TargetLowering.cpp revision e46f5508dcf45c7f4833913a49dfa3cd95ea74b3
1//===-- TargetLowering.cpp - Implement the TargetLowering class -----------===// 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 implements the TargetLowering class. 11// 12//===----------------------------------------------------------------------===// 13 14#include "llvm/Target/TargetAsmInfo.h" 15#include "llvm/Target/TargetLowering.h" 16#include "llvm/Target/TargetSubtarget.h" 17#include "llvm/Target/TargetData.h" 18#include "llvm/Target/TargetMachine.h" 19#include "llvm/Target/TargetRegisterInfo.h" 20#include "llvm/GlobalVariable.h" 21#include "llvm/DerivedTypes.h" 22#include "llvm/CodeGen/MachineFrameInfo.h" 23#include "llvm/CodeGen/SelectionDAG.h" 24#include "llvm/ADT/StringExtras.h" 25#include "llvm/ADT/STLExtras.h" 26#include "llvm/Support/MathExtras.h" 27using namespace llvm; 28 29namespace llvm { 30TLSModel::Model getTLSModel(const GlobalValue *GV, Reloc::Model reloc) { 31 bool isLocal = GV->hasLocalLinkage(); 32 bool isDeclaration = GV->isDeclaration(); 33 // FIXME: what should we do for protected and internal visibility? 34 // For variables, is internal different from hidden? 35 bool isHidden = GV->hasHiddenVisibility(); 36 37 if (reloc == Reloc::PIC_) { 38 if (isLocal || isHidden) 39 return TLSModel::LocalDynamic; 40 else 41 return TLSModel::GeneralDynamic; 42 } else { 43 if (!isDeclaration || isHidden) 44 return TLSModel::LocalExec; 45 else 46 return TLSModel::InitialExec; 47 } 48} 49} 50 51/// InitLibcallNames - Set default libcall names. 52/// 53static void InitLibcallNames(const char **Names) { 54 Names[RTLIB::SHL_I16] = "__ashli16"; 55 Names[RTLIB::SHL_I32] = "__ashlsi3"; 56 Names[RTLIB::SHL_I64] = "__ashldi3"; 57 Names[RTLIB::SHL_I128] = "__ashlti3"; 58 Names[RTLIB::SRL_I16] = "__lshri16"; 59 Names[RTLIB::SRL_I32] = "__lshrsi3"; 60 Names[RTLIB::SRL_I64] = "__lshrdi3"; 61 Names[RTLIB::SRL_I128] = "__lshrti3"; 62 Names[RTLIB::SRA_I16] = "__ashri16"; 63 Names[RTLIB::SRA_I32] = "__ashrsi3"; 64 Names[RTLIB::SRA_I64] = "__ashrdi3"; 65 Names[RTLIB::SRA_I128] = "__ashrti3"; 66 Names[RTLIB::MUL_I16] = "__muli16"; 67 Names[RTLIB::MUL_I32] = "__mulsi3"; 68 Names[RTLIB::MUL_I64] = "__muldi3"; 69 Names[RTLIB::MUL_I128] = "__multi3"; 70 Names[RTLIB::SDIV_I32] = "__divsi3"; 71 Names[RTLIB::SDIV_I64] = "__divdi3"; 72 Names[RTLIB::SDIV_I128] = "__divti3"; 73 Names[RTLIB::UDIV_I32] = "__udivsi3"; 74 Names[RTLIB::UDIV_I64] = "__udivdi3"; 75 Names[RTLIB::UDIV_I128] = "__udivti3"; 76 Names[RTLIB::SREM_I32] = "__modsi3"; 77 Names[RTLIB::SREM_I64] = "__moddi3"; 78 Names[RTLIB::SREM_I128] = "__modti3"; 79 Names[RTLIB::UREM_I32] = "__umodsi3"; 80 Names[RTLIB::UREM_I64] = "__umoddi3"; 81 Names[RTLIB::UREM_I128] = "__umodti3"; 82 Names[RTLIB::NEG_I32] = "__negsi2"; 83 Names[RTLIB::NEG_I64] = "__negdi2"; 84 Names[RTLIB::ADD_F32] = "__addsf3"; 85 Names[RTLIB::ADD_F64] = "__adddf3"; 86 Names[RTLIB::ADD_F80] = "__addxf3"; 87 Names[RTLIB::ADD_PPCF128] = "__gcc_qadd"; 88 Names[RTLIB::SUB_F32] = "__subsf3"; 89 Names[RTLIB::SUB_F64] = "__subdf3"; 90 Names[RTLIB::SUB_F80] = "__subxf3"; 91 Names[RTLIB::SUB_PPCF128] = "__gcc_qsub"; 92 Names[RTLIB::MUL_F32] = "__mulsf3"; 93 Names[RTLIB::MUL_F64] = "__muldf3"; 94 Names[RTLIB::MUL_F80] = "__mulxf3"; 95 Names[RTLIB::MUL_PPCF128] = "__gcc_qmul"; 96 Names[RTLIB::DIV_F32] = "__divsf3"; 97 Names[RTLIB::DIV_F64] = "__divdf3"; 98 Names[RTLIB::DIV_F80] = "__divxf3"; 99 Names[RTLIB::DIV_PPCF128] = "__gcc_qdiv"; 100 Names[RTLIB::REM_F32] = "fmodf"; 101 Names[RTLIB::REM_F64] = "fmod"; 102 Names[RTLIB::REM_F80] = "fmodl"; 103 Names[RTLIB::REM_PPCF128] = "fmodl"; 104 Names[RTLIB::POWI_F32] = "__powisf2"; 105 Names[RTLIB::POWI_F64] = "__powidf2"; 106 Names[RTLIB::POWI_F80] = "__powixf2"; 107 Names[RTLIB::POWI_PPCF128] = "__powitf2"; 108 Names[RTLIB::SQRT_F32] = "sqrtf"; 109 Names[RTLIB::SQRT_F64] = "sqrt"; 110 Names[RTLIB::SQRT_F80] = "sqrtl"; 111 Names[RTLIB::SQRT_PPCF128] = "sqrtl"; 112 Names[RTLIB::LOG_F32] = "logf"; 113 Names[RTLIB::LOG_F64] = "log"; 114 Names[RTLIB::LOG_F80] = "logl"; 115 Names[RTLIB::LOG_PPCF128] = "logl"; 116 Names[RTLIB::LOG2_F32] = "log2f"; 117 Names[RTLIB::LOG2_F64] = "log2"; 118 Names[RTLIB::LOG2_F80] = "log2l"; 119 Names[RTLIB::LOG2_PPCF128] = "log2l"; 120 Names[RTLIB::LOG10_F32] = "log10f"; 121 Names[RTLIB::LOG10_F64] = "log10"; 122 Names[RTLIB::LOG10_F80] = "log10l"; 123 Names[RTLIB::LOG10_PPCF128] = "log10l"; 124 Names[RTLIB::EXP_F32] = "expf"; 125 Names[RTLIB::EXP_F64] = "exp"; 126 Names[RTLIB::EXP_F80] = "expl"; 127 Names[RTLIB::EXP_PPCF128] = "expl"; 128 Names[RTLIB::EXP2_F32] = "exp2f"; 129 Names[RTLIB::EXP2_F64] = "exp2"; 130 Names[RTLIB::EXP2_F80] = "exp2l"; 131 Names[RTLIB::EXP2_PPCF128] = "exp2l"; 132 Names[RTLIB::SIN_F32] = "sinf"; 133 Names[RTLIB::SIN_F64] = "sin"; 134 Names[RTLIB::SIN_F80] = "sinl"; 135 Names[RTLIB::SIN_PPCF128] = "sinl"; 136 Names[RTLIB::COS_F32] = "cosf"; 137 Names[RTLIB::COS_F64] = "cos"; 138 Names[RTLIB::COS_F80] = "cosl"; 139 Names[RTLIB::COS_PPCF128] = "cosl"; 140 Names[RTLIB::POW_F32] = "powf"; 141 Names[RTLIB::POW_F64] = "pow"; 142 Names[RTLIB::POW_F80] = "powl"; 143 Names[RTLIB::POW_PPCF128] = "powl"; 144 Names[RTLIB::CEIL_F32] = "ceilf"; 145 Names[RTLIB::CEIL_F64] = "ceil"; 146 Names[RTLIB::CEIL_F80] = "ceill"; 147 Names[RTLIB::CEIL_PPCF128] = "ceill"; 148 Names[RTLIB::TRUNC_F32] = "truncf"; 149 Names[RTLIB::TRUNC_F64] = "trunc"; 150 Names[RTLIB::TRUNC_F80] = "truncl"; 151 Names[RTLIB::TRUNC_PPCF128] = "truncl"; 152 Names[RTLIB::RINT_F32] = "rintf"; 153 Names[RTLIB::RINT_F64] = "rint"; 154 Names[RTLIB::RINT_F80] = "rintl"; 155 Names[RTLIB::RINT_PPCF128] = "rintl"; 156 Names[RTLIB::NEARBYINT_F32] = "nearbyintf"; 157 Names[RTLIB::NEARBYINT_F64] = "nearbyint"; 158 Names[RTLIB::NEARBYINT_F80] = "nearbyintl"; 159 Names[RTLIB::NEARBYINT_PPCF128] = "nearbyintl"; 160 Names[RTLIB::FLOOR_F32] = "floorf"; 161 Names[RTLIB::FLOOR_F64] = "floor"; 162 Names[RTLIB::FLOOR_F80] = "floorl"; 163 Names[RTLIB::FLOOR_PPCF128] = "floorl"; 164 Names[RTLIB::FPEXT_F32_F64] = "__extendsfdf2"; 165 Names[RTLIB::FPROUND_F64_F32] = "__truncdfsf2"; 166 Names[RTLIB::FPROUND_F80_F32] = "__truncxfsf2"; 167 Names[RTLIB::FPROUND_PPCF128_F32] = "__trunctfsf2"; 168 Names[RTLIB::FPROUND_F80_F64] = "__truncxfdf2"; 169 Names[RTLIB::FPROUND_PPCF128_F64] = "__trunctfdf2"; 170 Names[RTLIB::FPTOSINT_F32_I32] = "__fixsfsi"; 171 Names[RTLIB::FPTOSINT_F32_I64] = "__fixsfdi"; 172 Names[RTLIB::FPTOSINT_F32_I128] = "__fixsfti"; 173 Names[RTLIB::FPTOSINT_F64_I32] = "__fixdfsi"; 174 Names[RTLIB::FPTOSINT_F64_I64] = "__fixdfdi"; 175 Names[RTLIB::FPTOSINT_F64_I128] = "__fixdfti"; 176 Names[RTLIB::FPTOSINT_F80_I32] = "__fixxfsi"; 177 Names[RTLIB::FPTOSINT_F80_I64] = "__fixxfdi"; 178 Names[RTLIB::FPTOSINT_F80_I128] = "__fixxfti"; 179 Names[RTLIB::FPTOSINT_PPCF128_I32] = "__fixtfsi"; 180 Names[RTLIB::FPTOSINT_PPCF128_I64] = "__fixtfdi"; 181 Names[RTLIB::FPTOSINT_PPCF128_I128] = "__fixtfti"; 182 Names[RTLIB::FPTOUINT_F32_I32] = "__fixunssfsi"; 183 Names[RTLIB::FPTOUINT_F32_I64] = "__fixunssfdi"; 184 Names[RTLIB::FPTOUINT_F32_I128] = "__fixunssfti"; 185 Names[RTLIB::FPTOUINT_F64_I32] = "__fixunsdfsi"; 186 Names[RTLIB::FPTOUINT_F64_I64] = "__fixunsdfdi"; 187 Names[RTLIB::FPTOUINT_F64_I128] = "__fixunsdfti"; 188 Names[RTLIB::FPTOUINT_F80_I32] = "__fixunsxfsi"; 189 Names[RTLIB::FPTOUINT_F80_I64] = "__fixunsxfdi"; 190 Names[RTLIB::FPTOUINT_F80_I128] = "__fixunsxfti"; 191 Names[RTLIB::FPTOUINT_PPCF128_I32] = "__fixunstfsi"; 192 Names[RTLIB::FPTOUINT_PPCF128_I64] = "__fixunstfdi"; 193 Names[RTLIB::FPTOUINT_PPCF128_I128] = "__fixunstfti"; 194 Names[RTLIB::SINTTOFP_I32_F32] = "__floatsisf"; 195 Names[RTLIB::SINTTOFP_I32_F64] = "__floatsidf"; 196 Names[RTLIB::SINTTOFP_I32_F80] = "__floatsixf"; 197 Names[RTLIB::SINTTOFP_I32_PPCF128] = "__floatsitf"; 198 Names[RTLIB::SINTTOFP_I64_F32] = "__floatdisf"; 199 Names[RTLIB::SINTTOFP_I64_F64] = "__floatdidf"; 200 Names[RTLIB::SINTTOFP_I64_F80] = "__floatdixf"; 201 Names[RTLIB::SINTTOFP_I64_PPCF128] = "__floatditf"; 202 Names[RTLIB::SINTTOFP_I128_F32] = "__floattisf"; 203 Names[RTLIB::SINTTOFP_I128_F64] = "__floattidf"; 204 Names[RTLIB::SINTTOFP_I128_F80] = "__floattixf"; 205 Names[RTLIB::SINTTOFP_I128_PPCF128] = "__floattitf"; 206 Names[RTLIB::UINTTOFP_I32_F32] = "__floatunsisf"; 207 Names[RTLIB::UINTTOFP_I32_F64] = "__floatunsidf"; 208 Names[RTLIB::UINTTOFP_I32_F80] = "__floatunsixf"; 209 Names[RTLIB::UINTTOFP_I32_PPCF128] = "__floatunsitf"; 210 Names[RTLIB::UINTTOFP_I64_F32] = "__floatundisf"; 211 Names[RTLIB::UINTTOFP_I64_F64] = "__floatundidf"; 212 Names[RTLIB::UINTTOFP_I64_F80] = "__floatundixf"; 213 Names[RTLIB::UINTTOFP_I64_PPCF128] = "__floatunditf"; 214 Names[RTLIB::UINTTOFP_I128_F32] = "__floatuntisf"; 215 Names[RTLIB::UINTTOFP_I128_F64] = "__floatuntidf"; 216 Names[RTLIB::UINTTOFP_I128_F80] = "__floatuntixf"; 217 Names[RTLIB::UINTTOFP_I128_PPCF128] = "__floatuntitf"; 218 Names[RTLIB::OEQ_F32] = "__eqsf2"; 219 Names[RTLIB::OEQ_F64] = "__eqdf2"; 220 Names[RTLIB::UNE_F32] = "__nesf2"; 221 Names[RTLIB::UNE_F64] = "__nedf2"; 222 Names[RTLIB::OGE_F32] = "__gesf2"; 223 Names[RTLIB::OGE_F64] = "__gedf2"; 224 Names[RTLIB::OLT_F32] = "__ltsf2"; 225 Names[RTLIB::OLT_F64] = "__ltdf2"; 226 Names[RTLIB::OLE_F32] = "__lesf2"; 227 Names[RTLIB::OLE_F64] = "__ledf2"; 228 Names[RTLIB::OGT_F32] = "__gtsf2"; 229 Names[RTLIB::OGT_F64] = "__gtdf2"; 230 Names[RTLIB::UO_F32] = "__unordsf2"; 231 Names[RTLIB::UO_F64] = "__unorddf2"; 232 Names[RTLIB::O_F32] = "__unordsf2"; 233 Names[RTLIB::O_F64] = "__unorddf2"; 234} 235 236/// getFPEXT - Return the FPEXT_*_* value for the given types, or 237/// UNKNOWN_LIBCALL if there is none. 238RTLIB::Libcall RTLIB::getFPEXT(MVT OpVT, MVT RetVT) { 239 if (OpVT == MVT::f32) { 240 if (RetVT == MVT::f64) 241 return FPEXT_F32_F64; 242 } 243 return UNKNOWN_LIBCALL; 244} 245 246/// getFPROUND - Return the FPROUND_*_* value for the given types, or 247/// UNKNOWN_LIBCALL if there is none. 248RTLIB::Libcall RTLIB::getFPROUND(MVT OpVT, MVT RetVT) { 249 if (RetVT == MVT::f32) { 250 if (OpVT == MVT::f64) 251 return FPROUND_F64_F32; 252 if (OpVT == MVT::f80) 253 return FPROUND_F80_F32; 254 if (OpVT == MVT::ppcf128) 255 return FPROUND_PPCF128_F32; 256 } else if (RetVT == MVT::f64) { 257 if (OpVT == MVT::f80) 258 return FPROUND_F80_F64; 259 if (OpVT == MVT::ppcf128) 260 return FPROUND_PPCF128_F64; 261 } 262 return UNKNOWN_LIBCALL; 263} 264 265/// getFPTOSINT - Return the FPTOSINT_*_* value for the given types, or 266/// UNKNOWN_LIBCALL if there is none. 267RTLIB::Libcall RTLIB::getFPTOSINT(MVT OpVT, MVT RetVT) { 268 if (OpVT == MVT::f32) { 269 if (RetVT == MVT::i32) 270 return FPTOSINT_F32_I32; 271 if (RetVT == MVT::i64) 272 return FPTOSINT_F32_I64; 273 if (RetVT == MVT::i128) 274 return FPTOSINT_F32_I128; 275 } else if (OpVT == MVT::f64) { 276 if (RetVT == MVT::i32) 277 return FPTOSINT_F64_I32; 278 if (RetVT == MVT::i64) 279 return FPTOSINT_F64_I64; 280 if (RetVT == MVT::i128) 281 return FPTOSINT_F64_I128; 282 } else if (OpVT == MVT::f80) { 283 if (RetVT == MVT::i32) 284 return FPTOSINT_F80_I32; 285 if (RetVT == MVT::i64) 286 return FPTOSINT_F80_I64; 287 if (RetVT == MVT::i128) 288 return FPTOSINT_F80_I128; 289 } else if (OpVT == MVT::ppcf128) { 290 if (RetVT == MVT::i32) 291 return FPTOSINT_PPCF128_I32; 292 if (RetVT == MVT::i64) 293 return FPTOSINT_PPCF128_I64; 294 if (RetVT == MVT::i128) 295 return FPTOSINT_PPCF128_I128; 296 } 297 return UNKNOWN_LIBCALL; 298} 299 300/// getFPTOUINT - Return the FPTOUINT_*_* value for the given types, or 301/// UNKNOWN_LIBCALL if there is none. 302RTLIB::Libcall RTLIB::getFPTOUINT(MVT OpVT, MVT RetVT) { 303 if (OpVT == MVT::f32) { 304 if (RetVT == MVT::i32) 305 return FPTOUINT_F32_I32; 306 if (RetVT == MVT::i64) 307 return FPTOUINT_F32_I64; 308 if (RetVT == MVT::i128) 309 return FPTOUINT_F32_I128; 310 } else if (OpVT == MVT::f64) { 311 if (RetVT == MVT::i32) 312 return FPTOUINT_F64_I32; 313 if (RetVT == MVT::i64) 314 return FPTOUINT_F64_I64; 315 if (RetVT == MVT::i128) 316 return FPTOUINT_F64_I128; 317 } else if (OpVT == MVT::f80) { 318 if (RetVT == MVT::i32) 319 return FPTOUINT_F80_I32; 320 if (RetVT == MVT::i64) 321 return FPTOUINT_F80_I64; 322 if (RetVT == MVT::i128) 323 return FPTOUINT_F80_I128; 324 } else if (OpVT == MVT::ppcf128) { 325 if (RetVT == MVT::i32) 326 return FPTOUINT_PPCF128_I32; 327 if (RetVT == MVT::i64) 328 return FPTOUINT_PPCF128_I64; 329 if (RetVT == MVT::i128) 330 return FPTOUINT_PPCF128_I128; 331 } 332 return UNKNOWN_LIBCALL; 333} 