ARMISelLowering.cpp revision b5fb4282cd2968be81a853c0f3199d167e402bdf
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 "ARMPerfectShuffle.h" 21#include "ARMRegisterInfo.h" 22#include "ARMSubtarget.h" 23#include "ARMTargetMachine.h" 24#include "ARMTargetObjectFile.h" 25#include "llvm/CallingConv.h" 26#include "llvm/Constants.h" 27#include "llvm/Function.h" 28#include "llvm/Instruction.h" 29#include "llvm/Intrinsics.h" 30#include "llvm/GlobalValue.h" 31#include "llvm/CodeGen/CallingConvLower.h" 32#include "llvm/CodeGen/MachineBasicBlock.h" 33#include "llvm/CodeGen/MachineFrameInfo.h" 34#include "llvm/CodeGen/MachineFunction.h" 35#include "llvm/CodeGen/MachineInstrBuilder.h" 36#include "llvm/CodeGen/MachineRegisterInfo.h" 37#include "llvm/CodeGen/PseudoSourceValue.h" 38#include "llvm/CodeGen/SelectionDAG.h" 39#include "llvm/Target/TargetOptions.h" 40#include "llvm/ADT/VectorExtras.h" 41#include "llvm/Support/ErrorHandling.h" 42#include "llvm/Support/MathExtras.h" 43using namespace llvm; 44 45static bool CC_ARM_APCS_Custom_f64(unsigned &ValNo, EVT &ValVT, EVT &LocVT, 46 CCValAssign::LocInfo &LocInfo, 47 ISD::ArgFlagsTy &ArgFlags, 48 CCState &State); 49static bool CC_ARM_AAPCS_Custom_f64(unsigned &ValNo, EVT &ValVT, EVT &LocVT, 50 CCValAssign::LocInfo &LocInfo, 51 ISD::ArgFlagsTy &ArgFlags, 52 CCState &State); 53static bool RetCC_ARM_APCS_Custom_f64(unsigned &ValNo, EVT &ValVT, EVT &LocVT, 54 CCValAssign::LocInfo &LocInfo, 55 ISD::ArgFlagsTy &ArgFlags, 56 CCState &State); 57static bool RetCC_ARM_AAPCS_Custom_f64(unsigned &ValNo, EVT &ValVT, EVT &LocVT, 58 CCValAssign::LocInfo &LocInfo, 59 ISD::ArgFlagsTy &ArgFlags, 60 CCState &State); 61 62void ARMTargetLowering::addTypeForNEON(EVT VT, EVT PromotedLdStVT, 63 EVT PromotedBitwiseVT) { 64 if (VT != PromotedLdStVT) { 65 setOperationAction(ISD::LOAD, VT.getSimpleVT(), Promote); 66 AddPromotedToType (ISD::LOAD, VT.getSimpleVT(), 67 PromotedLdStVT.getSimpleVT()); 68 69 setOperationAction(ISD::STORE, VT.getSimpleVT(), Promote); 70 AddPromotedToType (ISD::STORE, VT.getSimpleVT(), 71 PromotedLdStVT.getSimpleVT()); 72 } 73 74 EVT ElemTy = VT.getVectorElementType(); 75 if (ElemTy != MVT::i64 && ElemTy != MVT::f64) 76 setOperationAction(ISD::VSETCC, VT.getSimpleVT(), Custom); 77 if (ElemTy == MVT::i8 || ElemTy == MVT::i16) 78 setOperationAction(ISD::EXTRACT_VECTOR_ELT, VT.getSimpleVT(), Custom); 79 setOperationAction(ISD::BUILD_VECTOR, VT.getSimpleVT(), Custom); 80 setOperationAction(ISD::VECTOR_SHUFFLE, VT.getSimpleVT(), Custom); 81 setOperationAction(ISD::CONCAT_VECTORS, VT.getSimpleVT(), Custom); 82 setOperationAction(ISD::EXTRACT_SUBVECTOR, VT.getSimpleVT(), Expand); 83 if (VT.isInteger()) { 84 setOperationAction(ISD::SHL, VT.getSimpleVT(), Custom); 85 setOperationAction(ISD::SRA, VT.getSimpleVT(), Custom); 86 setOperationAction(ISD::SRL, VT.getSimpleVT(), Custom); 87 } 88 89 // Promote all bit-wise operations. 90 if (VT.isInteger() && VT != PromotedBitwiseVT) { 91 setOperationAction(ISD::AND, VT.getSimpleVT(), Promote); 92 AddPromotedToType (ISD::AND, VT.getSimpleVT(), 93 PromotedBitwiseVT.getSimpleVT()); 94 setOperationAction(ISD::OR, VT.getSimpleVT(), Promote); 95 AddPromotedToType (ISD::OR, VT.getSimpleVT(), 96 PromotedBitwiseVT.getSimpleVT()); 97 setOperationAction(ISD::XOR, VT.getSimpleVT(), Promote); 98 AddPromotedToType (ISD::XOR, VT.getSimpleVT(), 99 PromotedBitwiseVT.getSimpleVT()); 100 } 101} 102 103void ARMTargetLowering::addDRTypeForNEON(EVT VT) { 104 addRegisterClass(VT, ARM::DPRRegisterClass); 105 addTypeForNEON(VT, MVT::f64, MVT::v2i32); 106} 107 108void ARMTargetLowering::addQRTypeForNEON(EVT VT) { 109 addRegisterClass(VT, ARM::QPRRegisterClass); 110 addTypeForNEON(VT, MVT::v2f64, MVT::v4i32); 111} 112 113static TargetLoweringObjectFile *createTLOF(TargetMachine &TM) { 114 if (TM.getSubtarget<ARMSubtarget>().isTargetDarwin()) 115 return new TargetLoweringObjectFileMachO(); 116 return new ARMElfTargetObjectFile(); 117} 118 119ARMTargetLowering::ARMTargetLowering(TargetMachine &TM) 120 : TargetLowering(TM, createTLOF(TM)), ARMPCLabelIndex(0) { 121 Subtarget = &TM.getSubtarget<ARMSubtarget>(); 122 123 if (Subtarget->isTargetDarwin()) { 124 // Uses VFP for Thumb libfuncs if available. 125 if (Subtarget->isThumb() && Subtarget->hasVFP2()) { 126 // Single-precision floating-point arithmetic. 127 setLibcallName(RTLIB::ADD_F32, "__addsf3vfp"); 128 setLibcallName(RTLIB::SUB_F32, "__subsf3vfp"); 129 setLibcallName(RTLIB::MUL_F32, "__mulsf3vfp"); 130 setLibcallName(RTLIB::DIV_F32, "__divsf3vfp"); 131 132 // Double-precision floating-point arithmetic. 133 setLibcallName(RTLIB::ADD_F64, "__adddf3vfp"); 134 setLibcallName(RTLIB::SUB_F64, "__subdf3vfp"); 135 setLibcallName(RTLIB::MUL_F64, "__muldf3vfp"); 136 setLibcallName(RTLIB::DIV_F64, "__divdf3vfp"); 137 138 // Single-precision comparisons. 139 setLibcallName(RTLIB::OEQ_F32, "__eqsf2vfp"); 140 setLibcallName(RTLIB::UNE_F32, "__nesf2vfp"); 141 setLibcallName(RTLIB::OLT_F32, "__ltsf2vfp"); 142 setLibcallName(RTLIB::OLE_F32, "__lesf2vfp"); 143 setLibcallName(RTLIB::OGE_F32, "__gesf2vfp"); 144 setLibcallName(RTLIB::OGT_F32, "__gtsf2vfp"); 145 setLibcallName(RTLIB::UO_F32, "__unordsf2vfp"); 146 setLibcallName(RTLIB::O_F32, "__unordsf2vfp"); 147 148 setCmpLibcallCC(RTLIB::OEQ_F32, ISD::SETNE); 149 setCmpLibcallCC(RTLIB::UNE_F32, ISD::SETNE); 150 setCmpLibcallCC(RTLIB::OLT_F32, ISD::SETNE); 151 setCmpLibcallCC(RTLIB::OLE_F32, ISD::SETNE); 152 setCmpLibcallCC(RTLIB::OGE_F32, ISD::SETNE); 153 setCmpLibcallCC(RTLIB::OGT_F32, ISD::SETNE); 154 setCmpLibcallCC(RTLIB::UO_F32, ISD::SETNE); 155 setCmpLibcallCC(RTLIB::O_F32, ISD::SETEQ); 156 157 // Double-precision comparisons. 158 setLibcallName(RTLIB::OEQ_F64, "__eqdf2vfp"); 159 setLibcallName(RTLIB::UNE_F64, "__nedf2vfp"); 160 setLibcallName(RTLIB::OLT_F64, "__ltdf2vfp"); 161 setLibcallName(RTLIB::OLE_F64, "__ledf2vfp"); 162 setLibcallName(RTLIB::OGE_F64, "__gedf2vfp"); 163 setLibcallName(RTLIB::OGT_F64, "__gtdf2vfp"); 164 setLibcallName(RTLIB::UO_F64, "__unorddf2vfp"); 165 setLibcallName(RTLIB::O_F64, "__unorddf2vfp"); 166 167 setCmpLibcallCC(RTLIB::OEQ_F64, ISD::SETNE); 168 setCmpLibcallCC(RTLIB::UNE_F64, ISD::SETNE); 169 setCmpLibcallCC(RTLIB::OLT_F64, ISD::SETNE); 170 setCmpLibcallCC(RTLIB::OLE_F64, ISD::SETNE); 171 setCmpLibcallCC(RTLIB::OGE_F64, ISD::SETNE); 172 setCmpLibcallCC(RTLIB::OGT_F64, ISD::SETNE); 173 setCmpLibcallCC(RTLIB::UO_F64, ISD::SETNE); 174 setCmpLibcallCC(RTLIB::O_F64, ISD::SETEQ); 175 176 // Floating-point to integer conversions. 177 // i64 conversions are done via library routines even when generating VFP 178 // instructions, so use the same ones. 179 setLibcallName(RTLIB::FPTOSINT_F64_I32, "__fixdfsivfp"); 180 setLibcallName(RTLIB::FPTOUINT_F64_I32, "__fixunsdfsivfp"); 181 setLibcallName(RTLIB::FPTOSINT_F32_I32, "__fixsfsivfp"); 182 setLibcallName(RTLIB::FPTOUINT_F32_I32, "__fixunssfsivfp"); 183 184 // Conversions between floating types. 185 setLibcallName(RTLIB::FPROUND_F64_F32, "__truncdfsf2vfp"); 186 setLibcallName(RTLIB::FPEXT_F32_F64, "__extendsfdf2vfp"); 187 188 // Integer to floating-point conversions. 189 // i64 conversions are done via library routines even when generating VFP 190 // instructions, so use the same ones. 191 // FIXME: There appears to be some naming inconsistency in ARM libgcc: 192 // e.g., __floatunsidf vs. __floatunssidfvfp. 193 setLibcallName(RTLIB::SINTTOFP_I32_F64, "__floatsidfvfp"); 194 setLibcallName(RTLIB::UINTTOFP_I32_F64, "__floatunssidfvfp"); 195 setLibcallName(RTLIB::SINTTOFP_I32_F32, "__floatsisfvfp"); 196 setLibcallName(RTLIB::UINTTOFP_I32_F32, "__floatunssisfvfp"); 197 } 198 } 199 200 // These libcalls are not available in 32-bit. 201 setLibcallName(RTLIB::SHL_I128, 0); 202 setLibcallName(RTLIB::SRL_I128, 0); 203 setLibcallName(RTLIB::SRA_I128, 0); 204 205 // Libcalls should use the AAPCS base standard ABI, even if hard float 206 // is in effect, as per the ARM RTABI specification, section 4.1.2. 207 if (Subtarget->isAAPCS_ABI()) { 208 for (int i = 0; i < RTLIB::UNKNOWN_LIBCALL; ++i) { 209 setLibcallCallingConv(static_cast<RTLIB::Libcall>(i), 210 CallingConv::ARM_AAPCS); 211 } 212 } 213 214 if (Subtarget->isThumb1Only()) 215 addRegisterClass(MVT::i32, ARM::tGPRRegisterClass); 216 else 217 addRegisterClass(MVT::i32, ARM::GPRRegisterClass); 218 if (!UseSoftFloat && Subtarget->hasVFP2() && !Subtarget->isThumb1Only()) { 219 addRegisterClass(MVT::f32, ARM::SPRRegisterClass); 220 addRegisterClass(MVT::f64, ARM::DPRRegisterClass); 221 222 setTruncStoreAction(MVT::f64, MVT::f32, Expand); 223 } 224 225 if (Subtarget->hasNEON()) { 226 addDRTypeForNEON(MVT::v2f32); 227 addDRTypeForNEON(MVT::v8i8); 228 addDRTypeForNEON(MVT::v4i16); 229 addDRTypeForNEON(MVT::v2i32); 230 addDRTypeForNEON(MVT::v1i64); 231 232 addQRTypeForNEON(MVT::v4f32); 233 addQRTypeForNEON(MVT::v2f64); 234 addQRTypeForNEON(MVT::v16i8); 235 addQRTypeForNEON(MVT::v8i16); 236 addQRTypeForNEON(MVT::v4i32); 237 addQRTypeForNEON(MVT::v2i64); 238 239 setTargetDAGCombine(ISD::INTRINSIC_WO_CHAIN); 240 setTargetDAGCombine(ISD::SHL); 241 setTargetDAGCombine(ISD::SRL); 242 setTargetDAGCombine(ISD::SRA); 243 setTargetDAGCombine(ISD::SIGN_EXTEND); 244 setTargetDAGCombine(ISD::ZERO_EXTEND); 245 setTargetDAGCombine(ISD::ANY_EXTEND); 246 } 247 248 computeRegisterProperties(); 249 250 // ARM does not have f32 extending load. 251 setLoadExtAction(ISD::EXTLOAD, MVT::f32, Expand); 252 253 // ARM does not have i1 sign extending load. 254 setLoadExtAction(ISD::SEXTLOAD, MVT::i1, Promote); 255 256 // ARM supports all 4 flavors of integer indexed load / store. 257 if (!Subtarget->isThumb1Only()) { 258 for (unsigned im = (unsigned)ISD::PRE_INC; 259 im != (unsigned)ISD::LAST_INDEXED_MODE; ++im) { 260 setIndexedLoadAction(im, MVT::i1, Legal); 261 setIndexedLoadAction(im, MVT::i8, Legal); 262 setIndexedLoadAction(im, MVT::i16, Legal); 263 setIndexedLoadAction(im, MVT::i32, Legal); 264 setIndexedStoreAction(im, MVT::i1, Legal); 265 setIndexedStoreAction(im, MVT::i8, Legal); 266 setIndexedStoreAction(im, MVT::i16, Legal); 267 setIndexedStoreAction(im, MVT::i32, Legal); 268 } 269 } 270 271 // i64 operation support. 272 if (Subtarget->isThumb1Only()) { 273 setOperationAction(ISD::MUL, MVT::i64, Expand); 274 setOperationAction(ISD::MULHU, MVT::i32, Expand); 275 setOperationAction(ISD::MULHS, MVT::i32, Expand); 276 setOperationAction(ISD::UMUL_LOHI, MVT::i32, Expand); 277 setOperationAction(ISD::SMUL_LOHI, MVT::i32, Expand); 278 } else { 279 setOperationAction(ISD::MUL, MVT::i64, Expand); 280 setOperationAction(ISD::MULHU, MVT::i32, Expand); 281 if (!Subtarget->hasV6Ops()) 282 setOperationAction(ISD::MULHS, MVT::i32, Expand); 283 } 284 setOperationAction(ISD::SHL_PARTS, MVT::i32, Expand); 285 setOperationAction(ISD::SRA_PARTS, MVT::i32, Expand); 286 setOperationAction(ISD::SRL_PARTS, MVT::i32, Expand); 287 setOperationAction(ISD::SRL, MVT::i64, Custom); 288 setOperationAction(ISD::SRA, MVT::i64, Custom); 289 290 // ARM does not have ROTL. 291 setOperationAction(ISD::ROTL, MVT::i32, Expand); 292 setOperationAction(ISD::CTTZ, MVT::i32, Expand); 293 setOperationAction(ISD::CTPOP, MVT::i32, Expand); 294 if (!Subtarget->hasV5TOps() || Subtarget->isThumb1Only()) 295 setOperationAction(ISD::CTLZ, MVT::i32, Expand); 296 297 // Only ARMv6 has BSWAP. 298 if (!Subtarget->hasV6Ops()) 299 setOperationAction(ISD::BSWAP, MVT::i32, Expand); 300 301 // These are expanded into libcalls. 302 setOperationAction(ISD::SDIV, MVT::i32, Expand); 303 setOperationAction(ISD::UDIV, MVT::i32, Expand); 304 setOperationAction(ISD::SREM, MVT::i32, Expand); 305 setOperationAction(ISD::UREM, MVT::i32, Expand); 306 setOperationAction(ISD::SDIVREM, MVT::i32, Expand); 307 setOperationAction(ISD::UDIVREM, MVT::i32, Expand); 308 309 // Support label based line numbers. 310 setOperationAction(ISD::DBG_STOPPOINT, MVT::Other, Expand); 311 setOperationAction(ISD::DEBUG_LOC, MVT::Other, Expand); 312 313 setOperationAction(ISD::GlobalAddress, MVT::i32, Custom); 314 setOperationAction(ISD::ConstantPool, MVT::i32, Custom); 315 setOperationAction(ISD::GLOBAL_OFFSET_TABLE, MVT::i32, Custom); 316 setOperationAction(ISD::GlobalTLSAddress, MVT::i32, Custom); 317 318 // Use the default implementation. 319 setOperationAction(ISD::VASTART, MVT::Other, Custom); 320 setOperationAction(ISD::VAARG, MVT::Other, Expand); 321 setOperationAction(ISD::VACOPY, MVT::Other, Expand); 322 setOperationAction(ISD::VAEND, MVT::Other, Expand); 323 setOperationAction(ISD::STACKSAVE, MVT::Other, Expand); 324 setOperationAction(ISD::STACKRESTORE, MVT::Other, Expand); 325 setOperationAction(ISD::EHSELECTION, MVT::i32, Expand); 326 // FIXME: Shouldn't need this, since no register is used, but the legalizer 327 // doesn't yet know how to not do that for SjLj. 328 setExceptionSelectorRegister(ARM::R0); 329 if (Subtarget->isThumb()) 330 setOperationAction(ISD::DYNAMIC_STACKALLOC, MVT::i32, Custom); 331 else 332 setOperationAction(ISD::DYNAMIC_STACKALLOC, MVT::i32, Expand); 333 setOperationAction(ISD::MEMBARRIER, MVT::Other, Expand); 334 335 if (!Subtarget->hasV6Ops() && !Subtarget->isThumb2()) { 336 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i16, Expand); 337 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i8, Expand); 338 } 339 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i1, Expand); 340 341 if (!UseSoftFloat && Subtarget->hasVFP2() && !Subtarget->isThumb1Only()) 342 // Turn f64->i64 into FMRRD, i64 -> f64 to FMDRR iff target supports vfp2. 343 setOperationAction(ISD::BIT_CONVERT, MVT::i64, Custom); 344 345 // We want to custom lower some of our intrinsics. 346 setOperationAction(ISD::INTRINSIC_WO_CHAIN, MVT::Other, Custom); 347 setOperationAction(ISD::INTRINSIC_W_CHAIN, MVT::Other, Custom); 348 setOperationAction(ISD::INTRINSIC_VOID, MVT::Other, Custom); 349 350 setOperationAction(ISD::SETCC, MVT::i32, Expand); 351 setOperationAction(ISD::SETCC, MVT::f32, Expand); 352 setOperationAction(ISD::SETCC, MVT::f64, Expand); 353 setOperationAction(ISD::SELECT, MVT::i32, Expand); 354 setOperationAction(ISD::SELECT, MVT::f32, Expand); 355 setOperationAction(ISD::SELECT, MVT::f64, Expand); 356 setOperationAction(ISD::SELECT_CC, MVT::i32, Custom); 357 setOperationAction(ISD::SELECT_CC, MVT::f32, Custom); 358 setOperationAction(ISD::SELECT_CC, MVT::f64, Custom); 359 360 setOperationAction(ISD::BRCOND, MVT::Other, Expand); 361 setOperationAction(ISD::BR_CC, MVT::i32, Custom); 362 setOperationAction(ISD::BR_CC, MVT::f32, Custom); 363 setOperationAction(ISD::BR_CC, MVT::f64, Custom); 364 setOperationAction(ISD::BR_JT, MVT::Other, Custom); 365 366 // We don't support sin/cos/fmod/copysign/pow 367 setOperationAction(ISD::FSIN, MVT::f64, Expand); 368 setOperationAction(ISD::FSIN, MVT::f32, Expand); 369 setOperationAction(ISD::FCOS, MVT::f32, Expand); 370 setOperationAction(ISD::FCOS, MVT::f64, Expand); 371 setOperationAction(ISD::FREM, MVT::f64, Expand); 372 setOperationAction(ISD::FREM, MVT::f32, Expand); 373 if (!UseSoftFloat && Subtarget->hasVFP2() && !Subtarget->isThumb1Only()) { 374 setOperationAction(ISD::FCOPYSIGN, MVT::f64, Custom); 375 setOperationAction(ISD::FCOPYSIGN, MVT::f32, Custom); 376 } 377 setOperationAction(ISD::FPOW, MVT::f64, Expand); 378 setOperationAction(ISD::FPOW, MVT::f32, Expand); 379 380 // int <-> fp are custom expanded into bit_convert + ARMISD ops. 381 if (!UseSoftFloat && Subtarget->hasVFP2() && !Subtarget->isThumb1Only()) { 382 setOperationAction(ISD::SINT_TO_FP, MVT::i32, Custom); 383 setOperationAction(ISD::UINT_TO_FP, MVT::i32, Custom); 384 setOperationAction(ISD::FP_TO_UINT, MVT::i32, Custom); 385 setOperationAction(ISD::FP_TO_SINT, MVT::i32, Custom); 386 } 387 388 // We have target-specific dag combine patterns for the following nodes: 389 // ARMISD::FMRRD - No need to call setTargetDAGCombine 390 setTargetDAGCombine(ISD::ADD); 391 setTargetDAGCombine(ISD::SUB); 392 393 setStackPointerRegisterToSaveRestore(ARM::SP); 394 setSchedulingPreference(SchedulingForRegPressure); 395 396 // FIXME: If-converter should use instruction latency to determine 397 // profitability rather than relying on fixed limits. 398 if (Subtarget->getCPUString() == "generic") { 399 // Generic (and overly aggressive) if-conversion limits. 400 setIfCvtBlockSizeLimit(10); 401 setIfCvtDupBlockSizeLimit(2); 402 } else if (Subtarget->hasV6Ops()) { 403 setIfCvtBlockSizeLimit(2); 404 setIfCvtDupBlockSizeLimit(1); 405 } else { 406 setIfCvtBlockSizeLimit(3); 407 setIfCvtDupBlockSizeLimit(2); 408 } 409 410 maxStoresPerMemcpy = 1; //// temporary - rewrite interface to use type 411 // Do not enable CodePlacementOpt for now: it currently runs after the 412 // ARMConstantIslandPass and messes up branch relaxation and placement 413 // of constant islands. 414 // benefitFromCodePlacementOpt = true; 415} 416 417const char *ARMTargetLowering::getTargetNodeName(unsigned Opcode) const { 418 switch (Opcode) { 419 default: return 0; 420 case ARMISD::Wrapper: return "ARMISD::Wrapper"; 421 case ARMISD::WrapperJT: return "ARMISD::WrapperJT"; 422 case ARMISD::CALL: return "ARMISD::CALL"; 423 case ARMISD::CALL_PRED: return "ARMISD::CALL_PRED"; 424 case ARMISD::CALL_NOLINK: return "ARMISD::CALL_NOLINK"; 425 case ARMISD::tCALL: return "ARMISD::tCALL"; 426 case ARMISD::BRCOND: return "ARMISD::BRCOND"; 427 case ARMISD::BR_JT: return "ARMISD::BR_JT"; 428 case ARMISD::BR2_JT: return "ARMISD::BR2_JT"; 429 case ARMISD::RET_FLAG: return "ARMISD::RET_FLAG"; 430 case ARMISD::PIC_ADD: return "ARMISD::PIC_ADD"; 431 case ARMISD::CMP: return "ARMISD::CMP"; 432 case ARMISD::CMPZ: return "ARMISD::CMPZ"; 433 case ARMISD::CMPFP: return "ARMISD::CMPFP"; 434 case ARMISD::CMPFPw0: return "ARMISD::CMPFPw0"; 435 case ARMISD::FMSTAT: return "ARMISD::FMSTAT"; 436 case ARMISD::CMOV: return "ARMISD::CMOV"; 437 case ARMISD::CNEG: return "ARMISD::CNEG"; 438 439 case ARMISD::FTOSI: return "ARMISD::FTOSI"; 440 case ARMISD::FTOUI: return "ARMISD::FTOUI"; 441 case ARMISD::SITOF: return "ARMISD::SITOF"; 442 case ARMISD::UITOF: return "ARMISD::UITOF"; 443 444 case ARMISD::SRL_FLAG: return "ARMISD::SRL_FLAG"; 445 case ARMISD::SRA_FLAG: return "ARMISD::SRA_FLAG"; 446 case ARMISD::RRX: return "ARMISD::RRX"; 447 448 case ARMISD::FMRRD: return "ARMISD::FMRRD"; 449 case ARMISD::FMDRR: return "ARMISD::FMDRR"; 450 451 case ARMISD::THREAD_POINTER:return "ARMISD::THREAD_POINTER"; 452 453 case ARMISD::DYN_ALLOC: return "ARMISD::DYN_ALLOC"; 454 455 case ARMISD::VCEQ: return "ARMISD::VCEQ"; 456 case ARMISD::VCGE: return "ARMISD::VCGE"; 457 case ARMISD::VCGEU: return "ARMISD::VCGEU"; 458 case ARMISD::VCGT: return "ARMISD::VCGT"; 459 case ARMISD::VCGTU: return "ARMISD::VCGTU"; 460 case ARMISD::VTST: return "ARMISD::VTST"; 461 462 case ARMISD::VSHL: return "ARMISD::VSHL"; 463 case ARMISD::VSHRs: return "ARMISD::VSHRs"; 464 case ARMISD::VSHRu: return "ARMISD::VSHRu"; 465 case ARMISD::VSHLLs: return "ARMISD::VSHLLs"; 466 case ARMISD::VSHLLu: return "ARMISD::VSHLLu"; 467 case ARMISD::VSHLLi: return "ARMISD::VSHLLi"; 468 case ARMISD::VSHRN: return "ARMISD::VSHRN"; 469 case ARMISD::VRSHRs: return "ARMISD::VRSHRs"; 470 case ARMISD::VRSHRu: return "ARMISD::VRSHRu"; 471 case ARMISD::VRSHRN: return "ARMISD::VRSHRN"; 472 case ARMISD::VQSHLs: return "ARMISD::VQSHLs"; 473 case ARMISD::VQSHLu: return "ARMISD::VQSHLu"; 474 case ARMISD::VQSHLsu: return "ARMISD::VQSHLsu"; 475 case ARMISD::VQSHRNs: return "ARMISD::VQSHRNs"; 476 case ARMISD::VQSHRNu: return "ARMISD::VQSHRNu"; 477 case ARMISD::VQSHRNsu: return "ARMISD::VQSHRNsu"; 478 case ARMISD::VQRSHRNs: return "ARMISD::VQRSHRNs"; 479 case ARMISD::VQRSHRNu: return "ARMISD::VQRSHRNu"; 480 case ARMISD::VQRSHRNsu: return "ARMISD::VQRSHRNsu"; 481 case ARMISD::VGETLANEu: return "ARMISD::VGETLANEu"; 482 case ARMISD::VGETLANEs: return "ARMISD::VGETLANEs"; 483 case ARMISD::VDUP: return "ARMISD::VDUP"; 484 case ARMISD::VDUPLANE: return "ARMISD::VDUPLANE"; 485 case ARMISD::VEXT: return "ARMISD::VEXT"; 486 case ARMISD::VREV64: return "ARMISD::VREV64"; 487 case ARMISD::VREV32: return "ARMISD::VREV32"; 488 case ARMISD::VREV16: return "ARMISD::VREV16"; 489 case ARMISD::VZIP: return "ARMISD::VZIP"; 490 case ARMISD::VUZP: return "ARMISD::VUZP"; 491 case ARMISD::VTRN: return "ARMISD::VTRN"; 492 } 493} 494 495/// getFunctionAlignment - Return the Log2 alignment of this function. 