRL: 2308 case ISD::SRA: { 2309 SDValue Res = LowerShift(N, DAG, Subtarget); 2310 if (Res.getNode()) 2311 Results.push_back(Res); 2312 return; 2313 } 2314 } 2315} 2316 2317//===----------------------------------------------------------------------===// 2318// ARM Scheduler Hooks 2319//===----------------------------------------------------------------------===// 2320 2321MachineBasicBlock * 2322ARMTargetLowering::EmitInstrWithCustomInserter(MachineInstr *MI, 2323 MachineBasicBlock *BB) const { 2324 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo(); 2325 DebugLoc dl = MI->getDebugLoc(); 2326 switch (MI->getOpcode()) { 2327 default: assert(false && "Unexpected instr type to insert"); 2328 case ARM::tMOVCCr: { 2329 // To "insert" a SELECT_CC instruction, we actually have to insert the 2330 // diamond control-flow pattern. The incoming instruction knows the 2331 // destination vreg to set, the condition code register to branch on, the 2332 // true/false values to select between, and a branch opcode to use. 2333 const BasicBlock *LLVM_BB = BB->getBasicBlock(); 2334 MachineFunction::iterator It = BB; 2335 ++It; 2336 2337 // thisMBB: 2338 // ... 2339 // TrueVal = ... 2340 // cmpTY ccX, r1, r2 2341 // bCC copy1MBB 2342 // fallthrough --> copy0MBB 2343 MachineBasicBlock *thisMBB = BB; 2344 MachineFunction *F = BB->getParent(); 2345 MachineBasicBlock *copy0MBB = F->CreateMachineBasicBlock(LLVM_BB); 2346 MachineBasicBlock *sinkMBB = F->CreateMachineBasicBlock(LLVM_BB); 2347 BuildMI(BB, dl, TII->get(ARM::tBcc)).addMBB(sinkMBB) 2348 .addImm(MI->getOperand(3).getImm()).addReg(MI->getOperand(4).getReg()); 2349 F->insert(It, copy0MBB); 2350 F->insert(It, sinkMBB); 2351 // Update machine-CFG edges by first adding all successors of the current 2352 // block to the new block which will contain the Phi node for the select. 2353 for(MachineBasicBlock::succ_iterator i = BB->succ_begin(), 2354 e = BB->succ_end(); i != e; ++i) 2355 sinkMBB->addSuccessor(*i); 2356 // Next, remove all successors of the current block, and add the true 2357 // and fallthrough blocks as its successors. 2358 while(!BB->succ_empty()) 2359 BB->removeSuccessor(BB->succ_begin()); 2360 BB->addSuccessor(copy0MBB); 2361 BB->addSuccessor(sinkMBB); 2362 2363 // copy0MBB: 2364 // %FalseValue = ... 2365 // # fallthrough to sinkMBB 2366 BB = copy0MBB; 2367 2368 // Update machine-CFG edges 2369 BB->addSuccessor(sinkMBB); 2370 2371 // sinkMBB: 2372 // %Result = phi [ %FalseValue, copy0MBB ], [ %TrueValue, thisMBB ] 2373 // ... 2374 BB = sinkMBB; 2375 BuildMI(BB, dl, TII->get(ARM::PHI), MI->getOperand(0).getReg()) 2376 .addReg(MI->getOperand(1).getReg()).addMBB(copy0MBB) 2377 .addReg(MI->getOperand(2).getReg()).addMBB(thisMBB); 2378 2379 F->DeleteMachineInstr(MI); // The pseudo instruction is gone now. 2380 return BB; 2381 } 2382 } 2383} 2384 2385//===----------------------------------------------------------------------===// 2386// ARM Optimization Hooks 2387//===----------------------------------------------------------------------===// 2388 2389static 2390SDValue combineSelectAndUse(SDNode *N, SDValue Slct, SDValue OtherOp, 2391 TargetLowering::DAGCombinerInfo &DCI) { 2392 SelectionDAG &DAG = DCI.DAG; 2393 const TargetLowering &TLI = DAG.getTargetLoweringInfo(); 2394 MVT VT = N->getValueType(0); 2395 unsigned Opc = N->getOpcode(); 2396 bool isSlctCC = Slct.getOpcode() == ISD::SELECT_CC; 2397 SDValue LHS = isSlctCC ? Slct.getOperand(2) : Slct.getOperand(1); 2398 SDValue RHS = isSlctCC ? Slct.getOperand(3) : Slct.getOperand(2); 2399 ISD::CondCode CC = ISD::SETCC_INVALID; 2400 2401 if (isSlctCC) { 2402 CC = cast<CondCodeSDNode>(Slct.getOperand(4))->get(); 2403 } else { 2404 SDValue CCOp = Slct.getOperand(0); 2405 if (CCOp.getOpcode() == ISD::SETCC) 2406 CC = cast<CondCodeSDNode>(CCOp.getOperand(2))->get(); 2407 } 2408 2409 bool DoXform = false; 2410 bool InvCC = false; 2411 assert ((Opc == ISD::ADD || (Opc == ISD::SUB && Slct == N->getOperand(1))) && 2412 "Bad input!"); 2413 2414 if (LHS.getOpcode() == ISD::Constant && 2415 cast<ConstantSDNode>(LHS)->isNullValue()) { 2416 DoXform = true; 2417 } else if (CC != ISD::SETCC_INVALID && 2418 RHS.getOpcode() == ISD::Constant && 2419 cast<ConstantSDNode>(RHS)->isNullValue()) { 2420 std::swap(LHS, RHS); 2421 SDValue Op0 = Slct.getOperand(0); 2422 MVT OpVT = isSlctCC ? Op0.getValueType() : 2423 Op0.getOperand(0).getValueType(); 2424 bool isInt = OpVT.isInteger(); 2425 CC = ISD::getSetCCInverse(CC, isInt); 2426 2427 if (!TLI.isCondCodeLegal(CC, OpVT)) 2428 return SDValue(); // Inverse operator isn't legal. 