334 335/// getSINTTOFP - Return the SINTTOFP_*_* value for the given types, or 336/// UNKNOWN_LIBCALL if there is none. 337RTLIB::Libcall RTLIB::getSINTTOFP(MVT OpVT, MVT RetVT) { 338 if (OpVT == MVT::i32) { 339 if (RetVT == MVT::f32) 340 return SINTTOFP_I32_F32; 341 else if (RetVT == MVT::f64) 342 return SINTTOFP_I32_F64; 343 else if (RetVT == MVT::f80) 344 return SINTTOFP_I32_F80; 345 else if (RetVT == MVT::ppcf128) 346 return SINTTOFP_I32_PPCF128; 347 } else if (OpVT == MVT::i64) { 348 if (RetVT == MVT::f32) 349 return SINTTOFP_I64_F32; 350 else if (RetVT == MVT::f64) 351 return SINTTOFP_I64_F64; 352 else if (RetVT == MVT::f80) 353 return SINTTOFP_I64_F80; 354 else if (RetVT == MVT::ppcf128) 355 return SINTTOFP_I64_PPCF128; 356 } else if (OpVT == MVT::i128) { 357 if (RetVT == MVT::f32) 358 return SINTTOFP_I128_F32; 359 else if (RetVT == MVT::f64) 360 return SINTTOFP_I128_F64; 361 else if (RetVT == MVT::f80) 362 return SINTTOFP_I128_F80; 363 else if (RetVT == MVT::ppcf128) 364 return SINTTOFP_I128_PPCF128; 365 } 366 return UNKNOWN_LIBCALL; 367} 368 369/// getUINTTOFP - Return the UINTTOFP_*_* value for the given types, or 370/// UNKNOWN_LIBCALL if there is none. 371RTLIB::Libcall RTLIB::getUINTTOFP(MVT OpVT, MVT RetVT) { 372 if (OpVT == MVT::i32) { 373 if (RetVT == MVT::f32) 374 return UINTTOFP_I32_F32; 375 else if (RetVT == MVT::f64) 376 return UINTTOFP_I32_F64; 377 else if (RetVT == MVT::f80) 378 return UINTTOFP_I32_F80; 379 else if (RetVT == MVT::ppcf128) 380 return UINTTOFP_I32_PPCF128; 381 } else if (OpVT == MVT::i64) { 382 if (RetVT == MVT::f32) 383 return UINTTOFP_I64_F32; 384 else if (RetVT == MVT::f64) 385 return UINTTOFP_I64_F64; 386 else if (RetVT == MVT::f80) 387 return UINTTOFP_I64_F80; 388 else if (RetVT == MVT::ppcf128) 389 return UINTTOFP_I64_PPCF128; 390 } else if (OpVT == MVT::i128) { 391 if (RetVT == MVT::f32) 392 return UINTTOFP_I128_F32; 393 else if (RetVT == MVT::f64) 394 return UINTTOFP_I128_F64; 395 else if (RetVT == MVT::f80) 396 return UINTTOFP_I128_F80; 397 else if (RetVT == MVT::ppcf128) 398 return UINTTOFP_I128_PPCF128; 399 } 400 return UNKNOWN_LIBCALL; 401} 402 403/// InitCmpLibcallCCs - Set default comparison libcall CC. 404/// 405static void InitCmpLibcallCCs(ISD::CondCode *CCs) { 406 memset(CCs, ISD::SETCC_INVALID, sizeof(ISD::CondCode)*RTLIB::UNKNOWN_LIBCALL); 407 CCs[RTLIB::OEQ_F32] = ISD::SETEQ; 408 CCs[RTLIB::OEQ_F64] = ISD::SETEQ; 409 CCs[RTLIB::UNE_F32] = ISD::SETNE; 410 CCs[RTLIB::UNE_F64] = ISD::SETNE; 411 CCs[RTLIB::OGE_F32] = ISD::SETGE; 412 CCs[RTLIB::OGE_F64] = ISD::SETGE; 413 CCs[RTLIB::OLT_F32] = ISD::SETLT; 414 CCs[RTLIB::OLT_F64] = ISD::SETLT; 415 CCs[RTLIB::OLE_F32] = ISD::SETLE; 416 CCs[RTLIB::OLE_F64] = ISD::SETLE; 417 CCs[RTLIB::OGT_F32] = ISD::SETGT; 418 CCs[RTLIB::OGT_F64] = ISD::SETGT; 419 CCs[RTLIB::UO_F32] = ISD::SETNE; 420 CCs[RTLIB::UO_F64] = ISD::SETNE; 421 CCs[RTLIB::O_F32] = ISD::SETEQ; 422 CCs[RTLIB::O_F64] = ISD::SETEQ; 423} 424 425TargetLowering::TargetLowering(TargetMachine &tm) 426 : TM(tm), TD(TM.getTargetData()) { 427 // All operations default to being supported. 428 memset(OpActions, 0, sizeof(OpActions)); 429 memset(LoadExtActions, 0, sizeof(LoadExtActions)); 430 memset(TruncStoreActions, 0, sizeof(TruncStoreActions)); 431 memset(IndexedModeActions, 0, sizeof(IndexedModeActions)); 432 memset(ConvertActions, 0, sizeof(ConvertActions)); 433 memset(CondCodeActions, 0, sizeof(CondCodeActions)); 434 435 // Set default actions for various operations. 436 for (unsigned VT = 0; VT != (unsigned)MVT::LAST_VALUETYPE; ++VT) { 437 // Default all indexed load / store to expand. 438 for (unsigned IM = (unsigned)ISD::PRE_INC; 439 IM != (unsigned)ISD::LAST_INDEXED_MODE; ++IM) { 440 setIndexedLoadAction(IM, (MVT::SimpleValueType)VT, Expand); 441 setIndexedStoreAction(IM, (MVT::SimpleValueType)VT, Expand); 442 } 443 444 // These operations default to expand. 445 setOperationAction(ISD::FGETSIGN, (MVT::SimpleValueType)VT, Expand); 446 } 447 448 // Most targets ignore the @llvm.prefetch intrinsic. 449 setOperationAction(ISD::PREFETCH, MVT::Other, Expand); 450 451 // ConstantFP nodes default to expand. Targets can either change this to 452 // Legal, in which case all fp constants are legal, or use addLegalFPImmediate 453 // to optimize expansions for certain constants. 454 setOperationAction(ISD::ConstantFP, MVT::f32, Expand); 455 setOperationAction(ISD::ConstantFP, MVT::f64, Expand); 456 setOperationAction(ISD::ConstantFP, MVT::f80, Expand); 457 458 // These library functions default to expand. 459 setOperationAction(ISD::FLOG , MVT::f64, Expand); 460 setOperationAction(ISD::FLOG2, MVT::f64, Expand); 461 setOperationAction(ISD::FLOG10,MVT::f64, Expand); 462 setOperationAction(ISD::FEXP , MVT::f64, Expand); 463 setOperationAction(ISD::FEXP2, MVT::f64, Expand); 464 setOperationAction(ISD::FLOG , MVT::f32, Expand); 465 setOperationAction(ISD::FLOG2, MVT::f32, Expand); 466 setOperationAction(ISD::FLOG10,MVT::f32, Expand); 467 setOperationAction(ISD::FEXP , MVT::f32, Expand); 468 setOperationAction(ISD::FEXP2, MVT::f32, Expand); 469 470 // Default ISD::TRAP to expand (which turns it into abort). 471 setOperationAction(ISD::TRAP, MVT::Other, Expand); 472 473 IsLittleEndian = TD->isLittleEndian(); 474 UsesGlobalOffsetTable = false; 475 ShiftAmountTy = PointerTy = getValueType(TD->getIntPtrType()); 476 ShiftAmtHandling = Undefined; 477 memset(RegClassForVT, 0,MVT::LAST_VALUETYPE*sizeof(TargetRegisterClass*)); 478 memset(TargetDAGCombineArray, 0, array_lengthof(TargetDAGCombineArray)); 479 maxStoresPerMemset = maxStoresPerMemcpy = maxStoresPerMemmove = 8; 480 allowUnalignedMemoryAccesses = false; 481 UseUnderscoreSetJmp = false; 482 UseUnderscoreLongJmp = false; 483 SelectIsExpensive = false; 484 IntDivIsCheap = false; 485 Pow2DivIsCheap = false; 486 StackPointerRegisterToSaveRestore = 0; 487 ExceptionPointerRegister = 0; 488 ExceptionSelectorRegister = 0; 489 BooleanContents = UndefinedBooleanContent; 490 SchedPreferenceInfo = SchedulingForLatency; 491 JumpBufSize = 0; 492 JumpBufAlignment = 0; 493 IfCvtBlockSizeLimit = 2; 494 IfCvtDupBlockSizeLimit = 0; 495 PrefLoopAlignment = 0; 496 497 InitLibcallNames(LibcallRoutineNames); 498 InitCmpLibcallCCs(CmpLibcallCCs); 499 500 // Tell Legalize whether the assembler supports DEBUG_LOC. 501 const TargetAsmInfo *TASM = TM.getTargetAsmInfo(); 502 if (!TASM || !TASM->hasDotLocAndDotFile()) 503 setOperationAction(ISD::DEBUG_LOC, MVT::Other, Expand); 504} 505 506TargetLowering::~TargetLowering() {} 507 508/// computeRegisterProperties - Once all of the register classes are added, 509/// this allows us to compute derived properties we expose. 510void TargetLowering::computeRegisterProperties() { 511 assert(MVT::LAST_VALUETYPE <= 32 && 512 "Too many value types for ValueTypeActions to hold!"); 513 514 // Everything defaults to needing one register. 515 for (unsigned i = 0; i != MVT::LAST_VALUETYPE; ++i) { 516 NumRegistersForVT[i] = 1; 517 RegisterTypeForVT[i] = TransformToType[i] = (MVT::SimpleValueType)i; 518 } 519 // ...except isVoid, which doesn't need any registers. 520 NumRegistersForVT[MVT::isVoid] = 0; 521 522 // Find the largest integer register class. 523 unsigned LargestIntReg = MVT::LAST_INTEGER_VALUETYPE; 524 for (; RegClassForVT[LargestIntReg] == 0; --LargestIntReg) 525 assert(LargestIntReg != MVT::i1 && "No integer registers defined!"); 526 527 // Every integer value type larger than this largest register takes twice as 528 // many registers to represent as the previous ValueType. 529 for (unsigned ExpandedReg = LargestIntReg + 1; ; ++ExpandedReg) { 530 MVT EVT = (MVT::SimpleValueType)ExpandedReg; 531 if (!EVT.isInteger()) 532 break; 533 NumRegistersForVT[ExpandedReg] = 2*NumRegistersForVT[ExpandedReg-1]; 534 RegisterTypeForVT[ExpandedReg] = (MVT::SimpleValueType)LargestIntReg; 535 TransformToType[ExpandedReg] = (MVT::SimpleValueType)(ExpandedReg - 1); 536 ValueTypeActions.setTypeAction(EVT, Expand); 537 } 538 539 // Inspect all of the ValueType's smaller than the largest integer 540 // register to see which ones need promotion. 541 unsigned LegalIntReg = LargestIntReg; 542 for (unsigned IntReg = LargestIntReg - 1; 543 IntReg >= (unsigned)MVT::i1; --IntReg) { 544 MVT IVT = (MVT::SimpleValueType)IntReg; 545 if (isTypeLegal(IVT)) { 546 LegalIntReg = IntReg; 547 } else { 548 RegisterTypeForVT[IntReg] = TransformToType[IntReg] = 549 (MVT::SimpleValueType)LegalIntReg; 550 ValueTypeActions.setTypeAction(IVT, Promote); 551 } 552 } 553 554 // ppcf128 type is really two f64's. 555 if (!isTypeLegal(MVT::ppcf128)) { 556 NumRegistersForVT[MVT::ppcf128] = 2*NumRegistersForVT[MVT::f64]; 557 RegisterTypeForVT[MVT::ppcf128] = MVT::f64; 558 TransformToType[MVT::ppcf128] = MVT::f64; 559 ValueTypeActions.setTypeAction(MVT::ppcf128, Expand); 560 } 561 562 // Decide how to handle f64. If the target does not have native f64 support, 563 // expand it to i64 and we will be generating soft float library calls. 564 if (!isTypeLegal(MVT::f64)) { 565 NumRegistersForVT[MVT::f64] = NumRegistersForVT[MVT::i64]; 566 RegisterTypeForVT[MVT::f64] = RegisterTypeForVT[MVT::i64]; 567 TransformToType[MVT::f64] = MVT::i64; 568 ValueTypeActions.setTypeAction(MVT::f64, Expand); 569 } 570 571 // Decide how to handle f32. If the target does not have native support for 572 // f32, promote it to f64 if it is legal. Otherwise, expand it to i32. 573 if (!isTypeLegal(MVT::f32)) { 574 if (isTypeLegal(MVT::f64)) { 575 NumRegistersForVT[MVT::f32] = NumRegistersForVT[MVT::f64]; 576 RegisterTypeForVT[MVT::f32] = RegisterTypeForVT[MVT::f64]; 577 TransformToType[MVT::f32] = MVT::f64; 578 ValueTypeActions.setTypeAction(MVT::f32, Promote); 579 } else { 580 NumRegistersForVT[MVT::f32] = NumRegistersForVT[MVT::i32]; 581 RegisterTypeForVT[MVT::f32] = RegisterTypeForVT[MVT::i32]; 582 TransformToType[MVT::f32] = MVT::i32; 583 ValueTypeActions.setTypeAction(MVT::f32, Expand); 584 } 585 } 586 587 // Loop over all of the vector value types to see which need transformations. 588 for (unsigned i = MVT::FIRST_VECTOR_VALUETYPE; 589 i <= (unsigned)MVT::LAST_VECTOR_VALUETYPE; ++i) { 590 MVT VT = (MVT::SimpleValueType)i; 591 if (!isTypeLegal(VT)) { 592 MVT IntermediateVT, RegisterVT; 593 unsigned NumIntermediates; 594 NumRegistersForVT[i] = 595 getVectorTypeBreakdown(VT, 596 IntermediateVT, NumIntermediates, 597 RegisterVT); 598 RegisterTypeForVT[i] = RegisterVT; 599 600 // Determine if there is a legal wider type. 601 bool IsLegalWiderType = false; 602 MVT EltVT = VT.getVectorElementType(); 603 unsigned NElts = VT.getVectorNumElements(); 604 for (unsigned nVT = i+1; nVT <= MVT::LAST_VECTOR_VALUETYPE; ++nVT) { 605 MVT SVT = (MVT::SimpleValueType)nVT; 606 if (isTypeLegal(SVT) && SVT.getVectorElementType() == EltVT && 607 SVT.getVectorNumElements() > NElts) { 608 TransformToType[i] = SVT; 609 ValueTypeActions.setTypeAction(VT, Promote); 610 IsLegalWiderType = true; 611 break; 612 } 613 } 614 if (!IsLegalWiderType) { 615 MVT NVT = VT.getPow2VectorType(); 616 if (NVT == VT) { 617 // Type is already a power of 2. The default action is to split. 618 TransformToType[i] = MVT::Other; 619 ValueTypeActions.setTypeAction(VT, Expand); 620 } else { 621 TransformToType[i] = NVT; 622 ValueTypeActions.setTypeAction(VT, Promote); 623 } 624 } 625 } 626 } 627} 628 629const char *TargetLowering::getTargetNodeName(unsigned Opcode) const { 630 return NULL; 631} 632 633 634MVT TargetLowering::getSetCCResultType(MVT VT) const { 635 return getValueType(TD->getIntPtrType()); 636} 637 638 639/// getVectorTypeBreakdown - Vector types are broken down into some number of 640/// legal first class types. For example, MVT::v8f32 maps to 2 MVT::v4f32 641/// with Altivec or SSE1, or 8 promoted MVT::f64 values with the X86 FP stack. 642/// Similarly, MVT::v2i64 turns into 4 MVT::i32 values with both PPC and X86. 643/// 644/// This method returns the number of registers needed, and the VT for each 645/// register. It also returns the VT and quantity of the intermediate values 646/// before they are promoted/expanded. 647/// 648unsigned TargetLowering::getVectorTypeBreakdown(MVT VT, 649 MVT &IntermediateVT, 650 unsigned &NumIntermediates, 651 MVT &RegisterVT) const { 652 // Figure out the right, legal destination reg to copy into. 653 unsigned NumElts = VT.getVectorNumElements(); 654 MVT EltTy = VT.getVectorElementType(); 655 656 unsigned NumVectorRegs = 1; 657 658 // FIXME: We don't support non-power-of-2-sized vectors for now. Ideally we 659 // could break down into LHS/RHS like LegalizeDAG does. 660 if (!isPowerOf2_32(NumElts)) { 661 NumVectorRegs = NumElts; 662 NumElts = 1; 663 } 664 665 // Divide the input until we get to a supported size. This will always 666 // end with a scalar if the target doesn't support vectors. 