496unsigned ARMTargetLowering::getFunctionAlignment(const Function *F) const { 497 return getTargetMachine().getSubtarget<ARMSubtarget>().isThumb() ? 1 : 2; 498} 499 500//===----------------------------------------------------------------------===// 501// Lowering Code 502//===----------------------------------------------------------------------===// 503 504/// IntCCToARMCC - Convert a DAG integer condition code to an ARM CC 505static ARMCC::CondCodes IntCCToARMCC(ISD::CondCode CC) { 506 switch (CC) { 507 default: llvm_unreachable("Unknown condition code!"); 508 case ISD::SETNE: return ARMCC::NE; 509 case ISD::SETEQ: return ARMCC::EQ; 510 case ISD::SETGT: return ARMCC::GT; 511 case ISD::SETGE: return ARMCC::GE; 512 case ISD::SETLT: return ARMCC::LT; 513 case ISD::SETLE: return ARMCC::LE; 514 case ISD::SETUGT: return ARMCC::HI; 515 case ISD::SETUGE: return ARMCC::HS; 516 case ISD::SETULT: return ARMCC::LO; 517 case ISD::SETULE: return ARMCC::LS; 518 } 519} 520 521/// FPCCToARMCC - Convert a DAG fp condition code to an ARM CC. It 522/// returns true if the operands should be inverted to form the proper 523/// comparison. 524static bool FPCCToARMCC(ISD::CondCode CC, ARMCC::CondCodes &CondCode, 525 ARMCC::CondCodes &CondCode2) { 526 bool Invert = false; 527 CondCode2 = ARMCC::AL; 528 switch (CC) { 529 default: llvm_unreachable("Unknown FP condition!"); 530 case ISD::SETEQ: 531 case ISD::SETOEQ: CondCode = ARMCC::EQ; break; 532 case ISD::SETGT: 533 case ISD::SETOGT: CondCode = ARMCC::GT; break; 534 case ISD::SETGE: 535 case ISD::SETOGE: CondCode = ARMCC::GE; break; 536 case ISD::SETOLT: CondCode = ARMCC::MI; break; 537 case ISD::SETOLE: CondCode = ARMCC::GT; Invert = true; break; 538 case ISD::SETONE: CondCode = ARMCC::MI; CondCode2 = ARMCC::GT; break; 539 case ISD::SETO: CondCode = ARMCC::VC; break; 540 case ISD::SETUO: CondCode = ARMCC::VS; break; 541 case ISD::SETUEQ: CondCode = ARMCC::EQ; CondCode2 = ARMCC::VS; break; 542 case ISD::SETUGT: CondCode = ARMCC::HI; break; 543 case ISD::SETUGE: CondCode = ARMCC::PL; break; 544 case ISD::SETLT: 545 case ISD::SETULT: CondCode = ARMCC::LT; break; 546 case ISD::SETLE: 547 case ISD::SETULE: CondCode = ARMCC::LE; break; 548 case ISD::SETNE: 549 case ISD::SETUNE: CondCode = ARMCC::NE; break; 550 } 551 return Invert; 552} 553 554//===----------------------------------------------------------------------===// 555// Calling Convention Implementation 556//===----------------------------------------------------------------------===// 557 558#include "ARMGenCallingConv.inc" 559 560// APCS f64 is in register pairs, possibly split to stack 561static bool f64AssignAPCS(unsigned &ValNo, EVT &ValVT, EVT &LocVT, 562 CCValAssign::LocInfo &LocInfo, 563 CCState &State, bool CanFail) { 564 static const unsigned RegList[] = { ARM::R0, ARM::R1, ARM::R2, ARM::R3 }; 565 566 // Try to get the first register. 567 if (unsigned Reg = State.AllocateReg(RegList, 4)) 568 State.addLoc(CCValAssign::getCustomReg(ValNo, ValVT, Reg, LocVT, LocInfo)); 569 else { 570 // For the 2nd half of a v2f64, do not fail. 571 if (CanFail) 572 return false; 573 574 // Put the whole thing on the stack. 575 State.addLoc(CCValAssign::getCustomMem(ValNo, ValVT, 576 State.AllocateStack(8, 4), 577 LocVT, LocInfo)); 578 return true; 579 } 580 581 // Try to get the second register. 582 if (unsigned Reg = State.AllocateReg(RegList, 4)) 583 State.addLoc(CCValAssign::getCustomReg(ValNo, ValVT, Reg, LocVT, LocInfo)); 584 else 585 State.addLoc(CCValAssign::getCustomMem(ValNo, ValVT, 586 State.AllocateStack(4, 4), 587 LocVT, LocInfo)); 588 return true; 589} 590 591static bool CC_ARM_APCS_Custom_f64(unsigned &ValNo, EVT &ValVT, EVT &LocVT, 592 CCValAssign::LocInfo &LocInfo, 593 ISD::ArgFlagsTy &ArgFlags, 594 CCState &State) { 595 if (!f64AssignAPCS(ValNo, ValVT, LocVT, LocInfo, State, true)) 596 return false; 597 if (LocVT == MVT::v2f64 && 598 !f64AssignAPCS(ValNo, ValVT, LocVT, LocInfo, State, false)) 599 return false; 600 return true; // we handled it 601} 602 603// AAPCS f64 is in aligned register pairs 604static bool f64AssignAAPCS(unsigned &ValNo, EVT &ValVT, EVT &LocVT, 605 CCValAssign::LocInfo &LocInfo, 606 CCState &State, bool CanFail) { 607 static const unsigned HiRegList[] = { ARM::R0, ARM::R2 }; 608 static const unsigned LoRegList[] = { ARM::R1, ARM::R3 }; 609 610 unsigned Reg = State.AllocateReg(HiRegList, LoRegList, 2); 611 if (Reg == 0) { 612 // For the 2nd half of a v2f64, do not just fail. 613 if (CanFail) 614 return false; 615 616 // Put the whole thing on the stack. 617 State.addLoc(CCValAssign::getCustomMem(ValNo, ValVT, 618 State.AllocateStack(8, 8), 619 LocVT, LocInfo)); 620 return true; 621 } 622 623 unsigned i; 624 for (i = 0; i < 2; ++i) 625 if (HiRegList[i] == Reg) 626 break; 627 628 State.addLoc(CCValAssign::getCustomReg(ValNo, ValVT, Reg, LocVT, LocInfo)); 629 State.addLoc(CCValAssign::getCustomReg(ValNo, ValVT, LoRegList[i], 630 LocVT, LocInfo)); 631 return true; 632} 633 634static bool CC_ARM_AAPCS_Custom_f64(unsigned &ValNo, EVT &ValVT, EVT &LocVT, 635 CCValAssign::LocInfo &LocInfo, 636 ISD::ArgFlagsTy &ArgFlags, 637 CCState &State) { 638 if (!f64AssignAAPCS(ValNo, ValVT, LocVT, LocInfo, State, true)) 639 return false; 640 if (LocVT == MVT::v2f64 && 641 !f64AssignAAPCS(ValNo, ValVT, LocVT, LocInfo, State, false)) 642 return false; 643 return true; // we handled it 644} 645 646static bool f64RetAssign(unsigned &ValNo, EVT &ValVT, EVT &LocVT, 647 CCValAssign::LocInfo &LocInfo, CCState &State) { 648 static const unsigned HiRegList[] = { ARM::R0, ARM::R2 }; 649 static const unsigned LoRegList[] = { ARM::R1, ARM::R3 }; 650 651 unsigned Reg = State.AllocateReg(HiRegList, LoRegList, 2); 652 if (Reg == 0) 653 return false; // we didn't handle it 654 655 unsigned i; 656 for (i = 0; i < 2; ++i) 657 if (HiRegList[i] == Reg) 658 break; 659 660 State.addLoc(CCValAssign::getCustomReg(ValNo, ValVT, Reg, LocVT, LocInfo)); 661 State.addLoc(CCValAssign::getCustomReg(ValNo, ValVT, LoRegList[i], 662 LocVT, LocInfo)); 663 return true; 664} 665 666static bool RetCC_ARM_APCS_Custom_f64(unsigned &ValNo, EVT &ValVT, EVT &LocVT, 667 CCValAssign::LocInfo &LocInfo, 668 ISD::ArgFlagsTy &ArgFlags, 669 CCState &State) { 670 if (!f64RetAssign(ValNo, ValVT, LocVT, LocInfo, State)) 671 return false; 672 if (LocVT == MVT::v2f64 && !f64RetAssign(ValNo, ValVT, LocVT, LocInfo, State)) 673 return false; 674 return true; // we handled it 675} 676 677static bool RetCC_ARM_AAPCS_Custom_f64(unsigned &ValNo, EVT &ValVT, EVT &LocVT, 678 CCValAssign::LocInfo &LocInfo, 679 ISD::ArgFlagsTy &ArgFlags, 680 CCState &State) { 681 return RetCC_ARM_APCS_Custom_f64(ValNo, ValVT, LocVT, LocInfo, ArgFlags, 682 State); 683} 684 685/// CCAssignFnForNode - Selects the correct CCAssignFn for a the 686/// given CallingConvention value. 687CCAssignFn *ARMTargetLowering::CCAssignFnForNode(unsigned CC, 688 bool Return, 689 bool isVarArg) const { 690 switch (CC) { 691 default: 692 llvm_unreachable("Unsupported calling convention"); 693 case CallingConv::C: 694 case CallingConv::Fast: 695 // Use target triple & subtarget features to do actual dispatch. 696 if (Subtarget->isAAPCS_ABI()) { 697 if (Subtarget->hasVFP2() && 698 FloatABIType == FloatABI::Hard && !isVarArg) 699 return (Return ? RetCC_ARM_AAPCS_VFP: CC_ARM_AAPCS_VFP); 700 else 701 return (Return ? RetCC_ARM_AAPCS: CC_ARM_AAPCS); 702 } else 703 return (Return ? RetCC_ARM_APCS: CC_ARM_APCS); 704 case CallingConv::ARM_AAPCS_VFP: 705 return (Return ? RetCC_ARM_AAPCS_VFP: CC_ARM_AAPCS_VFP); 706 case CallingConv::ARM_AAPCS: 707 return (Return ? RetCC_ARM_AAPCS: CC_ARM_AAPCS); 708 case CallingConv::ARM_APCS: 709 return (Return ? RetCC_ARM_APCS: CC_ARM_APCS); 710 } 711} 712 713/// LowerCallResult - Lower the result values of a call into the 714/// appropriate copies out of appropriate physical registers. 715SDValue 716ARMTargetLowering::LowerCallResult(SDValue Chain, SDValue InFlag, 717 unsigned CallConv, bool isVarArg, 718 const SmallVectorImpl<ISD::InputArg> &Ins, 719 DebugLoc dl, SelectionDAG &DAG, 720 SmallVectorImpl<SDValue> &InVals) { 721 722 // Assign locations to each value returned by this call. 723 SmallVector<CCValAssign, 16> RVLocs; 724 CCState CCInfo(CallConv, isVarArg, getTargetMachine(), 725 RVLocs, *DAG.getContext()); 726 CCInfo.AnalyzeCallResult(Ins, 727 CCAssignFnForNode(CallConv, /* Return*/ true, 728 isVarArg)); 729 730 // Copy all of the result registers out of their specified physreg. 731 for (unsigned i = 0; i != RVLocs.size(); ++i) { 732 CCValAssign VA = RVLocs[i]; 733 734 SDValue Val; 735 if (VA.needsCustom()) { 736 // Handle f64 or half of a v2f64. 737 SDValue Lo = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(), MVT::i32, 738 InFlag); 739 Chain = Lo.getValue(1); 740 InFlag = Lo.getValue(2); 741 VA = RVLocs[++i]; // skip ahead to next loc 742 SDValue Hi = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(), MVT::i32, 743 InFlag); 744 Chain = Hi.getValue(1); 745 InFlag = Hi.getValue(2); 746 Val = DAG.getNode(ARMISD::FMDRR, dl, MVT::f64, Lo, Hi); 747 748 if (VA.getLocVT() == MVT::v2f64) { 749 SDValue Vec = DAG.getNode(ISD::UNDEF, dl, MVT::v2f64); 750 Vec = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v2f64, Vec, Val, 751 DAG.getConstant(0, MVT::i32)); 752 753 VA = RVLocs[++i]; // skip ahead to next loc 754 Lo = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(), MVT::i32, InFlag); 755 Chain = Lo.getValue(1); 756 InFlag = Lo.getValue(2); 757 VA = RVLocs[++i]; // skip ahead to next loc 758 Hi = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(), MVT::i32, InFlag); 759 Chain = Hi.getValue(1); 760 InFlag = Hi.getValue(2); 761 Val = DAG.getNode(ARMISD::FMDRR, dl, MVT::f64, Lo, Hi); 762 Val = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v2f64, Vec, Val, 763 DAG.getConstant(1, MVT::i32)); 764 } 765 } else { 766 Val = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(), VA.getLocVT(), 767 InFlag); 768 Chain = Val.getValue(1); 769 InFlag = Val.getValue(2); 770 } 771 772 switch (VA.getLocInfo()) { 773 default: llvm_unreachable("Unknown loc info!"); 774 case CCValAssign::Full: break; 775 case CCValAssign::BCvt: 776 Val = DAG.getNode(ISD::BIT_CONVERT, dl, VA.getValVT(), Val); 777 break; 778 } 779 780 InVals.push_back(Val); 781 } 782 783 return Chain; 784} 785 786/// CreateCopyOfByValArgument - Make a copy of an aggregate at address specified 787/// by "Src" to address "Dst" of size "Size". Alignment information is 788/// specified by the specific parameter attribute. The copy will be passed as 789/// a byval function parameter. 790/// Sometimes what we are copying is the end of a larger object, the part that 791/// does not fit in registers. 792static SDValue 793CreateCopyOfByValArgument(SDValue Src, SDValue Dst, SDValue Chain, 794 ISD::ArgFlagsTy Flags, SelectionDAG &DAG, 795 DebugLoc dl) { 796 SDValue SizeNode = DAG.getConstant(Flags.getByValSize(), MVT::i32); 797 return DAG.getMemcpy(Chain, dl, Dst, Src, SizeNode, Flags.getByValAlign(), 798 /*AlwaysInline=*/false, NULL, 0, NULL, 0); 799} 800 801/// LowerMemOpCallTo - Store the argument to the stack. 802SDValue 803ARMTargetLowering::LowerMemOpCallTo(SDValue Chain, 804 SDValue StackPtr, SDValue Arg, 805 DebugLoc dl, SelectionDAG &DAG, 806 const CCValAssign &VA, 807 ISD::ArgFlagsTy Flags) { 808 unsigned LocMemOffset = VA.getLocMemOffset(); 809 SDValue PtrOff = DAG.getIntPtrConstant(LocMemOffset); 810 PtrOff = DAG.getNode(ISD::ADD, dl, getPointerTy(), StackPtr, PtrOff); 811 if (Flags.isByVal()) { 812 return CreateCopyOfByValArgument(Arg, PtrOff, Chain, Flags, DAG, dl); 813 } 814 return DAG.getStore(Chain, dl, Arg, PtrOff, 815 PseudoSourceValue::getStack(), LocMemOffset); 816} 817 818void ARMTargetLowering::PassF64ArgInRegs(DebugLoc dl, SelectionDAG &DAG, 819 SDValue Chain, SDValue &Arg, 820 RegsToPassVector &RegsToPass, 821 CCValAssign &VA, CCValAssign &NextVA, 822 SDValue &StackPtr, 823 SmallVector<SDValue, 8> &MemOpChains, 824 ISD::ArgFlagsTy Flags) { 825 826 SDValue fmrrd = DAG.getNode(ARMISD::FMRRD, dl, 827 DAG.getVTList(MVT::i32, MVT::i32), Arg); 828 RegsToPass.push_back(std::make_pair(VA.getLocReg(), fmrrd)); 829 830 if (NextVA.isRegLoc()) 831 RegsToPass.push_back(std::make_pair(NextVA.getLocReg(), fmrrd.getValue(1))); 832 else { 833 assert(NextVA.isMemLoc()); 834 if (StackPtr.getNode() == 0) 835 StackPtr = DAG.getCopyFromReg(Chain, dl, ARM::SP, getPointerTy()); 836 837 MemOpChains.push_back(LowerMemOpCallTo(Chain, StackPtr, fmrrd.getValue(1), 838 dl, DAG, NextVA, 839 Flags)); 840 } 841} 842 843/// LowerCall - Lowering a call into a callseq_start <- 844/// ARMISD:CALL <- callseq_end chain. Also add input and output parameter 845/// nodes. 846SDValue 847ARMTargetLowering::LowerCall(SDValue Chain, SDValue Callee, 848 unsigned CallConv, bool isVarArg, 849 bool isTailCall, 850 const SmallVectorImpl<ISD::OutputArg> &Outs, 851 const SmallVectorImpl<ISD::InputArg> &Ins, 852 DebugLoc dl, SelectionDAG &DAG, 853 SmallVectorImpl<SDValue> &InVals) { 854 855 // Analyze operands of the call, assigning locations to each operand. 856 SmallVector<CCValAssign, 16> ArgLocs; 857 CCState CCInfo(CallConv, isVarArg, getTargetMachine(), ArgLocs, 858 *DAG.getContext()); 859 CCInfo.AnalyzeCallOperands(Outs, 860 CCAssignFnForNode(CallConv, /* Return*/ false, 861 isVarArg)); 862 863 // Get a count of how many bytes are to be pushed on the stack. 864 unsigned NumBytes = CCInfo.getNextStackOffset(); 865 866 // Adjust the stack pointer for the new arguments... 867 // These operations are automatically eliminated by the prolog/epilog pass 868 Chain = DAG.getCALLSEQ_START(Chain, DAG.getIntPtrConstant(NumBytes, true)); 869 870 SDValue StackPtr = DAG.getRegister(ARM::SP, MVT::i32); 871 872 RegsToPassVector RegsToPass; 873 SmallVector<SDValue, 8> MemOpChains; 874 875 // Walk the register/memloc assignments, inserting copies/loads. In the case 876 // of tail call optimization, arguments are handled later. 877 for (unsigned i = 0, realArgIdx = 0, e = ArgLocs.size(); 878 i != e; 879 ++i, ++realArgIdx) { 880 CCValAssign &VA = ArgLocs[i]; 881 SDValue Arg = Outs[realArgIdx].Val; 882 ISD::ArgFlagsTy Flags = Outs[realArgIdx].Flags; 883 884 // Promote the value if needed. 885 switch (VA.getLocInfo()) { 886 default: llvm_unreachable("Unknown loc info!"); 887 case CCValAssign::Full: break; 888 case CCValAssign::SExt: 889 Arg = DAG.getNode(ISD::SIGN_EXTEND, dl, VA.getLocVT(), Arg); 890 break; 891 case CCValAssign::ZExt: 892 Arg = DAG.getNode(ISD::ZERO_EXTEND, dl, VA.getLocVT(), Arg); 893 break; 894 case CCValAssign::AExt: 895 Arg = DAG.getNode(ISD::ANY_EXTEND, dl, VA.getLocVT(), Arg); 896 break; 897 case CCValAssign::BCvt: 898 Arg = DAG.getNode(ISD::BIT_CONVERT, dl, VA.getLocVT(), Arg); 899 break; 900 } 901 902 // f64 and v2f64 might be passed in i32 pairs and must be split into pieces 903 if (VA.needsCustom()) { 904 if (VA.getLocVT() == MVT::v2f64) { 905 SDValue Op0 = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64, Arg, 906 DAG.getConstant(0, MVT::i32)); 907 SDValue Op1 = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64, Arg, 908 DAG.getConstant(1, MVT::i32)); 909 910 PassF64ArgInRegs(dl, DAG, Chain, Op0, RegsToPass, 911 VA, ArgLocs[++i], StackPtr, MemOpChains, Flags); 912 913 VA = ArgLocs[++i]; // skip ahead to next loc 914 if (VA.isRegLoc()) { 915 PassF64ArgInRegs(dl, DAG, Chain, Op1, RegsToPass, 916 VA, ArgLocs[++i], StackPtr, MemOpChains, Flags); 917 } else { 918 assert(VA.isMemLoc()); 919 if (StackPtr.getNode() == 0) 920 StackPtr = DAG.getCopyFromReg(Chain, dl, ARM::SP, getPointerTy()); 921 922 MemOpChains.push_back(LowerMemOpCallTo(Chain, StackPtr, Op1, 923 dl, DAG, VA, Flags)); 924 } 925 } else { 926 PassF64ArgInRegs(dl, DAG, Chain, Arg, RegsToPass, VA, ArgLocs[++i], 927 StackPtr, MemOpChains, Flags); 928 } 929 } else if (VA.isRegLoc()) { 930 RegsToPass.push_back(std::make_pair(VA.getLocReg(), Arg)); 931 } else { 932 assert(VA.isMemLoc()); 933 if (StackPtr.getNode() == 0) 934 StackPtr = DAG.getCopyFromReg(Chain, dl, ARM::SP, getPointerTy()); 935 936 MemOpChains.push_back(LowerMemOpCallTo(Chain, StackPtr, Arg, 937 dl, DAG, VA, Flags)); 938 } 939 } 940 941 if (!MemOpChains.empty()) 942 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other, 943 &MemOpChains[0], MemOpChains.size()); 944 945 // Build a sequence of copy-to-reg nodes chained together with token chain 946 // and flag operands which copy the outgoing args into the appropriate regs. 947 SDValue InFlag; 948 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i) { 949 Chain = DAG.getCopyToReg(Chain, dl, RegsToPass[i].first, 950 RegsToPass[i].second, InFlag); 951 InFlag = Chain.getValue(1); 952 } 953 954 // If the callee is a GlobalAddress/ExternalSymbol node (quite common, every 955 // direct call is) turn it into a TargetGlobalAddress/TargetExternalSymbol 956 // node so that legalize doesn't hack it. 957 bool isDirect = false; 958 bool isARMFunc = false; 959 bool isLocalARMFunc = false; 960 if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee)) { 961 GlobalValue *GV = G->getGlobal(); 962 isDirect = true; 963 bool isExt = GV->isDeclaration() || GV->isWeakForLinker(); 964 bool isStub = (isExt && Subtarget->isTargetDarwin()) && 965 getTargetMachine().getRelocationModel() != Reloc::Static; 966 isARMFunc = !Subtarget->isThumb() || isStub; 967 // ARM call to a local ARM function is predicable. 