TargetLowering.cpp revision f190a030835622ec1d15e32f30939f9495827e39
1//===-- TargetLowering.cpp - Implement the TargetLowering class -----------===// 2// 3// The LLVM Compiler Infrastructure 4// 5// This file is distributed under the University of Illinois Open Source 6// License. See LICENSE.TXT for details. 7// 8//===----------------------------------------------------------------------===// 9// 10// This implements the TargetLowering class. 11// 12//===----------------------------------------------------------------------===// 13 14#include "llvm/Target/TargetAsmInfo.h" 15#include "llvm/Target/TargetLowering.h" 16#include "llvm/Target/TargetSubtarget.h" 17#include "llvm/Target/TargetData.h" 18#include "llvm/Target/TargetMachine.h" 19#include "llvm/Target/TargetRegisterInfo.h" 20#include "llvm/GlobalVariable.h" 21#include "llvm/DerivedTypes.h" 22#include "llvm/CodeGen/MachineFrameInfo.h" 23#include "llvm/CodeGen/SelectionDAG.h" 24#include "llvm/ADT/StringExtras.h" 25#include "llvm/ADT/STLExtras.h" 26#include "llvm/Support/MathExtras.h" 27using namespace llvm; 28 29/// InitLibcallNames - Set default libcall names. 30/// 31static void InitLibcallNames(const char **Names) { 32 Names[RTLIB::SHL_I16] = "__ashli16"; 33 Names[RTLIB::SHL_I32] = "__ashlsi3"; 34 Names[RTLIB::SHL_I64] = "__ashldi3"; 35 Names[RTLIB::SHL_I128] = "__ashlti3"; 36 Names[RTLIB::SRL_I16] = "__lshri16"; 37 Names[RTLIB::SRL_I32] = "__lshrsi3"; 38 Names[RTLIB::SRL_I64] = "__lshrdi3"; 39 Names[RTLIB::SRL_I128] = "__lshrti3"; 40 Names[RTLIB::SRA_I16] = "__ashri16"; 41 Names[RTLIB::SRA_I32] = "__ashrsi3"; 42 Names[RTLIB::SRA_I64] = "__ashrdi3"; 43 Names[RTLIB::SRA_I128] = "__ashrti3"; 44 Names[RTLIB::MUL_I16] = "__muli16"; 45 Names[RTLIB::MUL_I32] = "__mulsi3"; 46 Names[RTLIB::MUL_I64] = "__muldi3"; 47 Names[RTLIB::MUL_I128] = "__multi3"; 48 Names[RTLIB::SDIV_I32] = "__divsi3"; 49 Names[RTLIB::SDIV_I64] = "__divdi3"; 50 Names[RTLIB::SDIV_I128] = "__divti3"; 51 Names[RTLIB::UDIV_I32] = "__udivsi3"; 52 Names[RTLIB::UDIV_I64] = "__udivdi3"; 53 Names[RTLIB::UDIV_I128] = "__udivti3"; 54 Names[RTLIB::SREM_I32] = "__modsi3"; 55 Names[RTLIB::SREM_I64] = "__moddi3"; 56 Names[RTLIB::SREM_I128] = "__modti3"; 57 Names[RTLIB::UREM_I32] = "__umodsi3"; 58 Names[RTLIB::UREM_I64] = "__umoddi3"; 59 Names[RTLIB::UREM_I128] = "__umodti3"; 60 Names[RTLIB::NEG_I32] = "__negsi2"; 61 Names[RTLIB::NEG_I64] = "__negdi2"; 62 Names[RTLIB::ADD_F32] = "__addsf3"; 63 Names[RTLIB::ADD_F64] = "__adddf3"; 64 Names[RTLIB::ADD_F80] = "__addxf3"; 65 Names[RTLIB::ADD_PPCF128] = "__gcc_qadd"; 66 Names[RTLIB::SUB_F32] = "__subsf3"; 67 Names[RTLIB::SUB_F64] = "__subdf3"; 68 Names[RTLIB::SUB_F80] = "__subxf3"; 69 Names[RTLIB::SUB_PPCF128] = "__gcc_qsub"; 70 Names[RTLIB::MUL_F32] = "__mulsf3"; 71 Names[RTLIB::MUL_F64] = "__muldf3"; 72 Names[RTLIB::MUL_F80] = "__mulxf3"; 73 Names[RTLIB::MUL_PPCF128] = "__gcc_qmul"; 74 Names[RTLIB::DIV_F32] = "__divsf3"; 75 Names[RTLIB::DIV_F64] = "__divdf3"; 76 Names[RTLIB::DIV_F80] = "__divxf3"; 77 Names[RTLIB::DIV_PPCF128] = "__gcc_qdiv"; 78 Names[RTLIB::REM_F32] = "fmodf"; 79 Names[RTLIB::REM_F64] = "fmod"; 80 Names[RTLIB::REM_F80] = "fmodl"; 81 Names[RTLIB::REM_PPCF128] = "fmodl"; 82 Names[RTLIB::POWI_F32] = "__powisf2"; 83 Names[RTLIB::POWI_F64] = "__powidf2"; 84 Names[RTLIB::POWI_F80] = "__powixf2"; 85 Names[RTLIB::POWI_PPCF128] = "__powitf2"; 86 Names[RTLIB::SQRT_F32] = "sqrtf"; 87 Names[RTLIB::SQRT_F64] = "sqrt"; 88 Names[RTLIB::SQRT_F80] = "sqrtl"; 89 Names[RTLIB::SQRT_PPCF128] = "sqrtl"; 90 Names[RTLIB::LOG_F32] = "logf"; 91 Names[RTLIB::LOG_F64] = "log"; 92 Names[RTLIB::LOG_F80] = "logl"; 93 Names[RTLIB::LOG_PPCF128] = "logl"; 94 Names[RTLIB::LOG2_F32] = "log2f"; 95 Names[RTLIB::LOG2_F64] = "log2"; 96 Names[RTLIB::LOG2_F80] = "log2l"; 97 Names[RTLIB::LOG2_PPCF128] = "log2l"; 98 Names[RTLIB::LOG10_F32] = "log10f"; 99 Names[RTLIB::LOG10_F64] = "log10"; 100 Names[RTLIB::LOG10_F80] = "log10l"; 101 Names[RTLIB::LOG10_PPCF128] = "log10l"; 102 Names[RTLIB::EXP_F32] = "expf"; 103 Names[RTLIB::EXP_F64] = "exp"; 104 Names[RTLIB::EXP_F80] = "expl"; 105 Names[RTLIB::EXP_PPCF128] = "expl"; 106 Names[RTLIB::EXP2_F32] = "exp2f"; 107 Names[RTLIB::EXP2_F64] = "exp2"; 108 Names[RTLIB::EXP2_F80] = "exp2l"; 109 Names[RTLIB::EXP2_PPCF128] = "exp2l"; 110 Names[RTLIB::SIN_F32] = "sinf"; 111 Names[RTLIB::SIN_F64] = "sin"; 112 Names[RTLIB::SIN_F80] = "sinl"; 113 Names[RTLIB::SIN_PPCF128] = "sinl"; 114 Names[RTLIB::COS_F32] = "cosf"; 115 Names[RTLIB::COS_F64] = "cos"; 116 Names[RTLIB::COS_F80] = "cosl"; 117 Names[RTLIB::COS_PPCF128] = "cosl"; 118 Names[RTLIB::POW_F32] = "powf"; 119 Names[RTLIB::POW_F64] = "pow"; 120 Names[RTLIB::POW_F80] = "powl"; 121 Names[RTLIB::POW_PPCF128] = "powl"; 122 Names[RTLIB::CEIL_F32] = "ceilf"; 123 Names[RTLIB::CEIL_F64] = "ceil"; 124 Names[RTLIB::CEIL_F80] = "ceill"; 125 Names[RTLIB::CEIL_PPCF128] = "ceill"; 126 Names[RTLIB::TRUNC_F32] = "truncf"; 127 Names[RTLIB::TRUNC_F64] = "trunc"; 128 Names[RTLIB::TRUNC_F80] = "truncl"; 129 Names[RTLIB::TRUNC_PPCF128] = "truncl"; 130 Names[RTLIB::RINT_F32] = "rintf"; 131 Names[RTLIB::RINT_F64] = "rint"; 132 Names[RTLIB::RINT_F80] = "rintl"; 133 Names[RTLIB::RINT_PPCF128] = "rintl"; 134 Names[RTLIB::NEARBYINT_F32] = "nearbyintf"; 135 Names[RTLIB::NEARBYINT_F64] = "nearbyint"; 136 Names[RTLIB::NEARBYINT_F80] = "nearbyintl"; 137 Names[RTLIB::NEARBYINT_PPCF128] = "nearbyintl"; 138 Names[RTLIB::FLOOR_F32] = "floorf"; 139 Names[RTLIB::FLOOR_F64] = "floor"; 140 Names[RTLIB::FLOOR_F80] = "floorl"; 141 Names[RTLIB::FLOOR_PPCF128] = "floorl"; 142 Names[RTLIB::FPEXT_F32_F64] = "__extendsfdf2"; 143 Names[RTLIB::FPROUND_F64_F32] = "__truncdfsf2"; 144 Names[RTLIB::FPROUND_F80_F32] = "__truncxfsf2"; 145 Names[RTLIB::FPROUND_PPCF128_F32] = "__trunctfsf2"; 146 Names[RTLIB::FPROUND_F80_F64] = "__truncxfdf2"; 147 Names[RTLIB::FPROUND_PPCF128_F64] = "__trunctfdf2"; 148 Names[RTLIB::FPTOSINT_F32_I32] = "__fixsfsi"; 149 Names[RTLIB::FPTOSINT_F32_I64] = "__fixsfdi"; 150 Names[RTLIB::FPTOSINT_F32_I128] = "__fixsfti"; 151 Names[RTLIB::FPTOSINT_F64_I32] = "__fixdfsi"; 152 Names[RTLIB::FPTOSINT_F64_I64] = "__fixdfdi"; 153 Names[RTLIB::FPTOSINT_F64_I128] = "__fixdfti"; 154 Names[RTLIB::FPTOSINT_F80_I32] = "__fixxfsi"; 155 Names[RTLIB::FPTOSINT_F80_I64] = "__fixxfdi"; 156 Names[RTLIB::FPTOSINT_F80_I128] = "__fixxfti"; 157 Names[RTLIB::FPTOSINT_PPCF128_I32] = "__fixtfsi"; 158 Names[RTLIB::FPTOSINT_PPCF128_I64] = "__fixtfdi"; 159 Names[RTLIB::FPTOSINT_PPCF128_I128] = "__fixtfti"; 160 Names[RTLIB::FPTOUINT_F32_I32] = "__fixunssfsi"; 161 Names[RTLIB::FPTOUINT_F32_I64] = "__fixunssfdi"; 162 Names[RTLIB::FPTOUINT_F32_I128] = "__fixunssfti"; 163 Names[RTLIB::FPTOUINT_F64_I32] = "__fixunsdfsi"; 164 Names[RTLIB::FPTOUINT_F64_I64] = "__fixunsdfdi"; 165 Names[RTLIB::FPTOUINT_F64_I128] = "__fixunsdfti"; 166 Names[RTLIB::FPTOUINT_F80_I32] = "__fixunsxfsi"; 167 Names[RTLIB::FPTOUINT_F80_I64] = "__fixunsxfdi"; 168 Names[RTLIB::FPTOUINT_F80_I128] = "__fixunsxfti"; 169 Names[RTLIB::FPTOUINT_PPCF128_I32] = "__fixunstfsi"; 170 Names[RTLIB::FPTOUINT_PPCF128_I64] = "__fixunstfdi"; 171 Names[RTLIB::FPTOUINT_PPCF128_I128] = "__fixunstfti"; 172 Names[RTLIB::SINTTOFP_I32_F32] = "__floatsisf"; 173 Names[RTLIB::SINTTOFP_I32_F64] = "__floatsidf"; 174 Names[RTLIB::SINTTOFP_I32_F80] = "__floatsixf"; 175 Names[RTLIB::SINTTOFP_I32_PPCF128] = "__floatsitf"; 176 Names[RTLIB::SINTTOFP_I64_F32] = "__floatdisf"; 177 Names[RTLIB::SINTTOFP_I64_F64] = "__floatdidf"; 178 Names[RTLIB::SINTTOFP_I64_F80] = "__floatdixf"; 179 Names[RTLIB::SINTTOFP_I64_PPCF128] = "__floatditf"; 180 Names[RTLIB::SINTTOFP_I128_F32] = "__floattisf"; 181 Names[RTLIB::SINTTOFP_I128_F64] = "__floattidf"; 182 Names[RTLIB::SINTTOFP_I128_F80] = "__floattixf"; 183 Names[RTLIB::SINTTOFP_I128_PPCF128] = "__floattitf"; 184 Names[RTLIB::UINTTOFP_I32_F32] = "__floatunsisf"; 185 Names[RTLIB::UINTTOFP_I32_F64] = "__floatunsidf"; 186 Names[RTLIB::UINTTOFP_I32_F80] = "__floatunsixf"; 187 Names[RTLIB::UINTTOFP_I32_PPCF128] = "__floatunsitf"; 188 Names[RTLIB::UINTTOFP_I64_F32] = "__floatundisf"; 189 Names[RTLIB::UINTTOFP_I64_F64] = "__floatundidf"; 190 Names[RTLIB::UINTTOFP_I64_F80] = "__floatundixf"; 191 Names[RTLIB::UINTTOFP_I64_PPCF128] = "__floatunditf"; 192 Names[RTLIB::UINTTOFP_I128_F32] = "__floatuntisf"; 193 Names[RTLIB::UINTTOFP_I128_F64] = "__floatuntidf"; 194 Names[RTLIB::UINTTOFP_I128_F80] = "__floatuntixf"; 195 Names[RTLIB::UINTTOFP_I128_PPCF128] = "__floatuntitf"; 196 Names[RTLIB::OEQ_F32] = "__eqsf2"; 197 Names[RTLIB::OEQ_F64] = "__eqdf2"; 198 Names[RTLIB::UNE_F32] = "__nesf2"; 199 Names[RTLIB::UNE_F64] = "__nedf2"; 200 Names[RTLIB::OGE_F32] = "__gesf2"; 201 Names[RTLIB::OGE_F64] = "__gedf2"; 202 Names[RTLIB::OLT_F32] = "__ltsf2"; 203 Names[RTLIB::OLT_F64] = "__ltdf2"; 204 Names[RTLIB::OLE_F32] = "__lesf2"; 205 Names[RTLIB::OLE_F64] = "__ledf2"; 206 Names[RTLIB::OGT_F32] = "__gtsf2"; 207 Names[RTLIB::OGT_F64] = "__gtdf2"; 