667 while (NumElts > 1 && !isTypeLegal(MVT::getVectorVT(EltTy, NumElts))) { 668 NumElts >>= 1; 669 NumVectorRegs <<= 1; 670 } 671 672 NumIntermediates = NumVectorRegs; 673 674 MVT NewVT = MVT::getVectorVT(EltTy, NumElts); 675 if (!isTypeLegal(NewVT)) 676 NewVT = EltTy; 677 IntermediateVT = NewVT; 678 679 MVT DestVT = getTypeToTransformTo(NewVT); 680 RegisterVT = DestVT; 681 if (DestVT.bitsLT(NewVT)) { 682 // Value is expanded, e.g. i64 -> i16. 683 return NumVectorRegs*(NewVT.getSizeInBits()/DestVT.getSizeInBits()); 684 } else { 685 // Otherwise, promotion or legal types use the same number of registers as 686 // the vector decimated to the appropriate level. 687 return NumVectorRegs; 688 } 689 690 return 1; 691} 692 693/// getWidenVectorType: given a vector type, returns the type to widen to 694/// (e.g., v7i8 to v8i8). If the vector type is legal, it returns itself. 695/// If there is no vector type that we want to widen to, returns MVT::Other 696/// When and where to widen is target dependent based on the cost of 697/// scalarizing vs using the wider vector type. 698MVT TargetLowering::getWidenVectorType(MVT VT) const { 699 assert(VT.isVector()); 700 if (isTypeLegal(VT)) 701 return VT; 702 703 // Default is not to widen until moved to LegalizeTypes 704 return MVT::Other; 705} 706 707/// getByValTypeAlignment - Return the desired alignment for ByVal aggregate 708/// function arguments in the caller parameter area. This is the actual 709/// alignment, not its logarithm. 710unsigned TargetLowering::getByValTypeAlignment(const Type *Ty) const { 711 return TD->getCallFrameTypeAlignment(Ty); 712} 713 714SDValue TargetLowering::getPICJumpTableRelocBase(SDValue Table, 715 SelectionDAG &DAG) const { 716 if (usesGlobalOffsetTable()) 717 return DAG.getGLOBAL_OFFSET_TABLE(getPointerTy()); 718 return Table; 719} 720 721bool 722TargetLowering::isOffsetFoldingLegal(const GlobalAddressSDNode *GA) const { 723 // Assume that everything is safe in static mode. 724 if (getTargetMachine().getRelocationModel() == Reloc::Static) 725 return true; 726 727 // In dynamic-no-pic mode, assume that known defined values are safe. 728 if (getTargetMachine().getRelocationModel() == Reloc::DynamicNoPIC && 729 GA && 730 !GA->getGlobal()->isDeclaration() && 731 !GA->getGlobal()->isWeakForLinker()) 732 return true; 733 734 // Otherwise assume nothing is safe. 735 return false; 736} 737 738//===----------------------------------------------------------------------===// 739// Optimization Methods 740//===----------------------------------------------------------------------===// 741 742/// ShrinkDemandedConstant - Check to see if the specified operand of the 743/// specified instruction is a constant integer. If so, check to see if there 744/// are any bits set in the constant that are not demanded. If so, shrink the 745/// constant and return true. 746bool TargetLowering::TargetLoweringOpt::ShrinkDemandedConstant(SDValue Op, 747 const APInt &Demanded) { 748 DebugLoc dl = Op.getDebugLoc(); 749 750 // FIXME: ISD::SELECT, ISD::SELECT_CC 751 switch (Op.getOpcode()) { 752 default: break; 753 case ISD::XOR: 754 case ISD::AND: 755 case ISD::OR: { 756 ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op.getOperand(1)); 757 if (!C) return false; 758 759 if (Op.getOpcode() == ISD::XOR && 760 (C->getAPIntValue() | (~Demanded)).isAllOnesValue()) 761 return false; 762 763 // if we can expand it to have all bits set, do it 764 if (C->getAPIntValue().intersects(~Demanded)) { 765 MVT VT = Op.getValueType(); 766 SDValue New = DAG.getNode(Op.getOpcode(), dl, VT, Op.getOperand(0), 767 DAG.getConstant(Demanded & 768 C->getAPIntValue(), 769 VT)); 770 return CombineTo(Op, New); 771 } 772 773 break; 774 } 775 } 776 777 return false; 778} 779 780/// SimplifyDemandedBits - Look at Op. At this point, we know that only the 781/// DemandedMask bits of the result of Op are ever used downstream. If we can 782/// use this information to simplify Op, create a new simplified DAG node and 783/// return true, returning the original and new nodes in Old and New. Otherwise, 784/// analyze the expression and return a mask of KnownOne and KnownZero bits for 785/// the expression (used to simplify the caller). The KnownZero/One bits may 786/// only be accurate for those bits in the DemandedMask. 787bool TargetLowering::SimplifyDemandedBits(SDValue Op, 788 const APInt &DemandedMask, 789 APInt &KnownZero, 790 APInt &KnownOne, 791 TargetLoweringOpt &TLO, 792 unsigned Depth) const { 793 unsigned BitWidth = DemandedMask.getBitWidth(); 794 assert(Op.getValueSizeInBits() == BitWidth && 795 "Mask size mismatches value type size!"); 796 APInt NewMask = DemandedMask; 797 DebugLoc dl = Op.getDebugLoc(); 798 799 // Don't know anything. 800 KnownZero = KnownOne = APInt(BitWidth, 0); 801 802 // Other users may use these bits. 803 if (!Op.getNode()->hasOneUse()) { 804 if (Depth != 0) { 805 // If not at the root, Just compute the KnownZero/KnownOne bits to 806 // simplify things downstream. 807 TLO.DAG.ComputeMaskedBits(Op, DemandedMask, KnownZero, KnownOne, Depth); 808 return false; 809 } 810 // If this is the root being simplified, allow it to have multiple uses, 811 // just set the NewMask to all bits. 812 NewMask = APInt::getAllOnesValue(BitWidth); 813 } else if (DemandedMask == 0) { 814 // Not demanding any bits from Op. 815 if (Op.getOpcode() != ISD::UNDEF) 816 return TLO.CombineTo(Op, TLO.DAG.getUNDEF(Op.getValueType())); 817 return false; 818 } else if (Depth == 6) { // Limit search depth. 819 return false; 820 } 821 822 APInt KnownZero2, KnownOne2, KnownZeroOut, KnownOneOut; 823 switch (Op.getOpcode()) { 824 case ISD::Constant: 825 // We know all of the bits for a constant! 826 KnownOne = cast<ConstantSDNode>(Op)->getAPIntValue() & NewMask; 827 KnownZero = ~KnownOne & NewMask; 828 return false; // Don't fall through, will infinitely loop. 829 case ISD::AND: 830 // If the RHS is a constant, check to see if the LHS would be zero without 831 // using the bits from the RHS. Below, we use knowledge about the RHS to 832 // simplify the LHS, here we're using information from the LHS to simplify 833 // the RHS. 834 if (ConstantSDNode *RHSC = dyn_cast<ConstantSDNode>(Op.getOperand(1))) { 835 APInt LHSZero, LHSOne; 836 TLO.DAG.ComputeMaskedBits(Op.getOperand(0), NewMask, 837 LHSZero, LHSOne, Depth+1); 838 // If the LHS already has zeros where RHSC does, this and is dead. 839 if ((LHSZero & NewMask) == (~RHSC->getAPIntValue() & NewMask)) 840 return TLO.CombineTo(Op, Op.getOperand(0)); 841 // If any of the set bits in the RHS are known zero on the LHS, shrink 842 // the constant. 843 if (TLO.ShrinkDemandedConstant(Op, ~LHSZero & NewMask)) 844 return true; 845 } 846 847 if (SimplifyDemandedBits(Op.getOperand(1), NewMask, KnownZero, 848 KnownOne, TLO, Depth+1)) 849 return true; 850 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?"); 851 if (SimplifyDemandedBits(Op.getOperand(0), ~KnownZero & NewMask, 852 KnownZero2, KnownOne2, TLO, Depth+1)) 853 return true; 854 assert((KnownZero2 & KnownOne2) == 0 && "Bits known to be one AND zero?"); 855 856 // If all of the demanded bits are known one on one side, return the other. 857 // These bits cannot contribute to the result of the 'and'. 858 if ((NewMask & ~KnownZero2 & KnownOne) == (~KnownZero2 & NewMask)) 859 return TLO.CombineTo(Op, Op.getOperand(0)); 860 if ((NewMask & ~KnownZero & KnownOne2) == (~KnownZero & NewMask)) 861 return TLO.CombineTo(Op, Op.getOperand(1)); 862 // If all of the demanded bits in the inputs are known zeros, return zero. 863 if ((NewMask & (KnownZero|KnownZero2)) == NewMask) 864 return TLO.CombineTo(Op, TLO.DAG.getConstant(0, Op.getValueType())); 865 // If the RHS is a constant, see if we can simplify it. 866 if (TLO.ShrinkDemandedConstant(Op, ~KnownZero2 & NewMask)) 867 return true; 868 869 // Output known-1 bits are only known if set in both the LHS & RHS. 870 KnownOne &= KnownOne2; 871 // Output known-0 are known to be clear if zero in either the LHS | RHS. 872 KnownZero |= KnownZero2; 873 break; 874 case ISD::OR: 875 if (SimplifyDemandedBits(Op.getOperand(1), NewMask, KnownZero, 876 KnownOne, TLO, Depth+1)) 877 return true; 878 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?"); 879 if (SimplifyDemandedBits(Op.getOperand(0), ~KnownOne & NewMask, 880 KnownZero2, KnownOne2, TLO, Depth+1)) 881 return true; 882 assert((KnownZero2 & KnownOne2) == 0 && "Bits known to be one AND zero?"); 883 884 // If all of the demanded bits are known zero on one side, return the other. 885 // These bits cannot contribute to the result of the 'or'. 886 if ((NewMask & ~KnownOne2 & KnownZero) == (~KnownOne2 & NewMask)) 887 return TLO.CombineTo(Op, Op.getOperand(0)); 888 if ((NewMask & ~KnownOne & KnownZero2) == (~KnownOne & NewMask)) 889 return TLO.CombineTo(Op, Op.getOperand(1)); 890 // If all of the potentially set bits on one side are known to be set on 891 // the other side, just use the 'other' side. 892 if ((NewMask & ~KnownZero & KnownOne2) == (~KnownZero & NewMask)) 893 return TLO.CombineTo(Op, Op.getOperand(0)); 894 if ((NewMask & ~KnownZero2 & KnownOne) == (~KnownZero2 & NewMask)) 895 return TLO.CombineTo(Op, Op.getOperand(1)); 896 // If the RHS is a constant, see if we can simplify it. 897 if (TLO.ShrinkDemandedConstant(Op, NewMask)) 898 return true; 899 900 // Output known-0 bits are only known if clear in both the LHS & RHS. 901 KnownZero &= KnownZero2; 902 // Output known-1 are known to be set if set in either the LHS | RHS. 903 KnownOne |= KnownOne2; 904 break; 905 case ISD::XOR: 906 if (SimplifyDemandedBits(Op.getOperand(1), NewMask, KnownZero, 907 KnownOne, TLO, Depth+1)) 908 return true; 909 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?"); 910 if (SimplifyDemandedBits(Op.getOperand(0), NewMask, KnownZero2, 911 KnownOne2, TLO, Depth+1)) 912 return true; 913 assert((KnownZero2 & KnownOne2) == 0 && "Bits known to be one AND zero?"); 914 915 // If all of the demanded bits are known zero on one side, return the other. 916 // These bits cannot contribute to the result of the 'xor'. 917 if ((KnownZero & NewMask) == NewMask) 918 return TLO.CombineTo(Op, Op.getOperand(0)); 919 if ((KnownZero2 & NewMask) == NewMask) 920 return TLO.CombineTo(Op, Op.getOperand(1)); 921 922 // If all of the unknown bits are known to be zero on one side or the other 923 // (but not both) turn this into an *inclusive* or. 924 // e.g. (A & C1)^(B & C2) -> (A & C1)|(B & C2) iff C1&C2 == 0 925 if ((NewMask & ~KnownZero & ~KnownZero2) == 0) 926 return TLO.CombineTo(Op, TLO.DAG.getNode(ISD::OR, dl, Op.getValueType(), 927 Op.getOperand(0), 928 Op.getOperand(1))); 929 930 // Output known-0 bits are known if clear or set in both the LHS & RHS. 931 KnownZeroOut = (KnownZero & KnownZero2) | (KnownOne & KnownOne2); 932 // Output known-1 are known to be set if set in only one of the LHS, RHS. 933 KnownOneOut = (KnownZero & KnownOne2) | (KnownOne & KnownZero2); 934 935 // If all of the demanded bits on one side are known, and all of the set 936 // bits on that side are also known to be set on the other side, turn this 937 // into an AND, as we know the bits will be cleared. 938 // e.g. (X | C1) ^ C2 --> (X | C1) & ~C2 iff (C1&C2) == C2 939 if ((NewMask & (KnownZero|KnownOne)) == NewMask) { // all known 940 if ((KnownOne & KnownOne2) == KnownOne) { 941 MVT VT = Op.getValueType(); 942 SDValue ANDC = TLO.DAG.getConstant(~KnownOne & NewMask, VT); 943 return TLO.CombineTo(Op, TLO.DAG.getNode(ISD::AND, dl, VT, 944 Op.getOperand(0), ANDC)); 945 } 946 } 947 948 // If the RHS is a constant, see if we can simplify it. 949 // for XOR, we prefer to force bits to 1 if they will make a -1. 950 // if we can't force bits, try to shrink constant 951 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op.getOperand(1))) { 952 APInt Expanded = C->getAPIntValue() | (~NewMask); 953 // if we can expand it to have all bits set, do it 954 if (Expanded.isAllOnesValue()) { 955 if (Expanded != C->getAPIntValue()) { 956 MVT VT = Op.getValueType(); 957 SDValue New = TLO.DAG.getNode(Op.getOpcode(), dl,VT, Op.getOperand(0), 958 TLO.DAG.getConstant(Expanded, VT)); 959 return TLO.CombineTo(Op, New); 960 } 961 // if it already has all the bits set, nothing to change 962 // but don't shrink either! 