TargetLowering.cpp revision 11eab02b770086c119a18aae0fe214fbe5eed0d0
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/TargetLowering.h" 15#include "llvm/MC/MCAsmInfo.h" 16#include "llvm/Target/TargetData.h" 17#include "llvm/Target/TargetLoweringObjectFile.h" 18#include "llvm/Target/TargetMachine.h" 19#include "llvm/Target/TargetRegisterInfo.h" 20#include "llvm/Target/TargetSubtarget.h" 21#include "llvm/GlobalVariable.h" 22#include "llvm/DerivedTypes.h" 23#include "llvm/CodeGen/MachineFrameInfo.h" 24#include "llvm/CodeGen/SelectionDAG.h" 25#include "llvm/ADT/StringExtras.h" 26#include "llvm/ADT/STLExtras.h" 27#include "llvm/Support/ErrorHandling.h" 28#include "llvm/Support/MathExtras.h" 29using namespace llvm; 30 31namespace llvm { 32TLSModel::Model getTLSModel(const GlobalValue *GV, Reloc::Model reloc) { 33 bool isLocal = GV->hasLocalLinkage(); 34 bool isDeclaration = GV->isDeclaration(); 35 // FIXME: what should we do for protected and internal visibility? 36 // For variables, is internal different from hidden? 37 bool isHidden = GV->hasHiddenVisibility(); 38 39 if (reloc == Reloc::PIC_) { 40 if (isLocal || isHidden) 41 return TLSModel::LocalDynamic; 42 else 43 return TLSModel::GeneralDynamic; 44 } else { 45 if (!isDeclaration || isHidden) 46 return TLSModel::LocalExec; 47 else 48 return TLSModel::InitialExec; 49 } 50} 51} 52 53/// InitLibcallNames - Set default libcall names. 54/// 55static void InitLibcallNames(const char **Names) { 56 Names[RTLIB::SHL_I16] = "__ashlhi3"; 57 Names[RTLIB::SHL_I32] = "__ashlsi3"; 58 Names[RTLIB::SHL_I64] = "__ashldi3"; 59 Names[RTLIB::SHL_I128] = "__ashlti3"; 60 Names[RTLIB::SRL_I16] = "__lshrhi3"; 61 Names[RTLIB::SRL_I32] = "__lshrsi3"; 62 Names[RTLIB::SRL_I64] = "__lshrdi3"; 63 Names[RTLIB::SRL_I128] = "__lshrti3"; 64 Names[RTLIB::SRA_I16] = "__ashrhi3"; 65 Names[RTLIB::SRA_I32] = "__ashrsi3"; 66 Names[RTLIB::SRA_I64] = "__ashrdi3"; 67 Names[RTLIB::SRA_I128] = "__ashrti3"; 68 Names[RTLIB::MUL_I16] = "__mulhi3"; 69 Names[RTLIB::MUL_I32] = "__mulsi3"; 70 Names[RTLIB::MUL_I64] = "__muldi3"; 71 Names[RTLIB::MUL_I128] = "__multi3"; 72 Names[RTLIB::SDIV_I16] = "__divhi3"; 73 Names[RTLIB::SDIV_I32] = "__divsi3"; 74 Names[RTLIB::SDIV_I64] = "__divdi3"; 75 Names[RTLIB::SDIV_I128] = "__divti3"; 76 Names[RTLIB::UDIV_I16] = "__udivhi3"; 77 Names[RTLIB::UDIV_I32] = "__udivsi3"; 78 Names[RTLIB::UDIV_I64] = "__udivdi3"; 79 Names[RTLIB::UDIV_I128] = "__udivti3"; 80 Names[RTLIB::SREM_I16] = "__modhi3"; 81 Names[RTLIB::SREM_I32] = "__modsi3"; 82 Names[RTLIB::SREM_I64] = "__moddi3"; 83 Names[RTLIB::SREM_I128] = "__modti3"; 84 Names[RTLIB::UREM_I16] = "__umodhi3"; 85 Names[RTLIB::UREM_I32] = "__umodsi3"; 86 Names[RTLIB::UREM_I64] = "__umoddi3"; 87 Names[RTLIB::UREM_I128] = "__umodti3"; 88 Names[RTLIB::NEG_I32] = "__negsi2"; 89 Names[RTLIB::NEG_I64] = "__negdi2"; 90 Names[RTLIB::ADD_F32] = "__addsf3"; 91 Names[RTLIB::ADD_F64] = "__adddf3"; 92 Names[RTLIB::ADD_F80] = "__addxf3"; 93 Names[RTLIB::ADD_PPCF128] = "__gcc_qadd"; 94 Names[RTLIB::SUB_F32] = "__subsf3"; 95 Names[RTLIB::SUB_F64] = "__subdf3"; 96 Names[RTLIB::SUB_F80] = "__subxf3"; 97 Names[RTLIB::SUB_PPCF128] = "__gcc_qsub"; 98 Names[RTLIB::MUL_F32] = "__mulsf3"; 99 Names[RTLIB::MUL_F64] = "__muldf3"; 100 Names[RTLIB::MUL_F80] = "__mulxf3"; 101 Names[RTLIB::MUL_PPCF128] = "__gcc_qmul"; 102 Names[RTLIB::DIV_F32] = "__divsf3"; 103 Names[RTLIB::DIV_F64] = "__divdf3"; 104 Names[RTLIB::DIV_F80] = "__divxf3"; 105 Names[RTLIB::DIV_PPCF128] = "__gcc_qdiv"; 106 Names[RTLIB::REM_F32] = "fmodf"; 107 Names[RTLIB::REM_F64] = "fmod"; 108 Names[RTLIB::REM_F80] = "fmodl"; 109 Names[RTLIB::REM_PPCF128] = "fmodl"; 110 Names[RTLIB::POWI_F32] = "__powisf2"; 111 Names[RTLIB::POWI_F64] = "__powidf2"; 112 Names[RTLIB::POWI_F80] = "__powixf2"; 113 Names[RTLIB::POWI_PPCF128] = "__powitf2"; 114 Names[RTLIB::SQRT_F32] = "sqrtf"; 115 Names[RTLIB::SQRT_F64] = "sqrt"; 116 Names[RTLIB::SQRT_F80] = "sqrtl"; 117 Names[RTLIB::SQRT_PPCF128] = "sqrtl"; 118 Names[RTLIB::LOG_F32] = "logf"; 119 Names[RTLIB::LOG_F64] = "log"; 120 Names[RTLIB::LOG_F80] = "logl"; 121 Names[RTLIB::LOG_PPCF128] = "logl"; 122 Names[RTLIB::LOG2_F32] = "log2f"; 123 Names[RTLIB::LOG2_F64] = "log2"; 124 Names[RTLIB::LOG2_F80] = "log2l"; 125 Names[RTLIB::LOG2_PPCF128] = "log2l"; 126 Names[RTLIB::LOG10_F32] = "log10f"; 127 Names[RTLIB::LOG10_F64] = "log10"; 128 Names[RTLIB::LOG10_F80] = "log10l"; 129 Names[RTLIB::LOG10_PPCF128] = "log10l"; 130 Names[RTLIB::EXP_F32] = "expf"; 131 Names[RTLIB::EXP_F64] = "exp"; 132 Names[RTLIB::EXP_F80] = "expl"; 133 Names[RTLIB::EXP_PPCF128] = "expl"; 134 Names[RTLIB::EXP2_F32] = "exp2f"; 135 Names[RTLIB::EXP2_F64] = "exp2"; 136 Names[RTLIB::EXP2_F80] = "exp2l"; 137 Names[RTLIB::EXP2_PPCF128] = "exp2l"; 138 Names[RTLIB::SIN_F32] = "sinf"; 139 Names[RTLIB::SIN_F64] = "sin"; 140 Names[RTLIB::SIN_F80] = "sinl"; 141 Names[RTLIB::SIN_PPCF128] = "sinl"; 142 Names[RTLIB::COS_F32] = "cosf"; 143 Names[RTLIB::COS_F64] = "cos"; 144 Names[RTLIB::COS_F80] = "cosl"; 145 Names[RTLIB::COS_PPCF128] = "cosl"; 146 Names[RTLIB::POW_F32] = "powf"; 147 Names[RTLIB::POW_F64] = "pow"; 148 Names[RTLIB::POW_F80] = "powl"; 149 Names[RTLIB::POW_PPCF128] = "powl"; 150 Names[RTLIB::CEIL_F32] = "ceilf"; 151 Names[RTLIB::CEIL_F64] = "ceil"; 152 Names[RTLIB::CEIL_F80] = "ceill"; 153 Names[RTLIB::CEIL_PPCF128] = "ceill"; 154 Names[RTLIB::TRUNC_F32] = "truncf"; 155 Names[RTLIB::TRUNC_F64] = "trunc"; 156 Names[RTLIB::TRUNC_F80] = "truncl"; 157 Names[RTLIB::TRUNC_PPCF128] = "truncl"; 158 Names[RTLIB::RINT_F32] = "rintf"; 159 Names[RTLIB::RINT_F64] = "rint"; 160 Names[RTLIB::RINT_F80] = "rintl"; 161 Names[RTLIB::RINT_PPCF128] = "rintl"; 162 Names[RTLIB::NEARBYINT_F32] = "nearbyintf"; 163 Names[RTLIB::NEARBYINT_F64] = "nearbyint"; 164 Names[RTLIB::NEARBYINT_F80] = "nearbyintl"; 165 Names[RTLIB::NEARBYINT_PPCF128] = "nearbyintl"; 166 Names[RTLIB::FLOOR_F32] = "floorf"; 167 Names[RTLIB::FLOOR_F64] = "floor"; 168 Names[RTLIB::FLOOR_F80] = "floorl"; 169 Names[RTLIB::FLOOR_PPCF128] = "floorl"; 170 Names[RTLIB::FPEXT_F32_F64] = "__extendsfdf2"; 171 Names[RTLIB::FPROUND_F64_F32] = "__truncdfsf2"; 172 Names[RTLIB::FPROUND_F80_F32] = "__truncxfsf2"; 173 Names[RTLIB::FPROUND_PPCF128_F32] = "__trunctfsf2"; 174 Names[RTLIB::FPROUND_F80_F64] = "__truncxfdf2"; 175 Names[RTLIB::FPROUND_PPCF128_F64] = "__trunctfdf2"; 176 Names[RTLIB::FPTOSINT_F32_I8] = "__fixsfi8"; 177 Names[RTLIB::FPTOSINT_F32_I16] = "__fixsfi16"; 178 Names[RTLIB::FPTOSINT_F32_I32] = "__fixsfsi"; 179 Names[RTLIB::FPTOSINT_F32_I64] = "__fixsfdi"; 180 Names[RTLIB::FPTOSINT_F32_I128] = "__fixsfti"; 181 Names[RTLIB::FPTOSINT_F64_I32] = "__fixdfsi"; 182 Names[RTLIB::FPTOSINT_F64_I64] = "__fixdfdi"; 183 Names[RTLIB::FPTOSINT_F64_I128] = "__fixdfti"; 184 Names[RTLIB::FPTOSINT_F80_I32] = "__fixxfsi"; 185 Names[RTLIB::FPTOSINT_F80_I64] = "__fixxfdi"; 186 Names[RTLIB::FPTOSINT_F80_I128] = "__fixxfti"; 187 Names[RTLIB::FPTOSINT_PPCF128_I32] = "__fixtfsi"; 188 Names[RTLIB::FPTOSINT_PPCF128_I64] = "__fixtfdi"; 189 Names[RTLIB::FPTOSINT_PPCF128_I128] = "__fixtfti"; 190 Names[RTLIB::FPTOUINT_F32_I8] = "__fixunssfi8"; 191 Names[RTLIB::FPTOUINT_F32_I16] = "__fixunssfi16"; 192 Names[RTLIB::FPTOUINT_F32_I32] = "__fixunssfsi"; 193 Names[RTLIB::FPTOUINT_F32_I64] = "__fixunssfdi"; 194 Names[RTLIB::FPTOUINT_F32_I128] = "__fixunssfti"; 195 Names[RTLIB::FPTOUINT_F64_I32] = "__fixunsdfsi"; 196 Names[RTLIB::FPTOUINT_F64_I64] = "__fixunsdfdi"; 197 Names[RTLIB::FPTOUINT_F64_I128] = "__fixunsdfti"; 198 Names[RTLIB::FPTOUINT_F80_I32] = "__fixunsxfsi"; 199 Names[RTLIB::FPTOUINT_F80_I64] = "__fixunsxfdi"; 200 Names[RTLIB::FPTOUINT_F80_I128] = "__fixunsxfti"; 201 Names[RTLIB::FPTOUINT_PPCF128_I32] = "__fixunstfsi"; 202 Names[RTLIB::FPTOUINT_PPCF128_I64] = "__fixunstfdi"; 203 Names[RTLIB::FPTOUINT_PPCF128_I128] = "__fixunstfti"; 204 Names[RTLIB::SINTTOFP_I32_F32] = "__floatsisf"; 205 Names[RTLIB::SINTTOFP_I32_F64] = "__floatsidf"; 206 Names[RTLIB::SINTTOFP_I32_F80] = "__floatsixf"; 207 Names[RTLIB::SINTTOFP_I32_PPCF128] = "__floatsitf"; 208 Names[RTLIB::SINTTOFP_I64_F32] = "__floatdisf"; 209 Names[RTLIB::SINTTOFP_I64_F64] = "__floatdidf"; 210 Names[RTLIB::SINTTOFP_I64_F80] = "__floatdixf"; 211 Names[RTLIB::SINTTOFP_I64_PPCF128] = "__floatditf"; 212 Names[RTLIB::SINTTOFP_I128_F32] = "__floattisf"; 213 Names[RTLIB::SINTTOFP_I128_F64] = "__floattidf"; 214 Names[RTLIB::SINTTOFP_I128_F80] = "__floattixf"; 215 Names[RTLIB::SINTTOFP_I128_PPCF128] = "__floattitf"; 216 Names[RTLIB::UINTTOFP_I32_F32] = "__floatunsisf"; 217 Names[RTLIB::UINTTOFP_I32_F64] = "__floatunsidf"; 218 Names[RTLIB::UINTTOFP_I32_F80] = "__floatunsixf"; 219 Names[RTLIB::UINTTOFP_I32_PPCF128] = "__floatunsitf"; 220 Names[RTLIB::UINTTOFP_I64_F32] = "__floatundisf"; 221 Names[RTLIB::UINTTOFP_I64_F64] = "__floatundidf"; 222 Names[RTLIB::UINTTOFP_I64_F80] = "__floatundixf"; 223 Names[RTLIB::UINTTOFP_I64_PPCF128] = "__floatunditf"; 224 Names[RTLIB::UINTTOFP_I128_F32] = "__floatuntisf"; 225 Names[RTLIB::UINTTOFP_I128_F64] = "__floatuntidf"; 226 Names[RTLIB::UINTTOFP_I128_F80] = "__floatuntixf"; 227 Names[RTLIB::UINTTOFP_I128_PPCF128] = "__floatuntitf"; 228 Names[RTLIB::OEQ_F32] = "__eqsf2"; 229 Names[RTLIB::OEQ_F64] = "__eqdf2"; 230 Names[RTLIB::UNE_F32] = "__nesf2"; 231 Names[RTLIB::UNE_F64] = "__nedf2"; 232 Names[RTLIB::OGE_F32] = "__gesf2"; 233 Names[RTLIB::OGE_F64] = "__gedf2"; 234 Names[RTLIB::OLT_F32] = "__ltsf2"; 235 Names[RTLIB::OLT_F64] = "__ltdf2"; 236 Names[RTLIB::OLE_F32] = "__lesf2"; 237 Names[RTLIB::OLE_F64] = "__ledf2"; 238 Names[RTLIB::OGT_F32] = "__gtsf2"; 239 Names[RTLIB::OGT_F64] = "__gtdf2"; 240 Names[RTLIB::UO_F32] = "__unordsf2"; 241 Names[RTLIB::UO_F64] = "__unorddf2"; 242 Names[RTLIB::O_F32] = "__unordsf2"; 243 Names[RTLIB::O_F64] = "__unorddf2"; 244 Names[RTLIB::MEMCPY] = "memcpy"; 245 Names[RTLIB::MEMMOVE] = "memmove"; 246 Names[RTLIB::MEMSET] = "memset"; 247 Names[RTLIB::UNWIND_RESUME] = "_Unwind_Resume"; 248} 249 250/// InitLibcallCallingConvs - Set default libcall CallingConvs. 251/// 252static void InitLibcallCallingConvs(CallingConv::ID *CCs) { 253 for (int i = 0; i < RTLIB::UNKNOWN_LIBCALL; ++i) { 254 CCs[i] = CallingConv::C; 255 } 256} 257 258/// getFPEXT - Return the FPEXT_*_* value for the given types, or 259/// UNKNOWN_LIBCALL if there is none. 260RTLIB::Libcall RTLIB::getFPEXT(EVT OpVT, EVT RetVT) { 261 if (OpVT == MVT::f32) { 262 if (RetVT == MVT::f64) 263 return FPEXT_F32_F64; 264 } 265 return UNKNOWN_LIBCALL; 266} 267 268/// getFPROUND - Return the FPROUND_*_* value for the given types, or 269/// UNKNOWN_LIBCALL if there is none. 270RTLIB::Libcall RTLIB::getFPROUND(EVT OpVT, EVT RetVT) { 271 if (RetVT == MVT::f32) { 272 if (OpVT == MVT::f64) 273 return FPROUND_F64_F32; 274 if (OpVT == MVT::f80) 275 return FPROUND_F80_F32; 276 if (OpVT == MVT::ppcf128) 277 return FPROUND_PPCF128_F32; 278 } else if (RetVT == MVT::f64) { 279 if (OpVT == MVT::f80) 280 return FPROUND_F80_F64; 281 if (OpVT == MVT::ppcf128) 282 return FPROUND_PPCF128_F64; 283 } 284 return UNKNOWN_LIBCALL; 285} 286 287/// getFPTOSINT - Return the FPTOSINT_*_* value for the given types, or 288/// UNKNOWN_LIBCALL if there is none. 289RTLIB::Libcall RTLIB::getFPTOSINT(EVT OpVT, EVT RetVT) { 290 if (OpVT == MVT::f32) { 291 if (RetVT == MVT::i8) 292 return FPTOSINT_F32_I8; 293 if (RetVT == MVT::i16) 294 return FPTOSINT_F32_I16; 295 if (RetVT == MVT::i32) 296 return FPTOSINT_F32_I32; 297 if (RetVT == MVT::i64) 298 return FPTOSINT_F32_I64; 299 if (RetVT == MVT::i128) 300 return FPTOSINT_F32_I128; 301 } else if (OpVT == MVT::f64) { 302 if (RetVT == MVT::i32) 303 return FPTOSINT_F64_I32; 304 if (RetVT == MVT::i64) 305 return FPTOSINT_F64_I64; 306 if (RetVT == MVT::i128) 307 return FPTOSINT_F64_I128; 308 } else if (OpVT == MVT::f80) { 309 if (RetVT == MVT::i32) 310 return FPTOSINT_F80_I32; 311 if (RetVT == MVT::i64) 312 return FPTOSINT_F80_I64; 313 if (RetVT == MVT::i128) 314 return FPTOSINT_F80_I128; 315 } else if (OpVT == MVT::ppcf128) { 316 if (RetVT == MVT::i32) 317 return FPTOSINT_PPCF128_I32; 318 if (RetVT == MVT::i64) 319 return FPTOSINT_PPCF128_I64; 320 if (RetVT == MVT::i128) 321 return FPTOSINT_PPCF128_I128; 322 } 323 return UNKNOWN_LIBCALL; 324} 325 326/// getFPTOUINT - Return the FPTOUINT_*_* value for the given types, or 327/// UNKNOWN_LIBCALL if there is none. 328RTLIB::Libcall RTLIB::getFPTOUINT(EVT OpVT, EVT RetVT) { 329 if (OpVT == MVT::f32) { 330 if (RetVT == MVT::i8) 331 return FPTOUINT_F32_I8; 332 if (RetVT == MVT::i16) 333 return FPTOUINT_F32_I16; 334 if (RetVT == MVT::i32) 335 return FPTOUINT_F32_I32; 336 if (RetVT == MVT::i64) 337 return FPTOUINT_F32_I64; 338 if (RetVT == MVT::i128) 339 return FPTOUINT_F32_I128; 340 } else if (OpVT == MVT::f64) { 341 if (RetVT == MVT::i32) 342 return FPTOUINT_F64_I32; 343 if (RetVT == MVT::i64) 344 return FPTOUINT_F64_I64; 345 if (RetVT == MVT::i128) 346 return