2429 2430 DoXform = true; 2431 InvCC = true; 2432 } 2433 2434 if (DoXform) { 2435 SDValue Result = DAG.getNode(Opc, RHS.getDebugLoc(), VT, OtherOp, RHS); 2436 if (isSlctCC) 2437 return DAG.getSelectCC(N->getDebugLoc(), OtherOp, Result, 2438 Slct.getOperand(0), Slct.getOperand(1), CC); 2439 SDValue CCOp = Slct.getOperand(0); 2440 if (InvCC) 2441 CCOp = DAG.getSetCC(Slct.getDebugLoc(), CCOp.getValueType(), 2442 CCOp.getOperand(0), CCOp.getOperand(1), CC); 2443 return DAG.getNode(ISD::SELECT, N->getDebugLoc(), VT, 2444 CCOp, OtherOp, Result); 2445 } 2446 return SDValue(); 2447} 2448 2449/// PerformADDCombine - Target-specific dag combine xforms for ISD::ADD. 2450static SDValue PerformADDCombine(SDNode *N, 2451 TargetLowering::DAGCombinerInfo &DCI) { 2452 // added by evan in r37685 with no testcase. 2453 SDValue N0 = N->getOperand(0), N1 = N->getOperand(1); 2454 2455 // fold (add (select cc, 0, c), x) -> (select cc, x, (add, x, c)) 2456 if (N0.getOpcode() == ISD::SELECT && N0.getNode()->hasOneUse()) { 2457 SDValue Result = combineSelectAndUse(N, N0, N1, DCI); 2458 if (Result.getNode()) return Result; 2459 } 2460 if (N1.getOpcode() == ISD::SELECT && N1.getNode()->hasOneUse()) { 2461 SDValue Result = combineSelectAndUse(N, N1, N0, DCI); 2462 if (Result.getNode()) return Result; 2463 } 2464 2465 return SDValue(); 2466} 2467 2468/// PerformSUBCombine - Target-specific dag combine xforms for ISD::SUB. 2469static SDValue PerformSUBCombine(SDNode *N, 2470 TargetLowering::DAGCombinerInfo &DCI) { 2471 // added by evan in r37685 with no testcase. 2472 SDValue N0 = N->getOperand(0), N1 = N->getOperand(1); 2473 2474 // fold (sub x, (select cc, 0, c)) -> (select cc, x, (sub, x, c)) 2475 if (N1.getOpcode() == ISD::SELECT && N1.getNode()->hasOneUse()) { 2476 SDValue Result = combineSelectAndUse(N, N1, N0, DCI); 2477 if (Result.getNode()) return Result; 2478 } 2479 2480 return SDValue(); 2481} 2482 2483 2484/// PerformFMRRDCombine - Target-specific dag combine xforms for ARMISD::FMRRD. 2485static SDValue PerformFMRRDCombine(SDNode *N, 2486 TargetLowering::DAGCombinerInfo &DCI) { 2487 // fmrrd(fmdrr x, y) -> x,y 2488 SDValue InDouble = N->getOperand(0); 2489 if (InDouble.getOpcode() == ARMISD::FMDRR) 2490 return DCI.CombineTo(N, InDouble.getOperand(0), InDouble.getOperand(1)); 2491 return SDValue(); 2492} 2493 2494/// getVShiftImm - Check if this is a valid build_vector for the immediate 2495/// operand of a vector shift operation, where all the elements of the 2496/// build_vector must have the same constant integer value. 2497static bool getVShiftImm(SDValue Op, unsigned ElementBits, int64_t &Cnt) { 2498 // Ignore bit_converts. 2499 while (Op.getOpcode() == ISD::BIT_CONVERT) 2500 Op = Op.getOperand(0); 2501 BuildVectorSDNode *BVN = dyn_cast<BuildVectorSDNode>(Op.getNode()); 2502 APInt SplatBits, SplatUndef; 2503 unsigned SplatBitSize; 2504 bool HasAnyUndefs; 2505 if (! BVN || ! BVN->isConstantSplat(SplatBits, SplatUndef, SplatBitSize, 2506 HasAnyUndefs, ElementBits) || 2507 SplatBitSize > ElementBits) 2508 return false; 2509 Cnt = SplatBits.getSExtValue(); 2510 return true; 2511} 2512 2513/// isVShiftLImm - Check if this is a valid build_vector for the immediate 2514/// operand of a vector shift left operation. That value must be in the range: 2515/// 0 <= Value < ElementBits for a left shift; or 2516/// 0 <= Value <= ElementBits for a long left shift. 2517static bool isVShiftLImm(SDValue Op, MVT VT, bool isLong, int64_t &Cnt) { 2518 assert(VT.isVector() && "vector shift count is not a vector type"); 2519 unsigned ElementBits = VT.getVectorElementType().getSizeInBits(); 2520 if (! getVShiftImm(Op, ElementBits, Cnt)) 2521 return false; 2522 return (Cnt >= 0 && (isLong ? Cnt-1 : Cnt) < ElementBits); 2523} 2524 2525/// isVShiftRImm - Check if this is a valid build_vector for the immediate 2526/// operand of a vector shift right operation. For a shift opcode, the value 2527/// is positive, but for an intrinsic the value count must be negative. The 2528/// absolute value must be in the range: 2529/// 1 <= |Value| <= ElementBits for a right shift; or 2530/// 1 <= |Value| <= ElementBits/2 for a narrow right shift. 