208 Names[RTLIB::UO_F32] = "__unordsf2"; 209 Names[RTLIB::UO_F64] = "__unorddf2"; 210 Names[RTLIB::O_F32] = "__unordsf2"; 211 Names[RTLIB::O_F64] = "__unorddf2"; 212} 213 214/// getFPEXT - Return the FPEXT_*_* value for the given types, or 215/// UNKNOWN_LIBCALL if there is none. 216RTLIB::Libcall RTLIB::getFPEXT(MVT OpVT, MVT RetVT) { 217 if (OpVT == MVT::f32) { 218 if (RetVT == MVT::f64) 219 return FPEXT_F32_F64; 220 } 221 return UNKNOWN_LIBCALL; 222} 223 224/// getFPROUND - Return the FPROUND_*_* value for the given types, or 225/// UNKNOWN_LIBCALL if there is none. 226RTLIB::Libcall RTLIB::getFPROUND(MVT OpVT, MVT RetVT) { 227 if (RetVT == MVT::f32) { 228 if (OpVT == MVT::f64) 229 return FPROUND_F64_F32; 230 if (OpVT == MVT::f80) 231 return FPROUND_F80_F32; 232 if (OpVT == MVT::ppcf128) 233 return FPROUND_PPCF128_F32; 234 } else if (RetVT == MVT::f64) { 235 if (OpVT == MVT::f80) 236 return FPROUND_F80_F64; 237 if (OpVT == MVT::ppcf128) 238 return FPROUND_PPCF128_F64; 239 } 240 return UNKNOWN_LIBCALL; 241} 242 243/// getFPTOSINT - Return the FPTOSINT_*_* value for the given types, or 244/// UNKNOWN_LIBCALL if there is none. 245RTLIB::Libcall RTLIB::getFPTOSINT(MVT OpVT, MVT RetVT) { 246 if (OpVT == MVT::f32) { 247 if (RetVT == MVT::i32) 248 return FPTOSINT_F32_I32; 249 if (RetVT == MVT::i64) 250 return FPTOSINT_F32_I64; 251 if (RetVT == MVT::i128) 252 return FPTOSINT_F32_I128; 253 } else if (OpVT == MVT::f64) { 254 if (RetVT == MVT::i32) 255 return FPTOSINT_F64_I32; 256 if (RetVT == MVT::i64) 257 return FPTOSINT_F64_I64; 258 if (RetVT == MVT::i128) 259 return FPTOSINT_F64_I128; 260 } else if (OpVT == MVT::f80) { 261 if (RetVT == MVT::i32) 262 return FPTOSINT_F80_I32; 263 if (RetVT == MVT::i64) 264 return FPTOSINT_F80_I64; 265 if (RetVT == MVT::i128) 266 return FPTOSINT_F80_I128; 267 } else if (OpVT == MVT::ppcf128) { 268 if (RetVT == MVT::i32) 269 return FPTOSINT_PPCF128_I32; 270 if (RetVT == MVT::i64) 271 return FPTOSINT_PPCF128_I64; 272 if (RetVT == MVT::i128) 273 return FPTOSINT_PPCF128_I128; 274 } 275 return UNKNOWN_LIBCALL; 276} 277 278/// getFPTOUINT - Return the FPTOUINT_*_* value for the given types, or 279/// UNKNOWN_LIBCALL if there is none. 280RTLIB::Libcall RTLIB::getFPTOUINT(MVT OpVT, MVT RetVT) { 281 if (OpVT == MVT::f32) { 282 if (RetVT == MVT::i32) 283 return FPTOUINT_F32_I32; 284 if (RetVT == MVT::i64) 285 return FPTOUINT_F32_I64; 286 if (RetVT == MVT::i128) 287 return FPTOUINT_F32_I128; 288 } else if (OpVT == MVT::f64) { 289 if (RetVT == MVT::i32) 290 return FPTOUINT_F64_I32; 291 if (RetVT == MVT::i64) 292 return FPTOUINT_F64_I64; 293 if (RetVT == MVT::i128) 294 return FPTOUINT_F64_I128; 295 } else if (OpVT == MVT::f80) { 296 if (RetVT == MVT::i32) 297 return FPTOUINT_F80_I32; 298 if (RetVT == MVT::i64) 299 return FPTOUINT_F80_I64; 300 if (RetVT == MVT::i128) 301 return FPTOUINT_F80_I128; 302 } else if (OpVT == MVT::ppcf128) { 303 if (RetVT == MVT::i32) 304 return FPTOUINT_PPCF128_I32; 305 if (RetVT == MVT::i64) 306 return FPTOUINT_PPCF128_I64; 307 if (RetVT == MVT::i128) 308 return FPTOUINT_PPCF128_I128; 309 } 310 return UNKNOWN_LIBCALL; 311} 312 313/// getSINTTOFP - Return the SINTTOFP_*_* value for the given types, or 314/// UNKNOWN_LIBCALL if there is none. 315RTLIB::Libcall RTLIB::getSINTTOFP(MVT OpVT, MVT RetVT) { 316 if (OpVT == MVT::i32) { 317 if (RetVT == MVT::f32) 318 return SINTTOFP_I32_F32; 319 else if (RetVT == MVT::f64) 320 return SINTTOFP_I32_F64; 321 else if (RetVT == MVT::f80) 322 return SINTTOFP_I32_F80; 323 else if (RetVT == MVT::ppcf128) 324 return SINTTOFP_I32_PPCF128; 325 } else if (OpVT == MVT::i64) { 326 if (RetVT == MVT::f32) 327 return SINTTOFP_I64_F32; 328 else if (RetVT == MVT::f64) 329 return SINTTOFP_I64_F64; 330 else if (RetVT == MVT::f80) 331 return SINTTOFP_I64_F80; 332 else if (RetVT == MVT::ppcf128) 333 return SINTTOFP_I64_PPCF128; 334 } else if (OpVT == MVT::i128) { 335 if (RetVT == MVT::f32) 336 return SINTTOFP_I128_F32; 337 else if (RetVT == MVT::f64) 338 return SINTTOFP_I128_F64; 339 else if (RetVT == MVT::f80) 340 return SINTTOFP_I128_F80; 341 else if (RetVT == MVT::ppcf128) 342 return SINTTOFP_I128_PPCF128; 343 } 344 return UNKNOWN_LIBCALL; 345} 346 347/// getUINTTOFP - Return the UINTTOFP_*_* value for the given types, or 348/// UNKNOWN_LIBCALL if there is none. 349RTLIB::Libcall RTLIB::getUINTTOFP(MVT OpVT, MVT RetVT) { 350 if (OpVT == MVT::i32) { 351 if (RetVT == MVT::f32) 352 return UINTTOFP_I32_F32; 353 else if (RetVT == MVT::f64) 354 return UINTTOFP_I32_F64; 355 else if (RetVT == MVT::f80) 356 return UINTTOFP_I32_F80; 357 else if (RetVT == MVT::ppcf128) 358 return UINTTOFP_I32_PPCF128; 359 } else if (OpVT == MVT::i64) { 360 if (RetVT == MVT::f32) 361 return UINTTOFP_I64_F32; 362 else if (RetVT == MVT::f64) 363 return UINTTOFP_I64_F64; 364 else if (RetVT == MVT::f80) 365 return UINTTOFP_I64_F80; 366 else if (RetVT == MVT::ppcf128) 367 return UINTTOFP_I64_PPCF128; 368 } else if (OpVT == MVT::i128) { 369 if (RetVT == MVT::f32) 370 return UINTTOFP_I128_F32; 371 else if (RetVT == MVT::f64) 372 return UINTTOFP_I128_F64; 373 else if (RetVT == MVT::f80) 374 return UINTTOFP_I128_F80; 375 else if (RetVT == MVT::ppcf128) 376 return UINTTOFP_I128_PPCF128; 377 } 378 return UNKNOWN_LIBCALL; 379} 380 381/// InitCmpLibcallCCs - Set default comparison libcall CC. 382/// 383static void InitCmpLibcallCCs(ISD::CondCode *CCs) { 384 memset(CCs, ISD::SETCC_INVALID, sizeof(ISD::CondCode)*RTLIB::UNKNOWN_LIBCALL); 385 CCs[RTLIB::OEQ_F32] = ISD::SETEQ; 386 CCs[RTLIB::OEQ_F64] = ISD::SETEQ; 387 CCs[RTLIB::UNE_F32] = ISD::SETNE; 388 CCs[RTLIB::UNE_F64] = ISD::SETNE; 389 CCs[RTLIB::OGE_F32] = ISD::SETGE; 390 CCs[RTLIB::OGE_F64] = ISD::SETGE; 391 CCs[RTLIB::OLT_F32] = ISD::SETLT; 392 CCs[RTLIB::OLT_F64] = ISD::SETLT; 393 CCs[RTLIB::OLE_F32] = ISD::SETLE; 394 CCs[RTLIB::OLE_F64] = ISD::SETLE; 395 CCs[RTLIB::OGT_F32] = ISD::SETGT; 396 CCs[RTLIB::OGT_F64] = ISD::SETGT; 397 CCs[RTLIB::UO_F32] = ISD::SETNE; 398 CCs[RTLIB::UO_F64] = ISD::SETNE; 399 CCs[RTLIB::O_F32] = ISD::SETEQ; 400 CCs[RTLIB::O_F64] = ISD::SETEQ; 401} 402 403TargetLowering::TargetLowering(TargetMachine &tm) 404 : TM(tm), TD(TM.getTargetData()) { 405 // All operations default to being supported. 406 memset(OpActions, 0, sizeof(OpActions)); 407 memset(LoadExtActions, 0, sizeof(LoadExtActions)); 408 memset(TruncStoreActions, 0, sizeof(TruncStoreActions)); 409 memset(IndexedModeActions, 0, sizeof(IndexedModeActions)); 410 memset(ConvertActions, 0, sizeof(ConvertActions)); 411 memset(CondCodeActions, 0, sizeof(CondCodeActions)); 412 413 // Set default actions for various operations. 414 for (unsigned VT = 0; VT != (unsigned)MVT::LAST_VALUETYPE; ++VT) { 415 // Default all indexed load / store to expand. 416 for (unsigned IM = (unsigned)ISD::PRE_INC; 417 IM != (unsigned)ISD::LAST_INDEXED_MODE; ++IM) { 418 setIndexedLoadAction(IM, (MVT::SimpleValueType)VT, Expand); 419 setIndexedStoreAction(IM, (MVT::SimpleValueType)VT, Expand); 420 } 421 422 // These operations default to expand. 423 setOperationAction(ISD::FGETSIGN, (MVT::SimpleValueType)VT, Expand); 424 } 425 426 // Most targets ignore the @llvm.prefetch intrinsic. 427 setOperationAction(ISD::PREFETCH, MVT::Other, Expand); 428 429 // ConstantFP nodes default to expand. Targets can either change this to 430 // Legal, in which case all fp constants are legal, or use addLegalFPImmediate 431 // to optimize expansions for certain constants. 432 setOperationAction(ISD::ConstantFP, MVT::f32, Expand); 433 setOperationAction(ISD::ConstantFP, MVT::f64, Expand); 434 setOperationAction(ISD::ConstantFP, MVT::f80, Expand); 435 436 // These library functions default to expand. 437 setOperationAction(ISD::FLOG , MVT::f64, Expand); 438 setOperationAction(ISD::FLOG2, MVT::f64, Expand); 439 setOperationAction(ISD::FLOG10,MVT::f64, Expand); 440 setOperationAction(ISD::FEXP , MVT::f64, Expand); 441 setOperationAction(ISD::FEXP2, MVT::f64, Expand); 442 setOperationAction(ISD::FLOG , MVT::f32, Expand); 443 setOperationAction(ISD::FLOG2, MVT::f32, Expand); 444 setOperationAction(ISD::FLOG10,MVT::f32, Expand); 445 setOperationAction(ISD::FEXP , MVT::f32, Expand); 446 setOperationAction(ISD::FEXP2, MVT::f32, Expand); 447 448 // Default ISD::TRAP to expand (which turns it into abort). 