963 } else if (TLO.ShrinkDemandedConstant(Op, NewMask)) { 964 return true; 965 } 966 } 967 968 KnownZero = KnownZeroOut; 969 KnownOne = KnownOneOut; 970 break; 971 case ISD::SELECT: 972 if (SimplifyDemandedBits(Op.getOperand(2), NewMask, KnownZero, 973 KnownOne, TLO, Depth+1)) 974 return true; 975 if (SimplifyDemandedBits(Op.getOperand(1), NewMask, KnownZero2, 976 KnownOne2, TLO, Depth+1)) 977 return true; 978 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?"); 979 assert((KnownZero2 & KnownOne2) == 0 && "Bits known to be one AND zero?"); 980 981 // If the operands are constants, see if we can simplify them. 982 if (TLO.ShrinkDemandedConstant(Op, NewMask)) 983 return true; 984 985 // Only known if known in both the LHS and RHS. 986 KnownOne &= KnownOne2; 987 KnownZero &= KnownZero2; 988 break; 989 case ISD::SELECT_CC: 990 if (SimplifyDemandedBits(Op.getOperand(3), NewMask, KnownZero, 991 KnownOne, TLO, Depth+1)) 992 return true; 993 if (SimplifyDemandedBits(Op.getOperand(2), NewMask, KnownZero2, 994 KnownOne2, TLO, Depth+1)) 995 return true; 996 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?"); 997 assert((KnownZero2 & KnownOne2) == 0 && "Bits known to be one AND zero?"); 998 999 // If the operands are constants, see if we can simplify them. 1000 if (TLO.ShrinkDemandedConstant(Op, NewMask)) 1001 return true; 1002 1003 // Only known if known in both the LHS and RHS. 1004 KnownOne &= KnownOne2; 1005 KnownZero &= KnownZero2; 1006 break; 1007 case ISD::SHL: 1008 if (ConstantSDNode *SA = dyn_cast<ConstantSDNode>(Op.getOperand(1))) { 1009 unsigned ShAmt = SA->getZExtValue(); 1010 SDValue InOp = Op.getOperand(0); 1011 1012 // If the shift count is an invalid immediate, don't do anything. 1013 if (ShAmt >= BitWidth) 1014 break; 1015 1016 // If this is ((X >>u C1) << ShAmt), see if we can simplify this into a 1017 // single shift. We can do this if the bottom bits (which are shifted 1018 // out) are never demanded. 1019 if (InOp.getOpcode() == ISD::SRL && 1020 isa<ConstantSDNode>(InOp.getOperand(1))) { 1021 if (ShAmt && (NewMask & APInt::getLowBitsSet(BitWidth, ShAmt)) == 0) { 1022 unsigned C1= cast<ConstantSDNode>(InOp.getOperand(1))->getZExtValue(); 1023 unsigned Opc = ISD::SHL; 1024 int Diff = ShAmt-C1; 1025 if (Diff < 0) { 1026 Diff = -Diff; 1027 Opc = ISD::SRL; 1028 } 1029 1030 SDValue NewSA = 1031 TLO.DAG.getConstant(Diff, Op.getOperand(1).getValueType()); 1032 MVT VT = Op.getValueType(); 1033 return TLO.CombineTo(Op, TLO.DAG.getNode(Opc, dl, VT, 1034 InOp.getOperand(0), NewSA)); 1035 } 1036 } 1037 1038 if (SimplifyDemandedBits(Op.getOperand(0), NewMask.lshr(ShAmt), 1039 KnownZero, KnownOne, TLO, Depth+1)) 1040 return true; 1041 KnownZero <<= SA->getZExtValue(); 1042 KnownOne <<= SA->getZExtValue(); 1043 // low bits known zero. 1044 KnownZero |= APInt::getLowBitsSet(BitWidth, SA->getZExtValue()); 1045 } 1046 break; 1047 case ISD::SRL: 1048 if (ConstantSDNode *SA = dyn_cast<ConstantSDNode>(Op.getOperand(1))) { 1049 MVT VT = Op.getValueType(); 1050 unsigned ShAmt = SA->getZExtValue(); 1051 unsigned VTSize = VT.getSizeInBits(); 1052 SDValue InOp = Op.getOperand(0); 1053 1054 // If the shift count is an invalid immediate, don't do anything. 1055 if (ShAmt >= BitWidth) 1056 break; 1057 1058 // If this is ((X << C1) >>u ShAmt), see if we can simplify this into a 1059 // single shift. We can do this if the top bits (which are shifted out) 1060 // are never demanded. 1061 if (InOp.getOpcode() == ISD::SHL && 1062 isa<ConstantSDNode>(InOp.getOperand(1))) { 1063 if (ShAmt && (NewMask & APInt::getHighBitsSet(VTSize, ShAmt)) == 0) { 1064 unsigned C1= cast<ConstantSDNode>(InOp.getOperand(1))->getZExtValue(); 1065 unsigned Opc = ISD::SRL; 1066 int Diff = ShAmt-C1; 1067 if (Diff < 0) { 1068 Diff = -Diff; 1069 Opc = ISD::SHL; 1070 } 1071 1072 SDValue NewSA = 1073 TLO.DAG.getConstant(Diff, Op.getOperand(1).getValueType()); 1074 return TLO.CombineTo(Op, TLO.DAG.getNode(Opc, dl, VT, 1075 InOp.getOperand(0), NewSA)); 1076 } 1077 } 1078 1079 // Compute the new bits that are at the top now. 1080 if (SimplifyDemandedBits(InOp, (NewMask << ShAmt), 1081 KnownZero, KnownOne, TLO, Depth+1)) 1082 return true; 1083 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?"); 1084 KnownZero = KnownZero.lshr(ShAmt); 1085 KnownOne = KnownOne.lshr(ShAmt); 1086 1087 APInt HighBits = APInt::getHighBitsSet(BitWidth, ShAmt); 1088 KnownZero |= HighBits; // High bits known zero. 1089 } 1090 break; 1091 case ISD::SRA: 1092 // If this is an arithmetic shift right and only the low-bit is set, we can 1093 // always convert this into a logical shr, even if the shift amount is 1094 // variable. The low bit of the shift cannot be an input sign bit unless 1095 // the shift amount is >= the size of the datatype, which is undefined. 1096 if (DemandedMask == 1) 1097 return TLO.CombineTo(Op, TLO.DAG.getNode(ISD::SRL, dl, Op.getValueType(), 1098 Op.getOperand(0), Op.getOperand(1))); 1099 1100 if (ConstantSDNode *SA = dyn_cast<ConstantSDNode>(Op.getOperand(1))) { 1101 MVT VT = Op.getValueType(); 1102 unsigned ShAmt = SA->getZExtValue(); 1103 1104 // If the shift count is an invalid immediate, don't do anything. 1105 if (ShAmt >= BitWidth) 1106 break; 1107 1108 APInt InDemandedMask = (NewMask << ShAmt); 1109 1110 // If any of the demanded bits are produced by the sign extension, we also 1111 // demand the input sign bit. 1112 APInt HighBits = APInt::getHighBitsSet(BitWidth, ShAmt); 1113 if (HighBits.intersects(NewMask)) 1114 InDemandedMask |= APInt::getSignBit(VT.getSizeInBits()); 1115 1116 if (SimplifyDemandedBits(Op.getOperand(0), InDemandedMask, 1117 KnownZero, KnownOne, TLO, Depth+1)) 1118 return true; 1119 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?"); 1120 KnownZero = KnownZero.lshr(ShAmt); 1121 KnownOne = KnownOne.lshr(ShAmt); 1122 1123 // Handle the sign bit, adjusted to where it is now in the mask. 1124 APInt SignBit = APInt::getSignBit(BitWidth).lshr(ShAmt); 1125 1126 // If the input sign bit is known to be zero, or if none of the top bits 1127 // are demanded, turn this into an unsigned shift right. 1128 if (KnownZero.intersects(SignBit) || (HighBits & ~NewMask) == HighBits) { 1129 return TLO.CombineTo(Op, TLO.DAG.getNode(ISD::SRL, dl, VT, 1130 Op.getOperand(0), 1131 Op.getOperand(1))); 1132 } else if (KnownOne.intersects(SignBit)) { // New bits are known one. 1133 KnownOne |= HighBits; 1134 } 1135 } 1136 break; 1137 case ISD::SIGN_EXTEND_INREG: { 1138 MVT EVT = cast<VTSDNode>(Op.getOperand(1))->getVT(); 1139 1140 // Sign extension. Compute the demanded bits in the result that are not 1141 // present in the input. 1142 APInt NewBits = APInt::getHighBitsSet(BitWidth, 1143 BitWidth - EVT.getSizeInBits()) & 1144 NewMask; 1145 1146 // If none of the extended bits are demanded, eliminate the sextinreg. 1147 if (NewBits == 0) 1148 return TLO.CombineTo(Op, Op.getOperand(0)); 1149 1150 APInt InSignBit = APInt::getSignBit(EVT.getSizeInBits()); 1151 InSignBit.zext(BitWidth); 1152 APInt InputDemandedBits = APInt::getLowBitsSet(BitWidth, 1153 EVT.getSizeInBits()) & 1154 NewMask; 1155 1156 // Since the sign extended bits are demanded, we know that the sign 1157 // bit is demanded. 1158 InputDemandedBits |= InSignBit; 1159 1160 if (SimplifyDemandedBits(Op.getOperand(0), InputDemandedBits, 1161 KnownZero, KnownOne, TLO, Depth+1)) 1162 return true; 1163 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?"); 1164 1165 // If the sign bit of the input is known set or clear, then we know the 1166 // top bits of the result. 1167 1168 // If the input sign bit is known zero, convert this into a zero extension. 1169 if (KnownZero.intersects(InSignBit)) 1170 return TLO.CombineTo(Op, 1171 TLO.DAG.getZeroExtendInReg(Op.getOperand(0),dl,EVT)); 1172 1173 if (KnownOne.intersects(InSignBit)) { // Input sign bit known set 1174 KnownOne |= NewBits; 1175 KnownZero &= ~NewBits; 1176 } else { // Input sign bit unknown 1177 KnownZero &= ~NewBits; 1178 KnownOne &= ~NewBits; 1179 } 1180 break; 1181 } 1182 case ISD::ZERO_EXTEND: { 1183 unsigned OperandBitWidth = Op.getOperand(0).getValueSizeInBits(); 1184 APInt InMask = NewMask; 1185 InMask.trunc(OperandBitWidth); 1186 1187 // If none of the top bits are demanded, convert this into an any_extend. 1188 APInt NewBits = 1189 APInt::getHighBitsSet(BitWidth, BitWidth - OperandBitWidth) & NewMask; 1190 if (!NewBits.intersects(NewMask)) 1191 return TLO.CombineTo(Op, TLO.DAG.getNode(ISD::ANY_EXTEND, dl, 1192 Op.getValueType(), 1193 Op.getOperand(0))); 1194 1195 if (SimplifyDemandedBits(Op.getOperand(0), InMask, 1196 KnownZero, KnownOne, TLO, Depth+1)) 1197 return true; 1198 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?"); 1199 KnownZero.zext(BitWidth); 1200 KnownOne.zext(BitWidth); 1201 KnownZero |= NewBits; 1202 break; 1203 } 1204 case ISD::SIGN_EXTEND: { 1205 MVT InVT = Op.getOperand(0).getValueType(); 1206 unsigned InBits = InVT.getSizeInBits(); 1207 APInt InMask = APInt::getLowBitsSet(BitWidth, InBits); 1208 APInt InSignBit = APInt::getBitsSet(BitWidth, InBits - 1, InBits); 1209 APInt NewBits = ~InMask & NewMask; 1210 1211 // If none of the top bits are demanded, convert this into an any_extend. 1212 if (NewBits == 0) 1213 return TLO.CombineTo(Op,TLO.DAG.getNode(ISD::ANY_EXTEND, dl, 1214 Op.getValueType(), 1215 Op.getOperand(0))); 1216 1217 // Since some of the sign extended bits are demanded, we know that the sign 1218 // bit is demanded. 1219 APInt InDemandedBits = InMask & NewMask; 1220 InDemandedBits |= InSignBit; 1221 InDemandedBits.trunc(InBits); 1222 1223 if (SimplifyDemandedBits(Op.getOperand(0), InDemandedBits, KnownZero, 1224 KnownOne, TLO, Depth+1)) 1225 return true; 1226 KnownZero.zext(BitWidth); 1227 KnownOne.zext(BitWidth); 1228 1229 // If the sign bit is known zero, convert this to a zero extend. 1230 if (KnownZero.intersects(InSignBit)) 1231 return TLO.CombineTo(Op, TLO.DAG.getNode(ISD::ZERO_EXTEND, dl, 1232 Op.getValueType(), 1233 Op.getOperand(0))); 1234 1235 // If the sign bit is known one, the top bits match. 1236 if (KnownOne.intersects(InSignBit)) { 1237 KnownOne |= NewBits; 1238 KnownZero &= ~NewBits; 1239 } else { // Otherwise, top bits aren't known. 1240 KnownOne &= ~NewBits; 1241 KnownZero &= ~NewBits; 1242 } 1243 break; 1244 } 1245 case ISD::ANY_EXTEND: { 1246 unsigned OperandBitWidth = Op.getOperand(0).getValueSizeInBits(); 1247 APInt InMask = NewMask; 1248 InMask.trunc(OperandBitWidth); 1249 if (SimplifyDemandedBits(Op.getOperand(0), InMask, 1250 KnownZero, KnownOne, TLO, Depth+1)) 1251 return true; 1252 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?"); 1253 KnownZero.zext(BitWidth); 1254 KnownOne.zext(BitWidth); 1255 break; 1256 } 1257 case ISD::TRUNCATE: { 1258 // Simplify the input, using demanded bit information, and compute the known 1259 // zero/one bits live out. 1260 APInt TruncMask = NewMask; 1261 TruncMask.zext(Op.getOperand(0).getValueSizeInBits()); 1262 if (SimplifyDemandedBits(Op.getOperand(0), TruncMask, 1263 KnownZero, KnownOne, TLO, Depth+1)) 1264 return true; 1265 KnownZero.trunc(BitWidth); 1266 KnownOne.trunc(BitWidth); 1267 1268 // If the input is only used by this truncate, see if we can shrink it based 1269 // on the known demanded bits. 1270 if (Op.getOperand(0).getNode()->hasOneUse()) { 1271 SDValue In = Op.getOperand(0); 1272 unsigned InBitWidth = In.getValueSizeInBits(); 1273 switch (In.getOpcode()) { 1274 default: break; 1275 case ISD::SRL: 1276 // Shrink SRL by a constant if none of the high bits shifted in are 1277 // demanded. 1278 if (ConstantSDNode *ShAmt = dyn_cast<ConstantSDNode>(In.getOperand(1))){ 1279 APInt HighBits = APInt::getHighBitsSet(InBitWidth, 1280 InBitWidth - BitWidth); 1281 HighBits = HighBits.lshr(ShAmt->getZExtValue()); 1282 HighBits.trunc(BitWidth); 1283 1284 if (ShAmt->getZExtValue() < BitWidth && !(HighBits & NewMask)) { 1285 // None of the shifted in bits are needed. Add a truncate of the 1286 // shift input, then shift it. 1287 SDValue NewTrunc = TLO.DAG.getNode(ISD::TRUNCATE, dl, 1288 Op.getValueType(), 1289 In.getOperand(0)); 1290 return TLO.CombineTo(Op, TLO.DAG.getNode(ISD::SRL, dl, 1291 Op.getValueType(), 1292 NewTrunc, 1293 In.getOperand(1))); 1294 } 1295 } 1296 break; 1297 } 1298 } 1299 1300 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?"); 1301 break; 1302 } 1303 case ISD::AssertZext: { 1304 MVT VT = cast<VTSDNode>(Op.getOperand(1))->getVT(); 1305 APInt InMask = APInt::getLowBitsSet(BitWidth, 1306 VT.getSizeInBits()); 1307 if (SimplifyDemandedBits(Op.getOperand(0), InMask & NewMask, 1308 KnownZero, KnownOne, TLO, Depth+1)) 1309 return true; 1310 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?"); 1311 KnownZero |= ~InMask & NewMask; 1312 break; 1313 } 1314 case ISD::BIT_CONVERT: 1315#if 0 1316 // If this is an FP->Int bitcast and if the sign bit is the only thing that 1317 // is demanded, turn this into a FGETSIGN. 