FPTOUINT_F64_I128; 347 } else if (OpVT == MVT::f80) { 348 if (RetVT == MVT::i32) 349 return FPTOUINT_F80_I32; 350 if (RetVT == MVT::i64) 351 return FPTOUINT_F80_I64; 352 if (RetVT == MVT::i128) 353 return FPTOUINT_F80_I128; 354 } else if (OpVT == MVT::ppcf128) { 355 if (RetVT == MVT::i32) 356 return FPTOUINT_PPCF128_I32; 357 if (RetVT == MVT::i64) 358 return FPTOUINT_PPCF128_I64; 359 if (RetVT == MVT::i128) 360 return FPTOUINT_PPCF128_I128; 361 } 362 return UNKNOWN_LIBCALL; 363} 364 365/// getSINTTOFP - Return the SINTTOFP_*_* value for the given types, or 366/// UNKNOWN_LIBCALL if there is none. 367RTLIB::Libcall RTLIB::getSINTTOFP(EVT OpVT, EVT RetVT) { 368 if (OpVT == MVT::i32) { 369 if (RetVT == MVT::f32) 370 return SINTTOFP_I32_F32; 371 else if (RetVT == MVT::f64) 372 return SINTTOFP_I32_F64; 373 else if (RetVT == MVT::f80) 374 return SINTTOFP_I32_F80; 375 else if (RetVT == MVT::ppcf128) 376 return SINTTOFP_I32_PPCF128; 377 } else if (OpVT == MVT::i64) { 378 if (RetVT == MVT::f32) 379 return SINTTOFP_I64_F32; 380 else if (RetVT == MVT::f64) 381 return SINTTOFP_I64_F64; 382 else if (RetVT == MVT::f80) 383 return SINTTOFP_I64_F80; 384 else if (RetVT == MVT::ppcf128) 385 return SINTTOFP_I64_PPCF128; 386 } else if (OpVT == MVT::i128) { 387 if (RetVT == MVT::f32) 388 return SINTTOFP_I128_F32; 389 else if (RetVT == MVT::f64) 390 return SINTTOFP_I128_F64; 391 else if (RetVT == MVT::f80) 392 return SINTTOFP_I128_F80; 393 else if (RetVT == MVT::ppcf128) 394 return SINTTOFP_I128_PPCF128; 395 } 396 return UNKNOWN_LIBCALL; 397} 398 399/// getUINTTOFP - Return the UINTTOFP_*_* value for the given types, or 400/// UNKNOWN_LIBCALL if there is none. 401RTLIB::Libcall RTLIB::getUINTTOFP(EVT OpVT, EVT RetVT) { 402 if (OpVT == MVT::i32) { 403 if (RetVT == MVT::f32) 404 return UINTTOFP_I32_F32; 405 else if (RetVT == MVT::f64) 406 return UINTTOFP_I32_F64; 407 else if (RetVT == MVT::f80) 408 return UINTTOFP_I32_F80; 409 else if (RetVT == MVT::ppcf128) 410 return UINTTOFP_I32_PPCF128; 411 } else if (OpVT == MVT::i64) { 412 if (RetVT == MVT::f32) 413 return UINTTOFP_I64_F32; 414 else if (RetVT == MVT::f64) 415 return UINTTOFP_I64_F64; 416 else if (RetVT == MVT::f80) 417 return UINTTOFP_I64_F80; 418 else if (RetVT == MVT::ppcf128) 419 return UINTTOFP_I64_PPCF128; 420 } else if (OpVT == MVT::i128) { 421 if (RetVT == MVT::f32) 422 return UINTTOFP_I128_F32; 423 else if (RetVT == MVT::f64) 424 return UINTTOFP_I128_F64; 425 else if (RetVT == MVT::f80) 426 return UINTTOFP_I128_F80; 427 else if (RetVT == MVT::ppcf128) 428 return UINTTOFP_I128_PPCF128; 429 } 430 return UNKNOWN_LIBCALL; 431} 432 433/// InitCmpLibcallCCs - Set default comparison libcall CC. 434/// 435static void InitCmpLibcallCCs(ISD::CondCode *CCs) { 436 memset(CCs, ISD::SETCC_INVALID, sizeof(ISD::CondCode)*RTLIB::UNKNOWN_LIBCALL); 437 CCs[RTLIB::OEQ_F32] = ISD::SETEQ; 438 CCs[RTLIB::OEQ_F64] = ISD::SETEQ; 439 CCs[RTLIB::UNE_F32] = ISD::SETNE; 440 CCs[RTLIB::UNE_F64] = ISD::SETNE; 441 CCs[RTLIB::OGE_F32] = ISD::SETGE; 442 CCs[RTLIB::OGE_F64] = ISD::SETGE; 443 CCs[RTLIB::OLT_F32] = ISD::SETLT; 444 CCs[RTLIB::OLT_F64] = ISD::SETLT; 445 CCs[RTLIB::OLE_F32] = ISD::SETLE; 446 CCs[RTLIB::OLE_F64] = ISD::SETLE; 447 CCs[RTLIB::OGT_F32] = ISD::SETGT; 448 CCs[RTLIB::OGT_F64] = ISD::SETGT; 449 CCs[RTLIB::UO_F32] = ISD::SETNE; 450 CCs[RTLIB::UO_F64] = ISD::SETNE; 451 CCs[RTLIB::O_F32] = ISD::SETEQ; 452 CCs[RTLIB::O_F64] = ISD::SETEQ; 453} 454 455/// NOTE: The constructor takes ownership of TLOF. 456TargetLowering::TargetLowering(TargetMachine &tm,TargetLoweringObjectFile *tlof) 457 : TM(tm), TD(TM.getTargetData()), TLOF(*tlof) { 458 // All operations default to being supported. 459 memset(OpActions, 0, sizeof(OpActions)); 460 memset(LoadExtActions, 0, sizeof(LoadExtActions)); 461 memset(TruncStoreActions, 0, sizeof(TruncStoreActions)); 462 memset(IndexedModeActions, 0, sizeof(IndexedModeActions)); 463 memset(ConvertActions, 0, sizeof(ConvertActions)); 464 memset(CondCodeActions, 0, sizeof(CondCodeActions)); 465 466 // Set default actions for various operations. 467 for (unsigned VT = 0; VT != (unsigned)MVT::LAST_VALUETYPE; ++VT) { 468 // Default all indexed load / store to expand. 469 for (unsigned IM = (unsigned)ISD::PRE_INC; 470 IM != (unsigned)ISD::LAST_INDEXED_MODE; ++IM) { 471 setIndexedLoadAction(IM, (MVT::SimpleValueType)VT, Expand); 472 setIndexedStoreAction(IM, (MVT::SimpleValueType)VT, Expand); 473 } 474 475 // These operations default to expand. 476 setOperationAction(ISD::FGETSIGN, (MVT::SimpleValueType)VT, Expand); 477 setOperationAction(ISD::CONCAT_VECTORS, (MVT::SimpleValueType)VT, Expand); 478 } 479 480 // Most targets ignore the @llvm.prefetch intrinsic. 481 setOperationAction(ISD::PREFETCH, MVT::Other, Expand); 482 483 // ConstantFP nodes default to expand. Targets can either change this to 484 // Legal, in which case all fp constants are legal, or use addLegalFPImmediate 485 // to optimize expansions for certain constants. 486 setOperationAction(ISD::ConstantFP, MVT::f32, Expand); 487 setOperationAction(ISD::ConstantFP, MVT::f64, Expand); 488 setOperationAction(ISD::ConstantFP, MVT::f80, Expand); 489 490 // These library functions default to expand. 491 setOperationAction(ISD::FLOG , MVT::f64, Expand); 492 setOperationAction(ISD::FLOG2, MVT::f64, Expand); 493 setOperationAction(ISD::FLOG10,MVT::f64, Expand); 494 setOperationAction(ISD::FEXP , MVT::f64, Expand); 495 setOperationAction(ISD::FEXP2, MVT::f64, Expand); 496 setOperationAction(ISD::FLOG , MVT::f32, Expand); 497 setOperationAction(ISD::FLOG2, MVT::f32, Expand); 498 setOperationAction(ISD::FLOG10,MVT::f32, Expand); 499 setOperationAction(ISD::FEXP , MVT::f32, Expand); 500 setOperationAction(ISD::FEXP2, MVT::f32, Expand); 501 502 // Default ISD::TRAP to expand (which turns it into abort). 503 setOperationAction(ISD::TRAP, MVT::Other, Expand); 504 505 IsLittleEndian = TD->isLittleEndian(); 506 UsesGlobalOffsetTable = false; 507 ShiftAmountTy = PointerTy = MVT::getIntegerVT(8*TD->getPointerSize()); 508 memset(RegClassForVT, 0,MVT::LAST_VALUETYPE*sizeof(TargetRegisterClass*)); 509 memset(TargetDAGCombineArray, 0, array_lengthof(TargetDAGCombineArray)); 510 maxStoresPerMemset = maxStoresPerMemcpy = maxStoresPerMemmove = 8; 511 benefitFromCodePlacementOpt = false; 512 UseUnderscoreSetJmp = false; 513 UseUnderscoreLongJmp = false; 514 SelectIsExpensive = false; 515 IntDivIsCheap = false; 516 Pow2DivIsCheap = false; 517 StackPointerRegisterToSaveRestore = 0; 518 ExceptionPointerRegister = 0; 519 ExceptionSelectorRegister = 0; 520 BooleanContents = UndefinedBooleanContent; 521 SchedPreferenceInfo = SchedulingForLatency; 522 JumpBufSize = 0; 523 JumpBufAlignment = 0; 524 IfCvtBlockSizeLimit = 2; 525 IfCvtDupBlockSizeLimit = 0; 526 PrefLoopAlignment = 0; 527 528 InitLibcallNames(LibcallRoutineNames); 529 InitCmpLibcallCCs(CmpLibcallCCs); 530 InitLibcallCallingConvs(LibcallCallingConvs); 531 532 // Tell Legalize whether the assembler supports DEBUG_LOC. 533 const MCAsmInfo *TASM = TM.getMCAsmInfo(); 534 if (!TASM || !TASM->hasDotLocAndDotFile()) 535 setOperationAction(ISD::DEBUG_LOC, MVT::Other, Expand); 536} 537 538TargetLowering::~TargetLowering() { 539 delete &TLOF; 540} 541 542static unsigned getVectorTypeBreakdownMVT(MVT VT, MVT &IntermediateVT, 543 unsigned &NumIntermediates, 544 EVT &RegisterVT, 545 TargetLowering* TLI) { 546 // Figure out the right, legal destination reg to copy into. 547 unsigned NumElts = VT.getVectorNumElements(); 548 MVT EltTy = VT.getVectorElementType(); 549 550 unsigned NumVectorRegs = 1; 551 552 // FIXME: We don't support non-power-of-2-sized vectors for now. Ideally we 553 // could break down into LHS/RHS like LegalizeDAG does. 554 if (!isPowerOf2_32(NumElts)) { 555 NumVectorRegs = NumElts; 556 NumElts = 1; 557 } 558 559 // Divide the input until we get to a supported size. This will always 560 // end with a scalar if the target doesn't support vectors. 561 while (NumElts > 1 && !TLI->isTypeLegal(MVT::getVectorVT(EltTy, NumElts))) { 562 NumElts >>= 1; 563 NumVectorRegs <<= 1; 564 } 565 566 NumIntermediates = NumVectorRegs; 567 568 MVT NewVT = MVT::getVectorVT(EltTy, NumElts); 569 if (!TLI->isTypeLegal(NewVT)) 570 NewVT = EltTy; 571 IntermediateVT = NewVT; 572 573 EVT DestVT = TLI->getRegisterType(NewVT); 574 RegisterVT = DestVT; 575 if (EVT(DestVT).bitsLT(NewVT)) { 576 // Value is expanded, e.g. i64 -> i16. 577 return NumVectorRegs*(NewVT.getSizeInBits()/DestVT.getSizeInBits()); 578 } else { 579 // Otherwise, promotion or legal types use the same number of registers as 580 // the vector decimated to the appropriate level. 581 return NumVectorRegs; 582 } 583 584 return 1; 585} 586 587/// computeRegisterProperties - Once all of the register classes are added, 588/// this allows us to compute derived properties we expose. 589void TargetLowering::computeRegisterProperties() { 590 assert(MVT::LAST_VALUETYPE <= MVT::MAX_ALLOWED_VALUETYPE && 591 "Too many value types for ValueTypeActions to hold!"); 592 593 // Everything defaults to needing one register. 594 for (unsigned i = 0; i != MVT::LAST_VALUETYPE; ++i) { 595 NumRegistersForVT[i] = 1; 596 RegisterTypeForVT[i] = TransformToType[i] = (MVT::SimpleValueType)i; 597 } 598 // ...except isVoid, which doesn't need any registers. 599 NumRegistersForVT[MVT::isVoid] = 0; 600 601 // Find the largest integer register class. 602 unsigned LargestIntReg = MVT::LAST_INTEGER_VALUETYPE; 603 for (; RegClassForVT[LargestIntReg] == 0; --LargestIntReg) 604 assert(LargestIntReg != MVT::i1 && "No integer registers defined!"); 605 606 // Every integer value type larger than this largest register takes twice as 607 // many registers to represent as the previous ValueType. 608 for (unsigned ExpandedReg = LargestIntReg + 1; ; ++ExpandedReg) { 609 EVT ExpandedVT = (MVT::SimpleValueType)ExpandedReg; 610 if (!ExpandedVT.isInteger()) 611 break; 612 NumRegistersForVT[ExpandedReg] = 2*NumRegistersForVT[ExpandedReg-1]; 613 RegisterTypeForVT[ExpandedReg] = (MVT::SimpleValueType)LargestIntReg; 614 TransformToType[ExpandedReg] = (MVT::SimpleValueType)(ExpandedReg - 1); 615 ValueTypeActions.setTypeAction(ExpandedVT, Expand); 616 } 617 618 // Inspect all of the ValueType's smaller than the largest integer 619 // register to see which ones need promotion. 620 unsigned LegalIntReg = LargestIntReg; 621 for (unsigned IntReg = LargestIntReg - 1; 622 IntReg >= (unsigned)MVT::i1; --IntReg) { 623 EVT IVT = (MVT::SimpleValueType)IntReg; 624 if (isTypeLegal(IVT)) { 625 LegalIntReg = IntReg; 626 } else { 627 RegisterTypeForVT[IntReg] = TransformToType[IntReg] = 628 (MVT::SimpleValueType)LegalIntReg; 629 ValueTypeActions.setTypeAction(IVT, Promote); 630 } 631 } 632 633 // ppcf128 type is really two f64's. 634 if (!isTypeLegal(MVT::ppcf128)) { 635 NumRegistersForVT[MVT::ppcf128] = 2*NumRegistersForVT[MVT::f64]; 636 RegisterTypeForVT[MVT::ppcf128] = MVT::f64; 637 TransformToType[MVT::ppcf128] = MVT::f64; 638 ValueTypeActions.setTypeAction(MVT::ppcf128, Expand); 639 } 640 641 // Decide how to handle f64. If the target does not have native f64 support, 642 // expand it to i64 and we will be generating soft float library calls. 643 if (!isTypeLegal(MVT::f64)) { 644 NumRegistersForVT[MVT::f64] = NumRegistersForVT[MVT::i64]; 645 RegisterTypeForVT[MVT::f64] = RegisterTypeForVT[MVT::i64]; 646 TransformToType[MVT::f64] = MVT::i64; 647 ValueTypeActions.setTypeAction(MVT::f64, Expand); 648 } 649 650 // Decide how to handle f32. If the target does not have native support for 651 // f32, promote it to f64 if it is legal. Otherwise, expand it to i32. 652 if (!isTypeLegal(MVT::f32)) { 653 if (isTypeLegal(MVT::f64)) { 654 NumRegistersForVT[MVT::f32] = NumRegistersForVT[MVT::f64]; 655 RegisterTypeForVT[MVT::f32] = RegisterTypeForVT[MVT::f64]; 656 TransformToType[MVT::f32] = MVT::f64; 657 ValueTypeActions.setTypeAction(MVT::f32, Promote); 658 } else { 659 NumRegistersForVT[MVT::f32] = NumRegistersForVT[MVT::i32]; 660 RegisterTypeForVT[MVT::f32] = RegisterTypeForVT[MVT::i32]; 661 TransformToType[MVT::f32] = MVT::i32; 662 ValueTypeActions.setTypeAction(MVT::f32, Expand); 663 } 664 } 665 666 // Loop over all of the vector value types to see which need transformations. 667 for (unsigned i = MVT::FIRST_VECTOR_VALUETYPE; 668 i <= (unsigned)MVT::LAST_VECTOR_VALUETYPE; ++i) { 669 MVT VT = (MVT::SimpleValueType)i; 670 if (!isTypeLegal(VT)) { 671 MVT IntermediateVT; 672 EVT RegisterVT; 673 unsigned NumIntermediates; 674 NumRegistersForVT[i] = 675 getVectorTypeBreakdownMVT(VT, IntermediateVT, NumIntermediates, 676 RegisterVT, this); 677 RegisterTypeForVT[i] = RegisterVT; 678 679 // Determine if there is a legal wider type. 680 bool IsLegalWiderType = false; 681 EVT EltVT = VT.getVectorElementType(); 682 unsigned NElts = VT.getVectorNumElements(); 683 for (unsigned nVT = i+1; nVT <= MVT::LAST_VECTOR_VALUETYPE; ++nVT) { 684 EVT SVT = (MVT::SimpleValueType)nVT; 685 if (isTypeLegal(SVT) && SVT.getVectorElementType() == EltVT && 686 SVT.getVectorNumElements() > NElts) { 687 TransformToType[i] = SVT; 688 ValueTypeActions.setTypeAction(VT, Promote); 689 IsLegalWiderType = true; 690 break; 691 } 692 } 693 if (!IsLegalWiderType) { 694 EVT NVT = VT.getPow2VectorType(); 695 if (NVT == VT) { 696 // Type is already a power of 2. The default action is to split. 