968 isLocalARMFunc = !Subtarget->isThumb() && !isExt; 969 // tBX takes a register source operand. 970 if (isARMFunc && Subtarget->isThumb1Only() && !Subtarget->hasV5TOps()) { 971 ARMConstantPoolValue *CPV = new ARMConstantPoolValue(GV, ARMPCLabelIndex, 972 ARMCP::CPStub, 4); 973 SDValue CPAddr = DAG.getTargetConstantPool(CPV, getPointerTy(), 4); 974 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr); 975 Callee = DAG.getLoad(getPointerTy(), dl, 976 DAG.getEntryNode(), CPAddr, NULL, 0); 977 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex++, MVT::i32); 978 Callee = DAG.getNode(ARMISD::PIC_ADD, dl, 979 getPointerTy(), Callee, PICLabel); 980 } else 981 Callee = DAG.getTargetGlobalAddress(GV, getPointerTy()); 982 } else if (ExternalSymbolSDNode *S = dyn_cast<ExternalSymbolSDNode>(Callee)) { 983 isDirect = true; 984 bool isStub = Subtarget->isTargetDarwin() && 985 getTargetMachine().getRelocationModel() != Reloc::Static; 986 isARMFunc = !Subtarget->isThumb() || isStub; 987 // tBX takes a register source operand. 988 const char *Sym = S->getSymbol(); 989 if (isARMFunc && Subtarget->isThumb1Only() && !Subtarget->hasV5TOps()) { 990 ARMConstantPoolValue *CPV = new ARMConstantPoolValue(*DAG.getContext(), 991 Sym, ARMPCLabelIndex, 992 ARMCP::CPStub, 4); 993 SDValue CPAddr = DAG.getTargetConstantPool(CPV, getPointerTy(), 4); 994 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr); 995 Callee = DAG.getLoad(getPointerTy(), dl, 996 DAG.getEntryNode(), CPAddr, NULL, 0); 997 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex++, MVT::i32); 998 Callee = DAG.getNode(ARMISD::PIC_ADD, dl, 999 getPointerTy(), Callee, PICLabel); 1000 } else 1001 Callee = DAG.getTargetExternalSymbol(Sym, getPointerTy()); 1002 } 1003 1004 // FIXME: handle tail calls differently. 1005 unsigned CallOpc; 1006 if (Subtarget->isThumb()) { 1007 if ((!isDirect || isARMFunc) && !Subtarget->hasV5TOps()) 1008 CallOpc = ARMISD::CALL_NOLINK; 1009 else 1010 CallOpc = isARMFunc ? ARMISD::CALL : ARMISD::tCALL; 1011 } else { 1012 CallOpc = (isDirect || Subtarget->hasV5TOps()) 1013 ? (isLocalARMFunc ? ARMISD::CALL_PRED : ARMISD::CALL) 1014 : ARMISD::CALL_NOLINK; 1015 } 1016 if (CallOpc == ARMISD::CALL_NOLINK && !Subtarget->isThumb1Only()) { 1017 // implicit def LR - LR mustn't be allocated as GRP:$dst of CALL_NOLINK 1018 Chain = DAG.getCopyToReg(Chain, dl, ARM::LR, DAG.getUNDEF(MVT::i32),InFlag); 1019 InFlag = Chain.getValue(1); 1020 } 1021 1022 std::vector<SDValue> Ops; 1023 Ops.push_back(Chain); 1024 Ops.push_back(Callee); 1025 1026 // Add argument registers to the end of the list so that they are known live 1027 // into the call. 1028 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i) 1029 Ops.push_back(DAG.getRegister(RegsToPass[i].first, 1030 RegsToPass[i].second.getValueType())); 1031 1032 if (InFlag.getNode()) 1033 Ops.push_back(InFlag); 1034 // Returns a chain and a flag for retval copy to use. 1035 Chain = DAG.getNode(CallOpc, dl, DAG.getVTList(MVT::Other, MVT::Flag), 1036 &Ops[0], Ops.size()); 1037 InFlag = Chain.getValue(1); 1038 1039 Chain = DAG.getCALLSEQ_END(Chain, DAG.getIntPtrConstant(NumBytes, true), 1040 DAG.getIntPtrConstant(0, true), InFlag); 1041 if (!Ins.empty()) 1042 InFlag = Chain.getValue(1); 1043 1044 // Handle result values, copying them out of physregs into vregs that we 1045 // return. 1046 return LowerCallResult(Chain, InFlag, CallConv, isVarArg, Ins, 1047 dl, DAG, InVals); 1048} 1049 1050SDValue 1051ARMTargetLowering::LowerReturn(SDValue Chain, 1052 unsigned CallConv, bool isVarArg, 1053 const SmallVectorImpl<ISD::OutputArg> &Outs, 1054 DebugLoc dl, SelectionDAG &DAG) { 1055 1056 // CCValAssign - represent the assignment of the return value to a location. 1057 SmallVector<CCValAssign, 16> RVLocs; 1058 1059 // CCState - Info about the registers and stack slots. 1060 CCState CCInfo(CallConv, isVarArg, getTargetMachine(), RVLocs, 1061 *DAG.getContext()); 1062 1063 // Analyze outgoing return values. 1064 CCInfo.AnalyzeReturn(Outs, CCAssignFnForNode(CallConv, /* Return */ true, 1065 isVarArg)); 1066 1067 // If this is the first return lowered for this function, add 1068 // the regs to the liveout set for the function. 1069 if (DAG.getMachineFunction().getRegInfo().liveout_empty()) { 1070 for (unsigned i = 0; i != RVLocs.size(); ++i) 1071 if (RVLocs[i].isRegLoc()) 1072 DAG.getMachineFunction().getRegInfo().addLiveOut(RVLocs[i].getLocReg()); 1073 } 1074 1075 SDValue Flag; 1076 1077 // Copy the result values into the output registers. 1078 for (unsigned i = 0, realRVLocIdx = 0; 1079 i != RVLocs.size(); 1080 ++i, ++realRVLocIdx) { 1081 CCValAssign &VA = RVLocs[i]; 1082 assert(VA.isRegLoc() && "Can only return in registers!"); 1083 1084 SDValue Arg = Outs[realRVLocIdx].Val; 1085 1086 switch (VA.getLocInfo()) { 1087 default: llvm_unreachable("Unknown loc info!"); 1088 case CCValAssign::Full: break; 1089 case CCValAssign::BCvt: 1090 Arg = DAG.getNode(ISD::BIT_CONVERT, dl, VA.getLocVT(), Arg); 1091 break; 1092 } 1093 1094 if (VA.needsCustom()) { 1095 if (VA.getLocVT() == MVT::v2f64) { 1096 // Extract the first half and return it in two registers. 1097 SDValue Half = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64, Arg, 1098 DAG.getConstant(0, MVT::i32)); 1099 SDValue HalfGPRs = DAG.getNode(ARMISD::FMRRD, dl, 1100 DAG.getVTList(MVT::i32, MVT::i32), Half); 1101 1102 Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(), HalfGPRs, Flag); 1103 Flag = Chain.getValue(1); 1104 VA = RVLocs[++i]; // skip ahead to next loc 1105 Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(), 1106 HalfGPRs.getValue(1), Flag); 1107 Flag = Chain.getValue(1); 1108 VA = RVLocs[++i]; // skip ahead to next loc 1109 1110 // Extract the 2nd half and fall through to handle it as an f64 value. 1111 Arg = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64, Arg, 1112 DAG.getConstant(1, MVT::i32)); 1113 } 1114 // Legalize ret f64 -> ret 2 x i32. We always have fmrrd if f64 is 1115 // available. 1116 SDValue fmrrd = DAG.getNode(ARMISD::FMRRD, dl, 1117 DAG.getVTList(MVT::i32, MVT::i32), &Arg, 1); 1118 Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(), fmrrd, Flag); 1119 Flag = Chain.getValue(1); 1120 VA = RVLocs[++i]; // skip ahead to next loc 1121 Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(), fmrrd.getValue(1), 1122 Flag); 1123 } else 1124 Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(), Arg, Flag); 1125 1126 // Guarantee that all emitted copies are 1127 // stuck together, avoiding something bad. 1128 Flag = Chain.getValue(1); 1129 } 1130 1131 SDValue result; 1132 if (Flag.getNode()) 1133 result = DAG.getNode(ARMISD::RET_FLAG, dl, MVT::Other, Chain, Flag); 1134 else // Return Void 1135 result = DAG.getNode(ARMISD::RET_FLAG, dl, MVT::Other, Chain); 1136 1137 return result; 1138} 1139 1140// ConstantPool, JumpTable, GlobalAddress, and ExternalSymbol are lowered as 1141// their target counterpart wrapped in the ARMISD::Wrapper node. Suppose N is 1142// one of the above mentioned nodes. It has to be wrapped because otherwise 1143// Select(N) returns N. So the raw TargetGlobalAddress nodes, etc. can only 1144// be used to form addressing mode. These wrapped nodes will be selected 1145// into MOVi. 1146static SDValue LowerConstantPool(SDValue Op, SelectionDAG &DAG) { 1147 EVT PtrVT = Op.getValueType(); 1148 // FIXME there is no actual debug info here 1149 DebugLoc dl = Op.getDebugLoc(); 1150 ConstantPoolSDNode *CP = cast<ConstantPoolSDNode>(Op); 1151 SDValue Res; 1152 if (CP->isMachineConstantPoolEntry()) 1153 Res = DAG.getTargetConstantPool(CP->getMachineCPVal(), PtrVT, 1154 CP->getAlignment()); 1155 else 1156 Res = DAG.getTargetConstantPool(CP->getConstVal(), PtrVT, 1157 CP->getAlignment()); 1158 return DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, Res); 1159} 1160 1161// Lower ISD::GlobalTLSAddress using the "general dynamic" model 1162SDValue 1163ARMTargetLowering::LowerToTLSGeneralDynamicModel(GlobalAddressSDNode *GA, 1164 SelectionDAG &DAG) { 1165 DebugLoc dl = GA->getDebugLoc(); 1166 EVT PtrVT = getPointerTy(); 1167 unsigned char PCAdj = Subtarget->isThumb() ? 4 : 8; 1168 ARMConstantPoolValue *CPV = 1169 new ARMConstantPoolValue(GA->getGlobal(), ARMPCLabelIndex, ARMCP::CPValue, 1170 PCAdj, "tlsgd", true); 1171 SDValue Argument = DAG.getTargetConstantPool(CPV, PtrVT, 4); 1172 Argument = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, Argument); 1173 Argument = DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), Argument, NULL, 0); 1174 SDValue Chain = Argument.getValue(1); 1175 1176 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex++, MVT::i32); 1177 Argument = DAG.getNode(ARMISD::PIC_ADD, dl, PtrVT, Argument, PICLabel); 1178 1179 // call __tls_get_addr. 1180 ArgListTy Args; 1181 ArgListEntry Entry; 1182 Entry.Node = Argument; 1183 Entry.Ty = (const Type *) Type::getInt32Ty(*DAG.getContext()); 1184 Args.push_back(Entry); 1185 // FIXME: is there useful debug info available here? 1186 std::pair<SDValue, SDValue> CallResult = 1187 LowerCallTo(Chain, (const Type *) Type::getInt32Ty(*DAG.getContext()), 1188 false, false, false, false, 1189 0, CallingConv::C, false, /*isReturnValueUsed=*/true, 1190 DAG.getExternalSymbol("__tls_get_addr", PtrVT), Args, DAG, dl); 1191 return CallResult.first; 1192} 1193 1194// Lower ISD::GlobalTLSAddress using the "initial exec" or 1195// "local exec" model. 1196SDValue 1197ARMTargetLowering::LowerToTLSExecModels(GlobalAddressSDNode *GA, 1198 SelectionDAG &DAG) { 1199 GlobalValue *GV = GA->getGlobal(); 1200 DebugLoc dl = GA->getDebugLoc(); 1201 SDValue Offset; 1202 SDValue Chain = DAG.getEntryNode(); 1203 EVT PtrVT = getPointerTy(); 1204 // Get the Thread Pointer 1205 SDValue ThreadPointer = DAG.getNode(ARMISD::THREAD_POINTER, dl, PtrVT); 1206 1207 if (GV->isDeclaration()) { 1208 // initial exec model 1209 unsigned char PCAdj = Subtarget->isThumb() ? 4 : 8; 1210 ARMConstantPoolValue *CPV = 1211 new ARMConstantPoolValue(GA->getGlobal(), ARMPCLabelIndex, ARMCP::CPValue, 1212 PCAdj, "gottpoff", true); 1213 Offset = DAG.getTargetConstantPool(CPV, PtrVT, 4); 1214 Offset = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, Offset); 1215 Offset = DAG.getLoad(PtrVT, dl, Chain, Offset, NULL, 0); 1216 Chain = Offset.getValue(1); 1217 1218 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex++, MVT::i32); 1219 Offset = DAG.getNode(ARMISD::PIC_ADD, dl, PtrVT, Offset, PICLabel); 1220 1221 Offset = DAG.getLoad(PtrVT, dl, Chain, Offset, NULL, 0); 1222 } else { 1223 // local exec model 1224 ARMConstantPoolValue *CPV = 1225 new ARMConstantPoolValue(GV, ARMCP::CPValue, "tpoff"); 1226 Offset = DAG.getTargetConstantPool(CPV, PtrVT, 4); 1227 Offset = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, Offset); 1228 Offset = DAG.getLoad(PtrVT, dl, Chain, Offset, NULL, 0); 1229 } 1230 1231 // The address of the thread local variable is the add of the thread 1232 // pointer with the offset of the variable. 1233 return DAG.getNode(ISD::ADD, dl, PtrVT, ThreadPointer, Offset); 1234} 1235 1236SDValue 1237ARMTargetLowering::LowerGlobalTLSAddress(SDValue Op, SelectionDAG &DAG) { 1238 // TODO: implement the "local dynamic" model 1239 assert(Subtarget->isTargetELF() && 1240 "TLS not implemented for non-ELF targets"); 1241 GlobalAddressSDNode *GA = cast<GlobalAddressSDNode>(Op); 1242 // If the relocation model is PIC, use the "General Dynamic" TLS Model, 1243 // otherwise use the "Local Exec" TLS Model 1244 if (getTargetMachine().getRelocationModel() == Reloc::PIC_) 1245 return LowerToTLSGeneralDynamicModel(GA, DAG); 1246 else 1247 return LowerToTLSExecModels(GA, DAG); 1248} 1249 1250SDValue ARMTargetLowering::LowerGlobalAddressELF(SDValue Op, 1251 SelectionDAG &DAG) { 1252 EVT PtrVT = getPointerTy(); 1253 DebugLoc dl = Op.getDebugLoc(); 1254 GlobalValue *GV = cast<GlobalAddressSDNode>(Op)->getGlobal(); 1255 Reloc::Model RelocM = getTargetMachine().getRelocationModel(); 1256 if (RelocM == Reloc::PIC_) { 1257 bool UseGOTOFF = GV->hasLocalLinkage() || GV->hasHiddenVisibility(); 1258 ARMConstantPoolValue *CPV = 1259 new ARMConstantPoolValue(GV, ARMCP::CPValue, UseGOTOFF ? "GOTOFF":"GOT"); 1260 SDValue CPAddr = DAG.getTargetConstantPool(CPV, PtrVT, 4); 1261 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr); 1262 SDValue Result = DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), 1263 CPAddr, NULL, 0); 1264 SDValue Chain = Result.getValue(1); 1265 SDValue GOT = DAG.getGLOBAL_OFFSET_TABLE(PtrVT); 1266 Result = DAG.getNode(ISD::ADD, dl, PtrVT, Result, GOT); 1267 if (!UseGOTOFF) 1268 Result = DAG.getLoad(PtrVT, dl, Chain, Result, NULL, 0); 1269 return Result; 1270 } else { 1271 SDValue CPAddr = DAG.getTargetConstantPool(GV, PtrVT, 4); 1272 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr); 1273 return DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), CPAddr, NULL, 0); 1274 } 1275} 1276 1277/// GVIsIndirectSymbol - true if the GV will be accessed via an indirect symbol 1278/// even in non-static mode. 1279static bool GVIsIndirectSymbol(GlobalValue *GV, Reloc::Model RelocM) { 1280 // If symbol visibility is hidden, the extra load is not needed if 1281 // the symbol is definitely defined in the current translation unit. 1282 bool isDecl = GV->isDeclaration() || GV->hasAvailableExternallyLinkage(); 1283 if (GV->hasHiddenVisibility() && (!isDecl && !GV->hasCommonLinkage())) 1284 return false; 1285 return RelocM != Reloc::Static && (isDecl || GV->isWeakForLinker()); 1286} 1287 1288SDValue ARMTargetLowering::LowerGlobalAddressDarwin(SDValue Op, 1289 SelectionDAG &DAG) { 1290 EVT PtrVT = getPointerTy(); 1291 DebugLoc dl = Op.getDebugLoc(); 1292 GlobalValue *GV = cast<GlobalAddressSDNode>(Op)->getGlobal(); 1293 Reloc::Model RelocM = getTargetMachine().getRelocationModel(); 1294 bool IsIndirect = GVIsIndirectSymbol(GV, RelocM); 1295 SDValue CPAddr; 1296 if (RelocM == Reloc::Static) 1297 CPAddr = DAG.getTargetConstantPool(GV, PtrVT, 4); 1298 else { 1299 unsigned PCAdj = (RelocM != Reloc::PIC_) 1300 ? 0 : (Subtarget->isThumb() ? 4 : 8); 1301 ARMCP::ARMCPKind Kind = IsIndirect ? ARMCP::CPNonLazyPtr 1302 : ARMCP::CPValue; 1303 ARMConstantPoolValue *CPV = new ARMConstantPoolValue(GV, ARMPCLabelIndex, 1304 Kind, PCAdj); 1305 CPAddr = DAG.getTargetConstantPool(CPV, PtrVT, 4); 1306 } 1307 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr); 1308 1309 SDValue Result = DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), CPAddr, NULL, 0); 1310 SDValue Chain = Result.getValue(1); 1311 1312 if (RelocM == Reloc::PIC_) { 1313 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex++, MVT::i32); 1314 Result = DAG.getNode(ARMISD::PIC_ADD, dl, PtrVT, Result, PICLabel); 1315 } 1316 if (IsIndirect) 1317 Result = DAG.getLoad(PtrVT, dl, Chain, Result, NULL, 0); 1318 1319 return Result; 1320} 1321 1322SDValue ARMTargetLowering::LowerGLOBAL_OFFSET_TABLE(SDValue Op, 1323 SelectionDAG &DAG){ 1324 assert(Subtarget->isTargetELF() && 1325 "GLOBAL OFFSET TABLE not implemented for non-ELF targets"); 1326 EVT PtrVT = getPointerTy(); 1327 DebugLoc dl = Op.getDebugLoc(); 1328 unsigned PCAdj = Subtarget->isThumb() ? 4 : 8; 1329 ARMConstantPoolValue *CPV = new ARMConstantPoolValue(*DAG.getContext(), 1330 "_GLOBAL_OFFSET_TABLE_", 1331 ARMPCLabelIndex, 1332 ARMCP::CPValue, PCAdj); 1333 SDValue CPAddr = DAG.getTargetConstantPool(CPV, PtrVT, 4); 1334 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr); 1335 SDValue Result = DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), CPAddr, NULL, 0); 1336 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex++, MVT::i32); 1337 return DAG.getNode(ARMISD::PIC_ADD, dl, PtrVT, Result, PICLabel); 1338} 1339 1340static SDValue LowerNeonVLDIntrinsic(SDValue Op, SelectionDAG &DAG, 1341 unsigned NumVecs) { 1342 SDNode *Node = Op.getNode(); 1343 EVT VT = Node->getValueType(0); 1344 1345 // No expansion needed for 64-bit vectors. 1346 if (VT.is64BitVector()) 1347 return SDValue(); 1348 1349 // FIXME: We need to expand VLD3 and VLD4 of 128-bit vectors into separate 1350 // operations to load the even and odd registers. 1351 return SDValue(); 1352} 1353 1354static SDValue LowerNeonVSTIntrinsic(SDValue Op, SelectionDAG &DAG, 1355 unsigned NumVecs) { 1356 SDNode *Node = Op.getNode(); 1357 EVT VT = Node->getOperand(3).getValueType(); 1358 1359 // No expansion needed for 64-bit vectors. 1360 if (VT.is64BitVector()) 1361 return SDValue(); 1362 1363 // FIXME: We need to expand VST3 and VST4 of 128-bit vectors into separate 1364 // operations to store the even and odd registers. 1365 return SDValue(); 1366} 1367 1368SDValue 1369ARMTargetLowering::LowerINTRINSIC_W_CHAIN(SDValue Op, SelectionDAG &DAG) { 1370 unsigned IntNo = cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue(); 1371 switch (IntNo) { 1372 case Intrinsic::arm_neon_vld3: 1373 return LowerNeonVLDIntrinsic(Op, DAG, 3); 1374 case Intrinsic::arm_neon_vld4: 1375 return LowerNeonVLDIntrinsic(Op, DAG, 4); 1376 case Intrinsic::arm_neon_vst3: 1377 return LowerNeonVSTIntrinsic(Op, DAG, 3); 1378 case Intrinsic::arm_neon_vst4: 1379 return LowerNeonVSTIntrinsic(Op, DAG, 4); 1380 default: return SDValue(); // Don't custom lower most intrinsics. 1381 } 1382} 1383 1384SDValue 1385ARMTargetLowering::LowerINTRINSIC_WO_CHAIN(SDValue Op, SelectionDAG &DAG) { 1386 unsigned IntNo = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue(); 1387 DebugLoc dl = Op.getDebugLoc(); 1388 switch (IntNo) { 1389 default: return SDValue(); // Don't custom lower most intrinsics. 1390 case Intrinsic::arm_thread_pointer: { 1391 EVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy(); 1392 return DAG.getNode(ARMISD::THREAD_POINTER, dl, PtrVT); 1393 } 1394 case Intrinsic::eh_sjlj_lsda: { 1395 // blah. horrible, horrible hack with the forced magic name. 1396 // really need to clean this up. It belongs in the target-independent 1397 // layer somehow that doesn't require the coupling with the asm 1398 // printer. 1399 MachineFunction &MF = DAG.getMachineFunction(); 1400 EVT PtrVT = getPointerTy(); 1401 DebugLoc dl = Op.getDebugLoc(); 1402 Reloc::Model RelocM = getTargetMachine().getRelocationModel(); 1403 SDValue CPAddr; 1404 unsigned PCAdj = (RelocM != Reloc::PIC_) 1405 ? 0 : (Subtarget->isThumb() ? 