1318 if (NewMask == MVT::getIntegerVTSignBit(Op.getValueType()) && 1319 MVT::isFloatingPoint(Op.getOperand(0).getValueType()) && 1320 !MVT::isVector(Op.getOperand(0).getValueType())) { 1321 // Only do this xform if FGETSIGN is valid or if before legalize. 1322 if (!TLO.AfterLegalize || 1323 isOperationLegal(ISD::FGETSIGN, Op.getValueType())) { 1324 // Make a FGETSIGN + SHL to move the sign bit into the appropriate 1325 // place. We expect the SHL to be eliminated by other optimizations. 1326 SDValue Sign = TLO.DAG.getNode(ISD::FGETSIGN, Op.getValueType(), 1327 Op.getOperand(0)); 1328 unsigned ShVal = Op.getValueType().getSizeInBits()-1; 1329 SDValue ShAmt = TLO.DAG.getConstant(ShVal, getShiftAmountTy()); 1330 return TLO.CombineTo(Op, TLO.DAG.getNode(ISD::SHL, Op.getValueType(), 1331 Sign, ShAmt)); 1332 } 1333 } 1334#endif 1335 break; 1336 default: 1337 // Just use ComputeMaskedBits to compute output bits. 1338 TLO.DAG.ComputeMaskedBits(Op, NewMask, KnownZero, KnownOne, Depth); 1339 break; 1340 } 1341 1342 // If we know the value of all of the demanded bits, return this as a 1343 // constant. 1344 if ((NewMask & (KnownZero|KnownOne)) == NewMask) 1345 return TLO.CombineTo(Op, TLO.DAG.getConstant(KnownOne, Op.getValueType())); 1346 1347 return false; 1348} 1349 1350/// computeMaskedBitsForTargetNode - Determine which of the bits specified 1351/// in Mask are known to be either zero or one and return them in the 1352/// KnownZero/KnownOne bitsets. 1353void TargetLowering::computeMaskedBitsForTargetNode(const SDValue Op, 1354 const APInt &Mask, 1355 APInt &KnownZero, 1356 APInt &KnownOne, 1357 const SelectionDAG &DAG, 1358 unsigned Depth) const { 1359 assert((Op.getOpcode() >= ISD::BUILTIN_OP_END || 1360 Op.getOpcode() == ISD::INTRINSIC_WO_CHAIN || 1361 Op.getOpcode() == ISD::INTRINSIC_W_CHAIN || 1362 Op.getOpcode() == ISD::INTRINSIC_VOID) && 1363 "Should use MaskedValueIsZero if you don't know whether Op" 1364 " is a target node!"); 1365 KnownZero = KnownOne = APInt(Mask.getBitWidth(), 0); 1366} 1367 1368/// ComputeNumSignBitsForTargetNode - This method can be implemented by 1369/// targets that want to expose additional information about sign bits to the 1370/// DAG Combiner. 1371unsigned TargetLowering::ComputeNumSignBitsForTargetNode(SDValue Op, 1372 unsigned Depth) const { 1373 assert((Op.getOpcode() >= ISD::BUILTIN_OP_END || 1374 Op.getOpcode() == ISD::INTRINSIC_WO_CHAIN || 1375 Op.getOpcode() == ISD::INTRINSIC_W_CHAIN || 1376 Op.getOpcode() == ISD::INTRINSIC_VOID) && 1377 "Should use ComputeNumSignBits if you don't know whether Op" 1378 " is a target node!"); 1379 return 1; 1380} 1381 1382/// ValueHasExactlyOneBitSet - Test if the given value is known to have exactly 1383/// one bit set. This differs from ComputeMaskedBits in that it doesn't need to 1384/// determine which bit is set. 1385/// 1386static bool ValueHasExactlyOneBitSet(SDValue Val, const SelectionDAG &DAG) { 1387 // A left-shift of a constant one will have exactly one bit set, because 1388 // shifting the bit off the end is undefined. 1389 if (Val.getOpcode() == ISD::SHL) 1390 if (ConstantSDNode *C = 1391 dyn_cast<ConstantSDNode>(Val.getNode()->getOperand(0))) 1392 if (C->getAPIntValue() == 1) 1393 return true; 1394 1395 // Similarly, a right-shift of a constant sign-bit will have exactly 1396 // one bit set. 1397 if (Val.getOpcode() == ISD::SRL) 1398 if (ConstantSDNode *C = 1399 dyn_cast<ConstantSDNode>(Val.getNode()->getOperand(0))) 1400 if (C->getAPIntValue().isSignBit()) 1401 return true; 1402 1403 // More could be done here, though the above checks are enough 1404 // to handle some common cases. 1405 1406 // Fall back to ComputeMaskedBits to catch other known cases. 1407 MVT OpVT = Val.getValueType(); 1408 unsigned BitWidth = OpVT.getSizeInBits(); 1409 APInt Mask = APInt::getAllOnesValue(BitWidth); 1410 APInt KnownZero, KnownOne; 1411 DAG.ComputeMaskedBits(Val, Mask, KnownZero, KnownOne); 1412 return (KnownZero.countPopulation() == BitWidth - 1) && 1413 (KnownOne.countPopulation() == 1); 1414} 1415 1416/// SimplifySetCC - Try to simplify a setcc built with the specified operands 1417/// and cc. If it is unable to simplify it, return a null SDValue. 1418SDValue 1419TargetLowering::SimplifySetCC(MVT VT, SDValue N0, SDValue N1, 1420 ISD::CondCode Cond, bool foldBooleans, 1421 DAGCombinerInfo &DCI, DebugLoc dl) const { 1422 SelectionDAG &DAG = DCI.DAG; 1423 1424 // These setcc operations always fold. 1425 switch (Cond) { 1426 default: break; 1427 case ISD::SETFALSE: 1428 case ISD::SETFALSE2: return DAG.getConstant(0, VT); 1429 case ISD::SETTRUE: 1430 case ISD::SETTRUE2: return DAG.getConstant(1, VT); 1431 } 1432 1433 if (ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1.getNode())) { 1434 const APInt &C1 = N1C->getAPIntValue(); 1435 if (isa<ConstantSDNode>(N0.getNode())) { 1436 return DAG.FoldSetCC(VT, N0, N1, Cond, dl); 1437 } else { 1438 // If the LHS is '(srl (ctlz x), 5)', the RHS is 0/1, and this is an 1439 // equality comparison, then we're just comparing whether X itself is 1440 // zero. 1441 if (N0.getOpcode() == ISD::SRL && (C1 == 0 || C1 == 1) && 1442 N0.getOperand(0).getOpcode() == ISD::CTLZ && 1443 N0.getOperand(1).getOpcode() == ISD::Constant) { 1444 unsigned ShAmt = cast<ConstantSDNode>(N0.getOperand(1))->getZExtValue(); 1445 if ((Cond == ISD::SETEQ || Cond == ISD::SETNE) && 1446 ShAmt == Log2_32(N0.getValueType().getSizeInBits())) { 1447 if ((C1 == 0) == (Cond == ISD::SETEQ)) { 1448 // (srl (ctlz x), 5) == 0 -> X != 0 1449 // (srl (ctlz x), 5) != 1 -> X != 0 1450 Cond = ISD::SETNE; 1451 } else { 1452 // (srl (ctlz x), 5) != 0 -> X == 0 1453 // (srl (ctlz x), 5) == 1 -> X == 0 1454 Cond = ISD::SETEQ; 1455 } 1456 SDValue Zero = DAG.getConstant(0, N0.getValueType()); 1457 return DAG.getSetCC(dl, VT, N0.getOperand(0).getOperand(0), 1458 Zero, Cond); 1459 } 1460 } 1461 1462 // If the LHS is '(and load, const)', the RHS is 0, 1463 // the test is for equality or unsigned, and all 1 bits of the const are 1464 // in the same partial word, see if we can shorten the load. 1465 if (DCI.isBeforeLegalize() && 1466 N0.getOpcode() == ISD::AND && C1 == 0 && 1467 isa<LoadSDNode>(N0.getOperand(0)) && 1468 N0.getOperand(0).getNode()->hasOneUse() && 1469 isa<ConstantSDNode>(N0.getOperand(1))) { 1470 LoadSDNode *Lod = cast<LoadSDNode>(N0.getOperand(0)); 1471 uint64_t Mask = cast<ConstantSDNode>(N0.getOperand(1))->getZExtValue(); 1472 uint64_t bestMask = 0; 1473 unsigned bestWidth = 0, bestOffset = 0; 1474 if (!Lod->isVolatile() && Lod->isUnindexed()) { 1475 unsigned origWidth = N0.getValueType().getSizeInBits(); 1476 // We can narrow (e.g.) 16-bit extending loads on 32-bit target to 1477 // 8 bits, but have to be careful... 1478 if (Lod->getExtensionType() != ISD::NON_EXTLOAD) 1479 origWidth = Lod->getMemoryVT().getSizeInBits(); 1480 for (unsigned width = origWidth / 2; width>=8; width /= 2) { 1481 uint64_t newMask = (1ULL << width) - 1; 1482 for (unsigned offset=0; offset<origWidth/width; offset++) { 1483 if ((newMask & Mask)==Mask) { 1484 if (!TD->isLittleEndian()) 1485 bestOffset = (origWidth/width - offset - 1) * (width/8); 1486 else 1487 bestOffset = (uint64_t)offset * (width/8); 1488 bestMask = Mask >> (offset * (width/8) * 8); 1489 bestWidth = width; 1490 break; 1491 } 1492 newMask = newMask << width; 1493 } 1494 } 1495 } 1496 if (bestWidth) { 1497 MVT newVT = MVT::getIntegerVT(bestWidth); 1498 if (newVT.isRound()) { 1499 MVT PtrType = Lod->getOperand(1).getValueType(); 1500 SDValue Ptr = Lod->getBasePtr(); 1501 if (bestOffset != 0) 1502 Ptr = DAG.getNode(ISD::ADD, dl, PtrType, Lod->getBasePtr(), 1503 DAG.getConstant(bestOffset, PtrType)); 1504 unsigned NewAlign = MinAlign(Lod->getAlignment(), bestOffset); 1505 SDValue NewLoad = DAG.getLoad(newVT, dl, Lod->getChain(), Ptr, 1506 Lod->getSrcValue(), 1507 Lod->getSrcValueOffset() + bestOffset, 1508 false, NewAlign); 1509 return DAG.getSetCC(dl, VT, 1510 DAG.getNode(ISD::AND, dl, newVT, NewLoad, 1511 DAG.getConstant(bestMask, newVT)), 1512 DAG.getConstant(0LL, newVT), Cond); 1513 } 1514 } 1515 } 1516 1517 // If the LHS is a ZERO_EXTEND, perform the comparison on the input. 1518 if (N0.getOpcode() == ISD::ZERO_EXTEND) { 1519 unsigned InSize = N0.getOperand(0).getValueType().getSizeInBits(); 1520 1521 // If the comparison constant has bits in the upper part, the 1522 // zero-extended value could never match. 1523 if (C1.intersects(APInt::getHighBitsSet(C1.getBitWidth(), 1524 C1.getBitWidth() - InSize))) { 1525 switch (Cond) { 1526 case ISD::SETUGT: 1527 case ISD::SETUGE: 1528 case ISD::SETEQ: return DAG.getConstant(0, VT); 1529 case ISD::SETULT: 1530 case ISD::SETULE: 1531 case ISD::SETNE: return DAG.getConstant(1, VT); 1532 case ISD::SETGT: 1533 case ISD::SETGE: 1534 // True if the sign bit of C1 is set. 1535 return DAG.getConstant(C1.isNegative(), VT); 1536 case ISD::SETLT: 1537 case ISD::SETLE: 1538 // True if the sign bit of C1 isn't set. 1539 return DAG.getConstant(C1.isNonNegative(), VT); 1540 default: 1541 break; 1542 } 1543 } 1544 1545 // Otherwise, we can perform the comparison with the low bits. 1546 switch (Cond) { 1547 case ISD::SETEQ: 1548 case ISD::SETNE: 1549 case ISD::SETUGT: 1550 case ISD::SETUGE: 1551 case ISD::SETULT: 1552 case ISD::SETULE: 1553 return DAG.getSetCC(dl, VT, N0.getOperand(0), 1554 DAG.getConstant(APInt(C1).trunc(InSize), 1555 N0.getOperand(0).getValueType()), 1556 Cond); 1557 default: 1558 break; // todo, be more careful with signed comparisons 1559 } 1560 } else if (N0.getOpcode() == ISD::SIGN_EXTEND_INREG && 1561 (Cond == ISD::SETEQ || Cond == ISD::SETNE)) { 1562 MVT ExtSrcTy = cast<VTSDNode>(N0.getOperand(1))->getVT(); 1563 unsigned ExtSrcTyBits = ExtSrcTy.getSizeInBits(); 1564 MVT ExtDstTy = N0.getValueType(); 1565 unsigned ExtDstTyBits = ExtDstTy.getSizeInBits(); 1566 1567 // If the extended part has any inconsistent bits, it cannot ever 1568 // compare equal. In other words, they have to be all ones or all 1569 // zeros. 1570 APInt ExtBits = 1571 APInt::getHighBitsSet(ExtDstTyBits, ExtDstTyBits - ExtSrcTyBits); 1572 if ((C1 & ExtBits) != 0 && (C1 & ExtBits) != ExtBits) 1573 return DAG.getConstant(Cond == ISD::SETNE, VT); 1574 1575 SDValue ZextOp; 1576 MVT Op0Ty = N0.getOperand(0).getValueType(); 1577 if (Op0Ty == ExtSrcTy) { 1578 ZextOp = N0.getOperand(0); 1579 } else { 1580 APInt Imm = APInt::getLowBitsSet(ExtDstTyBits, ExtSrcTyBits); 1581 ZextOp = DAG.getNode(ISD::AND, dl, Op0Ty, N0.getOperand(0), 1582 DAG.getConstant(Imm, Op0Ty)); 1583 } 1584 if (!DCI.isCalledByLegalizer()) 1585 DCI.AddToWorklist(ZextOp.getNode()); 1586 // Otherwise, make this a use of a zext. 1587 return DAG.getSetCC(dl, VT, ZextOp, 1588 DAG.getConstant(C1 & APInt::getLowBitsSet( 1589 ExtDstTyBits, 1590 ExtSrcTyBits), 1591 ExtDstTy), 1592 Cond); 1593 } else if ((N1C->isNullValue() || N1C->getAPIntValue() == 1) && 1594 (Cond == ISD::SETEQ || Cond == ISD::SETNE)) { 1595 1596 // SETCC (SETCC), [0|1], [EQ|NE] -> SETCC 1597 if (N0.getOpcode() == ISD::SETCC) { 1598 bool TrueWhenTrue = (Cond == ISD::SETEQ) ^ (N1C->getZExtValue() != 1); 1599 if (TrueWhenTrue) 1600 return N0; 1601 1602 // Invert the condition. 1603 ISD::CondCode CC = cast<CondCodeSDNode>(N0.getOperand(2))->get(); 1604 CC = ISD::getSetCCInverse(CC, 1605 N0.getOperand(0).getValueType().isInteger()); 1606 return DAG.getSetCC(dl, VT, N0.getOperand(0), N0.getOperand(1), CC); 1607 } 1608 1609 if ((N0.getOpcode() == ISD::XOR || 1610 (N0.getOpcode() == ISD::AND && 1611 N0.getOperand(0).getOpcode() == ISD::XOR && 1612 N0.getOperand(1) == N0.getOperand(0).getOperand(1))) && 1613 isa<ConstantSDNode>(N0.getOperand(1)) && 1614 cast<ConstantSDNode>(N0.getOperand(1))->getAPIntValue() == 1) { 1615 // If this is (X^1) == 0/1, swap the RHS and eliminate the xor. We 1616 // can only do this if the top bits are known zero. 1617 unsigned BitWidth = N0.getValueSizeInBits(); 1618 if (DAG.MaskedValueIsZero(N0, 1619 APInt::getHighBitsSet(BitWidth, 1620 BitWidth-1))) { 1621 // Okay, get the un-inverted input value. 