697 TransformToType[i] = MVT::Other; 698 ValueTypeActions.setTypeAction(VT, Expand); 699 } else { 700 TransformToType[i] = NVT; 701 ValueTypeActions.setTypeAction(VT, Promote); 702 } 703 } 704 } 705 } 706} 707 708const char *TargetLowering::getTargetNodeName(unsigned Opcode) const { 709 return NULL; 710} 711 712 713MVT::SimpleValueType TargetLowering::getSetCCResultType(EVT VT) const { 714 return PointerTy.SimpleTy; 715} 716 717/// getVectorTypeBreakdown - Vector types are broken down into some number of 718/// legal first class types. For example, MVT::v8f32 maps to 2 MVT::v4f32 719/// with Altivec or SSE1, or 8 promoted MVT::f64 values with the X86 FP stack. 720/// Similarly, MVT::v2i64 turns into 4 MVT::i32 values with both PPC and X86. 721/// 722/// This method returns the number of registers needed, and the VT for each 723/// register. It also returns the VT and quantity of the intermediate values 724/// before they are promoted/expanded. 725/// 726unsigned TargetLowering::getVectorTypeBreakdown(LLVMContext &Context, EVT VT, 727 EVT &IntermediateVT, 728 unsigned &NumIntermediates, 729 EVT &RegisterVT) const { 730 // Figure out the right, legal destination reg to copy into. 731 unsigned NumElts = VT.getVectorNumElements(); 732 EVT EltTy = VT.getVectorElementType(); 733 734 unsigned NumVectorRegs = 1; 735 736 // FIXME: We don't support non-power-of-2-sized vectors for now. Ideally we 737 // could break down into LHS/RHS like LegalizeDAG does. 738 if (!isPowerOf2_32(NumElts)) { 739 NumVectorRegs = NumElts; 740 NumElts = 1; 741 } 742 743 // Divide the input until we get to a supported size. This will always 744 // end with a scalar if the target doesn't support vectors. 745 while (NumElts > 1 && !isTypeLegal( 746 EVT::getVectorVT(Context, EltTy, NumElts))) { 747 NumElts >>= 1; 748 NumVectorRegs <<= 1; 749 } 750 751 NumIntermediates = NumVectorRegs; 752 753 EVT NewVT = EVT::getVectorVT(Context, EltTy, NumElts); 754 if (!isTypeLegal(NewVT)) 755 NewVT = EltTy; 756 IntermediateVT = NewVT; 757 758 EVT DestVT = getRegisterType(Context, NewVT); 759 RegisterVT = DestVT; 760 if (DestVT.bitsLT(NewVT)) { 761 // Value is expanded, e.g. i64 -> i16. 762 return NumVectorRegs*(NewVT.getSizeInBits()/DestVT.getSizeInBits()); 763 } else { 764 // Otherwise, promotion or legal types use the same number of registers as 765 // the vector decimated to the appropriate level. 766 return NumVectorRegs; 767 } 768 769 return 1; 770} 771 772/// getWidenVectorType: given a vector type, returns the type to widen to 773/// (e.g., v7i8 to v8i8). If the vector type is legal, it returns itself. 774/// If there is no vector type that we want to widen to, returns MVT::Other 775/// When and where to widen is target dependent based on the cost of 776/// scalarizing vs using the wider vector type. 777EVT TargetLowering::getWidenVectorType(EVT VT) const { 778 assert(VT.isVector()); 779 if (isTypeLegal(VT)) 780 return VT; 781 782 // Default is not to widen until moved to LegalizeTypes 783 return MVT::Other; 784} 785 786/// getByValTypeAlignment - Return the desired alignment for ByVal aggregate 787/// function arguments in the caller parameter area. This is the actual 788/// alignment, not its logarithm. 789unsigned TargetLowering::getByValTypeAlignment(const Type *Ty) const { 790 return TD->getCallFrameTypeAlignment(Ty); 791} 792 793SDValue TargetLowering::getPICJumpTableRelocBase(SDValue Table, 794 SelectionDAG &DAG) const { 795 if (usesGlobalOffsetTable()) 796 return DAG.getGLOBAL_OFFSET_TABLE(getPointerTy()); 797 return Table; 798} 799 800bool 801TargetLowering::isOffsetFoldingLegal(const GlobalAddressSDNode *GA) const { 802 // Assume that everything is safe in static mode. 803 if (getTargetMachine().getRelocationModel() == Reloc::Static) 804 return true; 805 806 // In dynamic-no-pic mode, assume that known defined values are safe. 807 if (getTargetMachine().getRelocationModel() == Reloc::DynamicNoPIC && 808 GA && 809 !GA->getGlobal()->isDeclaration() && 810 !GA->getGlobal()->isWeakForLinker()) 811 return true; 812 813 // Otherwise assume nothing is safe. 814 return false; 815} 816 817//===----------------------------------------------------------------------===// 818// Optimization Methods 819//===----------------------------------------------------------------------===// 820 821/// ShrinkDemandedConstant - Check to see if the specified operand of the 822/// specified instruction is a constant integer. If so, check to see if there 823/// are any bits set in the constant that are not demanded. If so, shrink the 824/// constant and return true. 825bool TargetLowering::TargetLoweringOpt::ShrinkDemandedConstant(SDValue Op, 826 const APInt &Demanded) { 827 DebugLoc dl = Op.getDebugLoc(); 828 829 // FIXME: ISD::SELECT, ISD::SELECT_CC 830 switch (Op.getOpcode()) { 831 default: break; 832 case ISD::XOR: 833 case ISD::AND: 834 case ISD::OR: { 835 ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op.getOperand(1)); 836 if (!C) return false; 837 838 if (Op.getOpcode() == ISD::XOR && 839 (C->getAPIntValue() | (~Demanded)).isAllOnesValue()) 840 return false; 841 842 // if we can expand it to have all bits set, do it 843 if (C->getAPIntValue().intersects(~Demanded)) { 844 EVT VT = Op.getValueType(); 845 SDValue New = DAG.getNode(Op.getOpcode(), dl, VT, Op.getOperand(0), 846 DAG.getConstant(Demanded & 847 C->getAPIntValue(), 848 VT)); 849 return CombineTo(Op, New); 850 } 851 852 break; 853 } 854 } 855 856 return false; 857} 858 859/// ShrinkDemandedOp - Convert x+y to (VT)((SmallVT)x+(SmallVT)y) if the 860/// casts are free. This uses isZExtFree and ZERO_EXTEND for the widening 861/// cast, but it could be generalized for targets with other types of 862/// implicit widening casts. 863bool 864TargetLowering::TargetLoweringOpt::ShrinkDemandedOp(SDValue Op, 865 unsigned BitWidth, 866 const APInt &Demanded, 867 DebugLoc dl) { 868 assert(Op.getNumOperands() == 2 && 869 "ShrinkDemandedOp only supports binary operators!"); 870 assert(Op.getNode()->getNumValues() == 1 && 871 "ShrinkDemandedOp only supports nodes with one result!"); 872 873 // Don't do this if the node has another user, which may require the 874 // full value. 875 if (!Op.getNode()->hasOneUse()) 876 return false; 877 878 // Search for the smallest integer type with free casts to and from 879 // Op's type. For expedience, just check power-of-2 integer types. 880 const TargetLowering &TLI = DAG.getTargetLoweringInfo(); 881 unsigned SmallVTBits = BitWidth - Demanded.countLeadingZeros(); 882 if (!isPowerOf2_32(SmallVTBits)) 883 SmallVTBits = NextPowerOf2(SmallVTBits); 884 for (; SmallVTBits < BitWidth; SmallVTBits = NextPowerOf2(SmallVTBits)) { 885 EVT SmallVT = EVT::getIntegerVT(*DAG.getContext(), SmallVTBits); 886 if (TLI.isTruncateFree(Op.getValueType(), SmallVT) && 887 TLI.isZExtFree(SmallVT, Op.getValueType())) { 888 // We found a type with free casts. 889 SDValue X = DAG.getNode(Op.getOpcode(), dl, SmallVT, 890 DAG.getNode(ISD::TRUNCATE, dl, SmallVT, 891 Op.getNode()->getOperand(0)), 892 DAG.getNode(ISD::TRUNCATE, dl, SmallVT, 893 Op.getNode()->getOperand(1))); 894 SDValue Z = DAG.getNode(ISD::ZERO_EXTEND, dl, Op.getValueType(), X); 895 return CombineTo(Op, Z); 896 } 897 } 898 return false; 899} 900 901/// SimplifyDemandedBits - Look at Op. At this point, we know that only the 902/// DemandedMask bits of the result of Op are ever used downstream. If we can 903/// use this information to simplify Op, create a new simplified DAG node and 904/// return true, returning the original and new nodes in Old and New. Otherwise, 905/// analyze the expression and return a mask of KnownOne and KnownZero bits for 906/// the expression (used to simplify the caller). The KnownZero/One bits may 907/// only be accurate for those bits in the DemandedMask. 908bool TargetLowering::SimplifyDemandedBits(SDValue Op, 909 const APInt &DemandedMask, 910 APInt &KnownZero, 911 APInt &KnownOne, 912 TargetLoweringOpt &TLO, 913 unsigned Depth) const { 914 unsigned BitWidth = DemandedMask.getBitWidth(); 915 assert(Op.getValueSizeInBits() == BitWidth && 916 "Mask size mismatches value type size!"); 917 APInt NewMask = DemandedMask; 918 DebugLoc dl = Op.getDebugLoc(); 919 920 // Don't know anything. 921 KnownZero = KnownOne = APInt(BitWidth, 0); 922 923 // Other users may use these bits. 924 if (!Op.getNode()->hasOneUse()) { 925 if (Depth != 0) { 926 // If not at the root, Just compute the KnownZero/KnownOne bits to 927 // simplify things downstream. 928 TLO.DAG.ComputeMaskedBits(Op, DemandedMask, KnownZero, KnownOne, Depth); 929 return false; 930 } 931 // If this is the root being simplified, allow it to have multiple uses, 932 // just set the NewMask to all bits. 933 NewMask = APInt::getAllOnesValue(BitWidth); 934 } else if (DemandedMask == 0) { 935 // Not demanding any bits from Op. 936 if (Op.getOpcode() != ISD::UNDEF) 937 return TLO.CombineTo(Op, TLO.DAG.getUNDEF(Op.getValueType())); 938 return false; 939 } else if (Depth == 6) { // Limit search depth. 940 return false; 941 } 942 943 APInt KnownZero2, KnownOne2, KnownZeroOut, KnownOneOut; 944 switch (Op.getOpcode()) { 945 case ISD::Constant: 946 // We know all of the bits for a constant! 947 KnownOne = cast<ConstantSDNode>(Op)->getAPIntValue() & NewMask; 948 KnownZero = ~KnownOne & NewMask; 949 return false; // Don't fall through, will infinitely loop. 950 case ISD::AND: 951 // If the RHS is a constant, check to see if the LHS would be zero without 952 // using the bits from the RHS. Below, we use knowledge about the RHS to 953 // simplify the LHS, here we're using information from the LHS to simplify 954 // the RHS. 955 if (ConstantSDNode *RHSC = dyn_cast<ConstantSDNode>(Op.getOperand(1))) { 956 APInt LHSZero, LHSOne; 957 TLO.DAG.ComputeMaskedBits(Op.getOperand(0), NewMask, 958 LHSZero, LHSOne, Depth+1); 959 // If the LHS already has zeros where RHSC does, this and is dead. 960 if ((LHSZero & NewMask) == (~RHSC->getAPIntValue() & NewMask)) 961 return TLO.CombineTo(Op, Op.getOperand(0)); 962 // If any of the set bits in the RHS are known zero on the LHS, shrink 963 // the constant. 964 if (TLO.ShrinkDemandedConstant(Op, ~LHSZero & NewMask)) 965 return true; 966 } 967 968 if (SimplifyDemandedBits(Op.getOperand(1), NewMask, KnownZero, 969 KnownOne, TLO, Depth+1)) 970 return true; 971 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?"); 972 if (SimplifyDemandedBits(Op.getOperand(0), ~KnownZero & NewMask, 973 KnownZero2, KnownOne2, TLO, Depth+1)) 974 return true; 975 assert((KnownZero2 & KnownOne2) == 0 && "Bits known to be one AND zero?"); 976 977 // If all of the demanded bits are known one on one side, return the other. 978 // These bits cannot contribute to the result of the 'and'. 979 if ((NewMask & ~KnownZero2 & KnownOne) == (~KnownZero2 & NewMask)) 980 return TLO.CombineTo(Op, Op.getOperand(0)); 981 if ((NewMask & ~KnownZero & KnownOne2) == (~KnownZero & NewMask)) 982 return TLO.CombineTo(Op, Op.getOperand(1)); 983 // If all of the demanded bits in the inputs are known zeros, return zero. 984 if ((NewMask & (KnownZero|KnownZero2)) == NewMask) 985 return TLO.CombineTo(Op, TLO.DAG.getConstant(0, Op.getValueType())); 986 // If the RHS is a constant, see if we can simplify it. 987 if (TLO.ShrinkDemandedConstant(Op, ~KnownZero2 & NewMask)) 988 return true; 989 // If the operation can be done in a smaller type, do so. 990 if (TLO.ShrinkDemandedOp(Op, BitWidth, NewMask, dl)) 991 return true; 992 993 // Output known-1 bits are only known if set in both the LHS & RHS. 994 KnownOne &= KnownOne2; 995 // Output known-0 are known to be clear if zero in either the LHS | RHS. 996 KnownZero |= KnownZero2; 997 break; 998 case ISD::OR: 999 if (SimplifyDemandedBits(Op.getOperand(1), NewMask, KnownZero, 1000 KnownOne, TLO, Depth+1)) 1001 return true; 1002 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?"); 1003 if (SimplifyDemandedBits(Op.getOperand(0), ~KnownOne & NewMask, 1004 KnownZero2, KnownOne2, TLO, Depth+1)) 1005 return true; 1006 assert((KnownZero2 & KnownOne2) == 0 && "Bits known to be one AND zero?"); 1007 1008 // If all of the demanded bits are known zero on one side, return the other. 1009 // These bits cannot contribute to the result of the 'or'. 