449 setOperationAction(ISD::TRAP, MVT::Other, Expand); 450 451 IsLittleEndian = TD->isLittleEndian(); 452 UsesGlobalOffsetTable = false; 453 ShiftAmountTy = PointerTy = getValueType(TD->getIntPtrType()); 454 ShiftAmtHandling = Undefined; 455 memset(RegClassForVT, 0,MVT::LAST_VALUETYPE*sizeof(TargetRegisterClass*)); 456 memset(TargetDAGCombineArray, 0, array_lengthof(TargetDAGCombineArray)); 457 maxStoresPerMemset = maxStoresPerMemcpy = maxStoresPerMemmove = 8; 458 allowUnalignedMemoryAccesses = false; 459 UseUnderscoreSetJmp = false; 460 UseUnderscoreLongJmp = false; 461 SelectIsExpensive = false; 462 IntDivIsCheap = false; 463 Pow2DivIsCheap = false; 464 StackPointerRegisterToSaveRestore = 0; 465 ExceptionPointerRegister = 0; 466 ExceptionSelectorRegister = 0; 467 BooleanContents = UndefinedBooleanContent; 468 SchedPreferenceInfo = SchedulingForLatency; 469 JumpBufSize = 0; 470 JumpBufAlignment = 0; 471 IfCvtBlockSizeLimit = 2; 472 IfCvtDupBlockSizeLimit = 0; 473 PrefLoopAlignment = 0; 474 475 InitLibcallNames(LibcallRoutineNames); 476 InitCmpLibcallCCs(CmpLibcallCCs); 477 478 // Tell Legalize whether the assembler supports DEBUG_LOC. 479 const TargetAsmInfo *TASM = TM.getTargetAsmInfo(); 480 if (!TASM || !TASM->hasDotLocAndDotFile()) 481 setOperationAction(ISD::DEBUG_LOC, MVT::Other, Expand); 482} 483 484TargetLowering::~TargetLowering() {} 485 486/// computeRegisterProperties - Once all of the register classes are added, 487/// this allows us to compute derived properties we expose. 488void TargetLowering::computeRegisterProperties() { 489 assert(MVT::LAST_VALUETYPE <= 32 && 490 "Too many value types for ValueTypeActions to hold!"); 491 492 // Everything defaults to needing one register. 493 for (unsigned i = 0; i != MVT::LAST_VALUETYPE; ++i) { 494 NumRegistersForVT[i] = 1; 495 RegisterTypeForVT[i] = TransformToType[i] = (MVT::SimpleValueType)i; 496 } 497 // ...except isVoid, which doesn't need any registers. 498 NumRegistersForVT[MVT::isVoid] = 0; 499 500 // Find the largest integer register class. 501 unsigned LargestIntReg = MVT::LAST_INTEGER_VALUETYPE; 502 for (; RegClassForVT[LargestIntReg] == 0; --LargestIntReg) 503 assert(LargestIntReg != MVT::i1 && "No integer registers defined!"); 504 505 // Every integer value type larger than this largest register takes twice as 506 // many registers to represent as the previous ValueType. 507 for (unsigned ExpandedReg = LargestIntReg + 1; ; ++ExpandedReg) { 508 MVT EVT = (MVT::SimpleValueType)ExpandedReg; 509 if (!EVT.isInteger()) 510 break; 511 NumRegistersForVT[ExpandedReg] = 2*NumRegistersForVT[ExpandedReg-1]; 512 RegisterTypeForVT[ExpandedReg] = (MVT::SimpleValueType)LargestIntReg; 513 TransformToType[ExpandedReg] = (MVT::SimpleValueType)(ExpandedReg - 1); 514 ValueTypeActions.setTypeAction(EVT, Expand); 515 } 516 517 // Inspect all of the ValueType's smaller than the largest integer 518 // register to see which ones need promotion. 519 unsigned LegalIntReg = LargestIntReg; 520 for (unsigned IntReg = LargestIntReg - 1; 521 IntReg >= (unsigned)MVT::i1; --IntReg) { 522 MVT IVT = (MVT::SimpleValueType)IntReg; 523 if (isTypeLegal(IVT)) { 524 LegalIntReg = IntReg; 525 } else { 526 RegisterTypeForVT[IntReg] = TransformToType[IntReg] = 527 (MVT::SimpleValueType)LegalIntReg; 528 ValueTypeActions.setTypeAction(IVT, Promote); 529 } 530 } 531 532 // ppcf128 type is really two f64's. 533 if (!isTypeLegal(MVT::ppcf128)) { 534 NumRegistersForVT[MVT::ppcf128] = 2*NumRegistersForVT[MVT::f64]; 535 RegisterTypeForVT[MVT::ppcf128] = MVT::f64; 536 TransformToType[MVT::ppcf128] = MVT::f64; 537 ValueTypeActions.setTypeAction(MVT::ppcf128, Expand); 538 } 539 540 // Decide how to handle f64. If the target does not have native f64 support, 541 // expand it to i64 and we will be generating soft float library calls. 542 if (!isTypeLegal(MVT::f64)) { 543 NumRegistersForVT[MVT::f64] = NumRegistersForVT[MVT::i64]; 544 RegisterTypeForVT[MVT::f64] = RegisterTypeForVT[MVT::i64]; 545 TransformToType[MVT::f64] = MVT::i64; 546 ValueTypeActions.setTypeAction(MVT::f64, Expand); 547 } 548 549 // Decide how to handle f32. If the target does not have native support for 550 // f32, promote it to f64 if it is legal. Otherwise, expand it to i32. 551 if (!isTypeLegal(MVT::f32)) { 552 if (isTypeLegal(MVT::f64)) { 553 NumRegistersForVT[MVT::f32] = NumRegistersForVT[MVT::f64]; 554 RegisterTypeForVT[MVT::f32] = RegisterTypeForVT[MVT::f64]; 555 TransformToType[MVT::f32] = MVT::f64; 556 ValueTypeActions.setTypeAction(MVT::f32, Promote); 557 } else { 558 NumRegistersForVT[MVT::f32] = NumRegistersForVT[MVT::i32]; 559 RegisterTypeForVT[MVT::f32] = RegisterTypeForVT[MVT::i32]; 560 TransformToType[MVT::f32] = MVT::i32; 561 ValueTypeActions.setTypeAction(MVT::f32, Expand); 562 } 563 } 564 565 // Loop over all of the vector value types to see which need transformations. 566 for (unsigned i = MVT::FIRST_VECTOR_VALUETYPE; 567 i <= (unsigned)MVT::LAST_VECTOR_VALUETYPE; ++i) { 568 MVT VT = (MVT::SimpleValueType)i; 569 if (!isTypeLegal(VT)) { 570 MVT IntermediateVT, RegisterVT; 571 unsigned NumIntermediates; 572 NumRegistersForVT[i] = 573 getVectorTypeBreakdown(VT, 574 IntermediateVT, NumIntermediates, 575 RegisterVT); 576 RegisterTypeForVT[i] = RegisterVT; 577 578 // Determine if there is a legal wider type. 579 bool IsLegalWiderType = false; 580 MVT EltVT = VT.getVectorElementType(); 581 unsigned NElts = VT.getVectorNumElements(); 582 for (unsigned nVT = i+1; nVT <= MVT::LAST_VECTOR_VALUETYPE; ++nVT) { 583 MVT SVT = (MVT::SimpleValueType)nVT; 584 if (isTypeLegal(SVT) && SVT.getVectorElementType() == EltVT && 585 SVT.getVectorNumElements() > NElts) { 586 TransformToType[i] = SVT; 587 ValueTypeActions.setTypeAction(VT, Promote); 588 IsLegalWiderType = true; 589 break; 590 } 591 } 592 if (!IsLegalWiderType) { 593 MVT NVT = VT.getPow2VectorType(); 594 if (NVT == VT) { 595 // Type is already a power of 2. The default action is to split. 596 TransformToType[i] = MVT::Other; 597 ValueTypeActions.setTypeAction(VT, Expand); 598 } else { 599 TransformToType[i] = NVT; 600 ValueTypeActions.setTypeAction(VT, Promote); 601 } 602 } 603 } 604 } 605} 606 607const char *TargetLowering::getTargetNodeName(unsigned Opcode) const { 608 return NULL; 609} 610 611 612MVT TargetLowering::getSetCCResultType(MVT VT) const { 613 return getValueType(TD->getIntPtrType()); 614} 615 616 617/// getVectorTypeBreakdown - Vector types are broken down into some number of 618/// legal first class types. For example, MVT::v8f32 maps to 2 MVT::v4f32 619/// with Altivec or SSE1, or 8 promoted MVT::f64 values with the X86 FP stack. 620/// Similarly, MVT::v2i64 turns into 4 MVT::i32 values with both PPC and X86. 621/// 622/// This method returns the number of registers needed, and the VT for each 623/// register. It also returns the VT and quantity of the intermediate values 624/// before they are promoted/expanded. 625/// 626unsigned TargetLowering::getVectorTypeBreakdown(MVT VT, 627 MVT &IntermediateVT, 628 unsigned &NumIntermediates, 629 MVT &RegisterVT) const { 630 // Figure out the right, legal destination reg to copy into. 631 unsigned NumElts = VT.getVectorNumElements(); 632 MVT EltTy = VT.getVectorElementType(); 633 634 unsigned NumVectorRegs = 1; 635 636 // FIXME: We don't support non-power-of-2-sized vectors for now. Ideally we 637 // could break down into LHS/RHS like LegalizeDAG does. 638 if (!isPowerOf2_32(NumElts)) { 639 NumVectorRegs = NumElts; 640 NumElts = 1; 641 } 642 643 // Divide the input until we get to a supported size. This will always 644 // end with a scalar if the target doesn't support vectors. 645 while (NumElts > 1 && !isTypeLegal(MVT::getVectorVT(EltTy, NumElts))) { 646 NumElts >>= 1; 647 NumVectorRegs <<= 1; 648 } 649 650 NumIntermediates = NumVectorRegs; 651 652 MVT NewVT = MVT::getVectorVT(EltTy, NumElts); 653 if (!isTypeLegal(NewVT)) 654 NewVT = EltTy; 655 IntermediateVT = NewVT; 656 657 MVT DestVT = getTypeToTransformTo(NewVT); 658 RegisterVT = DestVT; 659 if (DestVT.bitsLT(NewVT)) { 660 // Value is expanded, e.g. i64 -> i16. 661 return NumVectorRegs*(NewVT.getSizeInBits()/DestVT.getSizeInBits()); 662 } else { 663 // Otherwise, promotion or legal types use the same number of registers as 664 // the vector decimated to the appropriate level. 665 return NumVectorRegs; 666 } 667 668 return 1; 669} 670 671/// getWidenVectorType: given a vector type, returns the type to widen to 672/// (e.g., v7i8 to v8i8). If the vector type is legal, it returns itself. 673/// If there is no vector type that we want to widen to, returns MVT::Other 674/// When and where to widen is target dependent based on the cost of 675/// scalarizing vs using the wider vector type. 676MVT TargetLowering::getWidenVectorType(MVT VT) const { 677 assert(VT.isVector()); 678 if (isTypeLegal(VT)) 679 return VT; 680 681 // Default is not to widen until moved to LegalizeTypes 682 return MVT::Other; 683} 684 685/// getByValTypeAlignment - Return the desired alignment for ByVal aggregate 686/// function arguments in the caller parameter area. This is the actual 687/// alignment, not its logarithm. 