1622 SDValue Val; 1623 if (N0.getOpcode() == ISD::XOR) 1624 Val = N0.getOperand(0); 1625 else { 1626 assert(N0.getOpcode() == ISD::AND && 1627 N0.getOperand(0).getOpcode() == ISD::XOR); 1628 // ((X^1)&1)^1 -> X & 1 1629 Val = DAG.getNode(ISD::AND, dl, N0.getValueType(), 1630 N0.getOperand(0).getOperand(0), 1631 N0.getOperand(1)); 1632 } 1633 return DAG.getSetCC(dl, VT, Val, N1, 1634 Cond == ISD::SETEQ ? ISD::SETNE : ISD::SETEQ); 1635 } 1636 } 1637 } 1638 1639 APInt MinVal, MaxVal; 1640 unsigned OperandBitSize = N1C->getValueType(0).getSizeInBits(); 1641 if (ISD::isSignedIntSetCC(Cond)) { 1642 MinVal = APInt::getSignedMinValue(OperandBitSize); 1643 MaxVal = APInt::getSignedMaxValue(OperandBitSize); 1644 } else { 1645 MinVal = APInt::getMinValue(OperandBitSize); 1646 MaxVal = APInt::getMaxValue(OperandBitSize); 1647 } 1648 1649 // Canonicalize GE/LE comparisons to use GT/LT comparisons. 1650 if (Cond == ISD::SETGE || Cond == ISD::SETUGE) { 1651 if (C1 == MinVal) return DAG.getConstant(1, VT); // X >= MIN --> true 1652 // X >= C0 --> X > (C0-1) 1653 return DAG.getSetCC(dl, VT, N0, 1654 DAG.getConstant(C1-1, N1.getValueType()), 1655 (Cond == ISD::SETGE) ? ISD::SETGT : ISD::SETUGT); 1656 } 1657 1658 if (Cond == ISD::SETLE || Cond == ISD::SETULE) { 1659 if (C1 == MaxVal) return DAG.getConstant(1, VT); // X <= MAX --> true 1660 // X <= C0 --> X < (C0+1) 1661 return DAG.getSetCC(dl, VT, N0, 1662 DAG.getConstant(C1+1, N1.getValueType()), 1663 (Cond == ISD::SETLE) ? ISD::SETLT : ISD::SETULT); 1664 } 1665 1666 if ((Cond == ISD::SETLT || Cond == ISD::SETULT) && C1 == MinVal) 1667 return DAG.getConstant(0, VT); // X < MIN --> false 1668 if ((Cond == ISD::SETGE || Cond == ISD::SETUGE) && C1 == MinVal) 1669 return DAG.getConstant(1, VT); // X >= MIN --> true 1670 if ((Cond == ISD::SETGT || Cond == ISD::SETUGT) && C1 == MaxVal) 1671 return DAG.getConstant(0, VT); // X > MAX --> false 1672 if ((Cond == ISD::SETLE || Cond == ISD::SETULE) && C1 == MaxVal) 1673 return DAG.getConstant(1, VT); // X <= MAX --> true 1674 1675 // Canonicalize setgt X, Min --> setne X, Min 1676 if ((Cond == ISD::SETGT || Cond == ISD::SETUGT) && C1 == MinVal) 1677 return DAG.getSetCC(dl, VT, N0, N1, ISD::SETNE); 1678 // Canonicalize setlt X, Max --> setne X, Max 1679 if ((Cond == ISD::SETLT || Cond == ISD::SETULT) && C1 == MaxVal) 1680 return DAG.getSetCC(dl, VT, N0, N1, ISD::SETNE); 1681 1682 // If we have setult X, 1, turn it into seteq X, 0 1683 if ((Cond == ISD::SETLT || Cond == ISD::SETULT) && C1 == MinVal+1) 1684 return DAG.getSetCC(dl, VT, N0, 1685 DAG.getConstant(MinVal, N0.getValueType()), 1686 ISD::SETEQ); 1687 // If we have setugt X, Max-1, turn it into seteq X, Max 1688 else if ((Cond == ISD::SETGT || Cond == ISD::SETUGT) && C1 == MaxVal-1) 1689 return DAG.getSetCC(dl, VT, N0, 1690 DAG.getConstant(MaxVal, N0.getValueType()), 1691 ISD::SETEQ); 1692 1693 // If we have "setcc X, C0", check to see if we can shrink the immediate 1694 // by changing cc. 1695 1696 // SETUGT X, SINTMAX -> SETLT X, 0 1697 if (Cond == ISD::SETUGT && 1698 C1 == APInt::getSignedMaxValue(OperandBitSize)) 1699 return DAG.getSetCC(dl, VT, N0, 1700 DAG.getConstant(0, N1.getValueType()), 1701 ISD::SETLT); 1702 1703 // SETULT X, SINTMIN -> SETGT X, -1 1704 if (Cond == ISD::SETULT && 1705 C1 == APInt::getSignedMinValue(OperandBitSize)) { 1706 SDValue ConstMinusOne = 1707 DAG.getConstant(APInt::getAllOnesValue(OperandBitSize), 1708 N1.getValueType()); 1709 return DAG.getSetCC(dl, VT, N0, ConstMinusOne, ISD::SETGT); 1710 } 1711 1712 // Fold bit comparisons when we can. 1713 if ((Cond == ISD::SETEQ || Cond == ISD::SETNE) && 1714 VT == N0.getValueType() && N0.getOpcode() == ISD::AND) 1715 if (ConstantSDNode *AndRHS = 1716 dyn_cast<ConstantSDNode>(N0.getOperand(1))) { 1717 MVT ShiftTy = DCI.isBeforeLegalize() ? 1718 getPointerTy() : getShiftAmountTy(); 1719 if (Cond == ISD::SETNE && C1 == 0) {// (X & 8) != 0 --> (X & 8) >> 3 1720 // Perform the xform if the AND RHS is a single bit. 1721 if (isPowerOf2_64(AndRHS->getZExtValue())) { 1722 return DAG.getNode(ISD::SRL, dl, VT, N0, 1723 DAG.getConstant(Log2_64(AndRHS->getZExtValue()), 1724 ShiftTy)); 1725 } 1726 } else if (Cond == ISD::SETEQ && C1 == AndRHS->getZExtValue()) { 1727 // (X & 8) == 8 --> (X & 8) >> 3 1728 // Perform the xform if C1 is a single bit. 1729 if (C1.isPowerOf2()) { 1730 return DAG.getNode(ISD::SRL, dl, VT, N0, 1731 DAG.getConstant(C1.logBase2(), ShiftTy)); 1732 } 1733 } 1734 } 1735 } 1736 } else if (isa<ConstantSDNode>(N0.getNode())) { 1737 // Ensure that the constant occurs on the RHS. 1738 return DAG.getSetCC(dl, VT, N1, N0, ISD::getSetCCSwappedOperands(Cond)); 1739 } 1740 1741 if (isa<ConstantFPSDNode>(N0.getNode())) { 1742 // Constant fold or commute setcc. 1743 SDValue O = DAG.FoldSetCC(VT, N0, N1, Cond, dl); 1744 if (O.getNode()) return O; 1745 } else if (ConstantFPSDNode *CFP = dyn_cast<ConstantFPSDNode>(N1.getNode())) { 1746 // If the RHS of an FP comparison is a constant, simplify it away in 1747 // some cases. 1748 if (CFP->getValueAPF().isNaN()) { 1749 // If an operand is known to be a nan, we can fold it. 1750 switch (ISD::getUnorderedFlavor(Cond)) { 1751 default: assert(0 && "Unknown flavor!"); 1752 case 0: // Known false. 1753 return DAG.getConstant(0, VT); 1754 case 1: // Known true. 1755 return DAG.getConstant(1, VT); 1756 case 2: // Undefined. 1757 return DAG.getUNDEF(VT); 1758 } 1759 } 1760 1761 // Otherwise, we know the RHS is not a NaN. Simplify the node to drop the 1762 // constant if knowing that the operand is non-nan is enough. We prefer to 1763 // have SETO(x,x) instead of SETO(x, 0.0) because this avoids having to 1764 // materialize 0.0. 1765 if (Cond == ISD::SETO || Cond == ISD::SETUO) 1766 return DAG.getSetCC(dl, VT, N0, N0, Cond); 1767 } 1768 1769 if (N0 == N1) { 1770 // We can always fold X == X for integer setcc's. 1771 if (N0.getValueType().isInteger()) 1772 return DAG.getConstant(ISD::isTrueWhenEqual(Cond), VT); 1773 unsigned UOF = ISD::getUnorderedFlavor(Cond); 1774 if (UOF == 2) // FP operators that are undefined on NaNs. 1775 return DAG.getConstant(ISD::isTrueWhenEqual(Cond), VT); 1776 if (UOF == unsigned(ISD::isTrueWhenEqual(Cond))) 1777 return DAG.getConstant(UOF, VT); 1778 // Otherwise, we can't fold it. However, we can simplify it to SETUO/SETO 1779 // if it is not already. 1780 ISD::CondCode NewCond = UOF == 0 ? ISD::SETO : ISD::SETUO; 1781 if (NewCond != Cond) 1782 return DAG.getSetCC(dl, VT, N0, N1, NewCond); 1783 } 1784 1785 if ((Cond == ISD::SETEQ || Cond == ISD::SETNE) && 1786 N0.getValueType().isInteger()) { 1787 if (N0.getOpcode() == ISD::ADD || N0.getOpcode() == ISD::SUB || 1788 N0.getOpcode() == ISD::XOR) { 1789 // Simplify (X+Y) == (X+Z) --> Y == Z 1790 if (N0.getOpcode() == N1.getOpcode()) { 1791 if (N0.getOperand(0) == N1.getOperand(0)) 1792 return DAG.getSetCC(dl, VT, N0.getOperand(1), N1.getOperand(1), Cond); 1793 if (N0.getOperand(1) == N1.getOperand(1)) 1794 return DAG.getSetCC(dl, VT, N0.getOperand(0), N1.getOperand(0), Cond); 1795 if (DAG.isCommutativeBinOp(N0.getOpcode())) { 1796 // If X op Y == Y op X, try other combinations. 1797 if (N0.getOperand(0) == N1.getOperand(1)) 1798 return DAG.getSetCC(dl, VT, N0.getOperand(1), N1.getOperand(0), 1799 Cond); 1800 if (N0.getOperand(1) == N1.getOperand(0)) 1801 return DAG.getSetCC(dl, VT, N0.getOperand(0), N1.getOperand(1), 1802 Cond); 1803 } 1804 } 1805 1806 if (ConstantSDNode *RHSC = dyn_cast<ConstantSDNode>(N1)) { 1807 if (ConstantSDNode *LHSR = dyn_cast<ConstantSDNode>(N0.getOperand(1))) { 1808 // Turn (X+C1) == C2 --> X == C2-C1 1809 if (N0.getOpcode() == ISD::ADD && N0.getNode()->hasOneUse()) { 1810 return DAG.getSetCC(dl, VT, N0.getOperand(0), 1811 DAG.getConstant(RHSC->getAPIntValue()- 1812 LHSR->getAPIntValue(), 1813 N0.getValueType()), Cond); 1814 } 1815 1816 // Turn (X^C1) == C2 into X == C1^C2 iff X&~C1 = 0. 1817 if (N0.getOpcode() == ISD::XOR) 1818 // If we know that all of the inverted bits are zero, don't bother 1819 // performing the inversion. 1820 if (DAG.MaskedValueIsZero(N0.getOperand(0), ~LHSR->getAPIntValue())) 1821 return 1822 DAG.getSetCC(dl, VT, N0.getOperand(0), 1823 DAG.getConstant(LHSR->getAPIntValue() ^ 1824 RHSC->getAPIntValue(), 1825 N0.getValueType()), 1826 Cond); 1827 } 1828 1829 // Turn (C1-X) == C2 --> X == C1-C2 1830 if (ConstantSDNode *SUBC = dyn_cast<ConstantSDNode>(N0.getOperand(0))) { 1831 if (N0.getOpcode() == ISD::SUB && N0.getNode()->hasOneUse()) { 1832 return 1833 DAG.getSetCC(dl, VT, N0.getOperand(1), 1834 DAG.getConstant(SUBC->getAPIntValue() - 1835 RHSC->getAPIntValue(), 1836 N0.getValueType()), 1837 Cond); 1838 } 1839 } 1840 } 1841 1842 // Simplify (X+Z) == X --> Z == 0 1843 if (N0.getOperand(0) == N1) 1844 return DAG.getSetCC(dl, VT, N0.getOperand(1), 1845 DAG.getConstant(0, N0.getValueType()), Cond); 1846 if (N0.getOperand(1) == N1) { 1847 if (DAG.isCommutativeBinOp(N0.getOpcode())) 1848 return DAG.getSetCC(dl, VT, N0.getOperand(0), 1849 DAG.getConstant(0, N0.getValueType()), Cond); 1850 else if (N0.getNode()->hasOneUse()) { 1851 assert(N0.getOpcode() == ISD::SUB && "Unexpected operation!"); 1852 // (Z-X) == X --> Z == X<<1 1853 SDValue SH = DAG.getNode(ISD::SHL, dl, N1.getValueType(), 1854 N1, 1855 DAG.getConstant(1, getShiftAmountTy())); 1856 if (!DCI.isCalledByLegalizer()) 1857 DCI.AddToWorklist(SH.getNode()); 1858 return DAG.getSetCC(dl, VT, N0.getOperand(0), SH, Cond); 1859 } 1860 } 1861 } 1862 1863 if (N1.getOpcode() == ISD::ADD || N1.getOpcode() == ISD::SUB || 1864 N1.getOpcode() == ISD::XOR) { 1865 // Simplify X == (X+Z) --> Z == 0 1866 if (N1.getOperand(0) == N0) { 1867 return DAG.getSetCC(dl, VT, N1.getOperand(1), 1868 DAG.getConstant(0, N1.getValueType()), Cond); 1869 } else if (N1.getOperand(1) == N0) { 1870 if (DAG.isCommutativeBinOp(N1.getOpcode())) { 1871 return DAG.getSetCC(dl, VT, N1.getOperand(0), 1872 DAG.getConstant(0, N1.getValueType()), Cond); 1873 } else if (N1.getNode()->hasOneUse()) { 1874 assert(N1.getOpcode() == ISD::SUB && "Unexpected operation!"); 1875 // X == (Z-X) --> X<<1 == Z 1876 SDValue SH = DAG.getNode(ISD::SHL, dl, N1.getValueType(), N0, 1877 DAG.getConstant(1, getShiftAmountTy())); 1878 if (!DCI.isCalledByLegalizer()) 1879 DCI.AddToWorklist(SH.getNode()); 1880 return DAG.getSetCC(dl, VT, SH, N1.getOperand(0), Cond); 1881 } 1882 } 1883 } 1884 1885 // Simplify x&y == y to x&y != 0 if y has exactly one bit set. 1886 // Note that where y is variable and is known to have at most 1887 // one bit set (for example, if it is z&1) we cannot do this; 1888 // the expressions are not equivalent when y==0. 1889 if (N0.getOpcode() == ISD::AND) 1890 if (N0.getOperand(0) == N1 || N0.getOperand(1) == N1) { 1891 if (ValueHasExactlyOneBitSet(N1, DAG)) { 1892 Cond = ISD::getSetCCInverse(Cond, /*isInteger=*/true); 1893 SDValue Zero = DAG.getConstant(0, N1.getValueType()); 1894 return DAG.getSetCC(dl, VT, N0, Zero, Cond); 1895 } 1896 } 1897 if (N1.getOpcode() == ISD::AND) 1898 if (N1.getOperand(0) == N0 || N1.getOperand(1) == N0) { 1899 if (ValueHasExactlyOneBitSet(N0, DAG)) { 1900 Cond = ISD::getSetCCInverse(Cond, /*isInteger=*/true); 1901 SDValue Zero = DAG.getConstant(0, N0.getValueType()); 1902 return DAG.getSetCC(dl, VT, N1, Zero, Cond); 1903 } 1904 } 1905 } 1906 1907 // Fold away ALL boolean setcc's. 1908 SDValue Temp; 1909 if (N0.getValueType() == MVT::i1 && foldBooleans) { 1910 switch (Cond) { 1911 default: assert(0 && "Unknown integer setcc!"); 1912 case ISD::SETEQ: // X == Y -> ~(X^Y) 1913 Temp = DAG.getNode(ISD::XOR, dl, MVT::i1, N0, N1); 1914 N0 = DAG.getNOT(dl, Temp, MVT::i1); 1915 if (!DCI.isCalledByLegalizer()) 1916 DCI.AddToWorklist(Temp.getNode()); 1917 break; 1918 case ISD::SETNE: // X != Y --> (X^Y) 1919 N0 = DAG.getNode(ISD::XOR, dl, MVT::i1, N0, N1); 1920 break; 1921 case ISD::SETGT: // X >s Y --> X == 0 & Y == 1 --> ~X & Y 1922 case ISD::SETULT: // X <u Y --> X == 0 & Y == 1 --> ~X & Y 1923 Temp = DAG.getNOT(dl, N0, MVT::i1); 1924 N0 = DAG.getNode(ISD::AND, dl, MVT::i1, N1, Temp); 1925 if (!DCI.isCalledByLegalizer()) 1926 DCI.AddToWorklist(Temp.getNode()); 1927 break; 1928 case ISD::SETLT: // X <s Y --> X == 1 & Y == 0 --> ~Y & X 1929 case ISD::SETUGT: // X >u Y --> X == 1 & Y == 0 --> ~Y & X 1930 Temp = DAG.getNOT(dl, N1, MVT::i1); 1931 N0 = DAG.getNode(ISD::AND, dl, MVT::i1, N0, Temp); 1932 if (!DCI.isCalledByLegalizer()) 1933 DCI.AddToWorklist(Temp.getNode()); 1934 break; 1935 case ISD::SETULE: // X <=u Y --> X == 0 | Y == 1 --> ~X | Y 1936 case ISD::SETGE: // X >=s Y --> X == 0 | Y == 1 --> ~X | Y 1937 Temp = DAG.getNOT(dl, N0, MVT::i1); 1938 N0 = DAG.getNode(ISD::OR, dl, MVT::i1, N1, Temp); 1939 if (!DCI.isCalledByLegalizer()) 1940 DCI.AddToWorklist(Temp.getNode()); 1941 break; 1942 case ISD::SETUGE: // X >=u Y --> X == 1 | Y == 0 --> ~Y | X 1943 case ISD::SETLE: // X <=s Y --> X == 1 | Y == 0 --> ~Y | X 1944 Temp = DAG.getNOT(dl, N1, MVT::i1); 1945 N0 = DAG.getNode(ISD::OR, dl, MVT::i1, N0, Temp); 1946 break; 1947 } 1948 if (VT != MVT::i1) { 1949 if (!DCI.isCalledByLegalizer()) 1950 DCI.AddToWorklist(N0.getNode()); 1951 // FIXME: If running after legalize, we probably can't do this. 1952 N0 = DAG.getNode(ISD::ZERO_EXTEND, dl, VT, N0); 1953 } 1954 return N0; 1955 } 1956 1957 // Could not fold it. 1958 return SDValue(); 1959} 1960 1961/// isGAPlusOffset - Returns true (and the GlobalValue and the offset) if the 1962/// node is a GlobalAddress + offset. 