4 : 8); 1406 ARMCP::ARMCPKind Kind = ARMCP::CPValue; 1407 // Save off the LSDA name for the AsmPrinter to use when it's time 1408 // to emit the table 1409 std::string LSDAName = "L_lsda_"; 1410 LSDAName += MF.getFunction()->getName(); 1411 ARMConstantPoolValue *CPV = 1412 new ARMConstantPoolValue(*DAG.getContext(), LSDAName.c_str(), 1413 ARMPCLabelIndex, Kind, PCAdj); 1414 CPAddr = DAG.getTargetConstantPool(CPV, PtrVT, 4); 1415 CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr); 1416 SDValue Result = 1417 DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), CPAddr, NULL, 0); 1418 SDValue Chain = Result.getValue(1); 1419 1420 if (RelocM == Reloc::PIC_) { 1421 SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex++, MVT::i32); 1422 Result = DAG.getNode(ARMISD::PIC_ADD, dl, PtrVT, Result, PICLabel); 1423 } 1424 return Result; 1425 } 1426 case Intrinsic::eh_sjlj_setjmp: 1427 return DAG.getNode(ARMISD::EH_SJLJ_SETJMP, dl, MVT::i32, Op.getOperand(1)); 1428 } 1429} 1430 1431static SDValue LowerVASTART(SDValue Op, SelectionDAG &DAG, 1432 unsigned VarArgsFrameIndex) { 1433 // vastart just stores the address of the VarArgsFrameIndex slot into the 1434 // memory location argument. 1435 DebugLoc dl = Op.getDebugLoc(); 1436 EVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy(); 1437 SDValue FR = DAG.getFrameIndex(VarArgsFrameIndex, PtrVT); 1438 const Value *SV = cast<SrcValueSDNode>(Op.getOperand(2))->getValue(); 1439 return DAG.getStore(Op.getOperand(0), dl, FR, Op.getOperand(1), SV, 0); 1440} 1441 1442SDValue 1443ARMTargetLowering::LowerDYNAMIC_STACKALLOC(SDValue Op, SelectionDAG &DAG) { 1444 SDNode *Node = Op.getNode(); 1445 DebugLoc dl = Node->getDebugLoc(); 1446 EVT VT = Node->getValueType(0); 1447 SDValue Chain = Op.getOperand(0); 1448 SDValue Size = Op.getOperand(1); 1449 SDValue Align = Op.getOperand(2); 1450 1451 // Chain the dynamic stack allocation so that it doesn't modify the stack 1452 // pointer when other instructions are using the stack. 1453 Chain = DAG.getCALLSEQ_START(Chain, DAG.getIntPtrConstant(0, true)); 1454 1455 unsigned AlignVal = cast<ConstantSDNode>(Align)->getZExtValue(); 1456 unsigned StackAlign = getTargetMachine().getFrameInfo()->getStackAlignment(); 1457 if (AlignVal > StackAlign) 1458 // Do this now since selection pass cannot introduce new target 1459 // independent node. 1460 Align = DAG.getConstant(-(uint64_t)AlignVal, VT); 1461 1462 // In Thumb1 mode, there isn't a "sub r, sp, r" instruction, we will end up 1463 // using a "add r, sp, r" instead. Negate the size now so we don't have to 1464 // do even more horrible hack later. 1465 MachineFunction &MF = DAG.getMachineFunction(); 1466 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>(); 1467 if (AFI->isThumb1OnlyFunction()) { 1468 bool Negate = true; 1469 ConstantSDNode *C = dyn_cast<ConstantSDNode>(Size); 1470 if (C) { 1471 uint32_t Val = C->getZExtValue(); 1472 if (Val <= 508 && ((Val & 3) == 0)) 1473 Negate = false; 1474 } 1475 if (Negate) 1476 Size = DAG.getNode(ISD::SUB, dl, VT, DAG.getConstant(0, VT), Size); 1477 } 1478 1479 SDVTList VTList = DAG.getVTList(VT, MVT::Other); 1480 SDValue Ops1[] = { Chain, Size, Align }; 1481 SDValue Res = DAG.getNode(ARMISD::DYN_ALLOC, dl, VTList, Ops1, 3); 1482 Chain = Res.getValue(1); 1483 Chain = DAG.getCALLSEQ_END(Chain, DAG.getIntPtrConstant(0, true), 1484 DAG.getIntPtrConstant(0, true), SDValue()); 1485 SDValue Ops2[] = { Res, Chain }; 1486 return DAG.getMergeValues(Ops2, 2, dl); 1487} 1488 1489SDValue 1490ARMTargetLowering::GetF64FormalArgument(CCValAssign &VA, CCValAssign &NextVA, 1491 SDValue &Root, SelectionDAG &DAG, 1492 DebugLoc dl) { 1493 MachineFunction &MF = DAG.getMachineFunction(); 1494 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>(); 1495 1496 TargetRegisterClass *RC; 1497 if (AFI->isThumb1OnlyFunction()) 1498 RC = ARM::tGPRRegisterClass; 1499 else 1500 RC = ARM::GPRRegisterClass; 1501 1502 // Transform the arguments stored in physical registers into virtual ones. 1503 unsigned Reg = MF.addLiveIn(VA.getLocReg(), RC); 1504 SDValue ArgValue = DAG.getCopyFromReg(Root, dl, Reg, MVT::i32); 1505 1506 SDValue ArgValue2; 1507 if (NextVA.isMemLoc()) { 1508 unsigned ArgSize = NextVA.getLocVT().getSizeInBits()/8; 1509 MachineFrameInfo *MFI = MF.getFrameInfo(); 1510 int FI = MFI->CreateFixedObject(ArgSize, NextVA.getLocMemOffset()); 1511 1512 // Create load node to retrieve arguments from the stack. 1513 SDValue FIN = DAG.getFrameIndex(FI, getPointerTy()); 1514 ArgValue2 = DAG.getLoad(MVT::i32, dl, Root, FIN, NULL, 0); 1515 } else { 1516 Reg = MF.addLiveIn(NextVA.getLocReg(), RC); 1517 ArgValue2 = DAG.getCopyFromReg(Root, dl, Reg, MVT::i32); 1518 } 1519 1520 return DAG.getNode(ARMISD::FMDRR, dl, MVT::f64, ArgValue, ArgValue2); 1521} 1522 1523SDValue 1524ARMTargetLowering::LowerFormalArguments(SDValue Chain, 1525 unsigned CallConv, bool isVarArg, 1526 const SmallVectorImpl<ISD::InputArg> 1527 &Ins, 1528 DebugLoc dl, SelectionDAG &DAG, 1529 SmallVectorImpl<SDValue> &InVals) { 1530 1531 MachineFunction &MF = DAG.getMachineFunction(); 1532 MachineFrameInfo *MFI = MF.getFrameInfo(); 1533 1534 ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>(); 1535 1536 // Assign locations to all of the incoming arguments. 1537 SmallVector<CCValAssign, 16> ArgLocs; 1538 CCState CCInfo(CallConv, isVarArg, getTargetMachine(), ArgLocs, 1539 *DAG.getContext()); 1540 CCInfo.AnalyzeFormalArguments(Ins, 1541 CCAssignFnForNode(CallConv, /* Return*/ false, 1542 isVarArg)); 1543 1544 SmallVector<SDValue, 16> ArgValues; 1545 1546 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) { 1547 CCValAssign &VA = ArgLocs[i]; 1548 1549 // Arguments stored in registers. 1550 if (VA.isRegLoc()) { 1551 EVT RegVT = VA.getLocVT(); 1552 1553 SDValue ArgValue; 1554 if (VA.needsCustom()) { 1555 // f64 and vector types are split up into multiple registers or 1556 // combinations of registers and stack slots. 1557 RegVT = MVT::i32; 1558 1559 if (VA.getLocVT() == MVT::v2f64) { 1560 SDValue ArgValue1 = GetF64FormalArgument(VA, ArgLocs[++i], 1561 Chain, DAG, dl); 1562 VA = ArgLocs[++i]; // skip ahead to next loc 1563 SDValue ArgValue2 = GetF64FormalArgument(VA, ArgLocs[++i], 1564 Chain, DAG, dl); 1565 ArgValue = DAG.getNode(ISD::UNDEF, dl, MVT::v2f64); 1566 ArgValue = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v2f64, 1567 ArgValue, ArgValue1, DAG.getIntPtrConstant(0)); 1568 ArgValue = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v2f64, 1569 ArgValue, ArgValue2, DAG.getIntPtrConstant(1)); 1570 } else 1571 ArgValue = GetF64FormalArgument(VA, ArgLocs[++i], Chain, DAG, dl); 1572 1573 } else { 1574 TargetRegisterClass *RC; 1575 1576 if (RegVT == MVT::f32) 1577 RC = ARM::SPRRegisterClass; 1578 else if (RegVT == MVT::f64) 1579 RC = ARM::DPRRegisterClass; 1580 else if (RegVT == MVT::v2f64) 1581 RC = ARM::QPRRegisterClass; 1582 else if (RegVT == MVT::i32) 1583 RC = (AFI->isThumb1OnlyFunction() ? 1584 ARM::tGPRRegisterClass : ARM::GPRRegisterClass); 1585 else 1586 llvm_unreachable("RegVT not supported by FORMAL_ARGUMENTS Lowering"); 1587 1588 // Transform the arguments in physical registers into virtual ones. 1589 unsigned Reg = MF.addLiveIn(VA.getLocReg(), RC); 1590 ArgValue = DAG.getCopyFromReg(Chain, dl, Reg, RegVT); 1591 } 1592 1593 // If this is an 8 or 16-bit value, it is really passed promoted 1594 // to 32 bits. Insert an assert[sz]ext to capture this, then 1595 // truncate to the right size. 1596 switch (VA.getLocInfo()) { 1597 default: llvm_unreachable("Unknown loc info!"); 1598 case CCValAssign::Full: break; 1599 case CCValAssign::BCvt: 1600 ArgValue = DAG.getNode(ISD::BIT_CONVERT, dl, VA.getValVT(), ArgValue); 1601 break; 1602 case CCValAssign::SExt: 1603 ArgValue = DAG.getNode(ISD::AssertSext, dl, RegVT, ArgValue, 1604 DAG.getValueType(VA.getValVT())); 1605 ArgValue = DAG.getNode(ISD::TRUNCATE, dl, VA.getValVT(), ArgValue); 1606 break; 1607 case CCValAssign::ZExt: 1608 ArgValue = DAG.getNode(ISD::AssertZext, dl, RegVT, ArgValue, 1609 DAG.getValueType(VA.getValVT())); 1610 ArgValue = DAG.getNode(ISD::TRUNCATE, dl, VA.getValVT(), ArgValue); 1611 break; 1612 } 1613 1614 InVals.push_back(ArgValue); 1615 1616 } else { // VA.isRegLoc() 1617 1618 // sanity check 1619 assert(VA.isMemLoc()); 1620 assert(VA.getValVT() != MVT::i64 && "i64 should already be lowered"); 1621 1622 unsigned ArgSize = VA.getLocVT().getSizeInBits()/8; 1623 int FI = MFI->CreateFixedObject(ArgSize, VA.getLocMemOffset()); 1624 1625 // Create load nodes to retrieve arguments from the stack. 1626 SDValue FIN = DAG.getFrameIndex(FI, getPointerTy()); 1627 InVals.push_back(DAG.getLoad(VA.getValVT(), dl, Chain, FIN, NULL, 0)); 1628 } 1629 } 1630 1631 // varargs 1632 if (isVarArg) { 1633 static const unsigned GPRArgRegs[] = { 1634 ARM::R0, ARM::R1, ARM::R2, ARM::R3 1635 }; 1636 1637 unsigned NumGPRs = CCInfo.getFirstUnallocated 1638 (GPRArgRegs, sizeof(GPRArgRegs) / sizeof(GPRArgRegs[0])); 1639 1640 unsigned Align = MF.getTarget().getFrameInfo()->getStackAlignment(); 1641 unsigned VARegSize = (4 - NumGPRs) * 4; 1642 unsigned VARegSaveSize = (VARegSize + Align - 1) & ~(Align - 1); 1643 unsigned ArgOffset = 0; 1644 if (VARegSaveSize) { 1645 // If this function is vararg, store any remaining integer argument regs 1646 // to their spots on the stack so that they may be loaded by deferencing 1647 // the result of va_next. 1648 AFI->setVarArgsRegSaveSize(VARegSaveSize); 1649 ArgOffset = CCInfo.getNextStackOffset(); 1650 VarArgsFrameIndex = MFI->CreateFixedObject(VARegSaveSize, ArgOffset + 1651 VARegSaveSize - VARegSize); 1652 SDValue FIN = DAG.getFrameIndex(VarArgsFrameIndex, getPointerTy()); 1653 1654 SmallVector<SDValue, 4> MemOps; 1655 for (; NumGPRs < 4; ++NumGPRs) { 1656 TargetRegisterClass *RC; 1657 if (AFI->isThumb1OnlyFunction()) 1658 RC = ARM::tGPRRegisterClass; 1659 else 1660 RC = ARM::GPRRegisterClass; 1661 1662 unsigned VReg = MF.addLiveIn(GPRArgRegs[NumGPRs], RC); 1663 SDValue Val = DAG.getCopyFromReg(Chain, dl, VReg, MVT::i32); 1664 SDValue Store = DAG.getStore(Val.getValue(1), dl, Val, FIN, NULL, 0); 1665 MemOps.push_back(Store); 1666 FIN = DAG.getNode(ISD::ADD, dl, getPointerTy(), FIN, 1667 DAG.getConstant(4, getPointerTy())); 1668 } 1669 if (!MemOps.empty()) 1670 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other, 1671 &MemOps[0], MemOps.size()); 1672 } else 1673 // This will point to the next argument passed via stack. 1674 VarArgsFrameIndex = MFI->CreateFixedObject(4, ArgOffset); 1675 } 1676 1677 return Chain; 1678} 1679 1680/// isFloatingPointZero - Return true if this is +0.0. 1681static bool isFloatingPointZero(SDValue Op) { 1682 if (ConstantFPSDNode *CFP = dyn_cast<ConstantFPSDNode>(Op)) 1683 return CFP->getValueAPF().isPosZero(); 1684 else if (ISD::isEXTLoad(Op.getNode()) || ISD::isNON_EXTLoad(Op.getNode())) { 1685 // Maybe this has already been legalized into the constant pool? 1686 if (Op.getOperand(1).getOpcode() == ARMISD::Wrapper) { 1687 SDValue WrapperOp = Op.getOperand(1).getOperand(0); 1688 if (ConstantPoolSDNode *CP = dyn_cast<ConstantPoolSDNode>(WrapperOp)) 1689 if (ConstantFP *CFP = dyn_cast<ConstantFP>(CP->getConstVal())) 1690 return CFP->getValueAPF().isPosZero(); 1691 } 1692 } 1693 return false; 1694} 1695 1696static bool isLegalCmpImmediate(unsigned C, bool isThumb1Only) { 1697 return ( isThumb1Only && (C & ~255U) == 0) || 1698 (!isThumb1Only && ARM_AM::getSOImmVal(C) != -1); 1699} 1700 1701/// Returns appropriate ARM CMP (cmp) and corresponding condition code for 1702/// the given operands. 1703static SDValue getARMCmp(SDValue LHS, SDValue RHS, ISD::CondCode CC, 1704 SDValue &ARMCC, SelectionDAG &DAG, bool isThumb1Only, 1705 DebugLoc dl) { 1706 if (ConstantSDNode *RHSC = dyn_cast<ConstantSDNode>(RHS.getNode())) { 1707 unsigned C = RHSC->getZExtValue(); 1708 if (!isLegalCmpImmediate(C, isThumb1Only)) { 1709 // Constant does not fit, try adjusting it by one? 1710 switch (CC) { 1711 default: break; 1712 case ISD::SETLT: 1713 case ISD::SETGE: 1714 if (isLegalCmpImmediate(C-1, isThumb1Only)) { 1715 CC = (CC == ISD::SETLT) ? ISD::SETLE : ISD::SETGT; 1716 RHS = DAG.getConstant(C-1, MVT::i32); 1717 } 1718 break; 1719 case ISD::SETULT: 1720 case ISD::SETUGE: 1721 if (C > 0 && isLegalCmpImmediate(C-1, isThumb1Only)) { 1722 CC = (CC == ISD::SETULT) ? ISD::SETULE : ISD::SETUGT; 1723 RHS = DAG.getConstant(C-1, MVT::i32); 1724 } 1725 break; 1726 case ISD::SETLE: 1727 case ISD::SETGT: 1728 if (isLegalCmpImmediate(C+1, isThumb1Only)) { 1729 CC = (CC == ISD::SETLE) ? ISD::SETLT : ISD::SETGE; 1730 RHS = DAG.getConstant(C+1, MVT::i32); 1731 } 1732 break; 1733 case ISD::SETULE: 1734 case ISD::SETUGT: 1735 if (C < 0xffffffff && isLegalCmpImmediate(C+1, isThumb1Only)) { 1736 CC = (CC == ISD::SETULE) ? ISD::SETULT : ISD::SETUGE; 1737 RHS = DAG.getConstant(C+1, MVT::i32); 1738 } 1739 break; 1740 } 1741 } 1742 } 1743 1744 ARMCC::CondCodes CondCode = IntCCToARMCC(CC); 1745 ARMISD::NodeType CompareType; 1746 switch (CondCode) { 1747 default: 1748 CompareType = ARMISD::CMP; 1749 break; 1750 case ARMCC::EQ: 1751 case ARMCC::NE: 1752 // Uses only Z Flag 1753 CompareType = ARMISD::CMPZ; 1754 break; 1755 } 1756 ARMCC = DAG.getConstant(CondCode, MVT::i32); 1757 return DAG.getNode(CompareType, dl, MVT::Flag, LHS, RHS); 1758} 1759 1760/// Returns a appropriate VFP CMP (fcmp{s|d}+fmstat) for the given operands. 1761static SDValue getVFPCmp(SDValue LHS, SDValue RHS, SelectionDAG &DAG, 1762 DebugLoc dl) { 1763 SDValue Cmp; 1764 if (!isFloatingPointZero(RHS)) 1765 Cmp = DAG.getNode(ARMISD::CMPFP, dl, MVT::Flag, LHS, RHS); 1766 else 1767 Cmp = DAG.getNode(ARMISD::CMPFPw0, dl, MVT::Flag, LHS); 1768 return DAG.getNode(ARMISD::FMSTAT, dl, MVT::Flag, Cmp); 1769} 1770 1771static SDValue LowerSELECT_CC(SDValue Op, SelectionDAG &DAG, 1772 const ARMSubtarget *ST) { 1773 EVT VT = Op.getValueType(); 1774 SDValue LHS = Op.getOperand(0); 1775 SDValue RHS = Op.getOperand(1); 1776 ISD::CondCode CC = cast<CondCodeSDNode>(Op.getOperand(4))->get(); 1777 SDValue TrueVal = Op.getOperand(2); 1778 SDValue FalseVal = Op.getOperand(3); 1779 DebugLoc dl = Op.getDebugLoc(); 1780 1781 if (LHS.getValueType() == MVT::i32) { 1782 SDValue ARMCC; 1783 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32); 1784 SDValue Cmp = getARMCmp(LHS, RHS, CC, ARMCC, DAG, ST->isThumb1Only(), dl); 1785 return DAG.getNode(ARMISD::CMOV, dl, VT, FalseVal, TrueVal, ARMCC, CCR,Cmp); 1786 } 1787 1788 ARMCC::CondCodes CondCode, CondCode2; 1789 if (FPCCToARMCC(CC, CondCode, CondCode2)) 1790 std::swap(TrueVal, FalseVal); 1791 1792 SDValue ARMCC = DAG.getConstant(CondCode, MVT::i32); 1793 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32); 1794 SDValue Cmp = getVFPCmp(LHS, RHS, DAG, dl); 1795 SDValue Result = DAG.getNode(ARMISD::CMOV, dl, VT, FalseVal, TrueVal, 1796 ARMCC, CCR, Cmp); 1797 if (CondCode2 != ARMCC::AL) { 1798 SDValue ARMCC2 = DAG.getConstant(CondCode2, MVT::i32); 1799 // FIXME: Needs another CMP because flag can have but one use. 1800 SDValue Cmp2 = getVFPCmp(LHS, RHS, DAG, dl); 1801 Result = DAG.getNode(ARMISD::CMOV, dl, VT, 1802 Result, TrueVal, ARMCC2, CCR, Cmp2); 1803 } 1804 return Result; 1805} 1806 1807static SDValue LowerBR_CC(SDValue Op, SelectionDAG &DAG, 1808 const ARMSubtarget *ST) { 1809 SDValue Chain = Op.getOperand(0); 1810 ISD::CondCode CC = cast<CondCodeSDNode>(Op.getOperand(1))->get(); 1811 SDValue LHS = Op.getOperand(2); 1812 SDValue RHS = Op.getOperand(3); 1813 SDValue Dest = Op.getOperand(4); 1814 DebugLoc dl = Op.getDebugLoc(); 1815 1816 if (LHS.getValueType() == MVT::i32) { 1817 SDValue ARMCC; 1818 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32); 1819 SDValue Cmp = getARMCmp(LHS, RHS, CC, ARMCC, DAG, ST->isThumb1Only(), dl); 1820 return DAG.getNode(ARMISD::BRCOND, dl, MVT::Other, 1821 Chain, Dest, ARMCC, CCR,Cmp); 1822 } 1823 1824 assert(LHS.getValueType() == MVT::f32 || LHS.getValueType() == MVT::f64); 1825 ARMCC::CondCodes CondCode, CondCode2; 1826 if (FPCCToARMCC(CC, CondCode, CondCode2)) 1827 // Swap the LHS/RHS of the comparison if needed. 1828 std::swap(LHS, RHS); 1829 1830 SDValue Cmp = getVFPCmp(LHS, RHS, DAG, dl); 1831 SDValue ARMCC = DAG.getConstant(CondCode, MVT::i32); 1832 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32); 1833 SDVTList VTList = DAG.getVTList(MVT::Other, MVT::Flag); 1834 SDValue Ops[] = { Chain, Dest, ARMCC, CCR, Cmp }; 1835 SDValue Res = DAG.getNode(ARMISD::BRCOND, dl, VTList, Ops, 5); 1836 if (CondCode2 != ARMCC::AL) { 1837 ARMCC = DAG.getConstant(CondCode2, MVT::i32); 1838 SDValue Ops[] = { Res, Dest, ARMCC, CCR, Res.getValue(1) }; 1839 Res = DAG.getNode(ARMISD::BRCOND, dl, VTList, Ops, 5); 1840 } 1841 return Res; 1842} 1843 1844SDValue ARMTargetLowering::LowerBR_JT(SDValue Op, SelectionDAG &DAG) { 1845 SDValue Chain = Op.getOperand(0); 1846 SDValue Table = Op.getOperand(1); 1847 SDValue Index = Op.getOperand(2); 1848 DebugLoc dl = Op.getDebugLoc(); 1849 1850 EVT PTy = getPointerTy(); 1851 JumpTableSDNode *JT = cast<JumpTableSDNode>(Table); 1852 ARMFunctionInfo *AFI = DAG.getMachineFunction().getInfo<ARMFunctionInfo>(); 1853 SDValue UId = DAG.getConstant(AFI->createJumpTableUId(), PTy); 1854 SDValue JTI = DAG.getTargetJumpTable(JT->getIndex(), PTy); 1855 Table = DAG.getNode(ARMISD::WrapperJT, dl, MVT::i32, JTI, UId); 1856 Index = DAG.getNode(ISD::MUL, dl, PTy, Index, DAG.getConstant(4, PTy)); 1857 SDValue Addr = DAG.getNode(ISD::ADD, dl, PTy, Index, Table); 1858 if (Subtarget->isThumb2()) { 1859 // Thumb2 uses a two-level jump. That is, it jumps into the jump table 1860 // which does another jump to the destination. This also makes it easier 1861 // to translate it to TBB / TBH later. 1862 // FIXME: This might not work if the function is extremely large. 