1010 if ((NewMask & ~KnownOne2 & KnownZero) == (~KnownOne2 & NewMask)) 1011 return TLO.CombineTo(Op, Op.getOperand(0)); 1012 if ((NewMask & ~KnownOne & KnownZero2) == (~KnownOne & NewMask)) 1013 return TLO.CombineTo(Op, Op.getOperand(1)); 1014 // If all of the potentially set bits on one side are known to be set on 1015 // the other side, just use the 'other' side. 1016 if ((NewMask & ~KnownZero & KnownOne2) == (~KnownZero & NewMask)) 1017 return TLO.CombineTo(Op, Op.getOperand(0)); 1018 if ((NewMask & ~KnownZero2 & KnownOne) == (~KnownZero2 & NewMask)) 1019 return TLO.CombineTo(Op, Op.getOperand(1)); 1020 // If the RHS is a constant, see if we can simplify it. 1021 if (TLO.ShrinkDemandedConstant(Op, NewMask)) 1022 return true; 1023 // If the operation can be done in a smaller type, do so. 1024 if (TLO.ShrinkDemandedOp(Op, BitWidth, NewMask, dl)) 1025 return true; 1026 1027 // Output known-0 bits are only known if clear in both the LHS & RHS. 1028 KnownZero &= KnownZero2; 1029 // Output known-1 are known to be set if set in either the LHS | RHS. 1030 KnownOne |= KnownOne2; 1031 break; 1032 case ISD::XOR: 1033 if (SimplifyDemandedBits(Op.getOperand(1), NewMask, KnownZero, 1034 KnownOne, TLO, Depth+1)) 1035 return true; 1036 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?"); 1037 if (SimplifyDemandedBits(Op.getOperand(0), NewMask, KnownZero2, 1038 KnownOne2, TLO, Depth+1)) 1039 return true; 1040 assert((KnownZero2 & KnownOne2) == 0 && "Bits known to be one AND zero?"); 1041 1042 // If all of the demanded bits are known zero on one side, return the other. 1043 // These bits cannot contribute to the result of the 'xor'. 1044 if ((KnownZero & NewMask) == NewMask) 1045 return TLO.CombineTo(Op, Op.getOperand(0)); 1046 if ((KnownZero2 & NewMask) == NewMask) 1047 return TLO.CombineTo(Op, Op.getOperand(1)); 1048 // If the operation can be done in a smaller type, do so. 1049 if (TLO.ShrinkDemandedOp(Op, BitWidth, NewMask, dl)) 1050 return true; 1051 1052 // If all of the unknown bits are known to be zero on one side or the other 1053 // (but not both) turn this into an *inclusive* or. 1054 // e.g. (A & C1)^(B & C2) -> (A & C1)|(B & C2) iff C1&C2 == 0 1055 if ((NewMask & ~KnownZero & ~KnownZero2) == 0) 1056 return TLO.CombineTo(Op, TLO.DAG.getNode(ISD::OR, dl, Op.getValueType(), 1057 Op.getOperand(0), 1058 Op.getOperand(1))); 1059 1060 // Output known-0 bits are known if clear or set in both the LHS & RHS. 1061 KnownZeroOut = (KnownZero & KnownZero2) | (KnownOne & KnownOne2); 1062 // Output known-1 are known to be set if set in only one of the LHS, RHS. 1063 KnownOneOut = (KnownZero & KnownOne2) | (KnownOne & KnownZero2); 1064 1065 // If all of the demanded bits on one side are known, and all of the set 1066 // bits on that side are also known to be set on the other side, turn this 1067 // into an AND, as we know the bits will be cleared. 1068 // e.g. (X | C1) ^ C2 --> (X | C1) & ~C2 iff (C1&C2) == C2 1069 if ((NewMask & (KnownZero|KnownOne)) == NewMask) { // all known 1070 if ((KnownOne & KnownOne2) == KnownOne) { 1071 EVT VT = Op.getValueType(); 1072 SDValue ANDC = TLO.DAG.getConstant(~KnownOne & NewMask, VT); 1073 return TLO.CombineTo(Op, TLO.DAG.getNode(ISD::AND, dl, VT, 1074 Op.getOperand(0), ANDC)); 1075 } 1076 } 1077 1078 // If the RHS is a constant, see if we can simplify it. 1079 // for XOR, we prefer to force bits to 1 if they will make a -1. 1080 // if we can't force bits, try to shrink constant 1081 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op.getOperand(1))) { 1082 APInt Expanded = C->getAPIntValue() | (~NewMask); 1083 // if we can expand it to have all bits set, do it 1084 if (Expanded.isAllOnesValue()) { 1085 if (Expanded != C->getAPIntValue()) { 1086 EVT VT = Op.getValueType(); 1087 SDValue New = TLO.DAG.getNode(Op.getOpcode(), dl,VT, Op.getOperand(0), 1088 TLO.DAG.getConstant(Expanded, VT)); 1089 return TLO.CombineTo(Op, New); 1090 } 1091 // if it already has all the bits set, nothing to change 1092 // but don't shrink either! 1093 } else if (TLO.ShrinkDemandedConstant(Op, NewMask)) { 1094 return true; 1095 } 1096 } 1097 1098 KnownZero = KnownZeroOut; 1099 KnownOne = KnownOneOut; 1100 break; 1101 case ISD::SELECT: 1102 if (SimplifyDemandedBits(Op.getOperand(2), NewMask, KnownZero, 1103 KnownOne, TLO, Depth+1)) 1104 return true; 1105 if (SimplifyDemandedBits(Op.getOperand(1), NewMask, KnownZero2, 1106 KnownOne2, TLO, Depth+1)) 1107 return true; 1108 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?"); 1109 assert((KnownZero2 & KnownOne2) == 0 && "Bits known to be one AND zero?"); 1110 1111 // If the operands are constants, see if we can simplify them. 1112 if (TLO.ShrinkDemandedConstant(Op, NewMask)) 1113 return true; 1114 1115 // Only known if known in both the LHS and RHS. 1116 KnownOne &= KnownOne2; 1117 KnownZero &= KnownZero2; 1118 break; 1119 case ISD::SELECT_CC: 1120 if (SimplifyDemandedBits(Op.getOperand(3), NewMask, KnownZero, 1121 KnownOne, TLO, Depth+1)) 1122 return true; 1123 if (SimplifyDemandedBits(Op.getOperand(2), NewMask, KnownZero2, 1124 KnownOne2, TLO, Depth+1)) 1125 return true; 1126 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?"); 1127 assert((KnownZero2 & KnownOne2) == 0 && "Bits known to be one AND zero?"); 1128 1129 // If the operands are constants, see if we can simplify them. 1130 if (TLO.ShrinkDemandedConstant(Op, NewMask)) 1131 return true; 1132 1133 // Only known if known in both the LHS and RHS. 1134 KnownOne &= KnownOne2; 1135 KnownZero &= KnownZero2; 1136 break; 1137 case ISD::SHL: 1138 if (ConstantSDNode *SA = dyn_cast<ConstantSDNode>(Op.getOperand(1))) { 1139 unsigned ShAmt = SA->getZExtValue(); 1140 SDValue InOp = Op.getOperand(0); 1141 1142 // If the shift count is an invalid immediate, don't do anything. 1143 if (ShAmt >= BitWidth) 1144 break; 1145 1146 // If this is ((X >>u C1) << ShAmt), see if we can simplify this into a 1147 // single shift. We can do this if the bottom bits (which are shifted 1148 // out) are never demanded. 1149 if (InOp.getOpcode() == ISD::SRL && 1150 isa<ConstantSDNode>(InOp.getOperand(1))) { 1151 if (ShAmt && (NewMask & APInt::getLowBitsSet(BitWidth, ShAmt)) == 0) { 1152 unsigned C1= cast<ConstantSDNode>(InOp.getOperand(1))->getZExtValue(); 1153 unsigned Opc = ISD::SHL; 1154 int Diff = ShAmt-C1; 1155 if (Diff < 0) { 1156 Diff = -Diff; 1157 Opc = ISD::SRL; 1158 } 1159 1160 SDValue NewSA = 1161 TLO.DAG.getConstant(Diff, Op.getOperand(1).getValueType()); 1162 EVT VT = Op.getValueType(); 1163 return TLO.CombineTo(Op, TLO.DAG.getNode(Opc, dl, VT, 1164 InOp.getOperand(0), NewSA)); 1165 } 1166 } 1167 1168 if (SimplifyDemandedBits(Op.getOperand(0), NewMask.lshr(ShAmt), 1169 KnownZero, KnownOne, TLO, Depth+1)) 1170 return true; 1171 KnownZero <<= SA->getZExtValue(); 1172 KnownOne <<= SA->getZExtValue(); 1173 // low bits known zero. 1174 KnownZero |= APInt::getLowBitsSet(BitWidth, SA->getZExtValue()); 1175 } 1176 break; 1177 case ISD::SRL: 1178 if (ConstantSDNode *SA = dyn_cast<ConstantSDNode>(Op.getOperand(1))) { 1179 EVT VT = Op.getValueType(); 1180 unsigned ShAmt = SA->getZExtValue(); 1181 unsigned VTSize = VT.getSizeInBits(); 1182 SDValue InOp = Op.getOperand(0); 1183 1184 // If the shift count is an invalid immediate, don't do anything. 1185 if (ShAmt >= BitWidth) 1186 break; 1187 1188 // If this is ((X << C1) >>u ShAmt), see if we can simplify this into a 1189 // single shift. We can do this if the top bits (which are shifted out) 1190 // are never demanded. 1191 if (InOp.getOpcode() == ISD::SHL && 1192 isa<ConstantSDNode>(InOp.getOperand(1))) { 1193 if (ShAmt && (NewMask & APInt::getHighBitsSet(VTSize, ShAmt)) == 0) { 1194 unsigned C1= cast<ConstantSDNode>(InOp.getOperand(1))->getZExtValue(); 1195 unsigned Opc = ISD::SRL; 1196 int Diff = ShAmt-C1; 1197 if (Diff < 0) { 1198 Diff = -Diff; 1199 Opc = ISD::SHL; 1200 } 1201 1202 SDValue NewSA = 1203 TLO.DAG.getConstant(Diff, Op.getOperand(1).getValueType()); 1204 return TLO.CombineTo(Op, TLO.DAG.getNode(Opc, dl, VT, 1205 InOp.getOperand(0), NewSA)); 1206 } 1207 } 1208 1209 // Compute the new bits that are at the top now. 1210 if (SimplifyDemandedBits(InOp, (NewMask << ShAmt), 1211 KnownZero, KnownOne, TLO, Depth+1)) 1212 return true; 1213 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?"); 1214 KnownZero = KnownZero.lshr(ShAmt); 1215 KnownOne = KnownOne.lshr(ShAmt); 1216 1217 APInt HighBits = APInt::getHighBitsSet(BitWidth, ShAmt); 1218 KnownZero |= HighBits; // High bits known zero. 1219 } 1220 break; 1221 case ISD::SRA: 1222 // If this is an arithmetic shift right and only the low-bit is set, we can 1223 // always convert this into a logical shr, even if the shift amount is 1224 // variable. The low bit of the shift cannot be an input sign bit unless 1225 // the shift amount is >= the size of the datatype, which is undefined. 1226 if (DemandedMask == 1) 1227 return TLO.CombineTo(Op, TLO.DAG.getNode(ISD::SRL, dl, Op.getValueType(), 1228 Op.getOperand(0), Op.getOperand(1))); 1229 1230 if (ConstantSDNode *SA = dyn_cast<ConstantSDNode>(Op.getOperand(1))) { 1231 EVT VT = Op.getValueType(); 1232 unsigned ShAmt = SA->getZExtValue(); 1233 1234 // If the shift count is an invalid immediate, don't do anything. 1235 if (ShAmt >= BitWidth) 1236 break; 1237 1238 APInt InDemandedMask = (NewMask << ShAmt); 1239 1240 // If any of the demanded bits are produced by the sign extension, we also 1241 // demand the input sign bit. 1242 APInt HighBits = APInt::getHighBitsSet(BitWidth, ShAmt); 1243 if (HighBits.intersects(NewMask)) 1244 InDemandedMask |= APInt::getSignBit(VT.getSizeInBits()); 1245 1246 if (SimplifyDemandedBits(Op.getOperand(0), InDemandedMask, 1247 KnownZero, KnownOne, TLO, Depth+1)) 1248 return true; 1249 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?"); 1250 KnownZero = KnownZero.lshr(ShAmt); 1251 KnownOne = KnownOne.lshr(ShAmt); 1252 1253 // Handle the sign bit, adjusted to where it is now in the mask. 1254 APInt SignBit = APInt::getSignBit(BitWidth).lshr(ShAmt); 1255 1256 // If the input sign bit is known to be zero, or if none of the top bits 1257 // are demanded, turn this into an unsigned shift right. 1258 if (KnownZero.intersects(SignBit) || (HighBits & ~NewMask) == HighBits) { 1259 return TLO.CombineTo(Op, TLO.DAG.getNode(ISD::SRL, dl, VT, 1260 Op.getOperand(0), 1261 Op.getOperand(1))); 1262 } else if (KnownOne.intersects(SignBit)) { // New bits are known one. 1263 KnownOne |= HighBits; 1264 } 1265 } 1266 break; 1267 case ISD::SIGN_EXTEND_INREG: { 1268 EVT EVT = cast<VTSDNode>(Op.getOperand(1))->getVT(); 1269 1270 // Sign extension. Compute the demanded bits in the result that are not 1271 // present in the input. 1272 APInt NewBits = APInt::getHighBitsSet(BitWidth, 1273 BitWidth - EVT.getSizeInBits()) & 1274 NewMask; 1275 1276 // If none of the extended bits are demanded, eliminate the sextinreg. 1277 if (NewBits == 0) 1278 return TLO.CombineTo(Op, Op.getOperand(0)); 1279 1280 APInt InSignBit = APInt::getSignBit(EVT.getSizeInBits()); 1281 InSignBit.zext(BitWidth); 1282 APInt InputDemandedBits = APInt::getLowBitsSet(BitWidth, 1283 EVT.getSizeInBits()) & 1284 NewMask; 1285 1286 // Since the sign extended bits are demanded, we know that the sign 1287 // bit is demanded. 1288 InputDemandedBits |= InSignBit; 1289 1290 if (SimplifyDemandedBits(Op.getOperand(0), InputDemandedBits, 1291 KnownZero, KnownOne, TLO, Depth+1)) 1292 return true; 1293 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?"); 1294 1295 // If the sign bit of the input is known set or clear, then we know the 1296 // top bits of the result. 1297 1298 // If the input sign bit is known zero, convert this into a zero extension. 1299 if (KnownZero.intersects(InSignBit)) 1300 return TLO.CombineTo(Op, 1301 TLO.DAG.getZeroExtendInReg(Op.getOperand(0),dl,EVT)); 1302 1303 if (KnownOne.intersects(InSignBit)) { // Input sign bit known set 1304 KnownOne |= NewBits; 1305 KnownZero &= ~NewBits; 1306 } else { // Input sign bit unknown 1307 KnownZero &= ~NewBits; 1308 KnownOne &= ~NewBits; 1309 } 1310 break; 1311 } 1312 case ISD::ZERO_EXTEND: { 1313 unsigned OperandBitWidth = Op.getOperand(0).getValueSizeInBits(); 1314 APInt InMask = NewMask; 1315 InMask.trunc(OperandBitWidth); 1316 1317 // If none of the top bits are demanded, convert this into an any_extend. 1318 APInt NewBits = 1319 APInt::getHighBitsSet(BitWidth, BitWidth - OperandBitWidth) & NewMask; 1320 if (!NewBits.intersects(NewMask)) 1321 return TLO.CombineTo(Op, TLO.DAG.getNode(ISD::ANY_EXTEND, dl, 1322 Op.getValueType(), 1323 Op.getOperand(0))); 1324 1325 if (SimplifyDemandedBits(Op.getOperand(0), InMask, 1326 KnownZero, KnownOne, TLO, Depth+1)) 1327 return true; 1328 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?"); 1329 KnownZero.zext(BitWidth); 1330 KnownOne.zext(BitWidth); 1331 KnownZero |= NewBits; 1332 break; 1333 } 1334 case ISD::SIGN_EXTEND: { 1335 EVT InVT = Op.getOperand(0).getValueType(); 1336 unsigned InBits = InVT.getSizeInBits(); 1337 APInt InMask = APInt::getLowBitsSet(BitWidth, InBits); 1338 APInt InSignBit = APInt::getBitsSet(BitWidth, InBits - 1, InBits); 1339 APInt NewBits = ~InMask & NewMask; 1340 1341 // If none of the top bits are demanded, convert this into an any_extend. 