688unsigned TargetLowering::getByValTypeAlignment(const Type *Ty) const { 689 return TD->getCallFrameTypeAlignment(Ty); 690} 691 692SDValue TargetLowering::getPICJumpTableRelocBase(SDValue Table, 693 SelectionDAG &DAG) const { 694 if (usesGlobalOffsetTable()) 695 return DAG.getGLOBAL_OFFSET_TABLE(getPointerTy()); 696 return Table; 697} 698 699bool 700TargetLowering::isOffsetFoldingLegal(const GlobalAddressSDNode *GA) const { 701 // Assume that everything is safe in static mode. 702 if (getTargetMachine().getRelocationModel() == Reloc::Static) 703 return true; 704 705 // In dynamic-no-pic mode, assume that known defined values are safe. 706 if (getTargetMachine().getRelocationModel() == Reloc::DynamicNoPIC && 707 GA && 708 !GA->getGlobal()->isDeclaration() && 709 !GA->getGlobal()->mayBeOverridden()) 710 return true; 711 712 // Otherwise assume nothing is safe. 713 return false; 714} 715 716//===----------------------------------------------------------------------===// 717// Optimization Methods 718//===----------------------------------------------------------------------===// 719 720/// ShrinkDemandedConstant - Check to see if the specified operand of the 721/// specified instruction is a constant integer. If so, check to see if there 722/// are any bits set in the constant that are not demanded. If so, shrink the 723/// constant and return true. 724bool TargetLowering::TargetLoweringOpt::ShrinkDemandedConstant(SDValue Op, 725 const APInt &Demanded) { 726 DebugLoc dl = Op.getDebugLoc(); 727 // FIXME: ISD::SELECT, ISD::SELECT_CC 728 switch (Op.getOpcode()) { 729 default: break; 730 case ISD::AND: 731 case ISD::OR: 732 case ISD::XOR: 733 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op.getOperand(1))) 734 if (C->getAPIntValue().intersects(~Demanded)) { 735 MVT VT = Op.getValueType(); 736 SDValue New = DAG.getNode(Op.getOpcode(), dl, VT, Op.getOperand(0), 737 DAG.getConstant(Demanded & 738 C->getAPIntValue(), 739 VT)); 740 return CombineTo(Op, New); 741 } 742 break; 743 } 744 return false; 745} 746 747/// SimplifyDemandedBits - Look at Op. At this point, we know that only the 748/// DemandedMask bits of the result of Op are ever used downstream. If we can 749/// use this information to simplify Op, create a new simplified DAG node and 750/// return true, returning the original and new nodes in Old and New. Otherwise, 751/// analyze the expression and return a mask of KnownOne and KnownZero bits for 752/// the expression (used to simplify the caller). The KnownZero/One bits may 753/// only be accurate for those bits in the DemandedMask. 754bool TargetLowering::SimplifyDemandedBits(SDValue Op, 755 const APInt &DemandedMask, 756 APInt &KnownZero, 757 APInt &KnownOne, 758 TargetLoweringOpt &TLO, 759 unsigned Depth) const { 760 unsigned BitWidth = DemandedMask.getBitWidth(); 761 assert(Op.getValueSizeInBits() == BitWidth && 762 "Mask size mismatches value type size!"); 763 APInt NewMask = DemandedMask; 764 DebugLoc dl = Op.getDebugLoc(); 765 766 // Don't know anything. 767 KnownZero = KnownOne = APInt(BitWidth, 0); 768 769 // Other users may use these bits. 770 if (!Op.getNode()->hasOneUse()) { 771 if (Depth != 0) { 772 // If not at the root, Just compute the KnownZero/KnownOne bits to 773 // simplify things downstream. 774 TLO.DAG.ComputeMaskedBits(Op, DemandedMask, KnownZero, KnownOne, Depth); 775 return false; 776 } 777 // If this is the root being simplified, allow it to have multiple uses, 778 // just set the NewMask to all bits. 779 NewMask = APInt::getAllOnesValue(BitWidth); 780 } else if (DemandedMask == 0) { 781 // Not demanding any bits from Op. 782 if (Op.getOpcode() != ISD::UNDEF) 783 return TLO.CombineTo(Op, TLO.DAG.getUNDEF(Op.getValueType())); 784 return false; 785 } else if (Depth == 6) { // Limit search depth. 786 return false; 787 } 788 789 APInt KnownZero2, KnownOne2, KnownZeroOut, KnownOneOut; 790 switch (Op.getOpcode()) { 791 case ISD::Constant: 792 // We know all of the bits for a constant! 793 KnownOne = cast<ConstantSDNode>(Op)->getAPIntValue() & NewMask; 794 KnownZero = ~KnownOne & NewMask; 795 return false; // Don't fall through, will infinitely loop. 796 case ISD::AND: 797 // If the RHS is a constant, check to see if the LHS would be zero without 798 // using the bits from the RHS. Below, we use knowledge about the RHS to 799 // simplify the LHS, here we're using information from the LHS to simplify 800 // the RHS. 801 if (ConstantSDNode *RHSC = dyn_cast<ConstantSDNode>(Op.getOperand(1))) { 802 APInt LHSZero, LHSOne; 803 TLO.DAG.ComputeMaskedBits(Op.getOperand(0), NewMask, 804 LHSZero, LHSOne, Depth+1); 805 // If the LHS already has zeros where RHSC does, this and is dead. 806 if ((LHSZero & NewMask) == (~RHSC->getAPIntValue() & NewMask)) 807 return TLO.CombineTo(Op, Op.getOperand(0)); 808 // If any of the set bits in the RHS are known zero on the LHS, shrink 809 // the constant. 810 if (TLO.ShrinkDemandedConstant(Op, ~LHSZero & NewMask)) 811 return true; 812 } 813 814 if (SimplifyDemandedBits(Op.getOperand(1), NewMask, KnownZero, 815 KnownOne, TLO, Depth+1)) 816 return true; 817 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?"); 818 if (SimplifyDemandedBits(Op.getOperand(0), ~KnownZero & NewMask, 819 KnownZero2, KnownOne2, TLO, Depth+1)) 820 return true; 821 assert((KnownZero2 & KnownOne2) == 0 && "Bits known to be one AND zero?"); 822 823 // If all of the demanded bits are known one on one side, return the other. 824 // These bits cannot contribute to the result of the 'and'. 825 if ((NewMask & ~KnownZero2 & KnownOne) == (~KnownZero2 & NewMask)) 826 return TLO.CombineTo(Op, Op.getOperand(0)); 827 if ((NewMask & ~KnownZero & KnownOne2) == (~KnownZero & NewMask)) 828 return TLO.CombineTo(Op, Op.getOperand(1)); 829 // If all of the demanded bits in the inputs are known zeros, return zero. 830 if ((NewMask & (KnownZero|KnownZero2)) == NewMask) 831 return TLO.CombineTo(Op, TLO.DAG.getConstant(0, Op.getValueType())); 832 // If the RHS is a constant, see if we can simplify it. 833 if (TLO.ShrinkDemandedConstant(Op, ~KnownZero2 & NewMask)) 834 return true; 835 836 // Output known-1 bits are only known if set in both the LHS & RHS. 837 KnownOne &= KnownOne2; 838 // Output known-0 are known to be clear if zero in either the LHS | RHS. 839 KnownZero |= KnownZero2; 840 break; 841 case ISD::OR: 842 if (SimplifyDemandedBits(Op.getOperand(1), NewMask, KnownZero, 843 KnownOne, TLO, Depth+1)) 844 return true; 845 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?"); 846 if (SimplifyDemandedBits(Op.getOperand(0), ~KnownOne & NewMask, 847 KnownZero2, KnownOne2, TLO, Depth+1)) 848 return true; 849 assert((KnownZero2 & KnownOne2) == 0 && "Bits known to be one AND zero?"); 850 851 // If all of the demanded bits are known zero on one side, return the other. 852 // These bits cannot contribute to the result of the 'or'. 853 if ((NewMask & ~KnownOne2 & KnownZero) == (~KnownOne2 & NewMask)) 854 return TLO.CombineTo(Op, Op.getOperand(0)); 855 if ((NewMask & ~KnownOne & KnownZero2) == (~KnownOne & NewMask)) 856 return TLO.CombineTo(Op, Op.getOperand(1)); 857 // If all of the potentially set bits on one side are known to be set on 858 // the other side, just use the 'other' side. 859 if ((NewMask & ~KnownZero & KnownOne2) == (~KnownZero & NewMask)) 860 return TLO.CombineTo(Op, Op.getOperand(0)); 861 if ((NewMask & ~KnownZero2 & KnownOne) == (~KnownZero2 & NewMask)) 862 return TLO.CombineTo(Op, Op.getOperand(1)); 863 // If the RHS is a constant, see if we can simplify it. 864 if (TLO.ShrinkDemandedConstant(Op, NewMask)) 865 return true; 866 867 // Output known-0 bits are only known if clear in both the LHS & RHS. 868 KnownZero &= KnownZero2; 869 // Output known-1 are known to be set if set in either the LHS | RHS. 870 KnownOne |= KnownOne2; 871 break; 872 case ISD::XOR: 873 if (SimplifyDemandedBits(Op.getOperand(1), NewMask, KnownZero, 874 KnownOne, TLO, Depth+1)) 875 return true; 876 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?"); 877 if (SimplifyDemandedBits(Op.getOperand(0), NewMask, KnownZero2, 878 KnownOne2, TLO, Depth+1)) 879 return true; 880 assert((KnownZero2 & KnownOne2) == 0 && "Bits known to be one AND zero?"); 881 882 // If all of the demanded bits are known zero on one side, return the other. 883 // These bits cannot contribute to the result of the 'xor'. 884 if ((KnownZero & NewMask) == NewMask) 885 return TLO.CombineTo(Op, Op.getOperand(0)); 886 if ((KnownZero2 & NewMask) == NewMask) 887 return TLO.CombineTo(Op, Op.getOperand(1)); 888 889 // If all of the unknown bits are known to be zero on one side or the other 890 // (but not both) turn this into an *inclusive* or. 891 // e.g. (A & C1)^(B & C2) -> (A & C1)|(B & C2) iff C1&C2 == 0 892 if ((NewMask & ~KnownZero & ~KnownZero2) == 0) 893 return TLO.CombineTo(Op, TLO.DAG.getNode(ISD::OR, dl, Op.getValueType(), 894 Op.getOperand(0), 895 Op.getOperand(1))); 896 897 // Output known-0 bits are known if clear or set in both the LHS & RHS. 898 KnownZeroOut = (KnownZero & KnownZero2) | (KnownOne & KnownOne2); 899 // Output known-1 are known to be set if set in only one of the LHS, RHS. 900 KnownOneOut = (KnownZero & KnownOne2) | (KnownOne & KnownZero2); 901 902 // If all of the demanded bits on one side are known, and all of the set 903 // bits on that side are also known to be set on the other side, turn this 904 // into an AND, as we know the bits will be cleared. 