2531static bool isVShiftRImm(SDValue Op, MVT VT, bool isNarrow, bool isIntrinsic, 2532 int64_t &Cnt) { 2533 assert(VT.isVector() && "vector shift count is not a vector type"); 2534 unsigned ElementBits = VT.getVectorElementType().getSizeInBits(); 2535 if (! getVShiftImm(Op, ElementBits, Cnt)) 2536 return false; 2537 if (isIntrinsic) 2538 Cnt = -Cnt; 2539 return (Cnt >= 1 && Cnt <= (isNarrow ? ElementBits/2 : ElementBits)); 2540} 2541 2542/// PerformIntrinsicCombine - ARM-specific DAG combining for intrinsics. 2543static SDValue PerformIntrinsicCombine(SDNode *N, SelectionDAG &DAG) { 2544 unsigned IntNo = cast<ConstantSDNode>(N->getOperand(0))->getZExtValue(); 2545 switch (IntNo) { 2546 default: 2547 // Don't do anything for most intrinsics. 2548 break; 2549 2550 // Vector shifts: check for immediate versions and lower them. 2551 // Note: This is done during DAG combining instead of DAG legalizing because 2552 // the build_vectors for 64-bit vector element shift counts are generally 2553 // not legal, and it is hard to see their values after they get legalized to 2554 // loads from a constant pool. 2555 case Intrinsic::arm_neon_vshifts: 2556 case Intrinsic::arm_neon_vshiftu: 2557 case Intrinsic::arm_neon_vshiftls: 2558 case Intrinsic::arm_neon_vshiftlu: 2559 case Intrinsic::arm_neon_vshiftn: 2560 case Intrinsic::arm_neon_vrshifts: 2561 case Intrinsic::arm_neon_vrshiftu: 2562 case Intrinsic::arm_neon_vrshiftn: 2563 case Intrinsic::arm_neon_vqshifts: 2564 case Intrinsic::arm_neon_vqshiftu: 2565 case Intrinsic::arm_neon_vqshiftsu: 2566 case Intrinsic::arm_neon_vqshiftns: 2567 case Intrinsic::arm_neon_vqshiftnu: 2568 case Intrinsic::arm_neon_vqshiftnsu: 2569 case Intrinsic::arm_neon_vqrshiftns: 2570 case Intrinsic::arm_neon_vqrshiftnu: 2571 case Intrinsic::arm_neon_vqrshiftnsu: { 2572 MVT VT = N->getOperand(1).getValueType(); 2573 int64_t Cnt; 2574 unsigned VShiftOpc = 0; 2575 2576 switch (IntNo) { 2577 case Intrinsic::arm_neon_vshifts: 2578 case Intrinsic::arm_neon_vshiftu: 2579 if (isVShiftLImm(N->getOperand(2), VT, false, Cnt)) { 2580 VShiftOpc = ARMISD::VSHL; 2581 break; 2582 } 2583 if (isVShiftRImm(N->getOperand(2), VT, false, true, Cnt)) { 2584 VShiftOpc = (IntNo == Intrinsic::arm_neon_vshifts ? 2585 ARMISD::VSHRs : ARMISD::VSHRu); 2586 break; 2587 } 2588 return SDValue(); 2589 2590 case Intrinsic::arm_neon_vshiftls: 2591 case Intrinsic::arm_neon_vshiftlu: 2592 if (isVShiftLImm(N->getOperand(2), VT, true, Cnt)) 2593 break; 2594 assert(0 && "invalid shift count for vshll intrinsic"); 2595 abort(); 2596 2597 case Intrinsic::arm_neon_vrshifts: 2598 case Intrinsic::arm_neon_vrshiftu: 2599 if (isVShiftRImm(N->getOperand(2), VT, false, true, Cnt)) 2600 break; 2601 return SDValue(); 2602 2603 case Intrinsic::arm_neon_vqshifts: 2604 case Intrinsic::arm_neon_vqshiftu: 2605 if (isVShiftLImm(N->getOperand(2), VT, false, Cnt)) 2606 break; 2607 return SDValue(); 2608 2609 case Intrinsic::arm_neon_vqshiftsu: 2610 if (isVShiftLImm(N->getOperand(2), VT, false, Cnt)) 2611 break; 2612 assert(0 && "invalid shift count for vqshlu intrinsic"); 2613 abort(); 2614 2615 case Intrinsic::arm_neon_vshiftn: 2616 case Intrinsic::arm_neon_vrshiftn: 2617 case Intrinsic::arm_neon_vqshiftns: 2618 case Intrinsic::arm_neon_vqshiftnu: 2619 case Intrinsic::arm_neon_vqshiftnsu: 2620 case Intrinsic::arm_neon_vqrshiftns: 2621 case Intrinsic::arm_neon_vqrshiftnu: 2622 case Intrinsic::arm_neon_vqrshiftnsu: 2623 // Narrowing shifts require an immediate right shift. 2624 if (isVShiftRImm(N->getOperand(2), VT, true, true, Cnt)) 2625 break; 2626 assert(0 && "invalid shift count for narrowing vector shift intrinsic"); 2627 abort(); 2628 2629 default: 2630 assert(0 && "unhandled vector shift"); 2631 } 2632 2633 switch (IntNo) { 2634 case Intrinsic::arm_neon_vshifts: 2635 case Intrinsic::arm_neon_vshiftu: 2636 // Opcode already set above. 