1342 if (NewBits == 0) 1343 return TLO.CombineTo(Op,TLO.DAG.getNode(ISD::ANY_EXTEND, dl, 1344 Op.getValueType(), 1345 Op.getOperand(0))); 1346 1347 // Since some of the sign extended bits are demanded, we know that the sign 1348 // bit is demanded. 1349 APInt InDemandedBits = InMask & NewMask; 1350 InDemandedBits |= InSignBit; 1351 InDemandedBits.trunc(InBits); 1352 1353 if (SimplifyDemandedBits(Op.getOperand(0), InDemandedBits, KnownZero, 1354 KnownOne, TLO, Depth+1)) 1355 return true; 1356 KnownZero.zext(BitWidth); 1357 KnownOne.zext(BitWidth); 1358 1359 // If the sign bit is known zero, convert this to a zero extend. 1360 if (KnownZero.intersects(InSignBit)) 1361 return TLO.CombineTo(Op, TLO.DAG.getNode(ISD::ZERO_EXTEND, dl, 1362 Op.getValueType(), 1363 Op.getOperand(0))); 1364 1365 // If the sign bit is known one, the top bits match. 1366 if (KnownOne.intersects(InSignBit)) { 1367 KnownOne |= NewBits; 1368 KnownZero &= ~NewBits; 1369 } else { // Otherwise, top bits aren't known. 1370 KnownOne &= ~NewBits; 1371 KnownZero &= ~NewBits; 1372 } 1373 break; 1374 } 1375 case ISD::ANY_EXTEND: { 1376 unsigned OperandBitWidth = Op.getOperand(0).getValueSizeInBits(); 1377 APInt InMask = NewMask; 1378 InMask.trunc(OperandBitWidth); 1379 if (SimplifyDemandedBits(Op.getOperand(0), InMask, 1380 KnownZero, KnownOne, TLO, Depth+1)) 1381 return true; 1382 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?"); 1383 KnownZero.zext(BitWidth); 1384 KnownOne.zext(BitWidth); 1385 break; 1386 } 1387 case ISD::TRUNCATE: { 1388 // Simplify the input, using demanded bit information, and compute the known 1389 // zero/one bits live out. 1390 APInt TruncMask = NewMask; 1391 TruncMask.zext(Op.getOperand(0).getValueSizeInBits()); 1392 if (SimplifyDemandedBits(Op.getOperand(0), TruncMask, 1393 KnownZero, KnownOne, TLO, Depth+1)) 1394 return true; 1395 KnownZero.trunc(BitWidth); 1396 KnownOne.trunc(BitWidth); 1397 1398 // If the input is only used by this truncate, see if we can shrink it based 1399 // on the known demanded bits. 1400 if (Op.getOperand(0).getNode()->hasOneUse()) { 1401 SDValue In = Op.getOperand(0); 1402 unsigned InBitWidth = In.getValueSizeInBits(); 1403 switch (In.getOpcode()) { 1404 default: break; 1405 case ISD::SRL: 1406 // Shrink SRL by a constant if none of the high bits shifted in are 1407 // demanded. 1408 if (ConstantSDNode *ShAmt = dyn_cast<ConstantSDNode>(In.getOperand(1))){ 1409 APInt HighBits = APInt::getHighBitsSet(InBitWidth, 1410 InBitWidth - BitWidth); 1411 HighBits = HighBits.lshr(ShAmt->getZExtValue()); 1412 HighBits.trunc(BitWidth); 1413 1414 if (ShAmt->getZExtValue() < BitWidth && !(HighBits & NewMask)) { 1415 // None of the shifted in bits are needed. Add a truncate of the 1416 // shift input, then shift it. 1417 SDValue NewTrunc = TLO.DAG.getNode(ISD::TRUNCATE, dl, 1418 Op.getValueType(), 1419 In.getOperand(0)); 1420 return TLO.CombineTo(Op, TLO.DAG.getNode(ISD::SRL, dl, 1421 Op.getValueType(), 1422 NewTrunc, 1423 In.getOperand(1))); 1424 } 1425 } 1426 break; 1427 } 1428 } 1429 1430 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?"); 1431 break; 1432 } 1433 case ISD::AssertZext: { 1434 EVT VT = cast<VTSDNode>(Op.getOperand(1))->getVT(); 1435 APInt InMask = APInt::getLowBitsSet(BitWidth, 1436 VT.getSizeInBits()); 1437 if (SimplifyDemandedBits(Op.getOperand(0), InMask & NewMask, 1438 KnownZero, KnownOne, TLO, Depth+1)) 1439 return true; 1440 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?"); 1441 KnownZero |= ~InMask & NewMask; 1442 break; 1443 } 1444 case ISD::BIT_CONVERT: 1445#if 0 1446 // If this is an FP->Int bitcast and if the sign bit is the only thing that 1447 // is demanded, turn this into a FGETSIGN. 1448 if (NewMask == EVT::getIntegerVTSignBit(Op.getValueType()) && 1449 MVT::isFloatingPoint(Op.getOperand(0).getValueType()) && 1450 !MVT::isVector(Op.getOperand(0).getValueType())) { 1451 // Only do this xform if FGETSIGN is valid or if before legalize. 1452 if (!TLO.AfterLegalize || 1453 isOperationLegal(ISD::FGETSIGN, Op.getValueType())) { 1454 // Make a FGETSIGN + SHL to move the sign bit into the appropriate 1455 // place. We expect the SHL to be eliminated by other optimizations. 1456 SDValue Sign = TLO.DAG.getNode(ISD::FGETSIGN, Op.getValueType(), 1457 Op.getOperand(0)); 1458 unsigned ShVal = Op.getValueType().getSizeInBits()-1; 1459 SDValue ShAmt = TLO.DAG.getConstant(ShVal, getShiftAmountTy()); 1460 return TLO.CombineTo(Op, TLO.DAG.getNode(ISD::SHL, Op.getValueType(), 1461 Sign, ShAmt)); 1462 } 1463 } 1464#endif 1465 break; 1466 case ISD::ADD: 1467 case ISD::MUL: 1468 case ISD::SUB: { 1469 // Add, Sub, and Mul don't demand any bits in positions beyond that 1470 // of the highest bit demanded of them. 1471 APInt LoMask = APInt::getLowBitsSet(BitWidth, 1472 BitWidth - NewMask.countLeadingZeros()); 1473 if (SimplifyDemandedBits(Op.getOperand(0), LoMask, KnownZero2, 1474 KnownOne2, TLO, Depth+1)) 1475 return true; 1476 if (SimplifyDemandedBits(Op.getOperand(1), LoMask, KnownZero2, 1477 KnownOne2, TLO, Depth+1)) 1478 return true; 1479 // See if the operation should be performed at a smaller bit width. 1480 if (TLO.ShrinkDemandedOp(Op, BitWidth, NewMask, dl)) 1481 return true; 1482 } 1483 // FALL THROUGH 1484 default: 1485 // Just use ComputeMaskedBits to compute output bits. 1486 TLO.DAG.ComputeMaskedBits(Op, NewMask, KnownZero, KnownOne, Depth); 1487 break; 1488 } 1489 1490 // If we know the value of all of the demanded bits, return this as a 1491 // constant. 1492 if ((NewMask & (KnownZero|KnownOne)) == NewMask) 1493 return TLO.CombineTo(Op, TLO.DAG.getConstant(KnownOne, Op.getValueType())); 1494 1495 return false; 1496} 1497 1498/// computeMaskedBitsForTargetNode - Determine which of the bits specified 1499/// in Mask are known to be either zero or one and return them in the 1500/// KnownZero/KnownOne bitsets. 1501void TargetLowering::computeMaskedBitsForTargetNode(const SDValue Op, 1502 const APInt &Mask, 1503 APInt &KnownZero, 1504 APInt &KnownOne, 1505 const SelectionDAG &DAG, 1506 unsigned Depth) const { 1507 assert((Op.getOpcode() >= ISD::BUILTIN_OP_END || 1508 Op.getOpcode() == ISD::INTRINSIC_WO_CHAIN || 1509 Op.getOpcode() == ISD::INTRINSIC_W_CHAIN || 1510 Op.getOpcode() == ISD::INTRINSIC_VOID) && 1511 "Should use MaskedValueIsZero if you don't know whether Op" 1512 " is a target node!"); 1513 KnownZero = KnownOne = APInt(Mask.getBitWidth(), 0); 1514} 1515 1516/// ComputeNumSignBitsForTargetNode - This method can be implemented by 1517/// targets that want to expose additional information about sign bits to the 1518/// DAG Combiner. 1519unsigned TargetLowering::ComputeNumSignBitsForTargetNode(SDValue Op, 1520 unsigned Depth) const { 1521 assert((Op.getOpcode() >= ISD::BUILTIN_OP_END || 1522 Op.getOpcode() == ISD::INTRINSIC_WO_CHAIN || 1523 Op.getOpcode() == ISD::INTRINSIC_W_CHAIN || 1524 Op.getOpcode() == ISD::INTRINSIC_VOID) && 1525 "Should use ComputeNumSignBits if you don't know whether Op" 1526 " is a target node!"); 1527 return 1; 1528} 1529 1530/// ValueHasExactlyOneBitSet - Test if the given value is known to have exactly 1531/// one bit set. This differs from ComputeMaskedBits in that it doesn't need to 1532/// determine which bit is set. 1533/// 1534static bool ValueHasExactlyOneBitSet(SDValue Val, const SelectionDAG &DAG) { 1535 // A left-shift of a constant one will have exactly one bit set, because 1536 // shifting the bit off the end is undefined. 1537 if (Val.getOpcode() == ISD::SHL) 1538 if (ConstantSDNode *C = 1539 dyn_cast<ConstantSDNode>(Val.getNode()->getOperand(0))) 1540 if (C->getAPIntValue() == 1) 1541 return true; 1542 1543 // Similarly, a right-shift of a constant sign-bit will have exactly 1544 // one bit set. 1545 if (Val.getOpcode() == ISD::SRL) 1546 if (ConstantSDNode *C = 1547 dyn_cast<ConstantSDNode>(Val.getNode()->getOperand(0))) 1548 if (C->getAPIntValue().isSignBit()) 1549 return true; 1550 1551 // More could be done here, though the above checks are enough 1552 // to handle some common cases. 1553 1554 // Fall back to ComputeMaskedBits to catch other known cases. 1555 EVT OpVT = Val.getValueType(); 1556 unsigned BitWidth = OpVT.getSizeInBits(); 1557 APInt Mask = APInt::getAllOnesValue(BitWidth); 1558 APInt KnownZero, KnownOne; 1559 DAG.ComputeMaskedBits(Val, Mask, KnownZero, KnownOne); 1560 return (KnownZero.countPopulation() == BitWidth - 1) && 1561 (KnownOne.countPopulation() == 1); 1562} 1563 1564/// SimplifySetCC - Try to simplify a setcc built with the specified operands 1565/// and cc. If it is unable to simplify it, return a null SDValue. 1566SDValue 1567TargetLowering::SimplifySetCC(EVT VT, SDValue N0, SDValue N1, 1568 ISD::CondCode Cond, bool foldBooleans, 1569 DAGCombinerInfo &DCI, DebugLoc dl) const { 1570 SelectionDAG &DAG = DCI.DAG; 1571 LLVMContext &Context = *DAG.getContext(); 1572 1573 // These setcc operations always fold. 1574 switch (Cond) { 1575 default: break; 1576 case ISD::SETFALSE: 1577 case ISD::SETFALSE2: return DAG.getConstant(0, VT); 1578 case ISD::SETTRUE: 1579 case ISD::SETTRUE2: return DAG.getConstant(1, VT); 1580 } 1581 1582 if (isa<ConstantSDNode>(N0.getNode())) { 1583 // Ensure that the constant occurs on the RHS, and fold constant 1584 // comparisons. 1585 return DAG.getSetCC(dl, VT, N1, N0, ISD::getSetCCSwappedOperands(Cond)); 1586 } 1587 1588 if (ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1.getNode())) { 1589 const APInt &C1 = N1C->getAPIntValue(); 1590 1591 // If the LHS is '(srl (ctlz x), 5)', the RHS is 0/1, and this is an 1592 // equality comparison, then we're just comparing whether X itself is 1593 // zero. 1594 if (N0.getOpcode() == ISD::SRL && (C1 == 0 || C1 == 1) && 1595 N0.getOperand(0).getOpcode() == ISD::CTLZ && 1596 N0.getOperand(1).getOpcode() == ISD::Constant) { 1597 unsigned ShAmt = cast<ConstantSDNode>(N0.getOperand(1))->getZExtValue(); 1598 if ((Cond == ISD::SETEQ || Cond == ISD::SETNE) && 1599 ShAmt == Log2_32(N0.getValueType().getSizeInBits())) { 1600 if ((C1 == 0) == (Cond == ISD::SETEQ)) { 1601 // (srl (ctlz x), 5) == 0 -> X != 0 1602 // (srl (ctlz x), 5) != 1 -> X != 0 1603 Cond = ISD::SETNE; 1604 } else { 1605 // (srl (ctlz x), 5) != 0 -> X == 0 1606 // (srl (ctlz x), 5) == 1 -> X == 0 1607 Cond = ISD::SETEQ; 1608 } 1609 SDValue Zero = DAG.getConstant(0, N0.getValueType()); 1610 return DAG.getSetCC(dl, VT, N0.getOperand(0).getOperand(0), 1611 Zero, Cond); 1612 } 1613 } 1614 1615 // If the LHS is '(and load, const)', the RHS is 0, 1616 // the test is for equality or unsigned, and all 1 bits of the const are 1617 // in the same partial word, see if we can shorten the load. 1618 if (DCI.isBeforeLegalize() && 1619 N0.getOpcode() == ISD::AND && C1 == 0 && 1620 N0.getNode()->hasOneUse() && 1621 isa<LoadSDNode>(N0.getOperand(0)) && 1622 N0.getOperand(0).getNode()->hasOneUse() && 1623 isa<ConstantSDNode>(N0.getOperand(1))) { 1624 LoadSDNode *Lod = cast<LoadSDNode>(N0.getOperand(0)); 1625 uint64_t bestMask = 0; 1626 unsigned bestWidth = 0, bestOffset = 0; 1627 if (!Lod->isVolatile() && Lod->isUnindexed() && 1628 // FIXME: This uses getZExtValue() below so it only works on i64 and 1629 // below. 1630 N0.getValueType().getSizeInBits() <= 64) { 1631 unsigned origWidth = N0.getValueType().getSizeInBits(); 1632 // We can narrow (e.g.) 16-bit extending loads on 32-bit target to 1633 // 8 bits, but have to be careful... 1634 if (Lod->getExtensionType() != ISD::NON_EXTLOAD) 1635 origWidth = Lod->getMemoryVT().getSizeInBits(); 1636 uint64_t Mask =cast<ConstantSDNode>(N0.getOperand(1))->getZExtValue(); 1637 for (unsigned width = origWidth / 2; width>=8; width /= 2) { 1638 uint64_t newMask = (1ULL << width) - 1; 1639 for (unsigned offset=0; offset<origWidth/width; offset++) { 1640 if ((newMask & Mask) == Mask) { 1641 if (!TD->isLittleEndian()) 1642 bestOffset = (origWidth/width - offset - 1) * (width/8); 1643 else 1644 bestOffset = (uint64_t)offset * (width/8); 1645 bestMask = Mask >> (offset * (width/8) * 8); 1646 bestWidth = width; 1647 break; 1648 } 1649 newMask = newMask << width; 1650 } 1651 } 1652 } 1653 if (bestWidth) { 1654 EVT newVT = EVT::getIntegerVT(Context, bestWidth); 1655 if (newVT.isRound()) { 1656 EVT PtrType = Lod->getOperand(1).getValueType(); 1657 SDValue Ptr = Lod->getBasePtr(); 1658 if (bestOffset != 0) 1659 Ptr = DAG.getNode(ISD::ADD, dl, PtrType, Lod->getBasePtr(), 1660 DAG.getConstant(bestOffset, PtrType)); 1661 unsigned NewAlign = MinAlign(Lod->getAlignment(), bestOffset); 1662 SDValue NewLoad = DAG.getLoad(newVT, dl, Lod->getChain(), Ptr, 1663 Lod->getSrcValue(), 1664 Lod->getSrcValueOffset() + bestOffset, 1665 false, NewAlign); 1666 return DAG.getSetCC(dl, VT, 1667 DAG.getNode(ISD::AND, dl, newVT, NewLoad, 1668 DAG.getConstant(bestMask, newVT)), 1669 DAG.getConstant(0LL, newVT), Cond); 1670 } 1671 } 1672 } 1673 1674 // If the LHS is a ZERO_EXTEND, perform the comparison on the input. 