905 // e.g. (X | C1) ^ C2 --> (X | C1) & ~C2 iff (C1&C2) == C2 906 if ((NewMask & (KnownZero|KnownOne)) == NewMask) { // all known 907 if ((KnownOne & KnownOne2) == KnownOne) { 908 MVT VT = Op.getValueType(); 909 SDValue ANDC = TLO.DAG.getConstant(~KnownOne & NewMask, VT); 910 return TLO.CombineTo(Op, TLO.DAG.getNode(ISD::AND, dl, VT, 911 Op.getOperand(0), ANDC)); 912 } 913 } 914 915 // If the RHS is a constant, see if we can simplify it. 916 // for XOR, we prefer to force bits to 1 if they will make a -1. 917 // if we can't force bits, try to shrink constant 918 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op.getOperand(1))) { 919 APInt Expanded = C->getAPIntValue() | (~NewMask); 920 // if we can expand it to have all bits set, do it 921 if (Expanded.isAllOnesValue()) { 922 if (Expanded != C->getAPIntValue()) { 923 MVT VT = Op.getValueType(); 924 SDValue New = TLO.DAG.getNode(Op.getOpcode(), dl,VT, Op.getOperand(0), 925 TLO.DAG.getConstant(Expanded, VT)); 926 return TLO.CombineTo(Op, New); 927 } 928 // if it already has all the bits set, nothing to change 929 // but don't shrink either! 930 } else if (TLO.ShrinkDemandedConstant(Op, NewMask)) { 931 return true; 932 } 933 } 934 935 KnownZero = KnownZeroOut; 936 KnownOne = KnownOneOut; 937 break; 938 case ISD::SELECT: 939 if (SimplifyDemandedBits(Op.getOperand(2), NewMask, KnownZero, 940 KnownOne, TLO, Depth+1)) 941 return true; 942 if (SimplifyDemandedBits(Op.getOperand(1), NewMask, KnownZero2, 943 KnownOne2, TLO, Depth+1)) 944 return true; 945 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?"); 946 assert((KnownZero2 & KnownOne2) == 0 && "Bits known to be one AND zero?"); 947 948 // If the operands are constants, see if we can simplify them. 949 if (TLO.ShrinkDemandedConstant(Op, NewMask)) 950 return true; 951 952 // Only known if known in both the LHS and RHS. 953 KnownOne &= KnownOne2; 954 KnownZero &= KnownZero2; 955 break; 956 case ISD::SELECT_CC: 957 if (SimplifyDemandedBits(Op.getOperand(3), NewMask, KnownZero, 958 KnownOne, TLO, Depth+1)) 959 return true; 960 if (SimplifyDemandedBits(Op.getOperand(2), NewMask, KnownZero2, 961 KnownOne2, TLO, Depth+1)) 962 return true; 963 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?"); 964 assert((KnownZero2 & KnownOne2) == 0 && "Bits known to be one AND zero?"); 965 966 // If the operands are constants, see if we can simplify them. 967 if (TLO.ShrinkDemandedConstant(Op, NewMask)) 968 return true; 969 970 // Only known if known in both the LHS and RHS. 971 KnownOne &= KnownOne2; 972 KnownZero &= KnownZero2; 973 break; 974 case ISD::SHL: 975 if (ConstantSDNode *SA = dyn_cast<ConstantSDNode>(Op.getOperand(1))) { 976 unsigned ShAmt = SA->getZExtValue(); 977 SDValue InOp = Op.getOperand(0); 978 979 // If the shift count is an invalid immediate, don't do anything. 980 if (ShAmt >= BitWidth) 981 break; 982 983 // If this is ((X >>u C1) << ShAmt), see if we can simplify this into a 984 // single shift. We can do this if the bottom bits (which are shifted 985 // out) are never demanded. 986 if (InOp.getOpcode() == ISD::SRL && 987 isa<ConstantSDNode>(InOp.getOperand(1))) { 988 if (ShAmt && (NewMask & APInt::getLowBitsSet(BitWidth, ShAmt)) == 0) { 989 unsigned C1= cast<ConstantSDNode>(InOp.getOperand(1))->getZExtValue(); 990 unsigned Opc = ISD::SHL; 991 int Diff = ShAmt-C1; 992 if (Diff < 0) { 993 Diff = -Diff; 994 Opc = ISD::SRL; 995 } 996 997 SDValue NewSA = 998 TLO.DAG.getConstant(Diff, Op.getOperand(1).getValueType()); 999 MVT VT = Op.getValueType(); 1000 return TLO.CombineTo(Op, TLO.DAG.getNode(Opc, dl, VT, 1001 InOp.getOperand(0), NewSA)); 1002 } 1003 } 1004 1005 if (SimplifyDemandedBits(Op.getOperand(0), NewMask.lshr(ShAmt), 1006 KnownZero, KnownOne, TLO, Depth+1)) 1007 return true; 1008 KnownZero <<= SA->getZExtValue(); 1009 KnownOne <<= SA->getZExtValue(); 1010 // low bits known zero. 1011 KnownZero |= APInt::getLowBitsSet(BitWidth, SA->getZExtValue()); 1012 } 1013 break; 1014 case ISD::SRL: 1015 if (ConstantSDNode *SA = dyn_cast<ConstantSDNode>(Op.getOperand(1))) { 1016 MVT VT = Op.getValueType(); 1017 unsigned ShAmt = SA->getZExtValue(); 1018 unsigned VTSize = VT.getSizeInBits(); 1019 SDValue InOp = Op.getOperand(0); 1020 1021 // If the shift count is an invalid immediate, don't do anything. 1022 if (ShAmt >= BitWidth) 1023 break; 1024 1025 // If this is ((X << C1) >>u ShAmt), see if we can simplify this into a 1026 // single shift. We can do this if the top bits (which are shifted out) 1027 // are never demanded. 1028 if (InOp.getOpcode() == ISD::SHL && 1029 isa<ConstantSDNode>(InOp.getOperand(1))) { 1030 if (ShAmt && (NewMask & APInt::getHighBitsSet(VTSize, ShAmt)) == 0) { 1031 unsigned C1= cast<ConstantSDNode>(InOp.getOperand(1))->getZExtValue(); 1032 unsigned Opc = ISD::SRL; 1033 int Diff = ShAmt-C1; 1034 if (Diff < 0) { 1035 Diff = -Diff; 1036 Opc = ISD::SHL; 1037 } 1038 1039 SDValue NewSA = 1040 TLO.DAG.getConstant(Diff, Op.getOperand(1).getValueType()); 1041 return TLO.CombineTo(Op, TLO.DAG.getNode(Opc, dl, VT, 1042 InOp.getOperand(0), NewSA)); 1043 } 1044 } 1045 1046 // Compute the new bits that are at the top now. 1047 if (SimplifyDemandedBits(InOp, (NewMask << ShAmt), 1048 KnownZero, KnownOne, TLO, Depth+1)) 1049 return true; 1050 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?"); 1051 KnownZero = KnownZero.lshr(ShAmt); 1052 KnownOne = KnownOne.lshr(ShAmt); 1053 1054 APInt HighBits = APInt::getHighBitsSet(BitWidth, ShAmt); 1055 KnownZero |= HighBits; // High bits known zero. 1056 } 1057 break; 1058 case ISD::SRA: 1059 // If this is an arithmetic shift right and only the low-bit is set, we can 1060 // always convert this into a logical shr, even if the shift amount is 1061 // variable. The low bit of the shift cannot be an input sign bit unless 1062 // the shift amount is >= the size of the datatype, which is undefined. 1063 if (DemandedMask == 1) 1064 return TLO.CombineTo(Op, TLO.DAG.getNode(ISD::SRL, dl, Op.getValueType(), 1065 Op.getOperand(0), Op.getOperand(1))); 1066 1067 if (ConstantSDNode *SA = dyn_cast<ConstantSDNode>(Op.getOperand(1))) { 1068 MVT VT = Op.getValueType(); 1069 unsigned ShAmt = SA->getZExtValue(); 1070 1071 // If the shift count is an invalid immediate, don't do anything. 1072 if (ShAmt >= BitWidth) 1073 break; 1074 1075 APInt InDemandedMask = (NewMask << ShAmt); 1076 1077 // If any of the demanded bits are produced by the sign extension, we also 1078 // demand the input sign bit. 1079 APInt HighBits = APInt::getHighBitsSet(BitWidth, ShAmt); 1080 if (HighBits.intersects(NewMask)) 1081 InDemandedMask |= APInt::getSignBit(VT.getSizeInBits()); 1082 1083 if (SimplifyDemandedBits(Op.getOperand(0), InDemandedMask, 1084 KnownZero, KnownOne, TLO, Depth+1)) 1085 return true; 1086 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?"); 1087 KnownZero = KnownZero.lshr(ShAmt); 1088 KnownOne = KnownOne.lshr(ShAmt); 1089 1090 // Handle the sign bit, adjusted to where it is now in the mask. 1091 APInt SignBit = APInt::getSignBit(BitWidth).lshr(ShAmt); 1092 1093 // If the input sign bit is known to be zero, or if none of the top bits 1094 // are demanded, turn this into an unsigned shift right. 1095 if (KnownZero.intersects(SignBit) || (HighBits & ~NewMask) == HighBits) { 1096 return TLO.CombineTo(Op, TLO.DAG.getNode(ISD::SRL, dl, VT, 1097 Op.getOperand(0), 1098 Op.getOperand(1))); 1099 } else if (KnownOne.intersects(SignBit)) { // New bits are known one. 1100 KnownOne |= HighBits; 1101 } 1102 } 1103 break; 1104 case ISD::SIGN_EXTEND_INREG: { 1105 MVT EVT = cast<VTSDNode>(Op.getOperand(1))->getVT(); 1106 1107 // Sign extension. Compute the demanded bits in the result that are not 1108 // present in the input. 1109 APInt NewBits = APInt::getHighBitsSet(BitWidth, 1110 BitWidth - EVT.getSizeInBits()) & 1111 NewMask; 1112 1113 // If none of the extended bits are demanded, eliminate the sextinreg. 1114 if (NewBits == 0) 1115 return TLO.CombineTo(Op, Op.getOperand(0)); 1116 1117 APInt InSignBit = APInt::getSignBit(EVT.getSizeInBits()); 1118 InSignBit.zext(BitWidth); 1119 APInt InputDemandedBits = APInt::getLowBitsSet(BitWidth, 1120 EVT.getSizeInBits()) & 1121 NewMask; 1122 1123 // Since the sign extended bits are demanded, we know that the sign 1124 // bit is demanded. 1125 InputDemandedBits |= InSignBit; 1126 1127 if (SimplifyDemandedBits(Op.getOperand(0), InputDemandedBits, 1128 KnownZero, KnownOne, TLO, Depth+1)) 1129 return true; 1130 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?"); 1131 1132 // If the sign bit of the input is known set or clear, then we know the 1133 // top bits of the result. 1134 1135 // If the input sign bit is known zero, convert this into a zero extension. 