2637 break; 2638 case Intrinsic::arm_neon_vshiftls: 2639 case Intrinsic::arm_neon_vshiftlu: 2640 if (Cnt == VT.getVectorElementType().getSizeInBits()) 2641 VShiftOpc = ARMISD::VSHLLi; 2642 else 2643 VShiftOpc = (IntNo == Intrinsic::arm_neon_vshiftls ? 2644 ARMISD::VSHLLs : ARMISD::VSHLLu); 2645 break; 2646 case Intrinsic::arm_neon_vshiftn: 2647 VShiftOpc = ARMISD::VSHRN; break; 2648 case Intrinsic::arm_neon_vrshifts: 2649 VShiftOpc = ARMISD::VRSHRs; break; 2650 case Intrinsic::arm_neon_vrshiftu: 2651 VShiftOpc = ARMISD::VRSHRu; break; 2652 case Intrinsic::arm_neon_vrshiftn: 2653 VShiftOpc = ARMISD::VRSHRN; break; 2654 case Intrinsic::arm_neon_vqshifts: 2655 VShiftOpc = ARMISD::VQSHLs; break; 2656 case Intrinsic::arm_neon_vqshiftu: 2657 VShiftOpc = ARMISD::VQSHLu; break; 2658 case Intrinsic::arm_neon_vqshiftsu: 2659 VShiftOpc = ARMISD::VQSHLsu; break; 2660 case Intrinsic::arm_neon_vqshiftns: 2661 VShiftOpc = ARMISD::VQSHRNs; break; 2662 case Intrinsic::arm_neon_vqshiftnu: 2663 VShiftOpc = ARMISD::VQSHRNu; break; 2664 case Intrinsic::arm_neon_vqshiftnsu: 2665 VShiftOpc = ARMISD::VQSHRNsu; break; 2666 case Intrinsic::arm_neon_vqrshiftns: 2667 VShiftOpc = ARMISD::VQRSHRNs; break; 2668 case Intrinsic::arm_neon_vqrshiftnu: 2669 VShiftOpc = ARMISD::VQRSHRNu; break; 2670 case Intrinsic::arm_neon_vqrshiftnsu: 2671 VShiftOpc = ARMISD::VQRSHRNsu; break; 2672 } 2673 2674 return DAG.getNode(VShiftOpc, N->getDebugLoc(), N->getValueType(0), 2675 N->getOperand(1), DAG.getConstant(Cnt, MVT::i32)); 2676 } 2677 2678 case Intrinsic::arm_neon_vshiftins: { 2679 MVT VT = N->getOperand(1).getValueType(); 2680 int64_t Cnt; 2681 unsigned VShiftOpc = 0; 2682 2683 if (isVShiftLImm(N->getOperand(3), VT, false, Cnt)) 2684 VShiftOpc = ARMISD::VSLI; 2685 else if (isVShiftRImm(N->getOperand(3), VT, false, true, Cnt)) 2686 VShiftOpc = ARMISD::VSRI; 2687 else { 2688 assert(0 && "invalid shift count for vsli/vsri intrinsic"); 2689 abort(); 2690 } 2691 2692 return DAG.getNode(VShiftOpc, N->getDebugLoc(), N->getValueType(0), 2693 N->getOperand(1), N->getOperand(2), 2694 DAG.getConstant(Cnt, MVT::i32)); 2695 } 2696 2697 case Intrinsic::arm_neon_vqrshifts: 2698 case Intrinsic::arm_neon_vqrshiftu: 2699 // No immediate versions of these to check for. 2700 break; 2701 } 2702 2703 return SDValue(); 2704} 2705 2706/// PerformShiftCombine - Checks for immediate versions of vector shifts and 2707/// lowers them. As with the vector shift intrinsics, this is done during DAG 2708/// combining instead of DAG legalizing because the build_vectors for 64-bit 2709/// vector element shift counts are generally not legal, and it is hard to see 2710/// their values after they get legalized to loads from a constant pool. 2711static SDValue PerformShiftCombine(SDNode *N, SelectionDAG &DAG, 2712 const ARMSubtarget *ST) { 2713 MVT VT = N->getValueType(0); 2714 2715 // Nothing to be done for scalar shifts. 2716 if (! VT.isVector()) 2717 return SDValue(); 2718 2719 assert(ST->hasNEON() && "unexpected vector shift"); 2720 int64_t Cnt; 2721 2722 switch (N->getOpcode()) { 2723 default: assert(0 && "unexpected shift opcode"); 2724 2725 case ISD::SHL: 2726 if (isVShiftLImm(N->getOperand(1), VT, false, Cnt)) 2727 return DAG.getNode(ARMISD::VSHL, N->getDebugLoc(), VT, N->getOperand(0), 2728 DAG.getConstant(Cnt, MVT::i32)); 2729 break; 2730 2731 case ISD::SRA: 2732 case ISD::SRL: 2733 if (isVShiftRImm(N->getOperand(1), VT, false, false, Cnt)) { 2734 unsigned VShiftOpc = (N->getOpcode() == ISD::SRA ? 2735 ARMISD::VSHRs : ARMISD::VSHRu); 2736 return DAG.getNode(VShiftOpc, N->getDebugLoc(), VT, N->getOperand(0), 2737 DAG.getConstant(Cnt, MVT::i32)); 2738 } 2739 } 2740 return SDValue(); 2741} 2742 2743/// PerformExtendCombine - Target-specific DAG combining for ISD::SIGN_EXTEND, 2744/// ISD::ZERO_EXTEND, and ISD::ANY_EXTEND. 