1675 if (N0.getOpcode() == ISD::ZERO_EXTEND) { 1676 unsigned InSize = N0.getOperand(0).getValueType().getSizeInBits(); 1677 1678 // If the comparison constant has bits in the upper part, the 1679 // zero-extended value could never match. 1680 if (C1.intersects(APInt::getHighBitsSet(C1.getBitWidth(), 1681 C1.getBitWidth() - InSize))) { 1682 switch (Cond) { 1683 case ISD::SETUGT: 1684 case ISD::SETUGE: 1685 case ISD::SETEQ: return DAG.getConstant(0, VT); 1686 case ISD::SETULT: 1687 case ISD::SETULE: 1688 case ISD::SETNE: return DAG.getConstant(1, VT); 1689 case ISD::SETGT: 1690 case ISD::SETGE: 1691 // True if the sign bit of C1 is set. 1692 return DAG.getConstant(C1.isNegative(), VT); 1693 case ISD::SETLT: 1694 case ISD::SETLE: 1695 // True if the sign bit of C1 isn't set. 1696 return DAG.getConstant(C1.isNonNegative(), VT); 1697 default: 1698 break; 1699 } 1700 } 1701 1702 // Otherwise, we can perform the comparison with the low bits. 1703 switch (Cond) { 1704 case ISD::SETEQ: 1705 case ISD::SETNE: 1706 case ISD::SETUGT: 1707 case ISD::SETUGE: 1708 case ISD::SETULT: 1709 case ISD::SETULE: { 1710 EVT newVT = N0.getOperand(0).getValueType(); 1711 if (DCI.isBeforeLegalizeOps() || 1712 (isOperationLegal(ISD::SETCC, newVT) && 1713 getCondCodeAction(Cond, newVT)==Legal)) 1714 return DAG.getSetCC(dl, VT, N0.getOperand(0), 1715 DAG.getConstant(APInt(C1).trunc(InSize), newVT), 1716 Cond); 1717 break; 1718 } 1719 default: 1720 break; // todo, be more careful with signed comparisons 1721 } 1722 } else if (N0.getOpcode() == ISD::SIGN_EXTEND_INREG && 1723 (Cond == ISD::SETEQ || Cond == ISD::SETNE)) { 1724 EVT ExtSrcTy = cast<VTSDNode>(N0.getOperand(1))->getVT(); 1725 unsigned ExtSrcTyBits = ExtSrcTy.getSizeInBits(); 1726 EVT ExtDstTy = N0.getValueType(); 1727 unsigned ExtDstTyBits = ExtDstTy.getSizeInBits(); 1728 1729 // If the extended part has any inconsistent bits, it cannot ever 1730 // compare equal. In other words, they have to be all ones or all 1731 // zeros. 1732 APInt ExtBits = 1733 APInt::getHighBitsSet(ExtDstTyBits, ExtDstTyBits - ExtSrcTyBits); 1734 if ((C1 & ExtBits) != 0 && (C1 & ExtBits) != ExtBits) 1735 return DAG.getConstant(Cond == ISD::SETNE, VT); 1736 1737 SDValue ZextOp; 1738 EVT Op0Ty = N0.getOperand(0).getValueType(); 1739 if (Op0Ty == ExtSrcTy) { 1740 ZextOp = N0.getOperand(0); 1741 } else { 1742 APInt Imm = APInt::getLowBitsSet(ExtDstTyBits, ExtSrcTyBits); 1743 ZextOp = DAG.getNode(ISD::AND, dl, Op0Ty, N0.getOperand(0), 1744 DAG.getConstant(Imm, Op0Ty)); 1745 } 1746 if (!DCI.isCalledByLegalizer()) 1747 DCI.AddToWorklist(ZextOp.getNode()); 1748 // Otherwise, make this a use of a zext. 1749 return DAG.getSetCC(dl, VT, ZextOp, 1750 DAG.getConstant(C1 & APInt::getLowBitsSet( 1751 ExtDstTyBits, 1752 ExtSrcTyBits), 1753 ExtDstTy), 1754 Cond); 1755 } else if ((N1C->isNullValue() || N1C->getAPIntValue() == 1) && 1756 (Cond == ISD::SETEQ || Cond == ISD::SETNE)) { 1757 1758 // SETCC (SETCC), [0|1], [EQ|NE] -> SETCC 1759 if (N0.getOpcode() == ISD::SETCC) { 1760 bool TrueWhenTrue = (Cond == ISD::SETEQ) ^ (N1C->getZExtValue() != 1); 1761 if (TrueWhenTrue) 1762 return N0; 1763 1764 // Invert the condition. 1765 ISD::CondCode CC = cast<CondCodeSDNode>(N0.getOperand(2))->get(); 1766 CC = ISD::getSetCCInverse(CC, 1767 N0.getOperand(0).getValueType().isInteger()); 1768 return DAG.getSetCC(dl, VT, N0.getOperand(0), N0.getOperand(1), CC); 1769 } 1770 1771 if ((N0.getOpcode() == ISD::XOR || 1772 (N0.getOpcode() == ISD::AND && 1773 N0.getOperand(0).getOpcode() == ISD::XOR && 1774 N0.getOperand(1) == N0.getOperand(0).getOperand(1))) && 1775 isa<ConstantSDNode>(N0.getOperand(1)) && 1776 cast<ConstantSDNode>(N0.getOperand(1))->getAPIntValue() == 1) { 1777 // If this is (X^1) == 0/1, swap the RHS and eliminate the xor. We 1778 // can only do this if the top bits are known zero. 1779 unsigned BitWidth = N0.getValueSizeInBits(); 1780 if (DAG.MaskedValueIsZero(N0, 1781 APInt::getHighBitsSet(BitWidth, 1782 BitWidth-1))) { 1783 // Okay, get the un-inverted input value. 1784 SDValue Val; 1785 if (N0.getOpcode() == ISD::XOR) 1786 Val = N0.getOperand(0); 1787 else { 1788 assert(N0.getOpcode() == ISD::AND && 1789 N0.getOperand(0).getOpcode() == ISD::XOR); 1790 // ((X^1)&1)^1 -> X & 1 1791 Val = DAG.getNode(ISD::AND, dl, N0.getValueType(), 1792 N0.getOperand(0).getOperand(0), 1793 N0.getOperand(1)); 1794 } 1795 return DAG.getSetCC(dl, VT, Val, N1, 1796 Cond == ISD::SETEQ ? ISD::SETNE : ISD::SETEQ); 1797 } 1798 } 1799 } 1800 1801 APInt MinVal, MaxVal; 1802 unsigned OperandBitSize = N1C->getValueType(0).getSizeInBits(); 1803 if (ISD::isSignedIntSetCC(Cond)) { 1804 MinVal = APInt::getSignedMinValue(OperandBitSize); 1805 MaxVal = APInt::getSignedMaxValue(OperandBitSize); 1806 } else { 1807 MinVal = APInt::getMinValue(OperandBitSize); 1808 MaxVal = APInt::getMaxValue(OperandBitSize); 1809 } 1810 1811 // Canonicalize GE/LE comparisons to use GT/LT comparisons. 1812 if (Cond == ISD::SETGE || Cond == ISD::SETUGE) { 1813 if (C1 == MinVal) return DAG.getConstant(1, VT); // X >= MIN --> true 1814 // X >= C0 --> X > (C0-1) 1815 return DAG.getSetCC(dl, VT, N0, 1816 DAG.getConstant(C1-1, N1.getValueType()), 1817 (Cond == ISD::SETGE) ? ISD::SETGT : ISD::SETUGT); 1818 } 1819 1820 if (Cond == ISD::SETLE || Cond == ISD::SETULE) { 1821 if (C1 == MaxVal) return DAG.getConstant(1, VT); // X <= MAX --> true 1822 // X <= C0 --> X < (C0+1) 1823 return DAG.getSetCC(dl, VT, N0, 1824 DAG.getConstant(C1+1, N1.getValueType()), 1825 (Cond == ISD::SETLE) ? ISD::SETLT : ISD::SETULT); 1826 } 1827 1828 if ((Cond == ISD::SETLT || Cond == ISD::SETULT) && C1 == MinVal) 1829 return DAG.getConstant(0, VT); // X < MIN --> false 1830 if ((Cond == ISD::SETGE || Cond == ISD::SETUGE) && C1 == MinVal) 1831 return DAG.getConstant(1, VT); // X >= MIN --> true 1832 if ((Cond == ISD::SETGT || Cond == ISD::SETUGT) && C1 == MaxVal) 1833 return DAG.getConstant(0, VT); // X > MAX --> false 1834 if ((Cond == ISD::SETLE || Cond == ISD::SETULE) && C1 == MaxVal) 1835 return DAG.getConstant(1, VT); // X <= MAX --> true 1836 1837 // Canonicalize setgt X, Min --> setne X, Min 1838 if ((Cond == ISD::SETGT || Cond == ISD::SETUGT) && C1 == MinVal) 1839 return DAG.getSetCC(dl, VT, N0, N1, ISD::SETNE); 1840 // Canonicalize setlt X, Max --> setne X, Max 1841 if ((Cond == ISD::SETLT || Cond == ISD::SETULT) && C1 == MaxVal) 1842 return DAG.getSetCC(dl, VT, N0, N1, ISD::SETNE); 1843 1844 // If we have setult X, 1, turn it into seteq X, 0 1845 if ((Cond == ISD::SETLT || Cond == ISD::SETULT) && C1 == MinVal+1) 1846 return DAG.getSetCC(dl, VT, N0, 1847 DAG.getConstant(MinVal, N0.getValueType()), 1848 ISD::SETEQ); 1849 // If we have setugt X, Max-1, turn it into seteq X, Max 1850 else if ((Cond == ISD::SETGT || Cond == ISD::SETUGT) && C1 == MaxVal-1) 1851 return DAG.getSetCC(dl, VT, N0, 1852 DAG.getConstant(MaxVal, N0.getValueType()), 1853 ISD::SETEQ); 1854 1855 // If we have "setcc X, C0", check to see if we can shrink the immediate 1856 // by changing cc. 1857 1858 // SETUGT X, SINTMAX -> SETLT X, 0 1859 if (Cond == ISD::SETUGT && 1860 C1 == APInt::getSignedMaxValue(OperandBitSize)) 1861 return DAG.getSetCC(dl, VT, N0, 1862 DAG.getConstant(0, N1.getValueType()), 1863 ISD::SETLT); 1864 1865 // SETULT X, SINTMIN -> SETGT X, -1 1866 if (Cond == ISD::SETULT && 1867 C1 == APInt::getSignedMinValue(OperandBitSize)) { 1868 SDValue ConstMinusOne = 1869 DAG.getConstant(APInt::getAllOnesValue(OperandBitSize), 1870 N1.getValueType()); 1871 return DAG.getSetCC(dl, VT, N0, ConstMinusOne, ISD::SETGT); 1872 } 1873 1874 // Fold bit comparisons when we can. 1875 if ((Cond == ISD::SETEQ || Cond == ISD::SETNE) && 1876 VT == N0.getValueType() && N0.getOpcode() == ISD::AND) 1877 if (ConstantSDNode *AndRHS = 1878 dyn_cast<ConstantSDNode>(N0.getOperand(1))) { 1879 EVT ShiftTy = DCI.isBeforeLegalize() ? 1880 getPointerTy() : getShiftAmountTy(); 1881 if (Cond == ISD::SETNE && C1 == 0) {// (X & 8) != 0 --> (X & 8) >> 3 1882 // Perform the xform if the AND RHS is a single bit. 1883 if (isPowerOf2_64(AndRHS->getZExtValue())) { 1884 return DAG.getNode(ISD::SRL, dl, VT, N0, 1885 DAG.getConstant(Log2_64(AndRHS->getZExtValue()), 1886 ShiftTy)); 1887 } 1888 } else if (Cond == ISD::SETEQ && C1 == AndRHS->getZExtValue()) { 1889 // (X & 8) == 8 --> (X & 8) >> 3 1890 // Perform the xform if C1 is a single bit. 1891 if (C1.isPowerOf2()) { 1892 return DAG.getNode(ISD::SRL, dl, VT, N0, 1893 DAG.getConstant(C1.logBase2(), ShiftTy)); 1894 } 1895 } 1896 } 1897 } 1898 1899 if (isa<ConstantFPSDNode>(N0.getNode())) { 1900 // Constant fold or commute setcc. 1901 SDValue O = DAG.FoldSetCC(VT, N0, N1, Cond, dl); 1902 if (O.getNode()) return O; 1903 } else if (ConstantFPSDNode *CFP = dyn_cast<ConstantFPSDNode>(N1.getNode())) { 1904 // If the RHS of an FP comparison is a constant, simplify it away in 1905 // some cases. 1906 if (CFP->getValueAPF().isNaN()) { 1907 // If an operand is known to be a nan, we can fold it. 1908 switch (ISD::getUnorderedFlavor(Cond)) { 1909 default: llvm_unreachable("Unknown flavor!"); 1910 case 0: // Known false. 1911 return DAG.getConstant(0, VT); 1912 case 1: // Known true. 1913 return DAG.getConstant(1, VT); 1914 case 2: // Undefined. 1915 return DAG.getUNDEF(VT); 1916 } 1917 } 1918 1919 // Otherwise, we know the RHS is not a NaN. Simplify the node to drop the 1920 // constant if knowing that the operand is non-nan is enough. We prefer to 1921 // have SETO(x,x) instead of SETO(x, 0.0) because this avoids having to 1922 // materialize 0.0. 1923 if (Cond == ISD::SETO || Cond == ISD::SETUO) 1924 return DAG.getSetCC(dl, VT, N0, N0, Cond); 1925 1926 // If the condition is not legal, see if we can find an equivalent one 1927 // which is legal. 1928 if (!isCondCodeLegal(Cond, N0.getValueType())) { 1929 // If the comparison was an awkward floating-point == or != and one of 1930 // the comparison operands is infinity or negative infinity, convert the 1931 // condition to a less-awkward <= or >=. 1932 if (CFP->getValueAPF().isInfinity()) { 1933 if (CFP->getValueAPF().isNegative()) { 1934 if (Cond == ISD::SETOEQ && 1935 isCondCodeLegal(ISD::SETOLE, N0.getValueType())) 1936 return DAG.getSetCC(dl, VT, N0, N1, ISD::SETOLE); 1937 if (Cond == ISD::SETUEQ && 1938 isCondCodeLegal(ISD::SETOLE, N0.getValueType())) 1939 return DAG.getSetCC(dl, VT, N0, N1, ISD::SETULE); 1940 if (Cond == ISD::SETUNE && 1941 isCondCodeLegal(ISD::SETUGT, N0.getValueType())) 1942 return DAG.getSetCC(dl, VT, N0, N1, ISD::SETUGT); 1943 if (Cond == ISD::SETONE && 1944 isCondCodeLegal(ISD::SETUGT, N0.getValueType())) 1945 return DAG.getSetCC(dl, VT, N0, N1, ISD::SETOGT); 1946 } else { 1947 if (Cond == ISD::SETOEQ && 1948 isCondCodeLegal(ISD::SETOGE, N0.getValueType())) 1949 return DAG.getSetCC(dl, VT, N0, N1, ISD::SETOGE); 1950 if (Cond == ISD::SETUEQ && 1951 isCondCodeLegal(ISD::SETOGE, N0.getValueType())) 1952 return DAG.getSetCC(dl, VT, N0, N1, ISD::SETUGE); 1953 if (Cond == ISD::SETUNE && 1954 isCondCodeLegal(ISD::SETULT, N0.getValueType())) 1955 return DAG.getSetCC(dl, VT, N0, N1, ISD::SETULT); 1956 if (Cond == ISD::SETONE && 1957 isCondCodeLegal(ISD::SETULT, N0.getValueType())) 1958 return DAG.getSetCC(dl, VT, N0, N1, ISD::SETOLT); 1959 } 1960 } 1961 } 1962 } 1963 1964 if (N0 == N1) { 1965 // We can always fold X == X for integer setcc's. 1966 if (N0.getValueType().isInteger()) 1967 return DAG.getConstant(ISD::isTrueWhenEqual(Cond), VT); 1968 unsigned UOF = ISD::getUnorderedFlavor(Cond); 1969 if (UOF == 2) // FP operators that are undefined on NaNs. 1970 return DAG.getConstant(ISD::isTrueWhenEqual(Cond), VT); 1971 if (UOF == unsigned(ISD::isTrueWhenEqual(Cond))) 1972 return DAG.getConstant(UOF, VT); 1973 // Otherwise, we can't fold it. However, we can simplify it to SETUO/SETO 1974 // if it is not already. 1975 ISD::CondCode NewCond = UOF == 0 ? ISD::SETO : ISD::SETUO; 1976 if (NewCond != Cond) 1977 return DAG.getSetCC(dl, VT, N0, N1, NewCond); 1978 } 1979 1980 if ((Cond == ISD::SETEQ || Cond == ISD::SETNE) && 1981 N0.getValueType().isInteger()) { 1982 if (N0.getOpcode() == ISD::ADD || N0.getOpcode() == ISD::SUB || 1983 N0.getOpcode() == ISD::XOR) { 1984 // Simplify (X+Y) == (X+Z) --> Y == Z 1985 if (N0.getOpcode() == N1.getOpcode()) { 1986 if (N0.getOperand(0) == N1.getOperand(0)) 1987 return DAG.getSetCC(dl, VT, N0.getOperand(1), N1.getOperand(1), Cond); 1988 if (N0.getOperand(1) == N1.getOperand(1)) 1989 return DAG.getSetCC(dl, VT, N0.getOperand(0), N1.getOperand(0), Cond); 1990 if (DAG.isCommutativeBinOp(N0.getOpcode())) { 1991 // If X op Y == Y op X, try other combinations. 