1863 return DAG.getNode(ARMISD::BR2_JT, dl, MVT::Other, Chain, 1864 Addr, Op.getOperand(2), JTI, UId); 1865 } 1866 if (getTargetMachine().getRelocationModel() == Reloc::PIC_) { 1867 Addr = DAG.getLoad((EVT)MVT::i32, dl, Chain, Addr, NULL, 0); 1868 Chain = Addr.getValue(1); 1869 Addr = DAG.getNode(ISD::ADD, dl, PTy, Addr, Table); 1870 return DAG.getNode(ARMISD::BR_JT, dl, MVT::Other, Chain, Addr, JTI, UId); 1871 } else { 1872 Addr = DAG.getLoad(PTy, dl, Chain, Addr, NULL, 0); 1873 Chain = Addr.getValue(1); 1874 return DAG.getNode(ARMISD::BR_JT, dl, MVT::Other, Chain, Addr, JTI, UId); 1875 } 1876} 1877 1878static SDValue LowerFP_TO_INT(SDValue Op, SelectionDAG &DAG) { 1879 DebugLoc dl = Op.getDebugLoc(); 1880 unsigned Opc = 1881 Op.getOpcode() == ISD::FP_TO_SINT ? ARMISD::FTOSI : ARMISD::FTOUI; 1882 Op = DAG.getNode(Opc, dl, MVT::f32, Op.getOperand(0)); 1883 return DAG.getNode(ISD::BIT_CONVERT, dl, MVT::i32, Op); 1884} 1885 1886static SDValue LowerINT_TO_FP(SDValue Op, SelectionDAG &DAG) { 1887 EVT VT = Op.getValueType(); 1888 DebugLoc dl = Op.getDebugLoc(); 1889 unsigned Opc = 1890 Op.getOpcode() == ISD::SINT_TO_FP ? ARMISD::SITOF : ARMISD::UITOF; 1891 1892 Op = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::f32, Op.getOperand(0)); 1893 return DAG.getNode(Opc, dl, VT, Op); 1894} 1895 1896static SDValue LowerFCOPYSIGN(SDValue Op, SelectionDAG &DAG) { 1897 // Implement fcopysign with a fabs and a conditional fneg. 1898 SDValue Tmp0 = Op.getOperand(0); 1899 SDValue Tmp1 = Op.getOperand(1); 1900 DebugLoc dl = Op.getDebugLoc(); 1901 EVT VT = Op.getValueType(); 1902 EVT SrcVT = Tmp1.getValueType(); 1903 SDValue AbsVal = DAG.getNode(ISD::FABS, dl, VT, Tmp0); 1904 SDValue Cmp = getVFPCmp(Tmp1, DAG.getConstantFP(0.0, SrcVT), DAG, dl); 1905 SDValue ARMCC = DAG.getConstant(ARMCC::LT, MVT::i32); 1906 SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32); 1907 return DAG.getNode(ARMISD::CNEG, dl, VT, AbsVal, AbsVal, ARMCC, CCR, Cmp); 1908} 1909 1910SDValue ARMTargetLowering::LowerFRAMEADDR(SDValue Op, SelectionDAG &DAG) { 1911 MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo(); 1912 MFI->setFrameAddressIsTaken(true); 1913 EVT VT = Op.getValueType(); 1914 DebugLoc dl = Op.getDebugLoc(); // FIXME probably not meaningful 1915 unsigned Depth = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue(); 1916 unsigned FrameReg = (Subtarget->isThumb() || Subtarget->isTargetDarwin()) 1917 ? ARM::R7 : ARM::R11; 1918 SDValue FrameAddr = DAG.getCopyFromReg(DAG.getEntryNode(), dl, FrameReg, VT); 1919 while (Depth--) 1920 FrameAddr = DAG.getLoad(VT, dl, DAG.getEntryNode(), FrameAddr, NULL, 0); 1921 return FrameAddr; 1922} 1923 1924SDValue 1925ARMTargetLowering::EmitTargetCodeForMemcpy(SelectionDAG &DAG, DebugLoc dl, 1926 SDValue Chain, 1927 SDValue Dst, SDValue Src, 1928 SDValue Size, unsigned Align, 1929 bool AlwaysInline, 1930 const Value *DstSV, uint64_t DstSVOff, 1931 const Value *SrcSV, uint64_t SrcSVOff){ 1932 // Do repeated 4-byte loads and stores. To be improved. 1933 // This requires 4-byte alignment. 1934 if ((Align & 3) != 0) 1935 return SDValue(); 1936 // This requires the copy size to be a constant, preferrably 1937 // within a subtarget-specific limit. 1938 ConstantSDNode *ConstantSize = dyn_cast<ConstantSDNode>(Size); 1939 if (!ConstantSize) 1940 return SDValue(); 1941 uint64_t SizeVal = ConstantSize->getZExtValue(); 1942 if (!AlwaysInline && SizeVal > getSubtarget()->getMaxInlineSizeThreshold()) 1943 return SDValue(); 1944 1945 unsigned BytesLeft = SizeVal & 3; 1946 unsigned NumMemOps = SizeVal >> 2; 1947 unsigned EmittedNumMemOps = 0; 1948 EVT VT = MVT::i32; 1949 unsigned VTSize = 4; 1950 unsigned i = 0; 1951 const unsigned MAX_LOADS_IN_LDM = 6; 1952 SDValue TFOps[MAX_LOADS_IN_LDM]; 1953 SDValue Loads[MAX_LOADS_IN_LDM]; 1954 uint64_t SrcOff = 0, DstOff = 0; 1955 1956 // Emit up to MAX_LOADS_IN_LDM loads, then a TokenFactor barrier, then the 1957 // same number of stores. The loads and stores will get combined into 1958 // ldm/stm later on. 1959 while (EmittedNumMemOps < NumMemOps) { 1960 for (i = 0; 1961 i < MAX_LOADS_IN_LDM && EmittedNumMemOps + i < NumMemOps; ++i) { 1962 Loads[i] = DAG.getLoad(VT, dl, Chain, 1963 DAG.getNode(ISD::ADD, dl, MVT::i32, Src, 1964 DAG.getConstant(SrcOff, MVT::i32)), 1965 SrcSV, SrcSVOff + SrcOff); 1966 TFOps[i] = Loads[i].getValue(1); 1967 SrcOff += VTSize; 1968 } 1969 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other, &TFOps[0], i); 1970 1971 for (i = 0; 1972 i < MAX_LOADS_IN_LDM && EmittedNumMemOps + i < NumMemOps; ++i) { 1973 TFOps[i] = DAG.getStore(Chain, dl, Loads[i], 1974 DAG.getNode(ISD::ADD, dl, MVT::i32, Dst, 1975 DAG.getConstant(DstOff, MVT::i32)), 1976 DstSV, DstSVOff + DstOff); 1977 DstOff += VTSize; 1978 } 1979 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other, &TFOps[0], i); 1980 1981 EmittedNumMemOps += i; 1982 } 1983 1984 if (BytesLeft == 0) 1985 return Chain; 1986 1987 // Issue loads / stores for the trailing (1 - 3) bytes. 1988 unsigned BytesLeftSave = BytesLeft; 1989 i = 0; 1990 while (BytesLeft) { 1991 if (BytesLeft >= 2) { 1992 VT = MVT::i16; 1993 VTSize = 2; 1994 } else { 1995 VT = MVT::i8; 1996 VTSize = 1; 1997 } 1998 1999 Loads[i] = DAG.getLoad(VT, dl, Chain, 2000 DAG.getNode(ISD::ADD, dl, MVT::i32, Src, 2001 DAG.getConstant(SrcOff, MVT::i32)), 2002 SrcSV, SrcSVOff + SrcOff); 2003 TFOps[i] = Loads[i].getValue(1); 2004 ++i; 2005 SrcOff += VTSize; 2006 BytesLeft -= VTSize; 2007 } 2008 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other, &TFOps[0], i); 2009 2010 i = 0; 2011 BytesLeft = BytesLeftSave; 2012 while (BytesLeft) { 2013 if (BytesLeft >= 2) { 2014 VT = MVT::i16; 2015 VTSize = 2; 2016 } else { 2017 VT = MVT::i8; 2018 VTSize = 1; 2019 } 2020 2021 TFOps[i] = DAG.getStore(Chain, dl, Loads[i], 2022 DAG.getNode(ISD::ADD, dl, MVT::i32, Dst, 2023 DAG.getConstant(DstOff, MVT::i32)), 2024 DstSV, DstSVOff + DstOff); 2025 ++i; 2026 DstOff += VTSize; 2027 BytesLeft -= VTSize; 2028 } 2029 return DAG.getNode(ISD::TokenFactor, dl, MVT::Other, &TFOps[0], i); 2030} 2031 2032static SDValue ExpandBIT_CONVERT(SDNode *N, SelectionDAG &DAG) { 2033 SDValue Op = N->getOperand(0); 2034 DebugLoc dl = N->getDebugLoc(); 2035 if (N->getValueType(0) == MVT::f64) { 2036 // Turn i64->f64 into FMDRR. 2037 SDValue Lo = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, Op, 2038 DAG.getConstant(0, MVT::i32)); 2039 SDValue Hi = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, Op, 2040 DAG.getConstant(1, MVT::i32)); 2041 return DAG.getNode(ARMISD::FMDRR, dl, MVT::f64, Lo, Hi); 2042 } 2043 2044 // Turn f64->i64 into FMRRD. 2045 SDValue Cvt = DAG.getNode(ARMISD::FMRRD, dl, 2046 DAG.getVTList(MVT::i32, MVT::i32), &Op, 1); 2047 2048 // Merge the pieces into a single i64 value. 2049 return DAG.getNode(ISD::BUILD_PAIR, dl, MVT::i64, Cvt, Cvt.getValue(1)); 2050} 2051 2052/// getZeroVector - Returns a vector of specified type with all zero elements. 2053/// 2054static SDValue getZeroVector(EVT VT, SelectionDAG &DAG, DebugLoc dl) { 2055 assert(VT.isVector() && "Expected a vector type"); 2056 2057 // Zero vectors are used to represent vector negation and in those cases 2058 // will be implemented with the NEON VNEG instruction. However, VNEG does 2059 // not support i64 elements, so sometimes the zero vectors will need to be 2060 // explicitly constructed. For those cases, and potentially other uses in 2061 // the future, always build zero vectors as <4 x i32> or <2 x i32> bitcasted 2062 // to their dest type. This ensures they get CSE'd. 2063 SDValue Vec; 2064 SDValue Cst = DAG.getTargetConstant(0, MVT::i32); 2065 if (VT.getSizeInBits() == 64) 2066 Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v2i32, Cst, Cst); 2067 else 2068 Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v4i32, Cst, Cst, Cst, Cst); 2069 2070 return DAG.getNode(ISD::BIT_CONVERT, dl, VT, Vec); 2071} 2072 2073/// getOnesVector - Returns a vector of specified type with all bits set. 2074/// 2075static SDValue getOnesVector(EVT VT, SelectionDAG &DAG, DebugLoc dl) { 2076 assert(VT.isVector() && "Expected a vector type"); 2077 2078 // Always build ones vectors as <4 x i32> or <2 x i32> bitcasted to their dest 2079 // type. This ensures they get CSE'd. 2080 SDValue Vec; 2081 SDValue Cst = DAG.getTargetConstant(~0U, MVT::i32); 2082 if (VT.getSizeInBits() == 64) 2083 Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v2i32, Cst, Cst); 2084 else 2085 Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v4i32, Cst, Cst, Cst, Cst); 2086 2087 return DAG.getNode(ISD::BIT_CONVERT, dl, VT, Vec); 2088} 2089 2090static SDValue LowerShift(SDNode *N, SelectionDAG &DAG, 2091 const ARMSubtarget *ST) { 2092 EVT VT = N->getValueType(0); 2093 DebugLoc dl = N->getDebugLoc(); 2094 2095 // Lower vector shifts on NEON to use VSHL. 2096 if (VT.isVector()) { 2097 assert(ST->hasNEON() && "unexpected vector shift"); 2098 2099 // Left shifts translate directly to the vshiftu intrinsic. 2100 if (N->getOpcode() == ISD::SHL) 2101 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT, 2102 DAG.getConstant(Intrinsic::arm_neon_vshiftu, MVT::i32), 2103 N->getOperand(0), N->getOperand(1)); 2104 2105 assert((N->getOpcode() == ISD::SRA || 2106 N->getOpcode() == ISD::SRL) && "unexpected vector shift opcode"); 2107 2108 // NEON uses the same intrinsics for both left and right shifts. For 2109 // right shifts, the shift amounts are negative, so negate the vector of 2110 // shift amounts. 2111 EVT ShiftVT = N->getOperand(1).getValueType(); 2112 SDValue NegatedCount = DAG.getNode(ISD::SUB, dl, ShiftVT, 2113 getZeroVector(ShiftVT, DAG, dl), 2114 N->getOperand(1)); 2115 Intrinsic::ID vshiftInt = (N->getOpcode() == ISD::SRA ? 2116 Intrinsic::arm_neon_vshifts : 2117 Intrinsic::arm_neon_vshiftu); 2118 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT, 2119 DAG.getConstant(vshiftInt, MVT::i32), 2120 N->getOperand(0), NegatedCount); 2121 } 2122 2123 // We can get here for a node like i32 = ISD::SHL i32, i64 2124 if (VT != MVT::i64) 2125 return SDValue(); 2126 2127 assert((N->getOpcode() == ISD::SRL || N->getOpcode() == ISD::SRA) && 2128 "Unknown shift to lower!"); 2129 2130 // We only lower SRA, SRL of 1 here, all others use generic lowering. 2131 if (!isa<ConstantSDNode>(N->getOperand(1)) || 2132 cast<ConstantSDNode>(N->getOperand(1))->getZExtValue() != 1) 2133 return SDValue(); 2134 2135 // If we are in thumb mode, we don't have RRX. 2136 if (ST->isThumb1Only()) return SDValue(); 2137 2138 // Okay, we have a 64-bit SRA or SRL of 1. Lower this to an RRX expr. 2139 SDValue Lo = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, N->getOperand(0), 2140 DAG.getConstant(0, MVT::i32)); 2141 SDValue Hi = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, N->getOperand(0), 2142 DAG.getConstant(1, MVT::i32)); 2143 2144 // First, build a SRA_FLAG/SRL_FLAG op, which shifts the top part by one and 2145 // captures the result into a carry flag. 2146 unsigned Opc = N->getOpcode() == ISD::SRL ? ARMISD::SRL_FLAG:ARMISD::SRA_FLAG; 2147 Hi = DAG.getNode(Opc, dl, DAG.getVTList(MVT::i32, MVT::Flag), &Hi, 1); 2148 2149 // The low part is an ARMISD::RRX operand, which shifts the carry in. 2150 Lo = DAG.getNode(ARMISD::RRX, dl, MVT::i32, Lo, Hi.getValue(1)); 2151 2152 // Merge the pieces into a single i64 value. 2153 return DAG.getNode(ISD::BUILD_PAIR, dl, MVT::i64, Lo, Hi); 2154} 2155 2156static SDValue LowerVSETCC(SDValue Op, SelectionDAG &DAG) { 2157 SDValue TmpOp0, TmpOp1; 2158 bool Invert = false; 2159 bool Swap = false; 2160 unsigned Opc = 0; 2161 2162 SDValue Op0 = Op.getOperand(0); 2163 SDValue Op1 = Op.getOperand(1); 2164 SDValue CC = Op.getOperand(2); 2165 EVT VT = Op.getValueType(); 2166 ISD::CondCode SetCCOpcode = cast<CondCodeSDNode>(CC)->get(); 2167 DebugLoc dl = Op.getDebugLoc(); 2168 2169 if (Op.getOperand(1).getValueType().isFloatingPoint()) { 2170 switch (SetCCOpcode) { 2171 default: llvm_unreachable("Illegal FP comparison"); break; 2172 case ISD::SETUNE: 2173 case ISD::SETNE: Invert = true; // Fallthrough 2174 case ISD::SETOEQ: 2175 case ISD::SETEQ: Opc = ARMISD::VCEQ; break; 2176 case ISD::SETOLT: 2177 case ISD::SETLT: Swap = true; // Fallthrough 2178 case ISD::SETOGT: 2179 case ISD::SETGT: Opc = ARMISD::VCGT; break; 2180 case ISD::SETOLE: 2181 case ISD::SETLE: Swap = true; // Fallthrough 2182 case ISD::SETOGE: 2183 case ISD::SETGE: Opc = ARMISD::VCGE; break; 2184 case ISD::SETUGE: Swap = true; // Fallthrough 2185 case ISD::SETULE: Invert = true; Opc = ARMISD::VCGT; break; 2186 case ISD::SETUGT: Swap = true; // Fallthrough 2187 case ISD::SETULT: Invert = true; Opc = ARMISD::VCGE; break; 2188 case ISD::SETUEQ: Invert = true; // Fallthrough 2189 case ISD::SETONE: 2190 // Expand this to (OLT | OGT). 2191 TmpOp0 = Op0; 2192 TmpOp1 = Op1; 2193 Opc = ISD::OR; 2194 Op0 = DAG.getNode(ARMISD::VCGT, dl, VT, TmpOp1, TmpOp0); 2195 Op1 = DAG.getNode(ARMISD::VCGT, dl, VT, TmpOp0, TmpOp1); 2196 break; 2197 case ISD::SETUO: Invert = true; // Fallthrough 2198 case ISD::SETO: 2199 // Expand this to (OLT | OGE). 2200 TmpOp0 = Op0; 2201 TmpOp1 = Op1; 2202 Opc = ISD::OR; 2203 Op0 = DAG.getNode(ARMISD::VCGT, dl, VT, TmpOp1, TmpOp0); 2204 Op1 = DAG.getNode(ARMISD::VCGE, dl, VT, TmpOp0, TmpOp1); 2205 break; 2206 } 2207 } else { 2208 // Integer comparisons. 2209 switch (SetCCOpcode) { 2210 default: llvm_unreachable("Illegal integer comparison"); break; 2211 case ISD::SETNE: Invert = true; 2212 case ISD::SETEQ: Opc = ARMISD::VCEQ; break; 2213 case ISD::SETLT: Swap = true; 2214 case ISD::SETGT: Opc = ARMISD::VCGT; break; 2215 case ISD::SETLE: Swap = true; 2216 case ISD::SETGE: Opc = ARMISD::VCGE; break; 2217 case ISD::SETULT: Swap = true; 2218 case ISD::SETUGT: Opc = ARMISD::VCGTU; break; 2219 case ISD::SETULE: Swap = true; 2220 case ISD::SETUGE: Opc = ARMISD::VCGEU; break; 2221 } 2222 2223 // Detect VTST (Vector Test Bits) = icmp ne (and (op0, op1), zero). 2224 if (Opc == ARMISD::VCEQ) { 2225 2226 SDValue AndOp; 2227 if (ISD::isBuildVectorAllZeros(Op1.getNode())) 2228 AndOp = Op0; 2229 else if (ISD::isBuildVectorAllZeros(Op0.getNode())) 2230 AndOp = Op1; 2231 2232 // Ignore bitconvert. 2233 if (AndOp.getNode() && AndOp.getOpcode() == ISD::BIT_CONVERT) 2234 AndOp = AndOp.getOperand(0); 2235 2236 if (AndOp.getNode() && AndOp.getOpcode() == ISD::AND) { 2237 Opc = ARMISD::VTST; 2238 Op0 = DAG.getNode(ISD::BIT_CONVERT, dl, VT, AndOp.getOperand(0)); 2239 Op1 = DAG.getNode(ISD::BIT_CONVERT, dl, VT, AndOp.getOperand(1)); 2240 Invert = !Invert; 2241 } 2242 } 2243 } 2244 2245 if (Swap) 2246 std::swap(Op0, Op1); 2247 2248 SDValue Result = DAG.getNode(Opc, dl, VT, Op0, Op1); 2249 2250 if (Invert) 2251 Result = DAG.getNOT(dl, Result, VT); 2252 2253 return Result; 2254} 2255 2256/// isVMOVSplat - Check if the specified splat value corresponds to an immediate 2257/// VMOV instruction, and if so, return the constant being splatted. 2258static SDValue isVMOVSplat(uint64_t SplatBits, uint64_t SplatUndef, 2259 unsigned SplatBitSize, SelectionDAG &DAG) { 2260 switch (SplatBitSize) { 2261 case 8: 2262 // Any 1-byte value is OK. 2263 assert((SplatBits & ~0xff) == 0 && "one byte splat value is too big"); 2264 return DAG.getTargetConstant(SplatBits, MVT::i8); 2265 2266 case 16: 2267 // NEON's 16-bit VMOV supports splat values where only one byte is nonzero. 2268 if ((SplatBits & ~0xff) == 0 || 2269 (SplatBits & ~0xff00) == 0) 2270 return DAG.getTargetConstant(SplatBits, MVT::i16); 2271 break; 2272 2273 case 32: 2274 // NEON's 32-bit VMOV supports splat values where: 2275 // * only one byte is nonzero, or 2276 // * the least significant byte is 0xff and the second byte is nonzero, or 2277 // * the least significant 2 bytes are 0xff and the third is nonzero. 2278 if ((SplatBits & ~0xff) == 0 || 2279 (SplatBits & ~0xff00) == 0 || 2280 (SplatBits & ~0xff0000) == 0 || 2281 (SplatBits & ~0xff000000) == 0) 2282 return DAG.getTargetConstant(SplatBits, MVT::i32); 2283 2284 if ((SplatBits & ~0xffff) == 0 && 2285 ((SplatBits | SplatUndef) & 0xff) == 0xff) 2286 return DAG.getTargetConstant(SplatBits | 0xff, MVT::i32); 2287 2288 if ((SplatBits & ~0xffffff) == 0 && 2289 ((SplatBits | SplatUndef) & 0xffff) == 0xffff) 2290 return DAG.getTargetConstant(SplatBits | 0xffff, MVT::i32); 2291 2292 // Note: there are a few 32-bit splat values (specifically: 00ffff00, 2293 // ff000000, ff0000ff, and ffff00ff) that are valid for VMOV.I64 but not 2294 // VMOV.I32. A (very) minor optimization would be to replicate the value 2295 // and fall through here to test for a valid 64-bit splat. But, then the 2296 // caller would also need to check and handle the change in size. 2297 break; 2298 2299 case 64: { 2300 // NEON has a 64-bit VMOV splat where each byte is either 0 or 0xff. 2301 uint64_t BitMask = 0xff; 2302 uint64_t Val = 0; 2303 for (int ByteNum = 0; ByteNum < 8; ++ByteNum) { 2304 if (((SplatBits | SplatUndef) & BitMask) == BitMask) 2305 Val |= BitMask; 2306 else if ((SplatBits & BitMask) != 0) 2307 return SDValue(); 2308 BitMask <<= 8; 2309 } 2310 return DAG.getTargetConstant(Val, MVT::i64); 2311 } 2312 2313 default: 2314 llvm_unreachable("unexpected size for isVMOVSplat"); 2315 break; 2316 } 2317 2318 return SDValue(); 2319} 2320 2321/// getVMOVImm - If this is a build_vector of constants which can be 2322/// formed by using a VMOV instruction of the specified element size, 2323/// return the constant being splatted. The ByteSize field indicates the 2324/// number of bytes of each element [1248]. 2325SDValue ARM::getVMOVImm(SDNode *N, unsigned ByteSize, SelectionDAG &DAG) { 2326 BuildVectorSDNode *BVN = dyn_cast<BuildVectorSDNode>(N); 2327 APInt SplatBits, SplatUndef; 2328 unsigned SplatBitSize; 2329 bool HasAnyUndefs; 2330 if (! BVN || ! BVN->isConstantSplat(SplatBits, SplatUndef, SplatBitSize, 2331 HasAnyUndefs, ByteSize * 8)) 2332 return SDValue(); 2333 2334 if (SplatBitSize > ByteSize * 8) 2335 return SDValue(); 2336 2337 return isVMOVSplat(SplatBits.getZExtValue(), SplatUndef.getZExtValue(), 2338 SplatBitSize, DAG); 2339} 2340 2341static bool isVEXTMask(const SmallVectorImpl<int> &M, EVT VT, 2342 bool &ReverseVEXT, unsigned &Imm) { 2343 unsigned NumElts = VT.getVectorNumElements(); 2344 ReverseVEXT = false; 2345 Imm = M[0]; 2346 2347 // If this is a VEXT shuffle, the immediate value is the index of the first 2348 // element. The other shuffle indices must be the successive elements after 2349 // the first one. 2350 unsigned ExpectedElt = Imm; 2351 for (unsigned i = 1; i < NumElts; ++i) { 2352 // Increment the expected index. If it wraps around, it may still be 2353 // a VEXT but the source vectors must be swapped. 2354 ExpectedElt += 1; 2355 if (ExpectedElt == NumElts * 2) { 2356 ExpectedElt = 0; 2357 ReverseVEXT = true; 2358 } 2359 2360 if (ExpectedElt != static_cast<unsigned>(M[i])) 2361 return false; 2362 } 2363 2364 // Adjust the index value if the source operands will be swapped. 2365 if (ReverseVEXT) 2366 Imm -= NumElts; 2367 2368 return true; 2369} 2370 2371/// isVREVMask - Check if a vector shuffle corresponds to a VREV 2372/// instruction with the specified blocksize. (The order of the elements 2373/// within each block of the vector is reversed.) 2374static bool isVREVMask(const SmallVectorImpl<int> &M, EVT VT, 2375 unsigned BlockSize) { 2376 assert((BlockSize==16 || BlockSize==32 || BlockSize==64) && 2377 "Only possible block sizes for VREV are: 16, 32, 64"); 2378 2379 unsigned NumElts = VT.getVectorNumElements(); 2380 unsigned EltSz = VT.getVectorElementType().getSizeInBits(); 2381 unsigned BlockElts = M[0] + 1; 2382 2383 if (BlockSize <= EltSz || BlockSize != BlockElts * EltSz) 2384 return false; 2385 2386 for (unsigned i = 0; i < NumElts; ++i) { 2387 if ((unsigned) M[i] != 2388 (i - i%BlockElts) + (BlockElts - 1 - i%BlockElts)) 2389 return false; 2390 } 2391 2392 return true; 2393} 2394 2395static bool isVTRNMask(const SmallVectorImpl<int> &M, EVT VT, 2396 unsigned &WhichResult) { 2397 unsigned NumElts = VT.getVectorNumElements(); 2398 WhichResult = (M[0] == 0 ? 