2745static SDValue PerformExtendCombine(SDNode *N, SelectionDAG &DAG, 2746 const ARMSubtarget *ST) { 2747 SDValue N0 = N->getOperand(0); 2748 2749 // Check for sign- and zero-extensions of vector extract operations of 8- 2750 // and 16-bit vector elements. NEON supports these directly. They are 2751 // handled during DAG combining because type legalization will promote them 2752 // to 32-bit types and it is messy to recognize the operations after that. 2753 if (ST->hasNEON() && N0.getOpcode() == ISD::EXTRACT_VECTOR_ELT) { 2754 SDValue Vec = N0.getOperand(0); 2755 SDValue Lane = N0.getOperand(1); 2756 MVT VT = N->getValueType(0); 2757 MVT EltVT = N0.getValueType(); 2758 const TargetLowering &TLI = DAG.getTargetLoweringInfo(); 2759 2760 if (VT == MVT::i32 && 2761 (EltVT == MVT::i8 || EltVT == MVT::i16) && 2762 TLI.isTypeLegal(Vec.getValueType())) { 2763 2764 unsigned Opc = 0; 2765 switch (N->getOpcode()) { 2766 default: assert(0 && "unexpected opcode"); 2767 case ISD::SIGN_EXTEND: 2768 Opc = ARMISD::VGETLANEs; 2769 break; 2770 case ISD::ZERO_EXTEND: 2771 case ISD::ANY_EXTEND: 2772 Opc = ARMISD::VGETLANEu; 2773 break; 2774 } 2775 return DAG.getNode(Opc, N->getDebugLoc(), VT, Vec, Lane); 2776 } 2777 } 2778 2779 return SDValue(); 2780} 2781 2782SDValue ARMTargetLowering::PerformDAGCombine(SDNode *N, 2783 DAGCombinerInfo &DCI) const { 2784 switch (N->getOpcode()) { 2785 default: break; 2786 case ISD::ADD: return PerformADDCombine(N, DCI); 2787 case ISD::SUB: return PerformSUBCombine(N, DCI); 2788 case ARMISD::FMRRD: return PerformFMRRDCombine(N, DCI); 2789 case ISD::INTRINSIC_WO_CHAIN: 2790 return PerformIntrinsicCombine(N, DCI.DAG); 2791 case ISD::SHL: 2792 case ISD::SRA: 2793 case ISD::SRL: 2794 return PerformShiftCombine(N, DCI.DAG, Subtarget); 2795 case ISD::SIGN_EXTEND: 2796 case ISD::ZERO_EXTEND: 2797 case ISD::ANY_EXTEND: 2798 return PerformExtendCombine(N, DCI.DAG, Subtarget); 2799 } 2800 return SDValue(); 2801} 2802 2803/// isLegalAddressImmediate - Return true if the integer value can be used 2804/// as the offset of the target addressing mode for load / store of the 2805/// given type. 2806static bool isLegalAddressImmediate(int64_t V, MVT VT, 2807 const ARMSubtarget *Subtarget) { 2808 if (V == 0) 2809 return true; 2810 2811 if (!VT.isSimple()) 2812 return false; 2813 2814 if (Subtarget->isThumb()) { 2815 if (V < 0) 2816 return false; 2817 2818 unsigned Scale = 1; 2819 switch (VT.getSimpleVT()) { 2820 default: return false; 2821 case MVT::i1: 2822 case MVT::i8: 2823 // Scale == 1; 2824 break; 2825 case MVT::i16: 2826 // Scale == 2; 2827 Scale = 2; 2828 break; 2829 case MVT::i32: 2830 // Scale == 4; 2831 Scale = 4; 2832 break; 2833 } 2834 2835 if ((V & (Scale - 1)) != 0) 2836 return false; 2837 V /= Scale; 2838 return V == (V & ((1LL << 5) - 1)); 2839 } 2840 2841 if (V < 0) 2842 V = - V; 2843 switch (VT.getSimpleVT()) { 2844 default: return false; 2845 case MVT::i1: 2846 case MVT::i8: 2847 case MVT::i32: 2848 // +- imm12 2849 return V == (V & ((1LL << 12) - 1)); 2850 case MVT::i16: 2851 // +- imm8 2852 return V == (V & ((1LL << 8) - 1)); 2853 case MVT::f32: 2854 case MVT::f64: 2855 if (!Subtarget->hasVFP2()) 2856 return false; 2857 if ((V & 3) != 0) 2858 return false; 2859 V >>= 2; 2860 return V == (V & ((1LL << 8) - 1)); 2861 } 2862} 2863 2864/// isLegalAddressingMode - Return true if the addressing mode represented 2865/// by AM is legal for this target, for a load/store of the specified type. 2866bool ARMTargetLowering::isLegalAddressingMode(const AddrMode &AM, 2867 const Type *Ty) const { 2868 MVT VT = getValueType(Ty, true); 2869 if (!isLegalAddressImmediate(AM.BaseOffs, VT, Subtarget)) 2870 return false; 2871 2872 // Can never fold addr of global into load/store. 2873 if (AM.BaseGV) 2874 return false; 2875 2876 switch (AM.Scale) { 2877 case 0: // no scale reg, must be "r+i" or "r", or "i". 2878 break; 2879 case 1: 2880 if (Subtarget->isThumb()) 2881 return false; 2882 // FALL THROUGH. 2883 default: 2884 // ARM doesn't support any R+R*scale+imm addr modes. 2885 if (AM.BaseOffs) 2886 return false; 2887 2888 if (!VT.isSimple()) 2889 return false; 2890 2891 int Scale = AM.Scale; 2892 switch (VT.getSimpleVT()) { 2893 default: return false; 2894 case MVT::i1: 2895 case MVT::i8: 2896 case MVT::i32: 2897 case MVT::i64: 2898 // This assumes i64 is legalized to a pair of i32. If not (i.e. 2899 // ldrd / strd are used, then its address mode is same as i16. 