1136 if (KnownZero.intersects(InSignBit)) 1137 return TLO.CombineTo(Op, 1138 TLO.DAG.getZeroExtendInReg(Op.getOperand(0),dl,EVT)); 1139 1140 if (KnownOne.intersects(InSignBit)) { // Input sign bit known set 1141 KnownOne |= NewBits; 1142 KnownZero &= ~NewBits; 1143 } else { // Input sign bit unknown 1144 KnownZero &= ~NewBits; 1145 KnownOne &= ~NewBits; 1146 } 1147 break; 1148 } 1149 case ISD::ZERO_EXTEND: { 1150 unsigned OperandBitWidth = Op.getOperand(0).getValueSizeInBits(); 1151 APInt InMask = NewMask; 1152 InMask.trunc(OperandBitWidth); 1153 1154 // If none of the top bits are demanded, convert this into an any_extend. 1155 APInt NewBits = 1156 APInt::getHighBitsSet(BitWidth, BitWidth - OperandBitWidth) & NewMask; 1157 if (!NewBits.intersects(NewMask)) 1158 return TLO.CombineTo(Op, TLO.DAG.getNode(ISD::ANY_EXTEND, dl, 1159 Op.getValueType(), 1160 Op.getOperand(0))); 1161 1162 if (SimplifyDemandedBits(Op.getOperand(0), InMask, 1163 KnownZero, KnownOne, TLO, Depth+1)) 1164 return true; 1165 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?"); 1166 KnownZero.zext(BitWidth); 1167 KnownOne.zext(BitWidth); 1168 KnownZero |= NewBits; 1169 break; 1170 } 1171 case ISD::SIGN_EXTEND: { 1172 MVT InVT = Op.getOperand(0).getValueType(); 1173 unsigned InBits = InVT.getSizeInBits(); 1174 APInt InMask = APInt::getLowBitsSet(BitWidth, InBits); 1175 APInt InSignBit = APInt::getBitsSet(BitWidth, InBits - 1, InBits); 1176 APInt NewBits = ~InMask & NewMask; 1177 1178 // If none of the top bits are demanded, convert this into an any_extend. 1179 if (NewBits == 0) 1180 return TLO.CombineTo(Op,TLO.DAG.getNode(ISD::ANY_EXTEND, dl, 1181 Op.getValueType(), 1182 Op.getOperand(0))); 1183 1184 // Since some of the sign extended bits are demanded, we know that the sign 1185 // bit is demanded. 1186 APInt InDemandedBits = InMask & NewMask; 1187 InDemandedBits |= InSignBit; 1188 InDemandedBits.trunc(InBits); 1189 1190 if (SimplifyDemandedBits(Op.getOperand(0), InDemandedBits, KnownZero, 1191 KnownOne, TLO, Depth+1)) 1192 return true; 1193 KnownZero.zext(BitWidth); 1194 KnownOne.zext(BitWidth); 1195 1196 // If the sign bit is known zero, convert this to a zero extend. 1197 if (KnownZero.intersects(InSignBit)) 1198 return TLO.CombineTo(Op, TLO.DAG.getNode(ISD::ZERO_EXTEND, dl, 1199 Op.getValueType(), 1200 Op.getOperand(0))); 1201 1202 // If the sign bit is known one, the top bits match. 1203 if (KnownOne.intersects(InSignBit)) { 1204 KnownOne |= NewBits; 1205 KnownZero &= ~NewBits; 1206 } else { // Otherwise, top bits aren't known. 1207 KnownOne &= ~NewBits; 1208 KnownZero &= ~NewBits; 1209 } 1210 break; 1211 } 1212 case ISD::ANY_EXTEND: { 1213 unsigned OperandBitWidth = Op.getOperand(0).getValueSizeInBits(); 1214 APInt InMask = NewMask; 1215 InMask.trunc(OperandBitWidth); 1216 if (SimplifyDemandedBits(Op.getOperand(0), InMask, 1217 KnownZero, KnownOne, TLO, Depth+1)) 1218 return true; 1219 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?"); 1220 KnownZero.zext(BitWidth); 1221 KnownOne.zext(BitWidth); 1222 break; 1223 } 1224 case ISD::TRUNCATE: { 1225 // Simplify the input, using demanded bit information, and compute the known 1226 // zero/one bits live out. 1227 APInt TruncMask = NewMask; 1228 TruncMask.zext(Op.getOperand(0).getValueSizeInBits()); 1229 if (SimplifyDemandedBits(Op.getOperand(0), TruncMask, 1230 KnownZero, KnownOne, TLO, Depth+1)) 1231 return true; 1232 KnownZero.trunc(BitWidth); 1233 KnownOne.trunc(BitWidth); 1234 1235 // If the input is only used by this truncate, see if we can shrink it based 1236 // on the known demanded bits. 1237 if (Op.getOperand(0).getNode()->hasOneUse()) { 1238 SDValue In = Op.getOperand(0); 1239 unsigned InBitWidth = In.getValueSizeInBits(); 1240 switch (In.getOpcode()) { 1241 default: break; 1242 case ISD::SRL: 1243 // Shrink SRL by a constant if none of the high bits shifted in are 1244 // demanded. 1245 if (ConstantSDNode *ShAmt = dyn_cast<ConstantSDNode>(In.getOperand(1))){ 1246 APInt HighBits = APInt::getHighBitsSet(InBitWidth, 1247 InBitWidth - BitWidth); 1248 HighBits = HighBits.lshr(ShAmt->getZExtValue()); 1249 HighBits.trunc(BitWidth); 1250 1251 if (ShAmt->getZExtValue() < BitWidth && !(HighBits & NewMask)) { 1252 // None of the shifted in bits are needed. Add a truncate of the 1253 // shift input, then shift it. 1254 SDValue NewTrunc = TLO.DAG.getNode(ISD::TRUNCATE, dl, 1255 Op.getValueType(), 1256 In.getOperand(0)); 1257 return TLO.CombineTo(Op, TLO.DAG.getNode(ISD::SRL, dl, 1258 Op.getValueType(), 1259 NewTrunc, 1260 In.getOperand(1))); 1261 } 1262 } 1263 break; 1264 } 1265 } 1266 1267 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?"); 1268 break; 1269 } 1270 case ISD::AssertZext: { 1271 MVT VT = cast<VTSDNode>(Op.getOperand(1))->getVT(); 1272 APInt InMask = APInt::getLowBitsSet(BitWidth, 1273 VT.getSizeInBits()); 1274 if (SimplifyDemandedBits(Op.getOperand(0), InMask & NewMask, 1275 KnownZero, KnownOne, TLO, Depth+1)) 1276 return true; 1277 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?"); 1278 KnownZero |= ~InMask & NewMask; 1279 break; 1280 } 1281 case ISD::BIT_CONVERT: 1282#if 0 1283 // If this is an FP->Int bitcast and if the sign bit is the only thing that 1284 // is demanded, turn this into a FGETSIGN. 1285 if (NewMask == MVT::getIntegerVTSignBit(Op.getValueType()) && 1286 MVT::isFloatingPoint(Op.getOperand(0).getValueType()) && 1287 !MVT::isVector(Op.getOperand(0).getValueType())) { 1288 // Only do this xform if FGETSIGN is valid or if before legalize. 1289 if (!TLO.AfterLegalize || 1290 isOperationLegal(ISD::FGETSIGN, Op.getValueType())) { 1291 // Make a FGETSIGN + SHL to move the sign bit into the appropriate 1292 // place. We expect the SHL to be eliminated by other optimizations. 1293 SDValue Sign = TLO.DAG.getNode(ISD::FGETSIGN, Op.getValueType(), 1294 Op.getOperand(0)); 1295 unsigned ShVal = Op.getValueType().getSizeInBits()-1; 1296 SDValue ShAmt = TLO.DAG.getConstant(ShVal, getShiftAmountTy()); 1297 return TLO.CombineTo(Op, TLO.DAG.getNode(ISD::SHL, Op.getValueType(), 1298 Sign, ShAmt)); 1299 } 1300 } 1301#endif 1302 break; 1303 default: 1304 // Just use ComputeMaskedBits to compute output bits. 1305 TLO.DAG.ComputeMaskedBits(Op, NewMask, KnownZero, KnownOne, Depth); 1306 break; 1307 } 1308 1309 // If we know the value of all of the demanded bits, return this as a 1310 // constant. 1311 if ((NewMask & (KnownZero|KnownOne)) == NewMask) 1312 return TLO.CombineTo(Op, TLO.DAG.getConstant(KnownOne, Op.getValueType())); 1313 1314 return false; 1315} 1316 1317/// computeMaskedBitsForTargetNode - Determine which of the bits specified 1318/// in Mask are known to be either zero or one and return them in the 1319/// KnownZero/KnownOne bitsets. 1320void TargetLowering::computeMaskedBitsForTargetNode(const SDValue Op, 1321 const APInt &Mask, 1322 APInt &KnownZero, 1323 APInt &KnownOne, 1324 const SelectionDAG &DAG, 1325 unsigned Depth) const { 1326 assert((Op.getOpcode() >= ISD::BUILTIN_OP_END || 1327 Op.getOpcode() == ISD::INTRINSIC_WO_CHAIN || 1328 Op.getOpcode() == ISD::INTRINSIC_W_CHAIN || 1329 Op.getOpcode() == ISD::INTRINSIC_VOID) && 1330 "Should use MaskedValueIsZero if you don't know whether Op" 1331 " is a target node!"); 1332 KnownZero = KnownOne = APInt(Mask.getBitWidth(), 0); 1333} 1334 1335/// ComputeNumSignBitsForTargetNode - This method can be implemented by 1336/// targets that want to expose additional information about sign bits to the 1337/// DAG Combiner. 1338unsigned TargetLowering::ComputeNumSignBitsForTargetNode(SDValue Op, 1339 unsigned Depth) const { 1340 assert((Op.getOpcode() >= ISD::BUILTIN_OP_END || 1341 Op.getOpcode() == ISD::INTRINSIC_WO_CHAIN || 1342 Op.getOpcode() == ISD::INTRINSIC_W_CHAIN || 1343 Op.getOpcode() == ISD::INTRINSIC_VOID) && 1344 "Should use ComputeNumSignBits if you don't know whether Op" 1345 " is a target node!"); 1346 return 1; 1347} 1348 1349static bool ValueHasExactlyOneBitSet(SDValue Val, const SelectionDAG &DAG) { 1350 // Logical shift right or left won't ever introduce new set bits. 1351 // We check for this case because we don't care which bits are 1352 // set, but ComputeMaskedBits won't know anything unless it can 1353 // determine which specific bits may be set. 1354 if (Val.getOpcode() == ISD::SHL || Val.getOpcode() == ISD::SRL) 1355 return ValueHasExactlyOneBitSet(Val.getOperand(0), DAG); 1356 1357 MVT OpVT = Val.getValueType(); 1358 unsigned BitWidth = OpVT.getSizeInBits(); 1359 APInt Mask = APInt::getAllOnesValue(BitWidth); 1360 APInt KnownZero, KnownOne; 1361 DAG.ComputeMaskedBits(Val, Mask, KnownZero, KnownOne); 1362 return (KnownZero.countPopulation() == BitWidth - 1) && 1363 (KnownOne.countPopulation() == 1); 1364} 1365 1366/// SimplifySetCC - Try to simplify a setcc built with the specified operands 1367/// and cc. If it is unable to simplify it, return a null SDValue. 