0 : 1); 2399 for (unsigned i = 0; i < NumElts; i += 2) { 2400 if ((unsigned) M[i] != i + WhichResult || 2401 (unsigned) M[i+1] != i + NumElts + WhichResult) 2402 return false; 2403 } 2404 return true; 2405} 2406 2407static bool isVUZPMask(const SmallVectorImpl<int> &M, EVT VT, 2408 unsigned &WhichResult) { 2409 unsigned NumElts = VT.getVectorNumElements(); 2410 WhichResult = (M[0] == 0 ? 0 : 1); 2411 for (unsigned i = 0; i != NumElts; ++i) { 2412 if ((unsigned) M[i] != 2 * i + WhichResult) 2413 return false; 2414 } 2415 2416 // VUZP.32 for 64-bit vectors is a pseudo-instruction alias for VTRN.32. 2417 if (VT.is64BitVector() && VT.getVectorElementType().getSizeInBits() == 32) 2418 return false; 2419 2420 return true; 2421} 2422 2423static bool isVZIPMask(const SmallVectorImpl<int> &M, EVT VT, 2424 unsigned &WhichResult) { 2425 unsigned NumElts = VT.getVectorNumElements(); 2426 WhichResult = (M[0] == 0 ? 0 : 1); 2427 unsigned Idx = WhichResult * NumElts / 2; 2428 for (unsigned i = 0; i != NumElts; i += 2) { 2429 if ((unsigned) M[i] != Idx || 2430 (unsigned) M[i+1] != Idx + NumElts) 2431 return false; 2432 Idx += 1; 2433 } 2434 2435 // VZIP.32 for 64-bit vectors is a pseudo-instruction alias for VTRN.32. 2436 if (VT.is64BitVector() && VT.getVectorElementType().getSizeInBits() == 32) 2437 return false; 2438 2439 return true; 2440} 2441 2442static SDValue BuildSplat(SDValue Val, EVT VT, SelectionDAG &DAG, DebugLoc dl) { 2443 // Canonicalize all-zeros and all-ones vectors. 2444 ConstantSDNode *ConstVal = cast<ConstantSDNode>(Val.getNode()); 2445 if (ConstVal->isNullValue()) 2446 return getZeroVector(VT, DAG, dl); 2447 if (ConstVal->isAllOnesValue()) 2448 return getOnesVector(VT, DAG, dl); 2449 2450 EVT CanonicalVT; 2451 if (VT.is64BitVector()) { 2452 switch (Val.getValueType().getSizeInBits()) { 2453 case 8: CanonicalVT = MVT::v8i8; break; 2454 case 16: CanonicalVT = MVT::v4i16; break; 2455 case 32: CanonicalVT = MVT::v2i32; break; 2456 case 64: CanonicalVT = MVT::v1i64; break; 2457 default: llvm_unreachable("unexpected splat element type"); break; 2458 } 2459 } else { 2460 assert(VT.is128BitVector() && "unknown splat vector size"); 2461 switch (Val.getValueType().getSizeInBits()) { 2462 case 8: CanonicalVT = MVT::v16i8; break; 2463 case 16: CanonicalVT = MVT::v8i16; break; 2464 case 32: CanonicalVT = MVT::v4i32; break; 2465 case 64: CanonicalVT = MVT::v2i64; break; 2466 default: llvm_unreachable("unexpected splat element type"); break; 2467 } 2468 } 2469 2470 // Build a canonical splat for this value. 2471 SmallVector<SDValue, 8> Ops; 2472 Ops.assign(CanonicalVT.getVectorNumElements(), Val); 2473 SDValue Res = DAG.getNode(ISD::BUILD_VECTOR, dl, CanonicalVT, &Ops[0], 2474 Ops.size()); 2475 return DAG.getNode(ISD::BIT_CONVERT, dl, VT, Res); 2476} 2477 2478// If this is a case we can't handle, return null and let the default 2479// expansion code take care of it. 2480static SDValue LowerBUILD_VECTOR(SDValue Op, SelectionDAG &DAG) { 2481 BuildVectorSDNode *BVN = cast<BuildVectorSDNode>(Op.getNode()); 2482 DebugLoc dl = Op.getDebugLoc(); 2483 EVT VT = Op.getValueType(); 2484 2485 APInt SplatBits, SplatUndef; 2486 unsigned SplatBitSize; 2487 bool HasAnyUndefs; 2488 if (BVN->isConstantSplat(SplatBits, SplatUndef, SplatBitSize, HasAnyUndefs)) { 2489 SDValue Val = isVMOVSplat(SplatBits.getZExtValue(), 2490 SplatUndef.getZExtValue(), SplatBitSize, DAG); 2491 if (Val.getNode()) 2492 return BuildSplat(Val, VT, DAG, dl); 2493 } 2494 2495 // If there are only 2 elements in a 128-bit vector, insert them into an 2496 // undef vector. This handles the common case for 128-bit vector argument 2497 // passing, where the insertions should be translated to subreg accesses 2498 // with no real instructions. 2499 if (VT.is128BitVector() && Op.getNumOperands() == 2) { 2500 SDValue Val = DAG.getUNDEF(VT); 2501 SDValue Op0 = Op.getOperand(0); 2502 SDValue Op1 = Op.getOperand(1); 2503 if (Op0.getOpcode() != ISD::UNDEF) 2504 Val = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, VT, Val, Op0, 2505 DAG.getIntPtrConstant(0)); 2506 if (Op1.getOpcode() != ISD::UNDEF) 2507 Val = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, VT, Val, Op1, 2508 DAG.getIntPtrConstant(1)); 2509 return Val; 2510 } 2511 2512 return SDValue(); 2513} 2514 2515/// isShuffleMaskLegal - Targets can use this to indicate that they only 2516/// support *some* VECTOR_SHUFFLE operations, those with specific masks. 2517/// By default, if a target supports the VECTOR_SHUFFLE node, all mask values 2518/// are assumed to be legal. 2519bool 2520ARMTargetLowering::isShuffleMaskLegal(const SmallVectorImpl<int> &M, 2521 EVT VT) const { 2522 if (VT.getVectorNumElements() == 4 && 2523 (VT.is128BitVector() || VT.is64BitVector())) { 2524 unsigned PFIndexes[4]; 2525 for (unsigned i = 0; i != 4; ++i) { 2526 if (M[i] < 0) 2527 PFIndexes[i] = 8; 2528 else 2529 PFIndexes[i] = M[i]; 2530 } 2531 2532 // Compute the index in the perfect shuffle table. 2533 unsigned PFTableIndex = 2534 PFIndexes[0]*9*9*9+PFIndexes[1]*9*9+PFIndexes[2]*9+PFIndexes[3]; 2535 unsigned PFEntry = PerfectShuffleTable[PFTableIndex]; 2536 unsigned Cost = (PFEntry >> 30); 2537 2538 if (Cost <= 4) 2539 return true; 2540 } 2541 2542 bool ReverseVEXT; 2543 unsigned Imm, WhichResult; 2544 2545 return (ShuffleVectorSDNode::isSplatMask(&M[0], VT) || 2546 isVREVMask(M, VT, 64) || 2547 isVREVMask(M, VT, 32) || 2548 isVREVMask(M, VT, 16) || 2549 isVEXTMask(M, VT, ReverseVEXT, Imm) || 2550 isVTRNMask(M, VT, WhichResult) || 2551 isVUZPMask(M, VT, WhichResult) || 2552 isVZIPMask(M, VT, WhichResult)); 2553} 2554 2555/// GeneratePerfectShuffle - Given an entry in the perfect-shuffle table, emit 2556/// the specified operations to build the shuffle. 2557static SDValue GeneratePerfectShuffle(unsigned PFEntry, SDValue LHS, 2558 SDValue RHS, SelectionDAG &DAG, 2559 DebugLoc dl) { 2560 unsigned OpNum = (PFEntry >> 26) & 0x0F; 2561 unsigned LHSID = (PFEntry >> 13) & ((1 << 13)-1); 2562 unsigned RHSID = (PFEntry >> 0) & ((1 << 13)-1); 2563 2564 enum { 2565 OP_COPY = 0, // Copy, used for things like <u,u,u,3> to say it is <0,1,2,3> 2566 OP_VREV, 2567 OP_VDUP0, 2568 OP_VDUP1, 2569 OP_VDUP2, 2570 OP_VDUP3, 2571 OP_VEXT1, 2572 OP_VEXT2, 2573 OP_VEXT3, 2574 OP_VUZPL, // VUZP, left result 2575 OP_VUZPR, // VUZP, right result 2576 OP_VZIPL, // VZIP, left result 2577 OP_VZIPR, // VZIP, right result 2578 OP_VTRNL, // VTRN, left result 2579 OP_VTRNR // VTRN, right result 2580 }; 2581 2582 if (OpNum == OP_COPY) { 2583 if (LHSID == (1*9+2)*9+3) return LHS; 2584 assert(LHSID == ((4*9+5)*9+6)*9+7 && "Illegal OP_COPY!"); 2585 return RHS; 2586 } 2587 2588 SDValue OpLHS, OpRHS; 2589 OpLHS = GeneratePerfectShuffle(PerfectShuffleTable[LHSID], LHS, RHS, DAG, dl); 2590 OpRHS = GeneratePerfectShuffle(PerfectShuffleTable[RHSID], LHS, RHS, DAG, dl); 2591 EVT VT = OpLHS.getValueType(); 2592 2593 switch (OpNum) { 2594 default: llvm_unreachable("Unknown shuffle opcode!"); 2595 case OP_VREV: 2596 return DAG.getNode(ARMISD::VREV64, dl, VT, OpLHS); 2597 case OP_VDUP0: 2598 case OP_VDUP1: 2599 case OP_VDUP2: 2600 case OP_VDUP3: 2601 return DAG.getNode(ARMISD::VDUPLANE, dl, VT, 2602 OpLHS, DAG.getConstant(OpNum-OP_VDUP0, MVT::i32)); 2603 case OP_VEXT1: 2604 case OP_VEXT2: 2605 case OP_VEXT3: 2606 return DAG.getNode(ARMISD::VEXT, dl, VT, 2607 OpLHS, OpRHS, 2608 DAG.getConstant(OpNum-OP_VEXT1+1, MVT::i32)); 2609 case OP_VUZPL: 2610 case OP_VUZPR: 2611 return DAG.getNode(ARMISD::VUZP, dl, DAG.getVTList(VT, VT), 2612 OpLHS, OpRHS).getValue(OpNum-OP_VUZPL); 2613 case OP_VZIPL: 2614 case OP_VZIPR: 2615 return DAG.getNode(ARMISD::VZIP, dl, DAG.getVTList(VT, VT), 2616 OpLHS, OpRHS).getValue(OpNum-OP_VZIPL); 2617 case OP_VTRNL: 2618 case OP_VTRNR: 2619 return DAG.getNode(ARMISD::VTRN, dl, DAG.getVTList(VT, VT), 2620 OpLHS, OpRHS).getValue(OpNum-OP_VTRNL); 2621 } 2622} 2623 2624static SDValue LowerVECTOR_SHUFFLE(SDValue Op, SelectionDAG &DAG) { 2625 SDValue V1 = Op.getOperand(0); 2626 SDValue V2 = Op.getOperand(1); 2627 DebugLoc dl = Op.getDebugLoc(); 2628 EVT VT = Op.getValueType(); 2629 ShuffleVectorSDNode *SVN = cast<ShuffleVectorSDNode>(Op.getNode()); 2630 SmallVector<int, 8> ShuffleMask; 2631 2632 // Convert shuffles that are directly supported on NEON to target-specific 2633 // DAG nodes, instead of keeping them as shuffles and matching them again 2634 // during code selection. This is more efficient and avoids the possibility 2635 // of inconsistencies between legalization and selection. 2636 // FIXME: floating-point vectors should be canonicalized to integer vectors 2637 // of the same time so that they get CSEd properly. 2638 SVN->getMask(ShuffleMask); 2639 2640 if (ShuffleVectorSDNode::isSplatMask(&ShuffleMask[0], VT)) { 2641 int Lane = SVN->getSplatIndex(); 2642 if (Lane == 0 && V1.getOpcode() == ISD::SCALAR_TO_VECTOR) { 2643 return DAG.getNode(ARMISD::VDUP, dl, VT, V1.getOperand(0)); 2644 } 2645 return DAG.getNode(ARMISD::VDUPLANE, dl, VT, V1, 2646 DAG.getConstant(Lane, MVT::i32)); 2647 } 2648 2649 bool ReverseVEXT; 2650 unsigned Imm; 2651 if (isVEXTMask(ShuffleMask, VT, ReverseVEXT, Imm)) { 2652 if (ReverseVEXT) 2653 std::swap(V1, V2); 2654 return DAG.getNode(ARMISD::VEXT, dl, VT, V1, V2, 2655 DAG.getConstant(Imm, MVT::i32)); 2656 } 2657 2658 if (isVREVMask(ShuffleMask, VT, 64)) 2659 return DAG.getNode(ARMISD::VREV64, dl, VT, V1); 2660 if (isVREVMask(ShuffleMask, VT, 32)) 2661 return DAG.getNode(ARMISD::VREV32, dl, VT, V1); 2662 if (isVREVMask(ShuffleMask, VT, 16)) 2663 return DAG.getNode(ARMISD::VREV16, dl, VT, V1); 2664 2665 // Check for Neon shuffles that modify both input vectors in place. 2666 // If both results are used, i.e., if there are two shuffles with the same 2667 // source operands and with masks corresponding to both results of one of 2668 // these operations, DAG memoization will ensure that a single node is 2669 // used for both shuffles. 2670 unsigned WhichResult; 2671 if (isVTRNMask(ShuffleMask, VT, WhichResult)) 2672 return DAG.getNode(ARMISD::VTRN, dl, DAG.getVTList(VT, VT), 2673 V1, V2).getValue(WhichResult); 2674 if (isVUZPMask(ShuffleMask, VT, WhichResult)) 2675 return DAG.getNode(ARMISD::VUZP, dl, DAG.getVTList(VT, VT), 2676 V1, V2).getValue(WhichResult); 2677 if (isVZIPMask(ShuffleMask, VT, WhichResult)) 2678 return DAG.getNode(ARMISD::VZIP, dl, DAG.getVTList(VT, VT), 2679 V1, V2).getValue(WhichResult); 2680 2681 // If the shuffle is not directly supported and it has 4 elements, use 2682 // the PerfectShuffle-generated table to synthesize it from other shuffles. 2683 if (VT.getVectorNumElements() == 4 && 2684 (VT.is128BitVector() || VT.is64BitVector())) { 2685 unsigned PFIndexes[4]; 2686 for (unsigned i = 0; i != 4; ++i) { 2687 if (ShuffleMask[i] < 0) 2688 PFIndexes[i] = 8; 2689 else 2690 PFIndexes[i] = ShuffleMask[i]; 2691 } 2692 2693 // Compute the index in the perfect shuffle table. 2694 unsigned PFTableIndex = 2695 PFIndexes[0]*9*9*9+PFIndexes[1]*9*9+PFIndexes[2]*9+PFIndexes[3]; 2696 2697 unsigned PFEntry = PerfectShuffleTable[PFTableIndex]; 2698 unsigned Cost = (PFEntry >> 30); 2699 2700 if (Cost <= 4) 2701 return GeneratePerfectShuffle(PFEntry, V1, V2, DAG, dl); 2702 } 2703 2704 return SDValue(); 2705} 2706 2707static SDValue LowerEXTRACT_VECTOR_ELT(SDValue Op, SelectionDAG &DAG) { 2708 EVT VT = Op.getValueType(); 2709 DebugLoc dl = Op.getDebugLoc(); 2710 assert((VT == MVT::i8 || VT == MVT::i16) && 2711 "unexpected type for custom-lowering vector extract"); 2712 SDValue Vec = Op.getOperand(0); 2713 SDValue Lane = Op.getOperand(1); 2714 Op = DAG.getNode(ARMISD::VGETLANEu, dl, MVT::i32, Vec, Lane); 2715 Op = DAG.getNode(ISD::AssertZext, dl, MVT::i32, Op, DAG.getValueType(VT)); 2716 return DAG.getNode(ISD::TRUNCATE, dl, VT, Op); 2717} 2718 2719static SDValue LowerCONCAT_VECTORS(SDValue Op, SelectionDAG &DAG) { 2720 // The only time a CONCAT_VECTORS operation can have legal types is when 2721 // two 64-bit vectors are concatenated to a 128-bit vector. 2722 assert(Op.getValueType().is128BitVector() && Op.getNumOperands() == 2 && 2723 "unexpected CONCAT_VECTORS"); 2724 DebugLoc dl = Op.getDebugLoc(); 2725 SDValue Val = DAG.getUNDEF(MVT::v2f64); 2726 SDValue Op0 = Op.getOperand(0); 2727 SDValue Op1 = Op.getOperand(1); 2728 if (Op0.getOpcode() != ISD::UNDEF) 2729 Val = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v2f64, Val, 2730 DAG.getNode(ISD::BIT_CONVERT, dl, MVT::f64, Op0), 2731 DAG.getIntPtrConstant(0)); 2732 if (Op1.getOpcode() != ISD::UNDEF) 2733 Val = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v2f64, Val, 2734 DAG.getNode(ISD::BIT_CONVERT, dl, MVT::f64, Op1), 2735 DAG.getIntPtrConstant(1)); 2736 return DAG.getNode(ISD::BIT_CONVERT, dl, Op.getValueType(), Val); 2737} 2738 2739SDValue ARMTargetLowering::LowerOperation(SDValue Op, SelectionDAG &DAG) { 2740 switch (Op.getOpcode()) { 2741 default: llvm_unreachable("Don't know how to custom lower this!"); 2742 case ISD::ConstantPool: return LowerConstantPool(Op, DAG); 2743 case ISD::GlobalAddress: 2744 return Subtarget->isTargetDarwin() ? LowerGlobalAddressDarwin(Op, DAG) : 2745 LowerGlobalAddressELF(Op, DAG); 2746 case ISD::GlobalTLSAddress: return LowerGlobalTLSAddress(Op, DAG); 2747 case ISD::SELECT_CC: return LowerSELECT_CC(Op, DAG, Subtarget); 2748 case ISD::BR_CC: return LowerBR_CC(Op, DAG, Subtarget); 2749 case ISD::BR_JT: return LowerBR_JT(Op, DAG); 2750 case ISD::DYNAMIC_STACKALLOC: return LowerDYNAMIC_STACKALLOC(Op, DAG); 2751 case ISD::VASTART: return LowerVASTART(Op, DAG, VarArgsFrameIndex); 2752 case ISD::SINT_TO_FP: 2753 case ISD::UINT_TO_FP: return LowerINT_TO_FP(Op, DAG); 2754 case ISD::FP_TO_SINT: 2755 case ISD::FP_TO_UINT: return LowerFP_TO_INT(Op, DAG); 2756 case ISD::FCOPYSIGN: return LowerFCOPYSIGN(Op, DAG); 2757 case ISD::RETURNADDR: break; 2758 case ISD::FRAMEADDR: return LowerFRAMEADDR(Op, DAG); 2759 case ISD::GLOBAL_OFFSET_TABLE: return LowerGLOBAL_OFFSET_TABLE(Op, DAG); 2760 case ISD::INTRINSIC_VOID: 2761 case ISD::INTRINSIC_W_CHAIN: return LowerINTRINSIC_W_CHAIN(Op, DAG); 2762 case ISD::INTRINSIC_WO_CHAIN: return LowerINTRINSIC_WO_CHAIN(Op, DAG); 2763 case ISD::BIT_CONVERT: return ExpandBIT_CONVERT(Op.getNode(), DAG); 2764 case ISD::SHL: 2765 case ISD::SRL: 2766 case ISD::SRA: return LowerShift(Op.getNode(), DAG, Subtarget); 2767 case ISD::VSETCC: return LowerVSETCC(Op, DAG); 2768 case ISD::BUILD_VECTOR: return LowerBUILD_VECTOR(Op, DAG); 2769 case ISD::VECTOR_SHUFFLE: return LowerVECTOR_SHUFFLE(Op, DAG); 2770 case ISD::EXTRACT_VECTOR_ELT: return LowerEXTRACT_VECTOR_ELT(Op, DAG); 2771 case ISD::CONCAT_VECTORS: return LowerCONCAT_VECTORS(Op, DAG); 2772 } 2773 return SDValue(); 2774} 2775 2776/// ReplaceNodeResults - Replace the results of node with an illegal result 2777/// type with new values built out of custom code. 2778void ARMTargetLowering::ReplaceNodeResults(SDNode *N, 2779 SmallVectorImpl<SDValue>&Results, 2780 SelectionDAG &DAG) { 2781 switch (N->getOpcode()) { 2782 default: 2783 llvm_unreachable("Don't know how to custom expand this!"); 2784 return; 2785 case ISD::BIT_CONVERT: 2786 Results.push_back(ExpandBIT_CONVERT(N, DAG)); 2787 return; 2788 case ISD::SRL: 2789 case ISD::SRA: { 2790 SDValue Res = LowerShift(N, DAG, Subtarget); 2791 if (Res.getNode()) 2792 Results.push_back(Res); 2793 return; 2794 } 2795 } 2796} 2797 2798//===----------------------------------------------------------------------===// 2799// ARM Scheduler Hooks 2800//===----------------------------------------------------------------------===// 2801 2802MachineBasicBlock * 2803ARMTargetLowering::EmitInstrWithCustomInserter(MachineInstr *MI, 2804 MachineBasicBlock *BB) const { 2805 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo(); 2806 DebugLoc dl = MI->getDebugLoc(); 2807 switch (MI->getOpcode()) { 2808 default: 2809 llvm_unreachable("Unexpected instr type to insert"); 2810 case ARM::tMOVCCr_pseudo: { 2811 // To "insert" a SELECT_CC instruction, we actually have to insert the 2812 // diamond control-flow pattern. The incoming instruction knows the 2813 // destination vreg to set, the condition code register to branch on, the 2814 // true/false values to select between, and a branch opcode to use. 2815 const BasicBlock *LLVM_BB = BB->getBasicBlock(); 2816 MachineFunction::iterator It = BB; 2817 ++It; 2818 2819 // thisMBB: 2820 // ... 2821 // TrueVal = ... 2822 // cmpTY ccX, r1, r2 2823 // bCC copy1MBB 2824 // fallthrough --> copy0MBB 2825 MachineBasicBlock *thisMBB = BB; 2826 MachineFunction *F = BB->getParent(); 2827 MachineBasicBlock *copy0MBB = F->CreateMachineBasicBlock(LLVM_BB); 2828 MachineBasicBlock *sinkMBB = F->CreateMachineBasicBlock(LLVM_BB); 2829 BuildMI(BB, dl, TII->get(ARM::tBcc)).addMBB(sinkMBB) 2830 .addImm(MI->getOperand(3).getImm()).addReg(MI->getOperand(4).getReg()); 2831 F->insert(It, copy0MBB); 2832 F->insert(It, sinkMBB); 2833 // Update machine-CFG edges by first adding all successors of the current 2834 // block to the new block which will contain the Phi node for the select. 2835 for(MachineBasicBlock::succ_iterator i = BB->succ_begin(), 2836 e = BB->succ_end(); i != e; ++i) 2837 sinkMBB->addSuccessor(*i); 2838 // Next, remove all successors of the current block, and add the true 2839 // and fallthrough blocks as its successors. 2840 while(!BB->succ_empty()) 2841 BB->removeSuccessor(BB->succ_begin()); 2842 BB->addSuccessor(copy0MBB); 2843 BB->addSuccessor(sinkMBB); 2844 2845 // copy0MBB: 2846 // %FalseValue = ... 2847 // # fallthrough to sinkMBB 2848 BB = copy0MBB; 2849 2850 // Update machine-CFG edges 2851 BB->addSuccessor(sinkMBB); 2852 2853 // sinkMBB: 2854 // %Result = phi [ %FalseValue, copy0MBB ], [ %TrueValue, thisMBB ] 2855 // ... 2856 BB = sinkMBB; 2857 BuildMI(BB, dl, TII->get(ARM::PHI), MI->getOperand(0).getReg()) 2858 .addReg(MI->getOperand(1).getReg()).addMBB(copy0MBB) 2859 .addReg(MI->getOperand(2).getReg()).addMBB(thisMBB); 2860 2861 F->DeleteMachineInstr(MI); // The pseudo instruction is gone now. 2862 return BB; 2863 } 2864 2865 case ARM::tANDsp: 2866 case ARM::tADDspr_: 2867 case ARM::tSUBspi_: 2868 case ARM::t2SUBrSPi_: 2869 case ARM::t2SUBrSPi12_: 2870 case ARM::t2SUBrSPs_: { 2871 MachineFunction *MF = BB->getParent(); 2872 unsigned DstReg = MI->getOperand(0).getReg(); 2873 unsigned SrcReg = MI->getOperand(1).getReg(); 2874 bool DstIsDead = MI->getOperand(0).isDead(); 2875 bool SrcIsKill = MI->getOperand(1).isKill(); 2876 2877 if (SrcReg != ARM::SP) { 2878 // Copy the source to SP from virtual register. 