1963bool TargetLowering::isGAPlusOffset(SDNode *N, GlobalValue* &GA, 1964 int64_t &Offset) const { 1965 if (isa<GlobalAddressSDNode>(N)) { 1966 GlobalAddressSDNode *GASD = cast<GlobalAddressSDNode>(N); 1967 GA = GASD->getGlobal(); 1968 Offset += GASD->getOffset(); 1969 return true; 1970 } 1971 1972 if (N->getOpcode() == ISD::ADD) { 1973 SDValue N1 = N->getOperand(0); 1974 SDValue N2 = N->getOperand(1); 1975 if (isGAPlusOffset(N1.getNode(), GA, Offset)) { 1976 ConstantSDNode *V = dyn_cast<ConstantSDNode>(N2); 1977 if (V) { 1978 Offset += V->getSExtValue(); 1979 return true; 1980 } 1981 } else if (isGAPlusOffset(N2.getNode(), GA, Offset)) { 1982 ConstantSDNode *V = dyn_cast<ConstantSDNode>(N1); 1983 if (V) { 1984 Offset += V->getSExtValue(); 1985 return true; 1986 } 1987 } 1988 } 1989 return false; 1990} 1991 1992 1993/// isConsecutiveLoad - Return true if LD (which must be a LoadSDNode) is 1994/// loading 'Bytes' bytes from a location that is 'Dist' units away from the 1995/// location that the 'Base' load is loading from. 1996bool TargetLowering::isConsecutiveLoad(SDNode *LD, SDNode *Base, 1997 unsigned Bytes, int Dist, 1998 const MachineFrameInfo *MFI) const { 1999 if (LD->getOperand(0).getNode() != Base->getOperand(0).getNode()) 2000 return false; 2001 MVT VT = LD->getValueType(0); 2002 if (VT.getSizeInBits() / 8 != Bytes) 2003 return false; 2004 2005 SDValue Loc = LD->getOperand(1); 2006 SDValue BaseLoc = Base->getOperand(1); 2007 if (Loc.getOpcode() == ISD::FrameIndex) { 2008 if (BaseLoc.getOpcode() != ISD::FrameIndex) 2009 return false; 2010 int FI = cast<FrameIndexSDNode>(Loc)->getIndex(); 2011 int BFI = cast<FrameIndexSDNode>(BaseLoc)->getIndex(); 2012 int FS = MFI->getObjectSize(FI); 2013 int BFS = MFI->getObjectSize(BFI); 2014 if (FS != BFS || FS != (int)Bytes) return false; 2015 return MFI->getObjectOffset(FI) == (MFI->getObjectOffset(BFI) + Dist*Bytes); 2016 } 2017 2018 GlobalValue *GV1 = NULL; 2019 GlobalValue *GV2 = NULL; 2020 int64_t Offset1 = 0; 2021 int64_t Offset2 = 0; 2022 bool isGA1 = isGAPlusOffset(Loc.getNode(), GV1, Offset1); 2023 bool isGA2 = isGAPlusOffset(BaseLoc.getNode(), GV2, Offset2); 2024 if (isGA1 && isGA2 && GV1 == GV2) 2025 return Offset1 == (Offset2 + Dist*Bytes); 2026 return false; 2027} 2028 2029 2030SDValue TargetLowering:: 2031PerformDAGCombine(SDNode *N, DAGCombinerInfo &DCI) const { 2032 // Default implementation: no optimization. 2033 return SDValue(); 2034} 2035 2036//===----------------------------------------------------------------------===// 2037// Inline Assembler Implementation Methods 2038//===----------------------------------------------------------------------===// 2039 2040 2041TargetLowering::ConstraintType 2042TargetLowering::getConstraintType(const std::string &Constraint) const { 2043 // FIXME: lots more standard ones to handle. 2044 if (Constraint.size() == 1) { 2045 switch (Constraint[0]) { 2046 default: break; 2047 case 'r': return C_RegisterClass; 2048 case 'm': // memory 2049 case 'o': // offsetable 2050 case 'V': // not offsetable 2051 return C_Memory; 2052 case 'i': // Simple Integer or Relocatable Constant 2053 case 'n': // Simple Integer 2054 case 's': // Relocatable Constant 2055 case 'X': // Allow ANY value. 2056 case 'I': // Target registers. 2057 case 'J': 2058 case 'K': 2059 case 'L': 2060 case 'M': 2061 case 'N': 2062 case 'O': 2063 case 'P': 2064 return C_Other; 2065 } 2066 } 2067 2068 if (Constraint.size() > 1 && Constraint[0] == '{' && 2069 Constraint[Constraint.size()-1] == '}') 2070 return C_Register; 2071 return C_Unknown; 2072} 2073 2074/// LowerXConstraint - try to replace an X constraint, which matches anything, 2075/// with another that has more specific requirements based on the type of the 2076/// corresponding operand. 2077const char *TargetLowering::LowerXConstraint(MVT ConstraintVT) const{ 2078 if (ConstraintVT.isInteger()) 2079 return "r"; 2080 if (ConstraintVT.isFloatingPoint()) 2081 return "f"; // works for many targets 2082 return 0; 2083} 2084 2085/// LowerAsmOperandForConstraint - Lower the specified operand into the Ops 2086/// vector. If it is invalid, don't add anything to Ops. 2087void TargetLowering::LowerAsmOperandForConstraint(SDValue Op, 2088 char ConstraintLetter, 2089 bool hasMemory, 2090 std::vector<SDValue> &Ops, 2091 SelectionDAG &DAG) const { 2092 switch (ConstraintLetter) { 2093 default: break; 2094 case 'X': // Allows any operand; labels (basic block) use this. 2095 if (Op.getOpcode() == ISD::BasicBlock) { 2096 Ops.push_back(Op); 2097 return; 2098 } 2099 // fall through 2100 case 'i': // Simple Integer or Relocatable Constant 2101 case 'n': // Simple Integer 2102 case 's': { // Relocatable Constant 2103 // These operands are interested in values of the form (GV+C), where C may 2104 // be folded in as an offset of GV, or it may be explicitly added. Also, it 2105 // is possible and fine if either GV or C are missing. 2106 ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op); 2107 GlobalAddressSDNode *GA = dyn_cast<GlobalAddressSDNode>(Op); 2108 2109 // If we have "(add GV, C)", pull out GV/C 2110 if (Op.getOpcode() == ISD::ADD) { 2111 C = dyn_cast<ConstantSDNode>(Op.getOperand(1)); 2112 GA = dyn_cast<GlobalAddressSDNode>(Op.getOperand(0)); 2113 if (C == 0 || GA == 0) { 2114 C = dyn_cast<ConstantSDNode>(Op.getOperand(0)); 2115 GA = dyn_cast<GlobalAddressSDNode>(Op.getOperand(1)); 2116 } 2117 if (C == 0 || GA == 0) 2118 C = 0, GA = 0; 2119 } 2120 2121 // If we find a valid operand, map to the TargetXXX version so that the 2122 // value itself doesn't get selected. 2123 if (GA) { // Either &GV or &GV+C 2124 if (ConstraintLetter != 'n') { 2125 int64_t Offs = GA->getOffset(); 2126 if (C) Offs += C->getZExtValue(); 2127 Ops.push_back(DAG.getTargetGlobalAddress(GA->getGlobal(), 2128 Op.getValueType(), Offs)); 2129 return; 2130 } 2131 } 2132 if (C) { // just C, no GV. 2133 // Simple constants are not allowed for 's'. 2134 if (ConstraintLetter != 's') { 2135 // gcc prints these as sign extended. Sign extend value to 64 bits 2136 // now; without this it would get ZExt'd later in 2137 // ScheduleDAGSDNodes::EmitNode, which is very generic. 2138 Ops.push_back(DAG.getTargetConstant(C->getAPIntValue().getSExtValue(), 2139 MVT::i64)); 2140 return; 2141 } 2142 } 2143 break; 2144 } 2145 } 2146} 2147 2148std::vector<unsigned> TargetLowering:: 2149getRegClassForInlineAsmConstraint(const std::string &Constraint, 2150 MVT VT) const { 2151 return std::vector<unsigned>(); 2152} 2153 2154 2155std::pair<unsigned, const TargetRegisterClass*> TargetLowering:: 2156getRegForInlineAsmConstraint(const std::string &Constraint, 2157 MVT VT) const { 2158 if (Constraint[0] != '{') 2159 return std::pair<unsigned, const TargetRegisterClass*>(0, 0); 2160 assert(*(Constraint.end()-1) == '}' && "Not a brace enclosed constraint?"); 2161 2162 // Remove the braces from around the name. 2163 std::string RegName(Constraint.begin()+1, Constraint.end()-1); 2164 2165 // Figure out which register class contains this reg. 2166 const TargetRegisterInfo *RI = TM.getRegisterInfo(); 2167 for (TargetRegisterInfo::regclass_iterator RCI = RI->regclass_begin(), 2168 E = RI->regclass_end(); RCI != E; ++RCI) { 2169 const TargetRegisterClass *RC = *RCI; 2170 2171 // If none of the the value types for this register class are valid, we 2172 // can't use it. For example, 64-bit reg classes on 32-bit targets. 2173 bool isLegal = false; 2174 for (TargetRegisterClass::vt_iterator I = RC->vt_begin(), E = RC->vt_end(); 2175 I != E; ++I) { 2176 if (isTypeLegal(*I)) { 2177 isLegal = true; 2178 break; 2179 } 2180 } 2181 2182 if (!isLegal) continue; 2183 2184 for (TargetRegisterClass::iterator I = RC->begin(), E = RC->end(); 2185 I != E; ++I) { 2186 if (StringsEqualNoCase(RegName, RI->get(*I).AsmName)) 2187 return std::make_pair(*I, RC); 2188 } 2189 } 2190 2191 return std::pair<unsigned, const TargetRegisterClass*>(0, 0); 2192} 2193 2194//===----------------------------------------------------------------------===// 2195// Constraint Selection. 2196 2197/// isMatchingInputConstraint - Return true of this is an input operand that is 2198/// a matching constraint like "4". 2199bool TargetLowering::AsmOperandInfo::isMatchingInputConstraint() const { 2200 assert(!ConstraintCode.empty() && "No known constraint!"); 2201 return isdigit(ConstraintCode[0]); 2202} 2203 2204/// getMatchedOperand - If this is an input matching constraint, this method 2205/// returns the output operand it matches. 2206unsigned TargetLowering::AsmOperandInfo::getMatchedOperand() const { 2207 assert(!ConstraintCode.empty() && "No known constraint!"); 2208 return atoi(ConstraintCode.c_str()); 2209} 2210 2211 2212/// getConstraintGenerality - Return an integer indicating how general CT 2213/// is. 2214static unsigned getConstraintGenerality(TargetLowering::ConstraintType CT) { 2215 switch (CT) { 2216 default: assert(0 && "Unknown constraint type!"); 2217 case TargetLowering::C_Other: 2218 case TargetLowering::C_Unknown: 2219 return 0; 2220 case TargetLowering::C_Register: 2221 return 1; 2222 case TargetLowering::C_RegisterClass: 2223 return 2; 2224 case TargetLowering::C_Memory: 2225 return 3; 2226 } 2227} 2228 2229/// ChooseConstraint - If there are multiple different constraints that we 2230/// could pick for this operand (e.g. "imr") try to pick the 'best' one. 2231/// This is somewhat tricky: constraints fall into four classes: 2232/// Other -> immediates and magic values 2233/// Register -> one specific register 2234/// RegisterClass -> a group of regs 2235/// Memory -> memory 2236/// Ideally, we would pick the most specific constraint possible: if we have 2237/// something that fits into a register, we would pick it. The problem here 2238/// is that if we have something that could either be in a register or in 2239/// memory that use of the register could cause selection of *other* 2240/// operands to fail: they might only succeed if we pick memory. Because of 2241/// this the heuristic we use is: 2242/// 2243/// 1) If there is an 'other' constraint, and if the operand is valid for 2244/// that constraint, use it. This makes us take advantage of 'i' 2245/// constraints when available. 2246/// 2) Otherwise, pick the most general constraint present. This prefers 2247/// 'm' over 'r', for example. 2248/// 2249static void ChooseConstraint(TargetLowering::AsmOperandInfo &OpInfo, 2250 bool hasMemory, const TargetLowering &TLI, 2251 SDValue Op, SelectionDAG *DAG) { 2252 assert(OpInfo.Codes.size() > 1 && "Doesn't have multiple constraint options"); 2253 unsigned BestIdx = 0; 2254 TargetLowering::ConstraintType BestType = TargetLowering::C_Unknown; 2255 int BestGenerality = -1; 2256 2257 // Loop over the options, keeping track of the most general one. 2258 for (unsigned i = 0, e = OpInfo.Codes.size(); i != e; ++i) { 2259 TargetLowering::ConstraintType CType = 2260 TLI.getConstraintType(OpInfo.Codes[i]); 2261 2262 // If this is an 'other' constraint, see if the operand is valid for it. 2263 // For example, on X86 we might have an 'rI' constraint. If the operand 2264 // is an integer in the range [0..31] we want to use I (saving a load 2265 // of a register), otherwise we must use 'r'. 