1992 if (N0.getOperand(0) == N1.getOperand(1)) 1993 return DAG.getSetCC(dl, VT, N0.getOperand(1), N1.getOperand(0), 1994 Cond); 1995 if (N0.getOperand(1) == N1.getOperand(0)) 1996 return DAG.getSetCC(dl, VT, N0.getOperand(0), N1.getOperand(1), 1997 Cond); 1998 } 1999 } 2000 2001 if (ConstantSDNode *RHSC = dyn_cast<ConstantSDNode>(N1)) { 2002 if (ConstantSDNode *LHSR = dyn_cast<ConstantSDNode>(N0.getOperand(1))) { 2003 // Turn (X+C1) == C2 --> X == C2-C1 2004 if (N0.getOpcode() == ISD::ADD && N0.getNode()->hasOneUse()) { 2005 return DAG.getSetCC(dl, VT, N0.getOperand(0), 2006 DAG.getConstant(RHSC->getAPIntValue()- 2007 LHSR->getAPIntValue(), 2008 N0.getValueType()), Cond); 2009 } 2010 2011 // Turn (X^C1) == C2 into X == C1^C2 iff X&~C1 = 0. 2012 if (N0.getOpcode() == ISD::XOR) 2013 // If we know that all of the inverted bits are zero, don't bother 2014 // performing the inversion. 2015 if (DAG.MaskedValueIsZero(N0.getOperand(0), ~LHSR->getAPIntValue())) 2016 return 2017 DAG.getSetCC(dl, VT, N0.getOperand(0), 2018 DAG.getConstant(LHSR->getAPIntValue() ^ 2019 RHSC->getAPIntValue(), 2020 N0.getValueType()), 2021 Cond); 2022 } 2023 2024 // Turn (C1-X) == C2 --> X == C1-C2 2025 if (ConstantSDNode *SUBC = dyn_cast<ConstantSDNode>(N0.getOperand(0))) { 2026 if (N0.getOpcode() == ISD::SUB && N0.getNode()->hasOneUse()) { 2027 return 2028 DAG.getSetCC(dl, VT, N0.getOperand(1), 2029 DAG.getConstant(SUBC->getAPIntValue() - 2030 RHSC->getAPIntValue(), 2031 N0.getValueType()), 2032 Cond); 2033 } 2034 } 2035 } 2036 2037 // Simplify (X+Z) == X --> Z == 0 2038 if (N0.getOperand(0) == N1) 2039 return DAG.getSetCC(dl, VT, N0.getOperand(1), 2040 DAG.getConstant(0, N0.getValueType()), Cond); 2041 if (N0.getOperand(1) == N1) { 2042 if (DAG.isCommutativeBinOp(N0.getOpcode())) 2043 return DAG.getSetCC(dl, VT, N0.getOperand(0), 2044 DAG.getConstant(0, N0.getValueType()), Cond); 2045 else if (N0.getNode()->hasOneUse()) { 2046 assert(N0.getOpcode() == ISD::SUB && "Unexpected operation!"); 2047 // (Z-X) == X --> Z == X<<1 2048 SDValue SH = DAG.getNode(ISD::SHL, dl, N1.getValueType(), 2049 N1, 2050 DAG.getConstant(1, getShiftAmountTy())); 2051 if (!DCI.isCalledByLegalizer()) 2052 DCI.AddToWorklist(SH.getNode()); 2053 return DAG.getSetCC(dl, VT, N0.getOperand(0), SH, Cond); 2054 } 2055 } 2056 } 2057 2058 if (N1.getOpcode() == ISD::ADD || N1.getOpcode() == ISD::SUB || 2059 N1.getOpcode() == ISD::XOR) { 2060 // Simplify X == (X+Z) --> Z == 0 2061 if (N1.getOperand(0) == N0) { 2062 return DAG.getSetCC(dl, VT, N1.getOperand(1), 2063 DAG.getConstant(0, N1.getValueType()), Cond); 2064 } else if (N1.getOperand(1) == N0) { 2065 if (DAG.isCommutativeBinOp(N1.getOpcode())) { 2066 return DAG.getSetCC(dl, VT, N1.getOperand(0), 2067 DAG.getConstant(0, N1.getValueType()), Cond); 2068 } else if (N1.getNode()->hasOneUse()) { 2069 assert(N1.getOpcode() == ISD::SUB && "Unexpected operation!"); 2070 // X == (Z-X) --> X<<1 == Z 2071 SDValue SH = DAG.getNode(ISD::SHL, dl, N1.getValueType(), N0, 2072 DAG.getConstant(1, getShiftAmountTy())); 2073 if (!DCI.isCalledByLegalizer()) 2074 DCI.AddToWorklist(SH.getNode()); 2075 return DAG.getSetCC(dl, VT, SH, N1.getOperand(0), Cond); 2076 } 2077 } 2078 } 2079 2080 // Simplify x&y == y to x&y != 0 if y has exactly one bit set. 2081 // Note that where y is variable and is known to have at most 2082 // one bit set (for example, if it is z&1) we cannot do this; 2083 // the expressions are not equivalent when y==0. 2084 if (N0.getOpcode() == ISD::AND) 2085 if (N0.getOperand(0) == N1 || N0.getOperand(1) == N1) { 2086 if (ValueHasExactlyOneBitSet(N1, DAG)) { 2087 Cond = ISD::getSetCCInverse(Cond, /*isInteger=*/true); 2088 SDValue Zero = DAG.getConstant(0, N1.getValueType()); 2089 return DAG.getSetCC(dl, VT, N0, Zero, Cond); 2090 } 2091 } 2092 if (N1.getOpcode() == ISD::AND) 2093 if (N1.getOperand(0) == N0 || N1.getOperand(1) == N0) { 2094 if (ValueHasExactlyOneBitSet(N0, DAG)) { 2095 Cond = ISD::getSetCCInverse(Cond, /*isInteger=*/true); 2096 SDValue Zero = DAG.getConstant(0, N0.getValueType()); 2097 return DAG.getSetCC(dl, VT, N1, Zero, Cond); 2098 } 2099 } 2100 } 2101 2102 // Fold away ALL boolean setcc's. 2103 SDValue Temp; 2104 if (N0.getValueType() == MVT::i1 && foldBooleans) { 2105 switch (Cond) { 2106 default: llvm_unreachable("Unknown integer setcc!"); 2107 case ISD::SETEQ: // X == Y -> ~(X^Y) 2108 Temp = DAG.getNode(ISD::XOR, dl, MVT::i1, N0, N1); 2109 N0 = DAG.getNOT(dl, Temp, MVT::i1); 2110 if (!DCI.isCalledByLegalizer()) 2111 DCI.AddToWorklist(Temp.getNode()); 2112 break; 2113 case ISD::SETNE: // X != Y --> (X^Y) 2114 N0 = DAG.getNode(ISD::XOR, dl, MVT::i1, N0, N1); 2115 break; 2116 case ISD::SETGT: // X >s Y --> X == 0 & Y == 1 --> ~X & Y 2117 case ISD::SETULT: // X <u Y --> X == 0 & Y == 1 --> ~X & Y 2118 Temp = DAG.getNOT(dl, N0, MVT::i1); 2119 N0 = DAG.getNode(ISD::AND, dl, MVT::i1, N1, Temp); 2120 if (!DCI.isCalledByLegalizer()) 2121 DCI.AddToWorklist(Temp.getNode()); 2122 break; 2123 case ISD::SETLT: // X <s Y --> X == 1 & Y == 0 --> ~Y & X 2124 case ISD::SETUGT: // X >u Y --> X == 1 & Y == 0 --> ~Y & X 2125 Temp = DAG.getNOT(dl, N1, MVT::i1); 2126 N0 = DAG.getNode(ISD::AND, dl, MVT::i1, N0, Temp); 2127 if (!DCI.isCalledByLegalizer()) 2128 DCI.AddToWorklist(Temp.getNode()); 2129 break; 2130 case ISD::SETULE: // X <=u Y --> X == 0 | Y == 1 --> ~X | Y 2131 case ISD::SETGE: // X >=s Y --> X == 0 | Y == 1 --> ~X | Y 2132 Temp = DAG.getNOT(dl, N0, MVT::i1); 2133 N0 = DAG.getNode(ISD::OR, dl, MVT::i1, N1, Temp); 2134 if (!DCI.isCalledByLegalizer()) 2135 DCI.AddToWorklist(Temp.getNode()); 2136 break; 2137 case ISD::SETUGE: // X >=u Y --> X == 1 | Y == 0 --> ~Y | X 2138 case ISD::SETLE: // X <=s Y --> X == 1 | Y == 0 --> ~Y | X 2139 Temp = DAG.getNOT(dl, N1, MVT::i1); 2140 N0 = DAG.getNode(ISD::OR, dl, MVT::i1, N0, Temp); 2141 break; 2142 } 2143 if (VT != MVT::i1) { 2144 if (!DCI.isCalledByLegalizer()) 2145 DCI.AddToWorklist(N0.getNode()); 2146 // FIXME: If running after legalize, we probably can't do this. 2147 N0 = DAG.getNode(ISD::ZERO_EXTEND, dl, VT, N0); 2148 } 2149 return N0; 2150 } 2151 2152 // Could not fold it. 2153 return SDValue(); 2154} 2155 2156/// isGAPlusOffset - Returns true (and the GlobalValue and the offset) if the 2157/// node is a GlobalAddress + offset. 2158bool TargetLowering::isGAPlusOffset(SDNode *N, GlobalValue* &GA, 2159 int64_t &Offset) const { 2160 if (isa<GlobalAddressSDNode>(N)) { 2161 GlobalAddressSDNode *GASD = cast<GlobalAddressSDNode>(N); 2162 GA = GASD->getGlobal(); 2163 Offset += GASD->getOffset(); 2164 return true; 2165 } 2166 2167 if (N->getOpcode() == ISD::ADD) { 2168 SDValue N1 = N->getOperand(0); 2169 SDValue N2 = N->getOperand(1); 2170 if (isGAPlusOffset(N1.getNode(), GA, Offset)) { 2171 ConstantSDNode *V = dyn_cast<ConstantSDNode>(N2); 2172 if (V) { 2173 Offset += V->getSExtValue(); 2174 return true; 2175 } 2176 } else if (isGAPlusOffset(N2.getNode(), GA, Offset)) { 2177 ConstantSDNode *V = dyn_cast<ConstantSDNode>(N1); 2178 if (V) { 2179 Offset += V->getSExtValue(); 2180 return true; 2181 } 2182 } 2183 } 2184 return false; 2185} 2186 2187 2188/// isConsecutiveLoad - Return true if LD is loading 'Bytes' bytes from a 2189/// location that is 'Dist' units away from the location that the 'Base' load 2190/// is loading from. 2191bool TargetLowering::isConsecutiveLoad(LoadSDNode *LD, LoadSDNode *Base, 2192 unsigned Bytes, int Dist, 2193 const MachineFrameInfo *MFI) const { 2194 if (LD->getChain() != Base->getChain()) 2195 return false; 2196 EVT VT = LD->getValueType(0); 2197 if (VT.getSizeInBits() / 8 != Bytes) 2198 return false; 2199 2200 SDValue Loc = LD->getOperand(1); 2201 SDValue BaseLoc = Base->getOperand(1); 2202 if (Loc.getOpcode() == ISD::FrameIndex) { 2203 if (BaseLoc.getOpcode() != ISD::FrameIndex) 2204 return false; 2205 int FI = cast<FrameIndexSDNode>(Loc)->getIndex(); 2206 int BFI = cast<FrameIndexSDNode>(BaseLoc)->getIndex(); 2207 int FS = MFI->getObjectSize(FI); 2208 int BFS = MFI->getObjectSize(BFI); 2209 if (FS != BFS || FS != (int)Bytes) return false; 2210 return MFI->getObjectOffset(FI) == (MFI->getObjectOffset(BFI) + Dist*Bytes); 2211 } 2212 if (Loc.getOpcode() == ISD::ADD && Loc.getOperand(0) == BaseLoc) { 2213 ConstantSDNode *V = dyn_cast<ConstantSDNode>(Loc.getOperand(1)); 2214 if (V && (V->getSExtValue() == Dist*Bytes)) 2215 return true; 2216 } 2217 2218 GlobalValue *GV1 = NULL; 2219 GlobalValue *GV2 = NULL; 2220 int64_t Offset1 = 0; 2221 int64_t Offset2 = 0; 2222 bool isGA1 = isGAPlusOffset(Loc.getNode(), GV1, Offset1); 2223 bool isGA2 = isGAPlusOffset(BaseLoc.getNode(), GV2, Offset2); 2224 if (isGA1 && isGA2 && GV1 == GV2) 2225 return Offset1 == (Offset2 + Dist*Bytes); 2226 return false; 2227} 2228 2229 2230SDValue TargetLowering:: 2231PerformDAGCombine(SDNode *N, DAGCombinerInfo &DCI) const { 2232 // Default implementation: no optimization. 2233 return SDValue(); 2234} 2235 2236//===----------------------------------------------------------------------===// 2237// Inline Assembler Implementation Methods 2238//===----------------------------------------------------------------------===// 2239 2240 2241TargetLowering::ConstraintType 2242TargetLowering::getConstraintType(const std::string &Constraint) const { 2243 // FIXME: lots more standard ones to handle. 2244 if (Constraint.size() == 1) { 2245 switch (Constraint[0]) { 2246 default: break; 2247 case 'r': return C_RegisterClass; 2248 case 'm': // memory 2249 case 'o': // offsetable 2250 case 'V': // not offsetable 2251 return C_Memory; 2252 case 'i': // Simple Integer or Relocatable Constant 2253 case 'n': // Simple Integer 2254 case 's': // Relocatable Constant 2255 case 'X': // Allow ANY value. 2256 case 'I': // Target registers. 2257 case 'J': 2258 case 'K': 2259 case 'L': 2260 case 'M': 2261 case 'N': 2262 case 'O': 2263 case 'P': 2264 return C_Other; 2265 } 2266 } 2267 2268 if (Constraint.size() > 1 && Constraint[0] == '{' && 2269 Constraint[Constraint.size()-1] == '}') 2270 return C_Register; 2271 return C_Unknown; 2272} 2273 2274/// LowerXConstraint - try to replace an X constraint, which matches anything, 2275/// with another that has more specific requirements based on the type of the 2276/// corresponding operand. 2277const char *TargetLowering::LowerXConstraint(EVT ConstraintVT) const{ 2278 if (ConstraintVT.isInteger()) 2279 return "r"; 2280 if (ConstraintVT.isFloatingPoint()) 2281 return "f"; // works for many targets 2282 return 0; 2283} 2284 2285/// LowerAsmOperandForConstraint - Lower the specified operand into the Ops 2286/// vector. If it is invalid, don't add anything to Ops. 2287void TargetLowering::LowerAsmOperandForConstraint(SDValue Op, 2288 char ConstraintLetter, 2289 bool hasMemory, 2290 std::vector<SDValue> &Ops, 2291 SelectionDAG &DAG) const { 2292 switch (ConstraintLetter) { 2293 default: break; 2294 case 'X': // Allows any operand; labels (basic block) use this. 2295 if (Op.getOpcode() == ISD::BasicBlock) { 2296 Ops.push_back(Op); 2297 return; 2298 } 2299 // fall through 2300 case 'i': // Simple Integer or Relocatable Constant 2301 case 'n': // Simple Integer 2302 case 's': { // Relocatable Constant 2303 // These operands are interested in values of the form (GV+C), where C may 2304 // be folded in as an offset of GV, or it may be explicitly added. Also, it 2305 // is possible and fine if either GV or C are missing. 2306 ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op); 2307 GlobalAddressSDNode *GA = dyn_cast<GlobalAddressSDNode>(Op); 2308 2309 // If we have "(add GV, C)", pull out GV/C 2310 if (Op.getOpcode() == ISD::ADD) { 2311 C = dyn_cast<ConstantSDNode>(Op.getOperand(1)); 2312 GA = dyn_cast<GlobalAddressSDNode>(Op.getOperand(0)); 2313 if (C == 0 || GA == 0) { 2314 C = dyn_cast<ConstantSDNode>(Op.getOperand(0)); 2315 GA = dyn_cast<GlobalAddressSDNode>(Op.getOperand(1)); 2316 } 2317 if (C == 0 || GA == 0) 2318 C = 0, GA = 0; 2319 } 2320 2321 // If we find a valid operand, map to the TargetXXX version so that the 2322 // value itself doesn't get selected. 2323 if (GA) { // Either &GV or &GV+C 2324 if (ConstraintLetter != 'n') { 2325 int64_t Offs = GA->getOffset(); 2326 if (C) Offs += C->getZExtValue(); 2327 Ops.push_back(DAG.getTargetGlobalAddress(GA->getGlobal(), 2328 Op.getValueType(), Offs)); 2329 return; 2330 } 2331 } 2332 if (C) { // just C, no GV. 2333 // Simple constants are not allowed for 's'. 2334 if (ConstraintLetter != 's') { 2335 // gcc prints these as sign extended. Sign extend value to 64 bits 2336 // now; without this it would get ZExt'd later in 2337 // ScheduleDAGSDNodes::EmitNode, which is very generic. 