2879 const TargetRegisterClass *RC = MF->getRegInfo().getRegClass(SrcReg); 2880 unsigned CopyOpc = (RC == ARM::tGPRRegisterClass) 2881 ? ARM::tMOVtgpr2gpr : ARM::tMOVgpr2gpr; 2882 BuildMI(BB, dl, TII->get(CopyOpc), ARM::SP) 2883 .addReg(SrcReg, getKillRegState(SrcIsKill)); 2884 } 2885 2886 unsigned OpOpc = 0; 2887 bool NeedPred = false, NeedCC = false, NeedOp3 = false; 2888 switch (MI->getOpcode()) { 2889 default: 2890 llvm_unreachable("Unexpected pseudo instruction!"); 2891 case ARM::tANDsp: 2892 OpOpc = ARM::tAND; 2893 NeedPred = true; 2894 break; 2895 case ARM::tADDspr_: 2896 OpOpc = ARM::tADDspr; 2897 break; 2898 case ARM::tSUBspi_: 2899 OpOpc = ARM::tSUBspi; 2900 break; 2901 case ARM::t2SUBrSPi_: 2902 OpOpc = ARM::t2SUBrSPi; 2903 NeedPred = true; NeedCC = true; 2904 break; 2905 case ARM::t2SUBrSPi12_: 2906 OpOpc = ARM::t2SUBrSPi12; 2907 NeedPred = true; 2908 break; 2909 case ARM::t2SUBrSPs_: 2910 OpOpc = ARM::t2SUBrSPs; 2911 NeedPred = true; NeedCC = true; NeedOp3 = true; 2912 break; 2913 } 2914 MachineInstrBuilder MIB = BuildMI(BB, dl, TII->get(OpOpc), ARM::SP); 2915 if (OpOpc == ARM::tAND) 2916 AddDefaultT1CC(MIB); 2917 MIB.addReg(ARM::SP); 2918 MIB.addOperand(MI->getOperand(2)); 2919 if (NeedOp3) 2920 MIB.addOperand(MI->getOperand(3)); 2921 if (NeedPred) 2922 AddDefaultPred(MIB); 2923 if (NeedCC) 2924 AddDefaultCC(MIB); 2925 2926 // Copy the result from SP to virtual register. 2927 const TargetRegisterClass *RC = MF->getRegInfo().getRegClass(DstReg); 2928 unsigned CopyOpc = (RC == ARM::tGPRRegisterClass) 2929 ? ARM::tMOVgpr2tgpr : ARM::tMOVgpr2gpr; 2930 BuildMI(BB, dl, TII->get(CopyOpc)) 2931 .addReg(DstReg, getDefRegState(true) | getDeadRegState(DstIsDead)) 2932 .addReg(ARM::SP); 2933 MF->DeleteMachineInstr(MI); // The pseudo instruction is gone now. 2934 return BB; 2935 } 2936 } 2937} 2938 2939//===----------------------------------------------------------------------===// 2940// ARM Optimization Hooks 2941//===----------------------------------------------------------------------===// 2942 2943static 2944SDValue combineSelectAndUse(SDNode *N, SDValue Slct, SDValue OtherOp, 2945 TargetLowering::DAGCombinerInfo &DCI) { 2946 SelectionDAG &DAG = DCI.DAG; 2947 const TargetLowering &TLI = DAG.getTargetLoweringInfo(); 2948 EVT VT = N->getValueType(0); 2949 unsigned Opc = N->getOpcode(); 2950 bool isSlctCC = Slct.getOpcode() == ISD::SELECT_CC; 2951 SDValue LHS = isSlctCC ? Slct.getOperand(2) : Slct.getOperand(1); 2952 SDValue RHS = isSlctCC ? Slct.getOperand(3) : Slct.getOperand(2); 2953 ISD::CondCode CC = ISD::SETCC_INVALID; 2954 2955 if (isSlctCC) { 2956 CC = cast<CondCodeSDNode>(Slct.getOperand(4))->get(); 2957 } else { 2958 SDValue CCOp = Slct.getOperand(0); 2959 if (CCOp.getOpcode() == ISD::SETCC) 2960 CC = cast<CondCodeSDNode>(CCOp.getOperand(2))->get(); 2961 } 2962 2963 bool DoXform = false; 2964 bool InvCC = false; 2965 assert ((Opc == ISD::ADD || (Opc == ISD::SUB && Slct == N->getOperand(1))) && 2966 "Bad input!"); 2967 2968 if (LHS.getOpcode() == ISD::Constant && 2969 cast<ConstantSDNode>(LHS)->isNullValue()) { 2970 DoXform = true; 2971 } else if (CC != ISD::SETCC_INVALID && 2972 RHS.getOpcode() == ISD::Constant && 2973 cast<ConstantSDNode>(RHS)->isNullValue()) { 2974 std::swap(LHS, RHS); 2975 SDValue Op0 = Slct.getOperand(0); 2976 EVT OpVT = isSlctCC ? Op0.getValueType() : 2977 Op0.getOperand(0).getValueType(); 2978 bool isInt = OpVT.isInteger(); 2979 CC = ISD::getSetCCInverse(CC, isInt); 2980 2981 if (!TLI.isCondCodeLegal(CC, OpVT)) 2982 return SDValue(); // Inverse operator isn't legal. 2983 2984 DoXform = true; 2985 InvCC = true; 2986 } 2987 2988 if (DoXform) { 2989 SDValue Result = DAG.getNode(Opc, RHS.getDebugLoc(), VT, OtherOp, RHS); 2990 if (isSlctCC) 2991 return DAG.getSelectCC(N->getDebugLoc(), OtherOp, Result, 2992 Slct.getOperand(0), Slct.getOperand(1), CC); 2993 SDValue CCOp = Slct.getOperand(0); 2994 if (InvCC) 2995 CCOp = DAG.getSetCC(Slct.getDebugLoc(), CCOp.getValueType(), 2996 CCOp.getOperand(0), CCOp.getOperand(1), CC); 2997 return DAG.getNode(ISD::SELECT, N->getDebugLoc(), VT, 2998 CCOp, OtherOp, Result); 2999 } 3000 return SDValue(); 3001} 3002 3003/// PerformADDCombine - Target-specific dag combine xforms for ISD::ADD. 3004static SDValue PerformADDCombine(SDNode *N, 3005 TargetLowering::DAGCombinerInfo &DCI) { 3006 // added by evan in r37685 with no testcase. 3007 SDValue N0 = N->getOperand(0), N1 = N->getOperand(1); 3008 3009 // fold (add (select cc, 0, c), x) -> (select cc, x, (add, x, c)) 3010 if (N0.getOpcode() == ISD::SELECT && N0.getNode()->hasOneUse()) { 3011 SDValue Result = combineSelectAndUse(N, N0, N1, DCI); 3012 if (Result.getNode()) return Result; 3013 } 3014 if (N1.getOpcode() == ISD::SELECT && N1.getNode()->hasOneUse()) { 3015 SDValue Result = combineSelectAndUse(N, N1, N0, DCI); 3016 if (Result.getNode()) return Result; 3017 } 3018 3019 return SDValue(); 3020} 3021 3022/// PerformSUBCombine - Target-specific dag combine xforms for ISD::SUB. 3023static SDValue PerformSUBCombine(SDNode *N, 3024 TargetLowering::DAGCombinerInfo &DCI) { 3025 // added by evan in r37685 with no testcase. 3026 SDValue N0 = N->getOperand(0), N1 = N->getOperand(1); 3027 3028 // fold (sub x, (select cc, 0, c)) -> (select cc, x, (sub, x, c)) 3029 if (N1.getOpcode() == ISD::SELECT && N1.getNode()->hasOneUse()) { 3030 SDValue Result = combineSelectAndUse(N, N1, N0, DCI); 3031 if (Result.getNode()) return Result; 3032 } 3033 3034 return SDValue(); 3035} 3036 3037 3038/// PerformFMRRDCombine - Target-specific dag combine xforms for ARMISD::FMRRD. 3039static SDValue PerformFMRRDCombine(SDNode *N, 3040 TargetLowering::DAGCombinerInfo &DCI) { 3041 // fmrrd(fmdrr x, y) -> x,y 3042 SDValue InDouble = N->getOperand(0); 3043 if (InDouble.getOpcode() == ARMISD::FMDRR) 3044 return DCI.CombineTo(N, InDouble.getOperand(0), InDouble.getOperand(1)); 3045 return SDValue(); 3046} 3047 3048/// getVShiftImm - Check if this is a valid build_vector for the immediate 3049/// operand of a vector shift operation, where all the elements of the 3050/// build_vector must have the same constant integer value. 3051static bool getVShiftImm(SDValue Op, unsigned ElementBits, int64_t &Cnt) { 3052 // Ignore bit_converts. 3053 while (Op.getOpcode() == ISD::BIT_CONVERT) 3054 Op = Op.getOperand(0); 3055 BuildVectorSDNode *BVN = dyn_cast<BuildVectorSDNode>(Op.getNode()); 3056 APInt SplatBits, SplatUndef; 3057 unsigned SplatBitSize; 3058 bool HasAnyUndefs; 3059 if (! BVN || ! BVN->isConstantSplat(SplatBits, SplatUndef, SplatBitSize, 3060 HasAnyUndefs, ElementBits) || 3061 SplatBitSize > ElementBits) 3062 return false; 3063 Cnt = SplatBits.getSExtValue(); 3064 return true; 3065} 3066 3067/// isVShiftLImm - Check if this is a valid build_vector for the immediate 3068/// operand of a vector shift left operation. That value must be in the range: 3069/// 0 <= Value < ElementBits for a left shift; or 3070/// 0 <= Value <= ElementBits for a long left shift. 3071static bool isVShiftLImm(SDValue Op, EVT VT, bool isLong, int64_t &Cnt) { 3072 assert(VT.isVector() && "vector shift count is not a vector type"); 3073 unsigned ElementBits = VT.getVectorElementType().getSizeInBits(); 3074 if (! getVShiftImm(Op, ElementBits, Cnt)) 3075 return false; 3076 return (Cnt >= 0 && (isLong ? Cnt-1 : Cnt) < ElementBits); 3077} 3078 3079/// isVShiftRImm - Check if this is a valid build_vector for the immediate 3080/// operand of a vector shift right operation. For a shift opcode, the value 3081/// is positive, but for an intrinsic the value count must be negative. The 3082/// absolute value must be in the range: 3083/// 1 <= |Value| <= ElementBits for a right shift; or 3084/// 1 <= |Value| <= ElementBits/2 for a narrow right shift. 3085static bool isVShiftRImm(SDValue Op, EVT VT, bool isNarrow, bool isIntrinsic, 3086 int64_t &Cnt) { 3087 assert(VT.isVector() && "vector shift count is not a vector type"); 3088 unsigned ElementBits = VT.getVectorElementType().getSizeInBits(); 3089 if (! getVShiftImm(Op, ElementBits, Cnt)) 3090 return false; 3091 if (isIntrinsic) 3092 Cnt = -Cnt; 3093 return (Cnt >= 1 && Cnt <= (isNarrow ? ElementBits/2 : ElementBits)); 3094} 3095 3096/// PerformIntrinsicCombine - ARM-specific DAG combining for intrinsics. 3097static SDValue PerformIntrinsicCombine(SDNode *N, SelectionDAG &DAG) { 3098 unsigned IntNo = cast<ConstantSDNode>(N->getOperand(0))->getZExtValue(); 3099 switch (IntNo) { 3100 default: 3101 // Don't do anything for most intrinsics. 3102 break; 3103 3104 // Vector shifts: check for immediate versions and lower them. 3105 // Note: This is done during DAG combining instead of DAG legalizing because 3106 // the build_vectors for 64-bit vector element shift counts are generally 3107 // not legal, and it is hard to see their values after they get legalized to 3108 // loads from a constant pool. 3109 case Intrinsic::arm_neon_vshifts: 3110 case Intrinsic::arm_neon_vshiftu: 3111 case Intrinsic::arm_neon_vshiftls: 3112 case Intrinsic::arm_neon_vshiftlu: 3113 case Intrinsic::arm_neon_vshiftn: 3114 case Intrinsic::arm_neon_vrshifts: 3115 case Intrinsic::arm_neon_vrshiftu: 3116 case Intrinsic::arm_neon_vrshiftn: 3117 case Intrinsic::arm_neon_vqshifts: 3118 case Intrinsic::arm_neon_vqshiftu: 3119 case Intrinsic::arm_neon_vqshiftsu: 3120 case Intrinsic::arm_neon_vqshiftns: 3121 case Intrinsic::arm_neon_vqshiftnu: 3122 case Intrinsic::arm_neon_vqshiftnsu: 3123 case Intrinsic::arm_neon_vqrshiftns: 3124 case Intrinsic::arm_neon_vqrshiftnu: 3125 case Intrinsic::arm_neon_vqrshiftnsu: { 3126 EVT VT = N->getOperand(1).getValueType(); 3127 int64_t Cnt; 3128 unsigned VShiftOpc = 0; 3129 3130 switch (IntNo) { 3131 case Intrinsic::arm_neon_vshifts: 3132 case Intrinsic::arm_neon_vshiftu: 3133 if (isVShiftLImm(N->getOperand(2), VT, false, Cnt)) { 3134 VShiftOpc = ARMISD::VSHL; 3135 break; 3136 } 3137 if (isVShiftRImm(N->getOperand(2), VT, false, true, Cnt)) { 3138 VShiftOpc = (IntNo == Intrinsic::arm_neon_vshifts ? 3139 ARMISD::VSHRs : ARMISD::VSHRu); 3140 break; 3141 } 3142 return SDValue(); 3143 3144 case Intrinsic::arm_neon_vshiftls: 3145 case Intrinsic::arm_neon_vshiftlu: 3146 if (isVShiftLImm(N->getOperand(2), VT, true, Cnt)) 3147 break; 3148 llvm_unreachable("invalid shift count for vshll intrinsic"); 3149 3150 case Intrinsic::arm_neon_vrshifts: 3151 case Intrinsic::arm_neon_vrshiftu: 3152 if (isVShiftRImm(N->getOperand(2), VT, false, true, Cnt)) 3153 break; 3154 return SDValue(); 3155 3156 case Intrinsic::arm_neon_vqshifts: 3157 case Intrinsic::arm_neon_vqshiftu: 3158 if (isVShiftLImm(N->getOperand(2), VT, false, Cnt)) 3159 break; 3160 return SDValue(); 3161 3162 case Intrinsic::arm_neon_vqshiftsu: 3163 if (isVShiftLImm(N->getOperand(2), VT, false, Cnt)) 3164 break; 3165 llvm_unreachable("invalid shift count for vqshlu intrinsic"); 3166 3167 case Intrinsic::arm_neon_vshiftn: 3168 case Intrinsic::arm_neon_vrshiftn: 3169 case Intrinsic::arm_neon_vqshiftns: 3170 case Intrinsic::arm_neon_vqshiftnu: 3171 case Intrinsic::arm_neon_vqshiftnsu: 3172 case Intrinsic::arm_neon_vqrshiftns: 3173 case Intrinsic::arm_neon_vqrshiftnu: 3174 case Intrinsic::arm_neon_vqrshiftnsu: 3175 // Narrowing shifts require an immediate right shift. 3176 if (isVShiftRImm(N->getOperand(2), VT, true, true, Cnt)) 3177 break; 3178 llvm_unreachable("invalid shift count for narrowing vector shift intrinsic"); 3179 3180 default: 3181 llvm_unreachable("unhandled vector shift"); 3182 } 3183 3184 switch (IntNo) { 3185 case Intrinsic::arm_neon_vshifts: 3186 case Intrinsic::arm_neon_vshiftu: 3187 // Opcode already set above. 3188 break; 3189 case Intrinsic::arm_neon_vshiftls: 3190 case Intrinsic::arm_neon_vshiftlu: 3191 if (Cnt == VT.getVectorElementType().getSizeInBits()) 3192 VShiftOpc = ARMISD::VSHLLi; 3193 else 3194 VShiftOpc = (IntNo == Intrinsic::arm_neon_vshiftls ? 3195 ARMISD::VSHLLs : ARMISD::VSHLLu); 3196 break; 3197 case Intrinsic::arm_neon_vshiftn: 3198 VShiftOpc = ARMISD::VSHRN; break; 3199 case Intrinsic::arm_neon_vrshifts: 3200 VShiftOpc = ARMISD::VRSHRs; break; 3201 case Intrinsic::arm_neon_vrshiftu: 3202 VShiftOpc = ARMISD::VRSHRu; break; 3203 case Intrinsic::arm_neon_vrshiftn: 3204 VShiftOpc = ARMISD::VRSHRN; break; 3205 case Intrinsic::arm_neon_vqshifts: 3206 VShiftOpc = ARMISD::VQSHLs; break; 3207 case Intrinsic::arm_neon_vqshiftu: 3208 VShiftOpc = ARMISD::VQSHLu; break; 3209 case Intrinsic::arm_neon_vqshiftsu: 3210 VShiftOpc = ARMISD::VQSHLsu; break; 3211 case Intrinsic::arm_neon_vqshiftns: 3212 VShiftOpc = ARMISD::VQSHRNs; break; 3213 case Intrinsic::arm_neon_vqshiftnu: 3214 VShiftOpc = ARMISD::VQSHRNu; break; 3215 case Intrinsic::arm_neon_vqshiftnsu: 3216 VShiftOpc = ARMISD::VQSHRNsu; break; 3217 case Intrinsic::arm_neon_vqrshiftns: 3218 VShiftOpc = ARMISD::VQRSHRNs; break; 3219 case Intrinsic::arm_neon_vqrshiftnu: 3220 VShiftOpc = ARMISD::VQRSHRNu; break; 3221 case Intrinsic::arm_neon_vqrshiftnsu: 3222 VShiftOpc = ARMISD::VQRSHRNsu; break; 3223 } 3224 3225 return DAG.getNode(VShiftOpc, N->getDebugLoc(), N->getValueType(0), 3226 N->getOperand(1), DAG.getConstant(Cnt, MVT::i32)); 3227 } 3228 3229 case Intrinsic::arm_neon_vshiftins: { 3230 EVT VT = N->getOperand(1).getValueType(); 3231 int64_t Cnt; 3232 unsigned VShiftOpc = 0; 3233 3234 if (isVShiftLImm(N->getOperand(3), VT, false, Cnt)) 3235 VShiftOpc = ARMISD::VSLI; 3236 else if (isVShiftRImm(N->getOperand(3), VT, false, true, Cnt)) 3237 VShiftOpc = ARMISD::VSRI; 3238 else { 3239 llvm_unreachable("invalid shift count for vsli/vsri intrinsic"); 3240 } 3241 3242 return DAG.getNode(VShiftOpc, N->getDebugLoc(), N->getValueType(0), 3243 N->getOperand(1), N->getOperand(2), 3244 DAG.getConstant(Cnt, MVT::i32)); 3245 } 3246 3247 case Intrinsic::arm_neon_vqrshifts: 3248 case Intrinsic::arm_neon_vqrshiftu: 3249 // No immediate versions of these to check for. 3250 break; 3251 } 3252 3253 return SDValue(); 3254} 3255 3256/// PerformShiftCombine - Checks for immediate versions of vector shifts and 3257/// lowers them. As with the vector shift intrinsics, this is done during DAG 3258/// combining instead of DAG legalizing because the build_vectors for 64-bit 3259/// vector element shift counts are generally not legal, and it is hard to see 3260/// their values after they get legalized to loads from a constant pool. 3261static SDValue PerformShiftCombine(SDNode *N, SelectionDAG &DAG, 3262 const ARMSubtarget *ST) { 3263 EVT VT = N->getValueType(0); 3264 3265 // Nothing to be done for scalar shifts. 3266 if (! VT.isVector()) 3267 return SDValue(); 3268 3269 assert(ST->hasNEON() && "unexpected vector shift"); 3270 int64_t Cnt; 3271 3272 switch (N->getOpcode()) { 3273 default: llvm_unreachable("unexpected shift opcode"); 3274 3275 case ISD::SHL: 3276 if (isVShiftLImm(N->getOperand(1), VT, false, Cnt)) 3277 return DAG.getNode(ARMISD::VSHL, N->getDebugLoc(), VT, N->getOperand(0), 3278 DAG.getConstant(Cnt, MVT::i32)); 3279 break; 3280 3281 case ISD::SRA: 3282 case ISD::SRL: 3283 if (isVShiftRImm(N->getOperand(1), VT, false, false, Cnt)) { 3284 unsigned VShiftOpc = (N->getOpcode() == ISD::SRA ? 3285 ARMISD::VSHRs : ARMISD::VSHRu); 3286 return DAG.getNode(VShiftOpc, N->getDebugLoc(), VT, N->getOperand(0), 3287 DAG.getConstant(Cnt, MVT::i32)); 3288 } 3289 } 3290 return SDValue(); 3291} 3292 3293/// PerformExtendCombine - Target-specific DAG combining for ISD::SIGN_EXTEND, 3294/// ISD::ZERO_EXTEND, and ISD::ANY_EXTEND. 3295static SDValue PerformExtendCombine(SDNode *N, SelectionDAG &DAG, 3296 const ARMSubtarget *ST) { 3297 SDValue N0 = N->getOperand(0); 3298 3299 // Check for sign- and zero-extensions of vector extract operations of 8- 3300 // and 16-bit vector elements. NEON supports these directly. They are 3301 // handled during DAG combining because type legalization will promote them 3302 // to 32-bit types and it is messy to recognize the operations after that. 3303 if (ST->hasNEON() && N0.getOpcode() == ISD::EXTRACT_VECTOR_ELT) { 3304 SDValue Vec = N0.getOperand(0); 3305 SDValue Lane = N0.getOperand(1); 3306 EVT VT = N->getValueType(0); 3307 EVT EltVT = N0.getValueType(); 3308 const TargetLowering &TLI = DAG.getTargetLoweringInfo(); 3309 3310 if (VT == MVT::i32 && 3311 (EltVT == MVT::i8 || EltVT == MVT::i16) && 3312 TLI.isTypeLegal(Vec.getValueType())) { 3313 3314 unsigned Opc = 0; 3315 switch (N->getOpcode()) { 3316 default: llvm_unreachable("unexpected opcode"); 3317 case ISD::SIGN_EXTEND: 3318 Opc = ARMISD::VGETLANEs; 3319 break; 3320 case ISD::ZERO_EXTEND: 3321 case ISD::ANY_EXTEND: 3322 Opc = ARMISD::VGETLANEu; 3323 break; 3324 } 3325 return DAG.getNode(Opc, N->getDebugLoc(), VT, Vec, Lane); 3326 } 3327 } 3328 3329 return SDValue(); 3330} 3331 3332SDValue ARMTargetLowering::PerformDAGCombine(SDNode *N, 3333 DAGCombinerInfo &DCI) const { 3334 switch (N->getOpcode()) { 3335 default: break; 3336 case ISD::ADD: return PerformADDCombine(N, DCI); 3337 case ISD::SUB: return PerformSUBCombine(N, DCI); 3338 case ARMISD::FMRRD: return PerformFMRRDCombine(N, DCI); 3339 case ISD::INTRINSIC_WO_CHAIN: 3340 return PerformIntrinsicCombine(N, DCI.DAG); 3341 case ISD::SHL: 3342 case ISD::SRA: 3343 case ISD::SRL: 3344 return PerformShiftCombine(N, DCI.DAG, Subtarget); 3345 case ISD::SIGN_EXTEND: 3346 case ISD::ZERO_EXTEND: 3347 case ISD::ANY_EXTEND: 3348 return PerformExtendCombine(N, DCI.DAG, Subtarget); 3349 } 3350 return SDValue(); 3351} 3352 3353bool ARMTargetLowering::allowsUnalignedMemoryAccesses(EVT VT) const { 3354 if (!Subtarget->hasV6Ops()) 3355 // Pre-v6 does not support unaligned mem access. 3356 return false; 3357 else if (!Subtarget->hasV6Ops()) { 3358 // v6 may or may not support unaligned mem access. 3359 if (!Subtarget->isTargetDarwin()) 3360 return false; 3361 } 3362 3363 switch (VT.getSimpleVT().SimpleTy) { 3364 default: 3365 return false; 3366 case MVT::i8: 3367 case MVT::i16: 3368 case MVT::i32: 3369 return true; 3370 // FIXME: VLD1 etc with standard alignment is legal. 