2266 if (CType == TargetLowering::C_Other && Op.getNode()) { 2267 assert(OpInfo.Codes[i].size() == 1 && 2268 "Unhandled multi-letter 'other' constraint"); 2269 std::vector<SDValue> ResultOps; 2270 TLI.LowerAsmOperandForConstraint(Op, OpInfo.Codes[i][0], hasMemory, 2271 ResultOps, *DAG); 2272 if (!ResultOps.empty()) { 2273 BestType = CType; 2274 BestIdx = i; 2275 break; 2276 } 2277 } 2278 2279 // This constraint letter is more general than the previous one, use it. 2280 int Generality = getConstraintGenerality(CType); 2281 if (Generality > BestGenerality) { 2282 BestType = CType; 2283 BestIdx = i; 2284 BestGenerality = Generality; 2285 } 2286 } 2287 2288 OpInfo.ConstraintCode = OpInfo.Codes[BestIdx]; 2289 OpInfo.ConstraintType = BestType; 2290} 2291 2292/// ComputeConstraintToUse - Determines the constraint code and constraint 2293/// type to use for the specific AsmOperandInfo, setting 2294/// OpInfo.ConstraintCode and OpInfo.ConstraintType. 2295void TargetLowering::ComputeConstraintToUse(AsmOperandInfo &OpInfo, 2296 SDValue Op, 2297 bool hasMemory, 2298 SelectionDAG *DAG) const { 2299 assert(!OpInfo.Codes.empty() && "Must have at least one constraint"); 2300 2301 // Single-letter constraints ('r') are very common. 2302 if (OpInfo.Codes.size() == 1) { 2303 OpInfo.ConstraintCode = OpInfo.Codes[0]; 2304 OpInfo.ConstraintType = getConstraintType(OpInfo.ConstraintCode); 2305 } else { 2306 ChooseConstraint(OpInfo, hasMemory, *this, Op, DAG); 2307 } 2308 2309 // 'X' matches anything. 2310 if (OpInfo.ConstraintCode == "X" && OpInfo.CallOperandVal) { 2311 // Labels and constants are handled elsewhere ('X' is the only thing 2312 // that matches labels). 2313 if (isa<BasicBlock>(OpInfo.CallOperandVal) || 2314 isa<ConstantInt>(OpInfo.CallOperandVal)) 2315 return; 2316 2317 // Otherwise, try to resolve it to something we know about by looking at 2318 // the actual operand type. 2319 if (const char *Repl = LowerXConstraint(OpInfo.ConstraintVT)) { 2320 OpInfo.ConstraintCode = Repl; 2321 OpInfo.ConstraintType = getConstraintType(OpInfo.ConstraintCode); 2322 } 2323 } 2324} 2325 2326//===----------------------------------------------------------------------===// 2327// Loop Strength Reduction hooks 2328//===----------------------------------------------------------------------===// 2329 2330/// isLegalAddressingMode - Return true if the addressing mode represented 2331/// by AM is legal for this target, for a load/store of the specified type. 2332bool TargetLowering::isLegalAddressingMode(const AddrMode &AM, 2333 const Type *Ty) const { 2334 // The default implementation of this implements a conservative RISCy, r+r and 2335 // r+i addr mode. 2336 2337 // Allows a sign-extended 16-bit immediate field. 2338 if (AM.BaseOffs <= -(1LL << 16) || AM.BaseOffs >= (1LL << 16)-1) 2339 return false; 2340 2341 // No global is ever allowed as a base. 2342 if (AM.BaseGV) 2343 return false; 2344 2345 // Only support r+r, 2346 switch (AM.Scale) { 2347 case 0: // "r+i" or just "i", depending on HasBaseReg. 2348 break; 2349 case 1: 2350 if (AM.HasBaseReg && AM.BaseOffs) // "r+r+i" is not allowed. 2351 return false; 2352 // Otherwise we have r+r or r+i. 2353 break; 2354 case 2: 2355 if (AM.HasBaseReg || AM.BaseOffs) // 2*r+r or 2*r+i is not allowed. 2356 return false; 2357 // Allow 2*r as r+r. 2358 break; 2359 } 2360 2361 return true; 2362} 2363 2364struct mu { 2365 APInt m; // magic number 2366 bool a; // add indicator 2367 unsigned s; // shift amount 2368}; 2369 2370/// magicu - calculate the magic numbers required to codegen an integer udiv as 2371/// a sequence of multiply, add and shifts. Requires that the divisor not be 0. 2372static mu magicu(const APInt& d) { 2373 unsigned p; 2374 APInt nc, delta, q1, r1, q2, r2; 2375 struct mu magu; 2376 magu.a = 0; // initialize "add" indicator 2377 APInt allOnes = APInt::getAllOnesValue(d.getBitWidth()); 2378 APInt signedMin = APInt::getSignedMinValue(d.getBitWidth()); 2379 APInt signedMax = APInt::getSignedMaxValue(d.getBitWidth()); 2380 2381 nc = allOnes - (-d).urem(d); 2382 p = d.getBitWidth() - 1; // initialize p 2383 q1 = signedMin.udiv(nc); // initialize q1 = 2p/nc 2384 r1 = signedMin - q1*nc; // initialize r1 = rem(2p,nc) 2385 q2 = signedMax.udiv(d); // initialize q2 = (2p-1)/d 2386 r2 = signedMax - q2*d; // initialize r2 = rem((2p-1),d) 2387 do { 2388 p = p + 1; 2389 if (r1.uge(nc - r1)) { 2390 q1 = q1 + q1 + 1; // update q1 2391 r1 = r1 + r1 - nc; // update r1 2392 } 2393 else { 2394 q1 = q1+q1; // update q1 2395 r1 = r1+r1; // update r1 2396 } 2397 if ((r2 + 1).uge(d - r2)) { 2398 if (q2.uge(signedMax)) magu.a = 1; 2399 q2 = q2+q2 + 1; // update q2 2400 r2 = r2+r2 + 1 - d; // update r2 2401 } 2402 else { 2403 if (q2.uge(signedMin)) magu.a = 1; 2404 q2 = q2+q2; // update q2 2405 r2 = r2+r2 + 1; // update r2 2406 } 2407 delta = d - 1 - r2; 2408 } while (p < d.getBitWidth()*2 && 2409 (q1.ult(delta) || (q1 == delta && r1 == 0))); 2410 magu.m = q2 + 1; // resulting magic number 2411 magu.s = p - d.getBitWidth(); // resulting shift 2412 return magu; 2413} 2414 2415// Magic for divide replacement 2416struct ms { 2417 APInt m; // magic number 2418 unsigned s; // shift amount 2419}; 2420 2421/// magic - calculate the magic numbers required to codegen an integer sdiv as 2422/// a sequence of multiply and shifts. Requires that the divisor not be 0, 1, 2423/// or -1. 2424static ms magic(const APInt& d) { 2425 unsigned p; 2426 APInt ad, anc, delta, q1, r1, q2, r2, t; 2427 APInt allOnes = APInt::getAllOnesValue(d.getBitWidth()); 2428 APInt signedMin = APInt::getSignedMinValue(d.getBitWidth()); 2429 APInt signedMax = APInt::getSignedMaxValue(d.getBitWidth()); 2430 struct ms mag; 2431 2432 ad = d.abs(); 2433 t = signedMin + (d.lshr(d.getBitWidth() - 1)); 2434 anc = t - 1 - t.urem(ad); // absolute value of nc 2435 p = d.getBitWidth() - 1; // initialize p 2436 q1 = signedMin.udiv(anc); // initialize q1 = 2p/abs(nc) 2437 r1 = signedMin - q1*anc; // initialize r1 = rem(2p,abs(nc)) 2438 q2 = signedMin.udiv(ad); // initialize q2 = 2p/abs(d) 2439 r2 = signedMin - q2*ad; // initialize r2 = rem(2p,abs(d)) 2440 do { 2441 p = p + 1; 2442 q1 = q1<<1; // update q1 = 2p/abs(nc) 2443 r1 = r1<<1; // update r1 = rem(2p/abs(nc)) 2444 if (r1.uge(anc)) { // must be unsigned comparison 2445 q1 = q1 + 1; 2446 r1 = r1 - anc; 2447 } 2448 q2 = q2<<1; // update q2 = 2p/abs(d) 2449 r2 = r2<<1; // update r2 = rem(2p/abs(d)) 2450 if (r2.uge(ad)) { // must be unsigned comparison 2451 q2 = q2 + 1; 2452 r2 = r2 - ad; 2453 } 2454 delta = ad - r2; 2455 } while (q1.ule(delta) || (q1 == delta && r1 == 0)); 2456 2457 mag.m = q2 + 1; 2458 if (d.isNegative()) mag.m = -mag.m; // resulting magic number 2459 mag.s = p - d.getBitWidth(); // resulting shift 2460 return mag; 2461} 2462 2463/// BuildSDIVSequence - Given an ISD::SDIV node expressing a divide by constant, 2464/// return a DAG expression to select that will generate the same value by 2465/// multiplying by a magic number. See: 2466/// <http://the.wall.riscom.net/books/proc/ppc/cwg/code2.html> 2467SDValue TargetLowering::BuildSDIV(SDNode *N, SelectionDAG &DAG, 2468 std::vector<SDNode*>* Created) const { 2469 MVT VT = N->getValueType(0); 2470 DebugLoc dl= N->getDebugLoc(); 2471 2472 // Check to see if we can do this. 2473 // FIXME: We should be more aggressive here. 2474 if (!isTypeLegal(VT)) 2475 return SDValue(); 2476 2477 APInt d = cast<ConstantSDNode>(N->getOperand(1))->getAPIntValue(); 2478 ms magics = magic(d); 2479 2480 // Multiply the numerator (operand 0) by the magic value 2481 // FIXME: We should support doing a MUL in a wider type 2482 SDValue Q; 2483 if (isOperationLegalOrCustom(ISD::MULHS, VT)) 2484 Q = DAG.getNode(ISD::MULHS, dl, VT, N->getOperand(0), 2485 DAG.getConstant(magics.m, VT)); 2486 else if (isOperationLegalOrCustom(ISD::SMUL_LOHI, VT)) 2487 Q = SDValue(DAG.getNode(ISD::SMUL_LOHI, dl, DAG.getVTList(VT, VT), 2488 N->getOperand(0), 2489 DAG.getConstant(magics.m, VT)).getNode(), 1); 2490 else 2491 return SDValue(); // No mulhs or equvialent 2492 // If d > 0 and m < 0, add the numerator 2493 if (d.isStrictlyPositive() && magics.m.isNegative()) { 2494 Q = DAG.getNode(ISD::ADD, dl, VT, Q, N->getOperand(0)); 2495 if (Created) 2496 Created->push_back(Q.getNode()); 2497 } 2498 // If d < 0 and m > 0, subtract the numerator. 2499 if (d.isNegative() && magics.m.isStrictlyPositive()) { 2500 Q = DAG.getNode(ISD::SUB, dl, VT, Q, N->getOperand(0)); 2501 if (Created) 2502 Created->push_back(Q.getNode()); 2503 } 2504 // Shift right algebraic if shift value is nonzero 2505 if (magics.s > 0) { 2506 Q = DAG.getNode(ISD::SRA, dl, VT, Q, 2507 DAG.getConstant(magics.s, getShiftAmountTy())); 2508 if (Created) 2509 Created->push_back(Q.getNode()); 2510 } 2511 // Extract the sign bit and add it to the quotient 2512 SDValue T = 2513 DAG.getNode(ISD::SRL, dl, VT, Q, DAG.getConstant(VT.getSizeInBits()-1, 2514 getShiftAmountTy())); 2515 if (Created) 2516 Created->push_back(T.getNode()); 2517 return DAG.getNode(ISD::ADD, dl, VT, Q, T); 2518} 2519 2520/// BuildUDIVSequence - Given an ISD::UDIV node expressing a divide by constant, 2521/// return a DAG expression to select that will generate the same value by 2522/// multiplying by a magic number. See: 2523/// <http://the.wall.riscom.net/books/proc/ppc/cwg/code2.html> 2524SDValue TargetLowering::BuildUDIV(SDNode *N, SelectionDAG &DAG, 2525 std::vector<SDNode*>* Created) const { 2526 MVT VT = N->getValueType(0); 2527 DebugLoc dl = N->getDebugLoc(); 2528 2529 // Check to see if we can do this. 2530 // FIXME: We should be more aggressive here. 2531 if (!isTypeLegal(VT)) 2532 return SDValue(); 2533 2534 // FIXME: We should use a narrower constant when the upper 2535 // bits are known to be zero. 2536 ConstantSDNode *N1C = cast<ConstantSDNode>(N->getOperand(1)); 2537 mu magics = magicu(N1C->getAPIntValue()); 2538 2539 // Multiply the numerator (operand 0) by the magic value 2540 // FIXME: We should support doing a MUL in a wider type 2541 SDValue Q; 2542 if (isOperationLegalOrCustom(ISD::MULHU, VT)) 2543 Q = DAG.getNode(ISD::MULHU, dl, VT, N->getOperand(0), 2544 DAG.getConstant(magics.m, VT)); 2545 else if (isOperationLegalOrCustom(ISD::UMUL_LOHI, VT)) 2546 Q = SDValue(DAG.getNode(ISD::UMUL_LOHI, dl, DAG.getVTList(VT, VT), 2547 N->getOperand(0), 2548 DAG.getConstant(magics.m, VT)).getNode(), 1); 2549 else 2550 return SDValue(); // No mulhu or equvialent 2551 if (Created) 2552 Created->push_back(Q.getNode()); 2553 2554 if (magics.a == 0) { 2555 assert(magics.s < N1C->getAPIntValue().getBitWidth() && 2556 "We shouldn't generate an undefined shift!"); 2557 return DAG.getNode(ISD::SRL, dl, VT, Q, 2558 DAG.getConstant(magics.s, getShiftAmountTy())); 2559 } else { 2560 SDValue NPQ = DAG.getNode(ISD::SUB, dl, VT, N->getOperand(0), Q); 2561 if (Created) 2562 Created->push_back(NPQ.getNode()); 2563 NPQ = DAG.getNode(ISD::SRL, dl, VT, NPQ, 2564 DAG.getConstant(1, getShiftAmountTy())); 2565 if (Created) 2566 Created->push_back(NPQ.getNode()); 2567 NPQ = DAG.getNode(ISD::ADD, dl, VT, NPQ, Q); 2568 if (Created) 2569 Created->push_back(NPQ.getNode()); 2570 return DAG.getNode(ISD::SRL, dl, VT, NPQ, 2571 DAG.getConstant(magics.s-1, getShiftAmountTy())); 2572 } 2573} 2574 2575/// IgnoreHarmlessInstructions - Ignore instructions between a CALL and RET 2576/// node that don't prevent tail call optimization. 2577static SDValue IgnoreHarmlessInstructions(SDValue node) { 2578 // Found call return. 2579 if (node.getOpcode() == ISD::CALL) return node; 2580 // Ignore MERGE_VALUES. Will have at least one operand. 2581 if (node.getOpcode() == ISD::MERGE_VALUES) 2582 return IgnoreHarmlessInstructions(node.getOperand(0)); 2583 // Ignore ANY_EXTEND node. 2584 if (node.getOpcode() == ISD::ANY_EXTEND) 2585 return IgnoreHarmlessInstructions(node.getOperand(0)); 2586 if (node.getOpcode() == ISD::TRUNCATE) 2587 return IgnoreHarmlessInstructions(node.getOperand(0)); 2588 // Any other node type. 2589 return node; 2590} 2591 2592bool TargetLowering::CheckTailCallReturnConstraints(CallSDNode *TheCall, 2593 SDValue Ret) { 2594 unsigned NumOps = Ret.getNumOperands(); 2595 // ISD::CALL results:(value0, ..., valuen, chain) 2596 // ISD::RET operands:(chain, value0, flag0, ..., valuen, flagn) 2597 // Value return: 2598 // Check that operand of the RET node sources from the CALL node. The RET node 2599 // has at least two operands. Operand 0 holds the chain. Operand 1 holds the 2600 // value. 2601 if (NumOps > 1 && 2602 IgnoreHarmlessInstructions(Ret.getOperand(1)) == SDValue(TheCall,0)) 2603 return true; 2604 // void return: The RET node has the chain result value of the CALL node as 2605 // input. 2606 if (NumOps == 1 && 2607 Ret.getOperand(0) == SDValue(TheCall, TheCall->getNumValues()-1)) 2608 return true; 2609 2610 return false; 2611} 2612