1368SDValue 1369TargetLowering::SimplifySetCC(MVT VT, SDValue N0, SDValue N1, 1370 ISD::CondCode Cond, bool foldBooleans, 1371 DAGCombinerInfo &DCI, DebugLoc dl) const { 1372 SelectionDAG &DAG = DCI.DAG; 1373 1374 // These setcc operations always fold. 1375 switch (Cond) { 1376 default: break; 1377 case ISD::SETFALSE: 1378 case ISD::SETFALSE2: return DAG.getConstant(0, VT); 1379 case ISD::SETTRUE: 1380 case ISD::SETTRUE2: return DAG.getConstant(1, VT); 1381 } 1382 1383 if (ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1.getNode())) { 1384 const APInt &C1 = N1C->getAPIntValue(); 1385 if (isa<ConstantSDNode>(N0.getNode())) { 1386 return DAG.FoldSetCC(VT, N0, N1, Cond, dl); 1387 } else { 1388 // If the LHS is '(srl (ctlz x), 5)', the RHS is 0/1, and this is an 1389 // equality comparison, then we're just comparing whether X itself is 1390 // zero. 1391 if (N0.getOpcode() == ISD::SRL && (C1 == 0 || C1 == 1) && 1392 N0.getOperand(0).getOpcode() == ISD::CTLZ && 1393 N0.getOperand(1).getOpcode() == ISD::Constant) { 1394 unsigned ShAmt = cast<ConstantSDNode>(N0.getOperand(1))->getZExtValue(); 1395 if ((Cond == ISD::SETEQ || Cond == ISD::SETNE) && 1396 ShAmt == Log2_32(N0.getValueType().getSizeInBits())) { 1397 if ((C1 == 0) == (Cond == ISD::SETEQ)) { 1398 // (srl (ctlz x), 5) == 0 -> X != 0 1399 // (srl (ctlz x), 5) != 1 -> X != 0 1400 Cond = ISD::SETNE; 1401 } else { 1402 // (srl (ctlz x), 5) != 0 -> X == 0 1403 // (srl (ctlz x), 5) == 1 -> X == 0 1404 Cond = ISD::SETEQ; 1405 } 1406 SDValue Zero = DAG.getConstant(0, N0.getValueType()); 1407 return DAG.getSetCC(dl, VT, N0.getOperand(0).getOperand(0), 1408 Zero, Cond); 1409 } 1410 } 1411 1412 // If the LHS is '(and load, const)', the RHS is 0, 1413 // the test is for equality or unsigned, and all 1 bits of the const are 1414 // in the same partial word, see if we can shorten the load. 1415 if (DCI.isBeforeLegalize() && 1416 N0.getOpcode() == ISD::AND && C1 == 0 && 1417 isa<LoadSDNode>(N0.getOperand(0)) && 1418 N0.getOperand(0).getNode()->hasOneUse() && 1419 isa<ConstantSDNode>(N0.getOperand(1))) { 1420 LoadSDNode *Lod = cast<LoadSDNode>(N0.getOperand(0)); 1421 uint64_t Mask = cast<ConstantSDNode>(N0.getOperand(1))->getZExtValue(); 1422 uint64_t bestMask = 0; 1423 unsigned bestWidth = 0, bestOffset = 0; 1424 if (!Lod->isVolatile() && Lod->isUnindexed()) { 1425 unsigned origWidth = N0.getValueType().getSizeInBits(); 1426 // We can narrow (e.g.) 16-bit extending loads on 32-bit target to 1427 // 8 bits, but have to be careful... 1428 if (Lod->getExtensionType() != ISD::NON_EXTLOAD) 1429 origWidth = Lod->getMemoryVT().getSizeInBits(); 1430 for (unsigned width = origWidth / 2; width>=8; width /= 2) { 1431 uint64_t newMask = (1ULL << width) - 1; 1432 for (unsigned offset=0; offset<origWidth/width; offset++) { 1433 if ((newMask & Mask)==Mask) { 1434 if (!TD->isLittleEndian()) 1435 bestOffset = (origWidth/width - offset - 1) * (width/8); 1436 else 1437 bestOffset = (uint64_t)offset * (width/8); 1438 bestMask = Mask >> (offset * (width/8) * 8); 1439 bestWidth = width; 1440 break; 1441 } 1442 newMask = newMask << width; 1443 } 1444 } 1445 } 1446 if (bestWidth) { 1447 MVT newVT = MVT::getIntegerVT(bestWidth); 1448 if (newVT.isRound()) { 1449 MVT PtrType = Lod->getOperand(1).getValueType(); 1450 SDValue Ptr = Lod->getBasePtr(); 1451 if (bestOffset != 0) 1452 Ptr = DAG.getNode(ISD::ADD, dl, PtrType, Lod->getBasePtr(), 1453 DAG.getConstant(bestOffset, PtrType)); 1454 unsigned NewAlign = MinAlign(Lod->getAlignment(), bestOffset); 1455 SDValue NewLoad = DAG.getLoad(newVT, dl, Lod->getChain(), Ptr, 1456 Lod->getSrcValue(), 1457 Lod->getSrcValueOffset() + bestOffset, 1458 false, NewAlign); 1459 return DAG.getSetCC(dl, VT, 1460 DAG.getNode(ISD::AND, dl, newVT, NewLoad, 1461 DAG.getConstant(bestMask, newVT)), 1462 DAG.getConstant(0LL, newVT), Cond); 1463 } 1464 } 1465 } 1466 1467 // If the LHS is a ZERO_EXTEND, perform the comparison on the input. 1468 if (N0.getOpcode() == ISD::ZERO_EXTEND) { 1469 unsigned InSize = N0.getOperand(0).getValueType().getSizeInBits(); 1470 1471 // If the comparison constant has bits in the upper part, the 1472 // zero-extended value could never match. 1473 if (C1.intersects(APInt::getHighBitsSet(C1.getBitWidth(), 1474 C1.getBitWidth() - InSize))) { 1475 switch (Cond) { 1476 case ISD::SETUGT: 1477 case ISD::SETUGE: 1478 case ISD::SETEQ: return DAG.getConstant(0, VT); 1479 case ISD::SETULT: 1480 case ISD::SETULE: 1481 case ISD::SETNE: return DAG.getConstant(1, VT); 1482 case ISD::SETGT: 1483 case ISD::SETGE: 1484 // True if the sign bit of C1 is set. 1485 return DAG.getConstant(C1.isNegative(), VT); 1486 case ISD::SETLT: 1487 case ISD::SETLE: 1488 // True if the sign bit of C1 isn't set. 1489 return DAG.getConstant(C1.isNonNegative(), VT); 1490 default: 1491 break; 1492 } 1493 } 1494 1495 // Otherwise, we can perform the comparison with the low bits. 1496 switch (Cond) { 1497 case ISD::SETEQ: 1498 case ISD::SETNE: 1499 case ISD::SETUGT: 1500 case ISD::SETUGE: 1501 case ISD::SETULT: 1502 case ISD::SETULE: 1503 return DAG.getSetCC(dl, VT, N0.getOperand(0), 1504 DAG.getConstant(APInt(C1).trunc(InSize), 1505 N0.getOperand(0).getValueType()), 1506 Cond); 1507 default: 1508 break; // todo, be more careful with signed comparisons 1509 } 1510 } else if (N0.getOpcode() == ISD::SIGN_EXTEND_INREG && 1511 (Cond == ISD::SETEQ || Cond == ISD::SETNE)) { 1512 MVT ExtSrcTy = cast<VTSDNode>(N0.getOperand(1))->getVT(); 1513 unsigned ExtSrcTyBits = ExtSrcTy.getSizeInBits(); 1514 MVT ExtDstTy = N0.getValueType(); 1515 unsigned ExtDstTyBits = ExtDstTy.getSizeInBits(); 1516 1517 // If the extended part has any inconsistent bits, it cannot ever 1518 // compare equal. In other words, they have to be all ones or all 1519 // zeros. 1520 APInt ExtBits = 1521 APInt::getHighBitsSet(ExtDstTyBits, ExtDstTyBits - ExtSrcTyBits); 1522 if ((C1 & ExtBits) != 0 && (C1 & ExtBits) != ExtBits) 1523 return DAG.getConstant(Cond == ISD::SETNE, VT); 1524 1525 SDValue ZextOp; 1526 MVT Op0Ty = N0.getOperand(0).getValueType(); 1527 if (Op0Ty == ExtSrcTy) { 1528 ZextOp = N0.getOperand(0); 1529 } else { 1530 APInt Imm = APInt::getLowBitsSet(ExtDstTyBits, ExtSrcTyBits); 1531 ZextOp = DAG.getNode(ISD::AND, dl, Op0Ty, N0.getOperand(0), 1532 DAG.getConstant(Imm, Op0Ty)); 1533 } 1534 if (!DCI.isCalledByLegalizer()) 1535 DCI.AddToWorklist(ZextOp.getNode()); 1536 // Otherwise, make this a use of a zext. 1537 return DAG.getSetCC(dl, VT, ZextOp, 1538 DAG.getConstant(C1 & APInt::getLowBitsSet( 1539 ExtDstTyBits, 1540 ExtSrcTyBits), 1541 ExtDstTy), 1542 Cond); 1543 } else if ((N1C->isNullValue() || N1C->getAPIntValue() == 1) && 1544 (Cond == ISD::SETEQ || Cond == ISD::SETNE)) { 1545 1546 // SETCC (SETCC), [0|1], [EQ|NE] -> SETCC 1547 if (N0.getOpcode() == ISD::SETCC) { 1548 bool TrueWhenTrue = (Cond == ISD::SETEQ) ^ (N1C->getZExtValue() != 1); 1549 if (TrueWhenTrue) 1550 return N0; 1551 1552 // Invert the condition. 1553 ISD::CondCode CC = cast<CondCodeSDNode>(N0.getOperand(2))->get(); 1554 CC = ISD::getSetCCInverse(CC, 1555 N0.getOperand(0).getValueType().isInteger()); 1556 return DAG.getSetCC(dl, VT, N0.getOperand(0), N0.getOperand(1), CC); 1557 } 1558 1559 if ((N0.getOpcode() == ISD::XOR || 1560 (N0.getOpcode() == ISD::AND && 1561 N0.getOperand(0).getOpcode() == ISD::XOR && 1562 N0.getOperand(1) == N0.getOperand(0).getOperand(1))) && 1563 isa<ConstantSDNode>(N0.getOperand(1)) && 1564 cast<ConstantSDNode>(N0.getOperand(1))->getAPIntValue() == 1) { 1565 // If this is (X^1) == 0/1, swap the RHS and eliminate the xor. We 1566 // can only do this if the top bits are known zero. 1567 unsigned BitWidth = N0.getValueSizeInBits(); 1568 if (DAG.MaskedValueIsZero(N0, 1569 APInt::getHighBitsSet(BitWidth, 1570 BitWidth-1))) { 1571 // Okay, get the un-inverted input value. 1572 SDValue Val; 1573 if (N0.getOpcode() == ISD::XOR) 1574 Val = N0.getOperand(0); 1575 else { 1576 assert(N0.getOpcode() == ISD::AND && 1577 N0.getOperand(0).getOpcode() == ISD::XOR); 1578 // ((X^1)&1)^1 -> X & 1 1579 Val = DAG.getNode(ISD::AND, dl, N0.getValueType(), 1580 N0.getOperand(0).getOperand(0), 1581 N0.getOperand(1)); 1582 } 1583 return DAG.getSetCC(dl, VT, Val, N1, 1584 Cond == ISD::SETEQ ? ISD::SETNE : ISD::SETEQ); 1585 } 1586 } 1587 } 1588 1589 APInt MinVal, MaxVal; 1590 unsigned OperandBitSize = N1C->getValueType(0).getSizeInBits(); 1591 if (ISD::isSignedIntSetCC(Cond)) { 1592 MinVal = APInt::getSignedMinValue(OperandBitSize); 1593 MaxVal = APInt::getSignedMaxValue(OperandBitSize); 1594 } else { 1595 MinVal = APInt::getMinValue(OperandBitSize); 1596 MaxVal = APInt::getMaxValue(OperandBitSize); 1597 } 1598 1599 // Canonicalize GE/LE comparisons to use GT/LT comparisons. 