2338 Ops.push_back(DAG.getTargetConstant(C->getAPIntValue().getSExtValue(), 2339 MVT::i64)); 2340 return; 2341 } 2342 } 2343 break; 2344 } 2345 } 2346} 2347 2348std::vector<unsigned> TargetLowering:: 2349getRegClassForInlineAsmConstraint(const std::string &Constraint, 2350 EVT VT) const { 2351 return std::vector<unsigned>(); 2352} 2353 2354 2355std::pair<unsigned, const TargetRegisterClass*> TargetLowering:: 2356getRegForInlineAsmConstraint(const std::string &Constraint, 2357 EVT VT) const { 2358 if (Constraint[0] != '{') 2359 return std::pair<unsigned, const TargetRegisterClass*>(0, 0); 2360 assert(*(Constraint.end()-1) == '}' && "Not a brace enclosed constraint?"); 2361 2362 // Remove the braces from around the name. 2363 std::string RegName(Constraint.begin()+1, Constraint.end()-1); 2364 2365 // Figure out which register class contains this reg. 2366 const TargetRegisterInfo *RI = TM.getRegisterInfo(); 2367 for (TargetRegisterInfo::regclass_iterator RCI = RI->regclass_begin(), 2368 E = RI->regclass_end(); RCI != E; ++RCI) { 2369 const TargetRegisterClass *RC = *RCI; 2370 2371 // If none of the the value types for this register class are valid, we 2372 // can't use it. For example, 64-bit reg classes on 32-bit targets. 2373 bool isLegal = false; 2374 for (TargetRegisterClass::vt_iterator I = RC->vt_begin(), E = RC->vt_end(); 2375 I != E; ++I) { 2376 if (isTypeLegal(*I)) { 2377 isLegal = true; 2378 break; 2379 } 2380 } 2381 2382 if (!isLegal) continue; 2383 2384 for (TargetRegisterClass::iterator I = RC->begin(), E = RC->end(); 2385 I != E; ++I) { 2386 if (StringsEqualNoCase(RegName, RI->getName(*I))) 2387 return std::make_pair(*I, RC); 2388 } 2389 } 2390 2391 return std::pair<unsigned, const TargetRegisterClass*>(0, 0); 2392} 2393 2394//===----------------------------------------------------------------------===// 2395// Constraint Selection. 2396 2397/// isMatchingInputConstraint - Return true of this is an input operand that is 2398/// a matching constraint like "4". 2399bool TargetLowering::AsmOperandInfo::isMatchingInputConstraint() const { 2400 assert(!ConstraintCode.empty() && "No known constraint!"); 2401 return isdigit(ConstraintCode[0]); 2402} 2403 2404/// getMatchedOperand - If this is an input matching constraint, this method 2405/// returns the output operand it matches. 2406unsigned TargetLowering::AsmOperandInfo::getMatchedOperand() const { 2407 assert(!ConstraintCode.empty() && "No known constraint!"); 2408 return atoi(ConstraintCode.c_str()); 2409} 2410 2411 2412/// getConstraintGenerality - Return an integer indicating how general CT 2413/// is. 2414static unsigned getConstraintGenerality(TargetLowering::ConstraintType CT) { 2415 switch (CT) { 2416 default: llvm_unreachable("Unknown constraint type!"); 2417 case TargetLowering::C_Other: 2418 case TargetLowering::C_Unknown: 2419 return 0; 2420 case TargetLowering::C_Register: 2421 return 1; 2422 case TargetLowering::C_RegisterClass: 2423 return 2; 2424 case TargetLowering::C_Memory: 2425 return 3; 2426 } 2427} 2428 2429/// ChooseConstraint - If there are multiple different constraints that we 2430/// could pick for this operand (e.g. "imr") try to pick the 'best' one. 2431/// This is somewhat tricky: constraints fall into four classes: 2432/// Other -> immediates and magic values 2433/// Register -> one specific register 2434/// RegisterClass -> a group of regs 2435/// Memory -> memory 2436/// Ideally, we would pick the most specific constraint possible: if we have 2437/// something that fits into a register, we would pick it. The problem here 2438/// is that if we have something that could either be in a register or in 2439/// memory that use of the register could cause selection of *other* 2440/// operands to fail: they might only succeed if we pick memory. Because of 2441/// this the heuristic we use is: 2442/// 2443/// 1) If there is an 'other' constraint, and if the operand is valid for 2444/// that constraint, use it. This makes us take advantage of 'i' 2445/// constraints when available. 2446/// 2) Otherwise, pick the most general constraint present. This prefers 2447/// 'm' over 'r', for example. 2448/// 2449static void ChooseConstraint(TargetLowering::AsmOperandInfo &OpInfo, 2450 bool hasMemory, const TargetLowering &TLI, 2451 SDValue Op, SelectionDAG *DAG) { 2452 assert(OpInfo.Codes.size() > 1 && "Doesn't have multiple constraint options"); 2453 unsigned BestIdx = 0; 2454 TargetLowering::ConstraintType BestType = TargetLowering::C_Unknown; 2455 int BestGenerality = -1; 2456 2457 // Loop over the options, keeping track of the most general one. 2458 for (unsigned i = 0, e = OpInfo.Codes.size(); i != e; ++i) { 2459 TargetLowering::ConstraintType CType = 2460 TLI.getConstraintType(OpInfo.Codes[i]); 2461 2462 // If this is an 'other' constraint, see if the operand is valid for it. 2463 // For example, on X86 we might have an 'rI' constraint. If the operand 2464 // is an integer in the range [0..31] we want to use I (saving a load 2465 // of a register), otherwise we must use 'r'. 2466 if (CType == TargetLowering::C_Other && Op.getNode()) { 2467 assert(OpInfo.Codes[i].size() == 1 && 2468 "Unhandled multi-letter 'other' constraint"); 2469 std::vector<SDValue> ResultOps; 2470 TLI.LowerAsmOperandForConstraint(Op, OpInfo.Codes[i][0], hasMemory, 2471 ResultOps, *DAG); 2472 if (!ResultOps.empty()) { 2473 BestType = CType; 2474 BestIdx = i; 2475 break; 2476 } 2477 } 2478 2479 // This constraint letter is more general than the previous one, use it. 2480 int Generality = getConstraintGenerality(CType); 2481 if (Generality > BestGenerality) { 2482 BestType = CType; 2483 BestIdx = i; 2484 BestGenerality = Generality; 2485 } 2486 } 2487 2488 OpInfo.ConstraintCode = OpInfo.Codes[BestIdx]; 2489 OpInfo.ConstraintType = BestType; 2490} 2491 2492/// ComputeConstraintToUse - Determines the constraint code and constraint 2493/// type to use for the specific AsmOperandInfo, setting 2494/// OpInfo.ConstraintCode and OpInfo.ConstraintType. 2495void TargetLowering::ComputeConstraintToUse(AsmOperandInfo &OpInfo, 2496 SDValue Op, 2497 bool hasMemory, 2498 SelectionDAG *DAG) const { 2499 assert(!OpInfo.Codes.empty() && "Must have at least one constraint"); 2500 2501 // Single-letter constraints ('r') are very common. 2502 if (OpInfo.Codes.size() == 1) { 2503 OpInfo.ConstraintCode = OpInfo.Codes[0]; 2504 OpInfo.ConstraintType = getConstraintType(OpInfo.ConstraintCode); 2505 } else { 2506 ChooseConstraint(OpInfo, hasMemory, *this, Op, DAG); 2507 } 2508 2509 // 'X' matches anything. 2510 if (OpInfo.ConstraintCode == "X" && OpInfo.CallOperandVal) { 2511 // Labels and constants are handled elsewhere ('X' is the only thing 2512 // that matches labels). For Functions, the type here is the type of 2513 // the result, which is not what we want to look at; leave them alone. 2514 Value *v = OpInfo.CallOperandVal; 2515 if (isa<BasicBlock>(v) || isa<ConstantInt>(v) || isa<Function>(v)) { 2516 OpInfo.CallOperandVal = v; 2517 return; 2518 } 2519 2520 // Otherwise, try to resolve it to something we know about by looking at 2521 // the actual operand type. 2522 if (const char *Repl = LowerXConstraint(OpInfo.ConstraintVT)) { 2523 OpInfo.ConstraintCode = Repl; 2524 OpInfo.ConstraintType = getConstraintType(OpInfo.ConstraintCode); 2525 } 2526 } 2527} 2528 2529//===----------------------------------------------------------------------===// 2530// Loop Strength Reduction hooks 2531//===----------------------------------------------------------------------===// 2532 2533/// isLegalAddressingMode - Return true if the addressing mode represented 2534/// by AM is legal for this target, for a load/store of the specified type. 2535bool TargetLowering::isLegalAddressingMode(const AddrMode &AM, 2536 const Type *Ty) const { 2537 // The default implementation of this implements a conservative RISCy, r+r and 2538 // r+i addr mode. 2539 2540 // Allows a sign-extended 16-bit immediate field. 2541 if (AM.BaseOffs <= -(1LL << 16) || AM.BaseOffs >= (1LL << 16)-1) 2542 return false; 2543 2544 // No global is ever allowed as a base. 2545 if (AM.BaseGV) 2546 return false; 2547 2548 // Only support r+r, 2549 switch (AM.Scale) { 2550 case 0: // "r+i" or just "i", depending on HasBaseReg. 2551 break; 2552 case 1: 2553 if (AM.HasBaseReg && AM.BaseOffs) // "r+r+i" is not allowed. 2554 return false; 2555 // Otherwise we have r+r or r+i. 2556 break; 2557 case 2: 2558 if (AM.HasBaseReg || AM.BaseOffs) // 2*r+r or 2*r+i is not allowed. 2559 return false; 2560 // Allow 2*r as r+r. 2561 break; 2562 } 2563 2564 return true; 2565} 2566 2567/// BuildSDIVSequence - Given an ISD::SDIV node expressing a divide by constant, 2568/// return a DAG expression to select that will generate the same value by 2569/// multiplying by a magic number. See: 2570/// <http://the.wall.riscom.net/books/proc/ppc/cwg/code2.html> 2571SDValue TargetLowering::BuildSDIV(SDNode *N, SelectionDAG &DAG, 2572 std::vector<SDNode*>* Created) const { 2573 EVT VT = N->getValueType(0); 2574 DebugLoc dl= N->getDebugLoc(); 2575 2576 // Check to see if we can do this. 2577 // FIXME: We should be more aggressive here. 2578 if (!isTypeLegal(VT)) 2579 return SDValue(); 2580 2581 APInt d = cast<ConstantSDNode>(N->getOperand(1))->getAPIntValue(); 2582 APInt::ms magics = d.magic(); 2583 2584 // Multiply the numerator (operand 0) by the magic value 2585 // FIXME: We should support doing a MUL in a wider type 2586 SDValue Q; 2587 if (isOperationLegalOrCustom(ISD::MULHS, VT)) 2588 Q = DAG.getNode(ISD::MULHS, dl, VT, N->getOperand(0), 2589 DAG.getConstant(magics.m, VT)); 2590 else if (isOperationLegalOrCustom(ISD::SMUL_LOHI, VT)) 2591 Q = SDValue(DAG.getNode(ISD::SMUL_LOHI, dl, DAG.getVTList(VT, VT), 2592 N->getOperand(0), 2593 DAG.getConstant(magics.m, VT)).getNode(), 1); 2594 else 2595 return SDValue(); // No mulhs or equvialent 2596 // If d > 0 and m < 0, add the numerator 2597 if (d.isStrictlyPositive() && magics.m.isNegative()) { 2598 Q = DAG.getNode(ISD::ADD, dl, VT, Q, N->getOperand(0)); 2599 if (Created) 2600 Created->push_back(Q.getNode()); 2601 } 2602 // If d < 0 and m > 0, subtract the numerator. 2603 if (d.isNegative() && magics.m.isStrictlyPositive()) { 2604 Q = DAG.getNode(ISD::SUB, dl, VT, Q, N->getOperand(0)); 2605 if (Created) 2606 Created->push_back(Q.getNode()); 2607 } 2608 // Shift right algebraic if shift value is nonzero 2609 if (magics.s > 0) { 2610 Q = DAG.getNode(ISD::SRA, dl, VT, Q, 2611 DAG.getConstant(magics.s, getShiftAmountTy())); 2612 if (Created) 2613 Created->push_back(Q.getNode()); 2614 } 2615 // Extract the sign bit and add it to the quotient 2616 SDValue T = 2617 DAG.getNode(ISD::SRL, dl, VT, Q, DAG.getConstant(VT.getSizeInBits()-1, 2618 getShiftAmountTy())); 2619 if (Created) 2620 Created->push_back(T.getNode()); 2621 return DAG.getNode(ISD::ADD, dl, VT, Q, T); 2622} 2623 2624/// BuildUDIVSequence - Given an ISD::UDIV node expressing a divide by constant, 2625/// return a DAG expression to select that will generate the same value by 2626/// multiplying by a magic number. See: 2627/// <http://the.wall.riscom.net/books/proc/ppc/cwg/code2.html> 2628SDValue TargetLowering::BuildUDIV(SDNode *N, SelectionDAG &DAG, 2629 std::vector<SDNode*>* Created) const { 2630 EVT VT = N->getValueType(0); 2631 DebugLoc dl = N->getDebugLoc(); 2632 2633 // Check to see if we can do this. 2634 // FIXME: We should be more aggressive here. 2635 if (!isTypeLegal(VT)) 2636 return SDValue(); 2637 2638 // FIXME: We should use a narrower constant when the upper 2639 // bits are known to be zero. 2640 ConstantSDNode *N1C = cast<ConstantSDNode>(N->getOperand(1)); 2641 APInt::mu magics = N1C->getAPIntValue().magicu(); 2642 2643 // Multiply the numerator (operand 0) by the magic value 2644 // FIXME: We should support doing a MUL in a wider type 2645 SDValue Q; 2646 if (isOperationLegalOrCustom(ISD::MULHU, VT)) 2647 Q = DAG.getNode(ISD::MULHU, dl, VT, N->getOperand(0), 2648 DAG.getConstant(magics.m, VT)); 2649 else if (isOperationLegalOrCustom(ISD::UMUL_LOHI, VT)) 2650 Q = SDValue(DAG.getNode(ISD::UMUL_LOHI, dl, DAG.getVTList(VT, VT), 2651 N->getOperand(0), 2652 DAG.getConstant(magics.m, VT)).getNode(), 1); 2653 else 2654 return SDValue(); // No mulhu or equvialent 2655 if (Created) 2656 Created->push_back(Q.getNode()); 2657 2658 if (magics.a == 0) { 2659 assert(magics.s < N1C->getAPIntValue().getBitWidth() && 2660 "We shouldn't generate an undefined shift!"); 2661 return DAG.getNode(ISD::SRL, dl, VT, Q, 2662 DAG.getConstant(magics.s, getShiftAmountTy())); 2663 } else { 2664 SDValue NPQ = DAG.getNode(ISD::SUB, dl, VT, N->getOperand(0), Q); 2665 if (Created) 2666 Created->push_back(NPQ.getNode()); 2667 NPQ = DAG.getNode(ISD::SRL, dl, VT, NPQ, 2668 DAG.getConstant(1, getShiftAmountTy())); 2669 if (Created) 2670 Created->push_back(NPQ.getNode()); 2671 NPQ = DAG.getNode(ISD::ADD, dl, VT, NPQ, Q); 2672 if (Created) 2673 Created->push_back(NPQ.getNode()); 2674 return DAG.getNode(ISD::SRL, dl, VT, NPQ, 2675 DAG.getConstant(magics.s-1, getShiftAmountTy())); 2676 } 2677} 2678