2900 // r + r 2901 if (Scale < 0) Scale = -Scale; 2902 if (Scale == 1) 2903 return true; 2904 // r + r << imm 2905 return isPowerOf2_32(Scale & ~1); 2906 case MVT::i16: 2907 // r + r 2908 if (((unsigned)AM.HasBaseReg + Scale) <= 2) 2909 return true; 2910 return false; 2911 2912 case MVT::isVoid: 2913 // Note, we allow "void" uses (basically, uses that aren't loads or 2914 // stores), because arm allows folding a scale into many arithmetic 2915 // operations. This should be made more precise and revisited later. 2916 2917 // Allow r << imm, but the imm has to be a multiple of two. 2918 if (AM.Scale & 1) return false; 2919 return isPowerOf2_32(AM.Scale); 2920 } 2921 break; 2922 } 2923 return true; 2924} 2925 2926static bool getIndexedAddressParts(SDNode *Ptr, MVT VT, 2927 bool isSEXTLoad, SDValue &Base, 2928 SDValue &Offset, bool &isInc, 2929 SelectionDAG &DAG) { 2930 if (Ptr->getOpcode() != ISD::ADD && Ptr->getOpcode() != ISD::SUB) 2931 return false; 2932 2933 if (VT == MVT::i16 || ((VT == MVT::i8 || VT == MVT::i1) && isSEXTLoad)) { 2934 // AddressingMode 3 2935 Base = Ptr->getOperand(0); 2936 if (ConstantSDNode *RHS = dyn_cast<ConstantSDNode>(Ptr->getOperand(1))) { 2937 int RHSC = (int)RHS->getZExtValue(); 2938 if (RHSC < 0 && RHSC > -256) { 2939 isInc = false; 2940 Offset = DAG.getConstant(-RHSC, RHS->getValueType(0)); 2941 return true; 2942 } 2943 } 2944 isInc = (Ptr->getOpcode() == ISD::ADD); 2945 Offset = Ptr->getOperand(1); 2946 return true; 2947 } else if (VT == MVT::i32 || VT == MVT::i8 || VT == MVT::i1) { 2948 // AddressingMode 2 2949 if (ConstantSDNode *RHS = dyn_cast<ConstantSDNode>(Ptr->getOperand(1))) { 2950 int RHSC = (int)RHS->getZExtValue(); 2951 if (RHSC < 0 && RHSC > -0x1000) { 2952 isInc = false; 2953 Offset = DAG.getConstant(-RHSC, RHS->getValueType(0)); 2954 Base = Ptr->getOperand(0); 2955 return true; 2956 } 2957 } 2958 2959 if (Ptr->getOpcode() == ISD::ADD) { 2960 isInc = true; 2961 ARM_AM::ShiftOpc ShOpcVal= ARM_AM::getShiftOpcForNode(Ptr->getOperand(0)); 2962 if (ShOpcVal != ARM_AM::no_shift) { 2963 Base = Ptr->getOperand(1); 2964 Offset = Ptr->getOperand(0); 2965 } else { 2966 Base = Ptr->getOperand(0); 2967 Offset = Ptr->getOperand(1); 2968 } 2969 return true; 2970 } 2971 2972 isInc = (Ptr->getOpcode() == ISD::ADD); 2973 Base = Ptr->getOperand(0); 2974 Offset = Ptr->getOperand(1); 2975 return true; 2976 } 2977 2978 // FIXME: Use FLDM / FSTM to emulate indexed FP load / store. 2979 return false; 2980} 2981 2982/// getPreIndexedAddressParts - returns true by value, base pointer and 2983/// offset pointer and addressing mode by reference if the node's address 2984/// can be legally represented as pre-indexed load / store address. 2985bool 2986ARMTargetLowering::getPreIndexedAddressParts(SDNode *N, SDValue &Base, 2987 SDValue &Offset, 2988 ISD::MemIndexedMode &AM, 2989 SelectionDAG &DAG) const { 2990 if (Subtarget->isThumb()) 2991 return false; 2992 2993 MVT VT; 2994 SDValue Ptr; 2995 bool isSEXTLoad = false; 2996 if (LoadSDNode *LD = dyn_cast<LoadSDNode>(N)) { 2997 Ptr = LD->getBasePtr(); 2998 VT = LD->getMemoryVT(); 2999 isSEXTLoad = LD->getExtensionType() == ISD::SEXTLOAD; 3000 } else if (StoreSDNode *ST = dyn_cast<StoreSDNode>(N)) { 3001 Ptr = ST->getBasePtr(); 3002 VT = ST->getMemoryVT(); 3003 } else 3004 return false; 3005 3006 bool isInc; 3007 bool isLegal = getIndexedAddressParts(Ptr.getNode(), VT, isSEXTLoad, Base, 3008 Offset, isInc, DAG); 3009 if (isLegal) { 3010 AM = isInc ? ISD::PRE_INC : ISD::PRE_DEC; 3011 return true; 3012 } 3013 return false; 3014} 3015 3016/// getPostIndexedAddressParts - returns true by value, base pointer and 3017/// offset pointer and addressing mode by reference if this node can be 3018/// combined with a load / store to form a post-indexed load / store. 