3371 } 3372} 3373 3374static bool isLegalT1AddressImmediate(int64_t V, EVT VT) { 3375 if (V < 0) 3376 return false; 3377 3378 unsigned Scale = 1; 3379 switch (VT.getSimpleVT().SimpleTy) { 3380 default: return false; 3381 case MVT::i1: 3382 case MVT::i8: 3383 // Scale == 1; 3384 break; 3385 case MVT::i16: 3386 // Scale == 2; 3387 Scale = 2; 3388 break; 3389 case MVT::i32: 3390 // Scale == 4; 3391 Scale = 4; 3392 break; 3393 } 3394 3395 if ((V & (Scale - 1)) != 0) 3396 return false; 3397 V /= Scale; 3398 return V == (V & ((1LL << 5) - 1)); 3399} 3400 3401static bool isLegalT2AddressImmediate(int64_t V, EVT VT, 3402 const ARMSubtarget *Subtarget) { 3403 bool isNeg = false; 3404 if (V < 0) { 3405 isNeg = true; 3406 V = - V; 3407 } 3408 3409 switch (VT.getSimpleVT().SimpleTy) { 3410 default: return false; 3411 case MVT::i1: 3412 case MVT::i8: 3413 case MVT::i16: 3414 case MVT::i32: 3415 // + imm12 or - imm8 3416 if (isNeg) 3417 return V == (V & ((1LL << 8) - 1)); 3418 return V == (V & ((1LL << 12) - 1)); 3419 case MVT::f32: 3420 case MVT::f64: 3421 // Same as ARM mode. FIXME: NEON? 3422 if (!Subtarget->hasVFP2()) 3423 return false; 3424 if ((V & 3) != 0) 3425 return false; 3426 V >>= 2; 3427 return V == (V & ((1LL << 8) - 1)); 3428 } 3429} 3430 3431/// isLegalAddressImmediate - Return true if the integer value can be used 3432/// as the offset of the target addressing mode for load / store of the 3433/// given type. 3434static bool isLegalAddressImmediate(int64_t V, EVT VT, 3435 const ARMSubtarget *Subtarget) { 3436 if (V == 0) 3437 return true; 3438 3439 if (!VT.isSimple()) 3440 return false; 3441 3442 if (Subtarget->isThumb1Only()) 3443 return isLegalT1AddressImmediate(V, VT); 3444 else if (Subtarget->isThumb2()) 3445 return isLegalT2AddressImmediate(V, VT, Subtarget); 3446 3447 // ARM mode. 3448 if (V < 0) 3449 V = - V; 3450 switch (VT.getSimpleVT().SimpleTy) { 3451 default: return false; 3452 case MVT::i1: 3453 case MVT::i8: 3454 case MVT::i32: 3455 // +- imm12 3456 return V == (V & ((1LL << 12) - 1)); 3457 case MVT::i16: 3458 // +- imm8 3459 return V == (V & ((1LL << 8) - 1)); 3460 case MVT::f32: 3461 case MVT::f64: 3462 if (!Subtarget->hasVFP2()) // FIXME: NEON? 3463 return false; 3464 if ((V & 3) != 0) 3465 return false; 3466 V >>= 2; 3467 return V == (V & ((1LL << 8) - 1)); 3468 } 3469} 3470 3471bool ARMTargetLowering::isLegalT2ScaledAddressingMode(const AddrMode &AM, 3472 EVT VT) const { 3473 int Scale = AM.Scale; 3474 if (Scale < 0) 3475 return false; 3476 3477 switch (VT.getSimpleVT().SimpleTy) { 3478 default: return false; 3479 case MVT::i1: 3480 case MVT::i8: 3481 case MVT::i16: 3482 case MVT::i32: 3483 if (Scale == 1) 3484 return true; 3485 // r + r << imm 3486 Scale = Scale & ~1; 3487 return Scale == 2 || Scale == 4 || Scale == 8; 3488 case MVT::i64: 3489 // r + r 3490 if (((unsigned)AM.HasBaseReg + Scale) <= 2) 3491 return true; 3492 return false; 3493 case MVT::isVoid: 3494 // Note, we allow "void" uses (basically, uses that aren't loads or 3495 // stores), because arm allows folding a scale into many arithmetic 3496 // operations. This should be made more precise and revisited later. 3497 3498 // Allow r << imm, but the imm has to be a multiple of two. 3499 if (Scale & 1) return false; 3500 return isPowerOf2_32(Scale); 3501 } 3502} 3503 3504/// isLegalAddressingMode - Return true if the addressing mode represented 3505/// by AM is legal for this target, for a load/store of the specified type. 3506bool ARMTargetLowering::isLegalAddressingMode(const AddrMode &AM, 3507 const Type *Ty) const { 3508 EVT VT = getValueType(Ty, true); 3509 if (!isLegalAddressImmediate(AM.BaseOffs, VT, Subtarget)) 3510 return false; 3511 3512 // Can never fold addr of global into load/store. 3513 if (AM.BaseGV) 3514 return false; 3515 3516 switch (AM.Scale) { 3517 case 0: // no scale reg, must be "r+i" or "r", or "i". 3518 break; 3519 case 1: 3520 if (Subtarget->isThumb1Only()) 3521 return false; 3522 // FALL THROUGH. 3523 default: 3524 // ARM doesn't support any R+R*scale+imm addr modes. 3525 if (AM.BaseOffs) 3526 return false; 3527 3528 if (!VT.isSimple()) 3529 return false; 3530 3531 if (Subtarget->isThumb2()) 3532 return isLegalT2ScaledAddressingMode(AM, VT); 3533 3534 int Scale = AM.Scale; 3535 switch (VT.getSimpleVT().SimpleTy) { 3536 default: return false; 3537 case MVT::i1: 3538 case MVT::i8: 3539 case MVT::i32: 3540 if (Scale < 0) Scale = -Scale; 3541 if (Scale == 1) 3542 return true; 3543 // r + r << imm 3544 return isPowerOf2_32(Scale & ~1); 3545 case MVT::i16: 3546 case MVT::i64: 3547 // r + r 3548 if (((unsigned)AM.HasBaseReg + Scale) <= 2) 3549 return true; 3550 return false; 3551 3552 case MVT::isVoid: 3553 // Note, we allow "void" uses (basically, uses that aren't loads or 3554 // stores), because arm allows folding a scale into many arithmetic 3555 // operations. This should be made more precise and revisited later. 3556 3557 // Allow r << imm, but the imm has to be a multiple of two. 3558 if (Scale & 1) return false; 3559 return isPowerOf2_32(Scale); 3560 } 3561 break; 3562 } 3563 return true; 3564} 3565 3566static bool getARMIndexedAddressParts(SDNode *Ptr, EVT VT, 3567 bool isSEXTLoad, SDValue &Base, 3568 SDValue &Offset, bool &isInc, 3569 SelectionDAG &DAG) { 3570 if (Ptr->getOpcode() != ISD::ADD && Ptr->getOpcode() != ISD::SUB) 3571 return false; 3572 3573 if (VT == MVT::i16 || ((VT == MVT::i8 || VT == MVT::i1) && isSEXTLoad)) { 3574 // AddressingMode 3 3575 Base = Ptr->getOperand(0); 3576 if (ConstantSDNode *RHS = dyn_cast<ConstantSDNode>(Ptr->getOperand(1))) { 3577 int RHSC = (int)RHS->getZExtValue(); 3578 if (RHSC < 0 && RHSC > -256) { 3579 assert(Ptr->getOpcode() == ISD::ADD); 3580 isInc = false; 3581 Offset = DAG.getConstant(-RHSC, RHS->getValueType(0)); 3582 return true; 3583 } 3584 } 3585 isInc = (Ptr->getOpcode() == ISD::ADD); 3586 Offset = Ptr->getOperand(1); 3587 return true; 3588 } else if (VT == MVT::i32 || VT == MVT::i8 || VT == MVT::i1) { 3589 // AddressingMode 2 3590 if (ConstantSDNode *RHS = dyn_cast<ConstantSDNode>(Ptr->getOperand(1))) { 3591 int RHSC = (int)RHS->getZExtValue(); 3592 if (RHSC < 0 && RHSC > -0x1000) { 3593 assert(Ptr->getOpcode() == ISD::ADD); 3594 isInc = false; 3595 Offset = DAG.getConstant(-RHSC, RHS->getValueType(0)); 3596 Base = Ptr->getOperand(0); 3597 return true; 3598 } 3599 } 3600 3601 if (Ptr->getOpcode() == ISD::ADD) { 3602 isInc = true; 3603 ARM_AM::ShiftOpc ShOpcVal= ARM_AM::getShiftOpcForNode(Ptr->getOperand(0)); 3604 if (ShOpcVal != ARM_AM::no_shift) { 3605 Base = Ptr->getOperand(1); 3606 Offset = Ptr->getOperand(0); 3607 } else { 3608 Base = Ptr->getOperand(0); 3609 Offset = Ptr->getOperand(1); 3610 } 3611 return true; 3612 } 3613 3614 isInc = (Ptr->getOpcode() == ISD::ADD); 3615 Base = Ptr->getOperand(0); 3616 Offset = Ptr->getOperand(1); 3617 return true; 3618 } 3619 3620 // FIXME: Use FLDM / FSTM to emulate indexed FP load / store. 3621 return false; 3622} 3623 3624static bool getT2IndexedAddressParts(SDNode *Ptr, EVT VT, 3625 bool isSEXTLoad, SDValue &Base, 3626 SDValue &Offset, bool &isInc, 3627 SelectionDAG &DAG) { 3628 if (Ptr->getOpcode() != ISD::ADD && Ptr->getOpcode() != ISD::SUB) 3629 return false; 3630 3631 Base = Ptr->getOperand(0); 3632 if (ConstantSDNode *RHS = dyn_cast<ConstantSDNode>(Ptr->getOperand(1))) { 3633 int RHSC = (int)RHS->getZExtValue(); 3634 if (RHSC < 0 && RHSC > -0x100) { // 8 bits. 3635 assert(Ptr->getOpcode() == ISD::ADD); 3636 isInc = false; 3637 Offset = DAG.getConstant(-RHSC, RHS->getValueType(0)); 3638 return true; 3639 } else if (RHSC > 0 && RHSC < 0x100) { // 8 bit, no zero. 3640 isInc = Ptr->getOpcode() == ISD::ADD; 3641 Offset = DAG.getConstant(RHSC, RHS->getValueType(0)); 3642 return true; 3643 } 3644 } 3645 3646 return false; 3647} 3648 3649/// getPreIndexedAddressParts - returns true by value, base pointer and 3650/// offset pointer and addressing mode by reference if the node's address 3651/// can be legally represented as pre-indexed load / store address. 3652bool 3653ARMTargetLowering::getPreIndexedAddressParts(SDNode *N, SDValue &Base, 3654 SDValue &Offset, 3655 ISD::MemIndexedMode &AM, 3656 SelectionDAG &DAG) const { 3657 if (Subtarget->isThumb1Only()) 3658 return false; 3659 3660 EVT VT; 3661 SDValue Ptr; 3662 bool isSEXTLoad = false; 3663 if (LoadSDNode *LD = dyn_cast<LoadSDNode>(N)) { 3664 Ptr = LD->getBasePtr(); 3665 VT = LD->getMemoryVT(); 3666 isSEXTLoad = LD->getExtensionType() == ISD::SEXTLOAD; 3667 } else if (StoreSDNode *ST = dyn_cast<StoreSDNode>(N)) { 3668 Ptr = ST->getBasePtr(); 3669 VT = ST->getMemoryVT(); 3670 } else 3671 return false; 3672 3673 bool isInc; 3674 bool isLegal = false; 3675 if (Subtarget->isThumb2()) 3676 isLegal = getT2IndexedAddressParts(Ptr.getNode(), VT, isSEXTLoad, Base, 3677 Offset, isInc, DAG); 3678 else 3679 isLegal = getARMIndexedAddressParts(Ptr.getNode(), VT, isSEXTLoad, Base, 3680 Offset, isInc, DAG); 3681 if (!isLegal) 3682 return false; 3683 3684 AM = isInc ? ISD::PRE_INC : ISD::PRE_DEC; 3685 return true; 3686} 3687 3688/// getPostIndexedAddressParts - returns true by value, base pointer and 3689/// offset pointer and addressing mode by reference if this node can be 3690/// combined with a load / store to form a post-indexed load / store. 3691bool ARMTargetLowering::getPostIndexedAddressParts(SDNode *N, SDNode *Op, 3692 SDValue &Base, 3693 SDValue &Offset, 3694 ISD::MemIndexedMode &AM, 3695 SelectionDAG &DAG) const { 3696 if (Subtarget->isThumb1Only()) 3697 return false; 3698 3699 EVT VT; 3700 SDValue Ptr; 3701 bool isSEXTLoad = false; 3702 if (LoadSDNode *LD = dyn_cast<LoadSDNode>(N)) { 3703 VT = LD->getMemoryVT(); 3704 isSEXTLoad = LD->getExtensionType() == ISD::SEXTLOAD; 3705 } else if (StoreSDNode *ST = dyn_cast<StoreSDNode>(N)) { 3706 VT = ST->getMemoryVT(); 3707 } else 3708 return false; 3709 3710 bool isInc; 3711 bool isLegal = false; 3712 if (Subtarget->isThumb2()) 3713 isLegal = getT2IndexedAddressParts(Op, VT, isSEXTLoad, Base, Offset, 3714 isInc, DAG); 3715 else 3716 isLegal = getARMIndexedAddressParts(Op, VT, isSEXTLoad, Base, Offset, 3717 isInc, DAG); 3718 if (!isLegal) 3719 return false; 3720 3721 AM = isInc ? ISD::POST_INC : ISD::POST_DEC; 3722 return true; 3723} 3724 3725void ARMTargetLowering::computeMaskedBitsForTargetNode(const SDValue Op, 3726 const APInt &Mask, 3727 APInt &KnownZero, 3728 APInt &KnownOne, 3729 const SelectionDAG &DAG, 3730 unsigned Depth) const { 3731 KnownZero = KnownOne = APInt(Mask.getBitWidth(), 0); 3732 switch (Op.getOpcode()) { 3733 default: break; 3734 case ARMISD::CMOV: { 3735 // Bits are known zero/one if known on the LHS and RHS. 3736 DAG.ComputeMaskedBits(Op.getOperand(0), Mask, KnownZero, KnownOne, Depth+1); 3737 if (KnownZero == 0 && KnownOne == 0) return; 3738 3739 APInt KnownZeroRHS, KnownOneRHS; 3740 DAG.ComputeMaskedBits(Op.getOperand(1), Mask, 3741 KnownZeroRHS, KnownOneRHS, Depth+1); 3742 KnownZero &= KnownZeroRHS; 3743 KnownOne &= KnownOneRHS; 3744 return; 3745 } 3746 } 3747} 3748 3749//===----------------------------------------------------------------------===// 3750// ARM Inline Assembly Support 3751//===----------------------------------------------------------------------===// 3752 3753/// getConstraintType - Given a constraint letter, return the type of 3754/// constraint it is for this target. 3755ARMTargetLowering::ConstraintType 3756ARMTargetLowering::getConstraintType(const std::string &Constraint) const { 3757 if (Constraint.size() == 1) { 3758 switch (Constraint[0]) { 3759 default: break; 3760 case 'l': return C_RegisterClass; 3761 case 'w': return C_RegisterClass; 3762 } 3763 } 3764 return TargetLowering::getConstraintType(Constraint); 3765} 3766 3767std::pair<unsigned, const TargetRegisterClass*> 3768ARMTargetLowering::getRegForInlineAsmConstraint(const std::string &Constraint, 3769 EVT VT) const { 3770 if (Constraint.size() == 1) { 3771 // GCC RS6000 Constraint Letters 3772 switch (Constraint[0]) { 3773 case 'l': 3774 if (Subtarget->isThumb1Only()) 3775 return std::make_pair(0U, ARM::tGPRRegisterClass); 3776 else 3777 return std::make_pair(0U, ARM::GPRRegisterClass); 3778 case 'r': 3779 return std::make_pair(0U, ARM::GPRRegisterClass); 3780 case 'w': 3781 if (VT == MVT::f32) 3782 return std::make_pair(0U, ARM::SPRRegisterClass); 3783 if (VT == MVT::f64) 3784 return std::make_pair(0U, ARM::DPRRegisterClass); 3785 break; 3786 } 3787 } 3788 return TargetLowering::getRegForInlineAsmConstraint(Constraint, VT); 3789} 3790 3791std::vector<unsigned> ARMTargetLowering:: 3792getRegClassForInlineAsmConstraint(const std::string &Constraint, 3793 EVT VT) const { 3794 if (Constraint.size() != 1) 3795 return std::vector<unsigned>(); 3796 3797 switch (Constraint[0]) { // GCC ARM Constraint Letters 3798 default: break; 3799 case 'l': 3800 return make_vector<unsigned>(ARM::R0, ARM::R1, ARM::R2, ARM::R3, 3801 ARM::R4, ARM::R5, ARM::R6, ARM::R7, 3802 0); 3803 case 'r': 3804 return make_vector<unsigned>(ARM::R0, ARM::R1, ARM::R2, ARM::R3, 3805 ARM::R4, ARM::R5, ARM::R6, ARM::R7, 3806 ARM::R8, ARM::R9, ARM::R10, ARM::R11, 3807 ARM::R12, ARM::LR, 0); 3808 case 'w': 3809 if (VT == MVT::f32) 3810 return make_vector<unsigned>(ARM::S0, ARM::S1, ARM::S2, ARM::S3, 3811 ARM::S4, ARM::S5, ARM::S6, ARM::S7, 3812 ARM::S8, ARM::S9, ARM::S10, ARM::S11, 3813 ARM::S12,ARM::S13,ARM::S14,ARM::S15, 3814 ARM::S16,ARM::S17,ARM::S18,ARM::S19, 3815 ARM::S20,ARM::S21,ARM::S22,ARM::S23, 3816 ARM::S24,ARM::S25,ARM::S26,ARM::S27, 3817 ARM::S28,ARM::S29,ARM::S30,ARM::S31, 0); 3818 if (VT == MVT::f64) 3819 return make_vector<unsigned>(ARM::D0, ARM::D1, ARM::D2, ARM::D3, 3820 ARM::D4, ARM::D5, ARM::D6, ARM::D7, 3821 ARM::D8, ARM::D9, ARM::D10,ARM::D11, 3822 ARM::D12,ARM::D13,ARM::D14,ARM::D15, 0); 3823 break; 3824 } 3825 3826 return std::vector<unsigned>(); 3827} 3828 3829/// LowerAsmOperandForConstraint - Lower the specified operand into the Ops 3830/// vector. If it is invalid, don't add anything to Ops. 3831void ARMTargetLowering::LowerAsmOperandForConstraint(SDValue Op, 3832 char Constraint, 3833 bool hasMemory, 3834 std::vector<SDValue>&Ops, 3835 SelectionDAG &DAG) const { 3836 SDValue Result(0, 0); 3837 3838 switch (Constraint) { 3839 default: break; 3840 case 'I': case 'J': case 'K': case 'L': 3841 case 'M': case 'N': case 'O': 3842 ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op); 3843 if (!C) 3844 return; 3845 3846 int64_t CVal64 = C->getSExtValue(); 3847 int CVal = (int) CVal64; 3848 // None of these constraints allow values larger than 32 bits. Check 3849 // that the value fits in an int. 3850 if (CVal != CVal64) 3851 return; 3852 3853 switch (Constraint) { 3854 case 'I': 3855 if (Subtarget->isThumb1Only()) { 3856 // This must be a constant between 0 and 255, for ADD 3857 // immediates. 3858 if (CVal >= 0 && CVal <= 255) 3859 break; 3860 } else if (Subtarget->isThumb2()) { 3861 // A constant that can be used as an immediate value in a 3862 // data-processing instruction. 3863 if (ARM_AM::getT2SOImmVal(CVal) != -1) 3864 break; 3865 } else { 3866 // A constant that can be used as an immediate value in a 3867 // data-processing instruction. 3868 if (ARM_AM::getSOImmVal(CVal) != -1) 3869 break; 3870 } 3871 return; 3872 3873 case 'J': 3874 if (Subtarget->isThumb()) { // FIXME thumb2 3875 // This must be a constant between -255 and -1, for negated ADD 3876 // immediates. This can be used in GCC with an "n" modifier that 3877 // prints the negated value, for use with SUB instructions. It is 3878 // not useful otherwise but is implemented for compatibility. 3879 if (CVal >= -255 && CVal <= -1) 3880 break; 3881 } else { 3882 // This must be a constant between -4095 and 4095. It is not clear 3883 // what this constraint is intended for. Implemented for 3884 // compatibility with GCC. 3885 if (CVal >= -4095 && CVal <= 4095) 3886 break; 3887 } 3888 return; 3889 3890 case 'K': 3891 if (Subtarget->isThumb1Only()) { 3892 // A 32-bit value where only one byte has a nonzero value. Exclude 3893 // zero to match GCC. This constraint is used by GCC internally for 3894 // constants that can be loaded with a move/shift combination. 3895 // It is not useful otherwise but is implemented for compatibility. 3896 if (CVal != 0 && ARM_AM::isThumbImmShiftedVal(CVal)) 3897 break; 3898 } else if (Subtarget->isThumb2()) { 3899 // A constant whose bitwise inverse can be used as an immediate 3900 // value in a data-processing instruction. This can be used in GCC 3901 // with a "B" modifier that prints the inverted value, for use with 3902 // BIC and MVN instructions. It is not useful otherwise but is 3903 // implemented for compatibility. 3904 if (ARM_AM::getT2SOImmVal(~CVal) != -1) 3905 break; 3906 } else { 3907 // A constant whose bitwise inverse can be used as an immediate 3908 // value in a data-processing instruction. This can be used in GCC 3909 // with a "B" modifier that prints the inverted value, for use with 3910 // BIC and MVN instructions. It is not useful otherwise but is 3911 // implemented for compatibility. 3912 if (ARM_AM::getSOImmVal(~CVal) != -1) 3913 break; 3914 } 3915 return; 3916 3917 case 'L': 3918 if (Subtarget->isThumb1Only()) { 3919 // This must be a constant between -7 and 7, 3920 // for 3-operand ADD/SUB immediate instructions. 3921 if (CVal >= -7 && CVal < 7) 3922 break; 3923 } else if (Subtarget->isThumb2()) { 3924 // A constant whose negation can be used as an immediate value in a 3925 // data-processing instruction. This can be used in GCC with an "n" 3926 // modifier that prints the negated value, for use with SUB 3927 // instructions. It is not useful otherwise but is implemented for 3928 // compatibility. 3929 if (ARM_AM::getT2SOImmVal(-CVal) != -1) 3930 break; 3931 } else { 3932 // A constant whose negation can be used as an immediate value in a 3933 // data-processing instruction. This can be used in GCC with an "n" 3934 // modifier that prints the negated value, for use with SUB 3935 // instructions. It is not useful otherwise but is implemented for 3936 // compatibility. 3937 if (ARM_AM::getSOImmVal(-CVal) != -1) 3938 break; 3939 } 3940 return; 3941 3942 case 'M': 3943 if (Subtarget->isThumb()) { // FIXME thumb2 3944 // This must be a multiple of 4 between 0 and 1020, for 3945 // ADD sp + immediate. 3946 if ((CVal >= 0 && CVal <= 1020) && ((CVal & 3) == 0)) 3947 break; 3948 } else { 3949 // A power of two or a constant between 0 and 32. This is used in 3950 // GCC for the shift amount on shifted register operands, but it is 3951 // useful in general for any shift amounts. 3952 if ((CVal >= 0 && CVal <= 32) || ((CVal & (CVal - 1)) == 0)) 3953 break; 3954 } 3955 return; 3956 3957 case 'N': 3958 if (Subtarget->isThumb()) { // FIXME thumb2 3959 // This must be a constant between 0 and 31, for shift amounts. 3960 if (CVal >= 0 && CVal <= 31) 3961 break; 3962 } 3963 return; 3964 3965 case 'O': 3966 if (Subtarget->isThumb()) { // FIXME thumb2 3967 // This must be a multiple of 4 between -508 and 508, for 3968 // ADD/SUB sp = sp + immediate. 3969 if ((CVal >= -508 && CVal <= 508) && ((CVal & 3) == 0)) 3970 break; 3971 } 3972 return; 3973 } 3974 Result = DAG.getTargetConstant(CVal, Op.getValueType()); 3975 break; 3976 } 3977 3978 if (Result.getNode()) { 3979 Ops.push_back(Result); 3980 return; 3981 } 3982 return TargetLowering::LowerAsmOperandForConstraint(Op, Constraint, hasMemory, 3983 Ops, DAG); 3984} 3985