3019bool ARMTargetLowering::getPostIndexedAddressParts(SDNode *N, SDNode *Op, 3020 SDValue &Base, 3021 SDValue &Offset, 3022 ISD::MemIndexedMode &AM, 3023 SelectionDAG &DAG) const { 3024 if (Subtarget->isThumb()) 3025 return false; 3026 3027 MVT VT; 3028 SDValue Ptr; 3029 bool isSEXTLoad = false; 3030 if (LoadSDNode *LD = dyn_cast<LoadSDNode>(N)) { 3031 VT = LD->getMemoryVT(); 3032 isSEXTLoad = LD->getExtensionType() == ISD::SEXTLOAD; 3033 } else if (StoreSDNode *ST = dyn_cast<StoreSDNode>(N)) { 3034 VT = ST->getMemoryVT(); 3035 } else 3036 return false; 3037 3038 bool isInc; 3039 bool isLegal = getIndexedAddressParts(Op, VT, isSEXTLoad, Base, Offset, 3040 isInc, DAG); 3041 if (isLegal) { 3042 AM = isInc ? ISD::POST_INC : ISD::POST_DEC; 3043 return true; 3044 } 3045 return false; 3046} 3047 3048void ARMTargetLowering::computeMaskedBitsForTargetNode(const SDValue Op, 3049 const APInt &Mask, 3050 APInt &KnownZero, 3051 APInt &KnownOne, 3052 const SelectionDAG &DAG, 3053 unsigned Depth) const { 3054 KnownZero = KnownOne = APInt(Mask.getBitWidth(), 0); 3055 switch (Op.getOpcode()) { 3056 default: break; 3057 case ARMISD::CMOV: { 3058 // Bits are known zero/one if known on the LHS and RHS. 3059 DAG.ComputeMaskedBits(Op.getOperand(0), Mask, KnownZero, KnownOne, Depth+1); 3060 if (KnownZero == 0 && KnownOne == 0) return; 3061 3062 APInt KnownZeroRHS, KnownOneRHS; 3063 DAG.ComputeMaskedBits(Op.getOperand(1), Mask, 3064 KnownZeroRHS, KnownOneRHS, Depth+1); 3065 KnownZero &= KnownZeroRHS; 3066 KnownOne &= KnownOneRHS; 3067 return; 3068 } 3069 } 3070} 3071 3072//===----------------------------------------------------------------------===// 3073// ARM Inline Assembly Support 3074//===----------------------------------------------------------------------===// 3075 3076/// getConstraintType - Given a constraint letter, return the type of 3077/// constraint it is for this target. 3078ARMTargetLowering::ConstraintType 3079ARMTargetLowering::getConstraintType(const std::string &Constraint) const { 3080 if (Constraint.size() == 1) { 3081 switch (Constraint[0]) { 3082 default: break; 3083 case 'l': return C_RegisterClass; 3084 case 'w': return C_RegisterClass; 3085 } 3086 } 3087 return TargetLowering::getConstraintType(Constraint); 3088} 3089 3090std::pair<unsigned, const TargetRegisterClass*> 3091ARMTargetLowering::getRegForInlineAsmConstraint(const std::string &Constraint, 3092 MVT VT) const { 3093 if (Constraint.size() == 1) { 3094 // GCC RS6000 Constraint Letters 3095 switch (Constraint[0]) { 3096 case 'l': 3097 if (Subtarget->isThumb()) 3098 return std::make_pair(0U, ARM::tGPRRegisterClass); 3099 else 3100 return std::make_pair(0U, ARM::GPRRegisterClass); 3101 case 'r': 3102 return std::make_pair(0U, ARM::GPRRegisterClass); 3103 case 'w': 3104 if (VT == MVT::f32) 3105 return std::make_pair(0U, ARM::SPRRegisterClass); 3106 if (VT == MVT::f64) 3107 return std::make_pair(0U, ARM::DPRRegisterClass); 3108 break; 3109 } 3110 } 3111 return TargetLowering::getRegForInlineAsmConstraint(Constraint, VT); 3112} 3113 3114std::vector<unsigned> ARMTargetLowering:: 3115getRegClassForInlineAsmConstraint(const std::string &Constraint, 3116 MVT VT) const { 3117 if (Constraint.size() != 1) 3118 return std::vector<unsigned>(); 3119 3120 switch (Constraint[0]) { // GCC ARM Constraint Letters 3121 default: break; 3122 case 'l': 3123 return make_vector<unsigned>(ARM::R0, ARM::R1, ARM::R2, ARM::R3, 3124 ARM::R4, ARM::R5, ARM::R6, ARM::R7, 3125 0); 3126 case 'r': 3127 return make_vector<unsigned>(ARM::R0, ARM::R1, ARM::R2, ARM::R3, 3128 ARM::R4, ARM::R5, ARM::R6, ARM::R7, 3129 ARM::R8, ARM::R9, ARM::R10, ARM::R11, 3130 ARM::R12, ARM::LR, 0); 3131 case 'w': 3132 if (VT == MVT::f32) 3133 return make_vector<unsigned>(ARM::S0, ARM::S1, ARM::S2, ARM::S3, 3134 ARM::S4, ARM::S5, ARM::S6, ARM::S7, 3135 ARM::S8, ARM::S9, ARM::S10, ARM::S11, 3136 ARM::S12,ARM::S13,ARM::S14,ARM::S15, 3137 ARM::S16,ARM::S17,ARM::S18,ARM::S19, 3138 ARM::S20,ARM::S21,ARM::S22,ARM::S23, 3139 ARM::S24,ARM::S25,ARM::S26,ARM::S27, 3140 ARM::S28,ARM::S29,ARM::S30,ARM::S31, 0); 3141 if (VT == MVT::f64) 3142 return make_vector<unsigned>(ARM::D0, ARM::D1, ARM::D2, ARM::D3, 3143 ARM::D4, ARM::D5, ARM::D6, ARM::D7, 3144 ARM::D8, ARM::D9, ARM::D10,ARM::D11, 3145 ARM::D12,ARM::D13,ARM::D14,ARM::D15, 0); 3146 break; 3147 } 3148 3149 return std::vector<unsigned>(); 3150} 3151 3152/// LowerAsmOperandForConstraint - Lower the specified operand into the Ops 3153/// vector. If it is invalid, don't add anything to Ops. 3154void ARMTargetLowering::LowerAsmOperandForConstraint(SDValue Op, 3155 char Constraint, 3156 bool hasMemory, 3157 std::vector<SDValue>&Ops, 3158 SelectionDAG &DAG) const { 3159 SDValue Result(0, 0); 3160 3161 switch (Constraint) { 3162 default: break; 3163 case 'I': case 'J': case 'K': case 'L': 3164 case 'M': case 'N': case 'O': 3165 ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op); 3166 if (!C) 3167 return; 3168 3169 int64_t CVal64 = C->getSExtValue(); 3170 int CVal = (int) CVal64; 3171 // None of these constraints allow values larger than 32 bits. Check 3172 // that the value fits in an int. 3173 if (CVal != CVal64) 3174 return; 3175 3176 switch (Constraint) { 3177 case 'I': 3178 if (Subtarget->isThumb()) { 3179 // This must be a constant between 0 and 255, for ADD immediates. 3180 if (CVal >= 0 && CVal <= 255) 3181 break; 3182 } else { 3183 // A constant that can be used as an immediate value in a 3184 // data-processing instruction. 3185 if (ARM_AM::getSOImmVal(CVal) != -1) 3186 break; 3187 } 3188 return; 3189 3190 case 'J': 3191 if (Subtarget->isThumb()) { 3192 // This must be a constant between -255 and -1, for negated ADD 3193 // immediates. This can be used in GCC with an "n" modifier that 3194 // prints the negated value, for use with SUB instructions. It is 3195 // not useful otherwise but is implemented for compatibility. 3196 if (CVal >= -255 && CVal <= -1) 3197 break; 3198 } else { 3199 // This must be a constant between -4095 and 4095. It is not clear 3200 // what this constraint is intended for. Implemented for 3201 // compatibility with GCC. 3202 if (CVal >= -4095 && CVal <= 4095) 3203 break; 3204 } 3205 return; 3206 3207 case 'K': 3208 if (Subtarget->isThumb()) { 3209 // A 32-bit value where only one byte has a nonzero value. Exclude 3210 // zero to match GCC. This constraint is used by GCC internally for 3211 // constants that can be loaded with a move/shift combination. 3212 // It is not useful otherwise but is implemented for compatibility. 3213 if (CVal != 0 && ARM_AM::isThumbImmShiftedVal(CVal)) 3214 break; 3215 } else { 3216 // A constant whose bitwise inverse can be used as an immediate 3217 // value in a data-processing instruction. This can be used in GCC 3218 // with a "B" modifier that prints the inverted value, for use with 3219 // BIC and MVN instructions. It is not useful otherwise but is 3220 // implemented for compatibility. 3221 if (ARM_AM::getSOImmVal(~CVal) != -1) 3222 break; 3223 } 3224 return; 3225 3226 case 'L': 3227 if (Subtarget->isThumb()) { 3228 // This must be a constant between -7 and 7, 3229 // for 3-operand ADD/SUB immediate instructions. 3230 if (CVal >= -7 && CVal < 7) 3231 break; 3232 } else { 3233 // A constant whose negation can be used as an immediate value in a 3234 // data-processing instruction. This can be used in GCC with an "n" 3235 // modifier that prints the negated value, for use with SUB 3236 // instructions. It is not useful otherwise but is implemented for 3237 // compatibility. 3238 if (ARM_AM::getSOImmVal(-CVal) != -1) 3239 break; 3240 } 3241 return; 3242 3243 case 'M': 3244 if (Subtarget->isThumb()) { 3245 // This must be a multiple of 4 between 0 and 1020, for 3246 // ADD sp + immediate. 3247 if ((CVal >= 0 && CVal <= 1020) && ((CVal & 3) == 0)) 3248 break; 3249 } else { 3250 // A power of two or a constant between 0 and 32. This is used in 3251 // GCC for the shift amount on shifted register operands, but it is 3252 // useful in general for any shift amounts. 3253 if ((CVal >= 0 && CVal <= 32) || ((CVal & (CVal - 1)) == 0)) 3254 break; 3255 } 3256 return; 3257 3258 case 'N': 3259 if (Subtarget->isThumb()) { 3260 // This must be a constant between 0 and 31, for shift amounts. 3261 if (CVal >= 0 && CVal <= 31) 3262 break; 3263 } 3264 return; 3265 3266 case 'O': 3267 if (Subtarget->isThumb()) { 3268 // This must be a multiple of 4 between -508 and 508, for 3269 // ADD/SUB sp = sp + immediate. 3270 if ((CVal >= -508 && CVal <= 508) && ((CVal & 3) == 0)) 3271 break; 3272 } 3273 return; 3274 } 3275 Result = DAG.getTargetConstant(CVal, Op.getValueType()); 3276 break; 3277 } 3278 3279 if (Result.getNode()) { 3280 Ops.push_back(Result); 3281 return; 3282 } 3283 return TargetLowering::LowerAsmOperandForConstraint(Op, Constraint, hasMemory, 3284 Ops, DAG); 3285} 3286