1600 if (Cond == ISD::SETGE || Cond == ISD::SETUGE) { 1601 if (C1 == MinVal) return DAG.getConstant(1, VT); // X >= MIN --> true 1602 // X >= C0 --> X > (C0-1) 1603 return DAG.getSetCC(dl, VT, N0, 1604 DAG.getConstant(C1-1, N1.getValueType()), 1605 (Cond == ISD::SETGE) ? ISD::SETGT : ISD::SETUGT); 1606 } 1607 1608 if (Cond == ISD::SETLE || Cond == ISD::SETULE) { 1609 if (C1 == MaxVal) return DAG.getConstant(1, VT); // X <= MAX --> true 1610 // X <= C0 --> X < (C0+1) 1611 return DAG.getSetCC(dl, VT, N0, 1612 DAG.getConstant(C1+1, N1.getValueType()), 1613 (Cond == ISD::SETLE) ? ISD::SETLT : ISD::SETULT); 1614 } 1615 1616 if ((Cond == ISD::SETLT || Cond == ISD::SETULT) && C1 == MinVal) 1617 return DAG.getConstant(0, VT); // X < MIN --> false 1618 if ((Cond == ISD::SETGE || Cond == ISD::SETUGE) && C1 == MinVal) 1619 return DAG.getConstant(1, VT); // X >= MIN --> true 1620 if ((Cond == ISD::SETGT || Cond == ISD::SETUGT) && C1 == MaxVal) 1621 return DAG.getConstant(0, VT); // X > MAX --> false 1622 if ((Cond == ISD::SETLE || Cond == ISD::SETULE) && C1 == MaxVal) 1623 return DAG.getConstant(1, VT); // X <= MAX --> true 1624 1625 // Canonicalize setgt X, Min --> setne X, Min 1626 if ((Cond == ISD::SETGT || Cond == ISD::SETUGT) && C1 == MinVal) 1627 return DAG.getSetCC(dl, VT, N0, N1, ISD::SETNE); 1628 // Canonicalize setlt X, Max --> setne X, Max 1629 if ((Cond == ISD::SETLT || Cond == ISD::SETULT) && C1 == MaxVal) 1630 return DAG.getSetCC(dl, VT, N0, N1, ISD::SETNE); 1631 1632 // If we have setult X, 1, turn it into seteq X, 0 1633 if ((Cond == ISD::SETLT || Cond == ISD::SETULT) && C1 == MinVal+1) 1634 return DAG.getSetCC(dl, VT, N0, 1635 DAG.getConstant(MinVal, N0.getValueType()), 1636 ISD::SETEQ); 1637 // If we have setugt X, Max-1, turn it into seteq X, Max 1638 else if ((Cond == ISD::SETGT || Cond == ISD::SETUGT) && C1 == MaxVal-1) 1639 return DAG.getSetCC(dl, VT, N0, 1640 DAG.getConstant(MaxVal, N0.getValueType()), 1641 ISD::SETEQ); 1642 1643 // If we have "setcc X, C0", check to see if we can shrink the immediate 1644 // by changing cc. 1645 1646 // SETUGT X, SINTMAX -> SETLT X, 0 1647 if (Cond == ISD::SETUGT && 1648 C1 == APInt::getSignedMaxValue(OperandBitSize)) 1649 return DAG.getSetCC(dl, VT, N0, 1650 DAG.getConstant(0, N1.getValueType()), 1651 ISD::SETLT); 1652 1653 // SETULT X, SINTMIN -> SETGT X, -1 1654 if (Cond == ISD::SETULT && 1655 C1 == APInt::getSignedMinValue(OperandBitSize)) { 1656 SDValue ConstMinusOne = 1657 DAG.getConstant(APInt::getAllOnesValue(OperandBitSize), 1658 N1.getValueType()); 1659 return DAG.getSetCC(dl, VT, N0, ConstMinusOne, ISD::SETGT); 1660 } 1661 1662 // Fold bit comparisons when we can. 1663 if ((Cond == ISD::SETEQ || Cond == ISD::SETNE) && 1664 VT == N0.getValueType() && N0.getOpcode() == ISD::AND) 1665 if (ConstantSDNode *AndRHS = 1666 dyn_cast<ConstantSDNode>(N0.getOperand(1))) { 1667 MVT ShiftTy = DCI.isBeforeLegalize() ? 1668 getPointerTy() : getShiftAmountTy(); 1669 if (Cond == ISD::SETNE && C1 == 0) {// (X & 8) != 0 --> (X & 8) >> 3 1670 // Perform the xform if the AND RHS is a single bit. 1671 if (isPowerOf2_64(AndRHS->getZExtValue())) { 1672 return DAG.getNode(ISD::SRL, dl, VT, N0, 1673 DAG.getConstant(Log2_64(AndRHS->getZExtValue()), 1674 ShiftTy)); 1675 } 1676 } else if (Cond == ISD::SETEQ && C1 == AndRHS->getZExtValue()) { 1677 // (X & 8) == 8 --> (X & 8) >> 3 1678 // Perform the xform if C1 is a single bit. 1679 if (C1.isPowerOf2()) { 1680 return DAG.getNode(ISD::SRL, dl, VT, N0, 1681 DAG.getConstant(C1.logBase2(), ShiftTy)); 1682 } 1683 } 1684 } 1685 } 1686 } else if (isa<ConstantSDNode>(N0.getNode())) { 1687 // Ensure that the constant occurs on the RHS. 1688 return DAG.getSetCC(dl, VT, N1, N0, ISD::getSetCCSwappedOperands(Cond)); 1689 } 1690 1691 if (isa<ConstantFPSDNode>(N0.getNode())) { 1692 // Constant fold or commute setcc. 1693 SDValue O = DAG.FoldSetCC(VT, N0, N1, Cond, dl); 1694 if (O.getNode()) return O; 1695 } else if (ConstantFPSDNode *CFP = dyn_cast<ConstantFPSDNode>(N1.getNode())) { 1696 // If the RHS of an FP comparison is a constant, simplify it away in 1697 // some cases. 1698 if (CFP->getValueAPF().isNaN()) { 1699 // If an operand is known to be a nan, we can fold it. 1700 switch (ISD::getUnorderedFlavor(Cond)) { 1701 default: assert(0 && "Unknown flavor!"); 1702 case 0: // Known false. 1703 return DAG.getConstant(0, VT); 1704 case 1: // Known true. 1705 return DAG.getConstant(1, VT); 1706 case 2: // Undefined. 1707 return DAG.getUNDEF(VT); 1708 } 1709 } 1710 1711 // Otherwise, we know the RHS is not a NaN. Simplify the node to drop the 1712 // constant if knowing that the operand is non-nan is enough. We prefer to 1713 // have SETO(x,x) instead of SETO(x, 0.0) because this avoids having to 1714 // materialize 0.0. 1715 if (Cond == ISD::SETO || Cond == ISD::SETUO) 1716 return DAG.getSetCC(dl, VT, N0, N0, Cond); 1717 } 1718 1719 if (N0 == N1) { 1720 // We can always fold X == X for integer setcc's. 1721 if (N0.getValueType().isInteger()) 1722 return DAG.getConstant(ISD::isTrueWhenEqual(Cond), VT); 1723 unsigned UOF = ISD::getUnorderedFlavor(Cond); 1724 if (UOF == 2) // FP operators that are undefined on NaNs. 1725 return DAG.getConstant(ISD::isTrueWhenEqual(Cond), VT); 1726 if (UOF == unsigned(ISD::isTrueWhenEqual(Cond))) 1727 return DAG.getConstant(UOF, VT); 1728 // Otherwise, we can't fold it. However, we can simplify it to SETUO/SETO 1729 // if it is not already. 1730 ISD::CondCode NewCond = UOF == 0 ? ISD::SETO : ISD::SETUO; 1731 if (NewCond != Cond) 1732 return DAG.getSetCC(dl, VT, N0, N1, NewCond); 1733 } 1734 1735 if ((Cond == ISD::SETEQ || Cond == ISD::SETNE) && 1736 N0.getValueType().isInteger()) { 1737 if (N0.getOpcode() == ISD::ADD || N0.getOpcode() == ISD::SUB || 1738 N0.getOpcode() == ISD::XOR) { 1739 // Simplify (X+Y) == (X+Z) --> Y == Z 1740 if (N0.getOpcode() == N1.getOpcode()) { 1741 if (N0.getOperand(0) == N1.getOperand(0)) 1742 return DAG.getSetCC(dl, VT, N0.getOperand(1), N1.getOperand(1), Cond); 1743 if (N0.getOperand(1) == N1.getOperand(1)) 1744 return DAG.getSetCC(dl, VT, N0.getOperand(0), N1.getOperand(0), Cond); 1745 if (DAG.isCommutativeBinOp(N0.getOpcode())) { 1746 // If X op Y == Y op X, try other combinations. 1747 if (N0.getOperand(0) == N1.getOperand(1)) 1748 return DAG.getSetCC(dl, VT, N0.getOperand(1), N1.getOperand(0), 1749 Cond); 1750 if (N0.getOperand(1) == N1.getOperand(0)) 1751 return DAG.getSetCC(dl, VT, N0.getOperand(0), N1.getOperand(1), 1752 Cond); 1753 } 1754 } 1755 1756 if (ConstantSDNode *RHSC = dyn_cast<ConstantSDNode>(N1)) { 1757 if (ConstantSDNode *LHSR = dyn_cast<ConstantSDNode>(N0.getOperand(1))) { 1758 // Turn (X+C1) == C2 --> X == C2-C1 1759 if (N0.getOpcode() == ISD::ADD && N0.getNode()->hasOneUse()) { 1760 return DAG.getSetCC(dl, VT, N0.getOperand(0), 1761 DAG.getConstant(RHSC->getAPIntValue()- 1762 LHSR->getAPIntValue(), 1763 N0.getValueType()), Cond); 1764 } 1765 1766 // Turn (X^C1) == C2 into X == C1^C2 iff X&~C1 = 0. 1767 if (N0.getOpcode() == ISD::XOR) 1768 // If we know that all of the inverted bits are zero, don't bother 1769 // performing the inversion. 1770 if (DAG.MaskedValueIsZero(N0.getOperand(0), ~LHSR->getAPIntValue())) 1771 return 1772 DAG.getSetCC(dl, VT, N0.getOperand(0), 1773 DAG.getConstant(LHSR->getAPIntValue() ^ 1774 RHSC->getAPIntValue(), 1775 N0.getValueType()), 1776 Cond); 1777 } 1778 1779 // Turn (C1-X) == C2 --> X == C1-C2 1780 if (ConstantSDNode *SUBC = dyn_cast<ConstantSDNode>(N0.getOperand(0))) { 1781 if (N0.getOpcode() == ISD::SUB && N0.getNode()->hasOneUse()) { 1782 return 1783 DAG.getSetCC(dl, VT, N0.getOperand(1), 1784 DAG.getConstant(SUBC->getAPIntValue() - 1785 RHSC->getAPIntValue(), 1786 N0.getValueType()), 1787 Cond); 1788 } 1789 } 1790 } 1791 1792 // Simplify (X+Z) == X --> Z == 0 1793 if (N0.getOperand(0) == N1) 1794 return DAG.getSetCC(dl, VT, N0.getOperand(1), 1795 DAG.getConstant(0, N0.getValueType()), Cond); 1796 if (N0.getOperand(1) == N1) { 1797 if (DAG.isCommutativeBinOp(N0.getOpcode())) 1798 return DAG.getSetCC(dl, VT, N0.getOperand(0), 1799 DAG.getConstant(0, N0.getValueType()), Cond); 1800 else if (N0.getNode()->hasOneUse()) { 1801 assert(N0.getOpcode() == ISD::SUB && "Unexpected operation!"); 1802 // (Z-X) == X --> Z == X<<1 1803 SDValue SH = DAG.getNode(ISD::SHL, dl, N1.getValueType(), 1804 N1, 1805 DAG.getConstant(1, getShiftAmountTy())); 1806 if (!DCI.isCalledByLegalizer()) 1807 DCI.AddToWorklist(SH.getNode()); 1808 return DAG.getSetCC(dl, VT, N0.getOperand(0), SH, Cond); 1809 } 1810 } 1811 } 1812 1813 if (N1.getOpcode() == ISD::ADD || N1.getOpcode() == ISD::SUB || 1814 N1.getOpcode() == ISD::XOR) { 1815 // Simplify X == (X+Z) --> Z == 0 1816 if (N1.getOperand(0) == N0) { 1817 return DAG.getSetCC(dl, VT, N1.getOperand(1), 1818 DAG.getConstant(0, N1.getValueType()), Cond); 1819 } else if (N1.getOperand(1) == N0) { 1820 if (DAG.isCommutativeBinOp(N1.getOpcode())) { 1821 return DAG.getSetCC(dl, VT, N1.getOperand(0), 1822 DAG.getConstant(0, N1.getValueType()), Cond); 1823 } else if (N1.getNode()->hasOneUse()) { 1824 assert(N1.getOpcode() == ISD::SUB && "Unexpected operation!"); 1825 // X == (Z-X) --> X<<1 == Z 1826 SDValue SH = DAG.getNode(ISD::SHL, dl, N1.getValueType(), N0, 1827 DAG.getConstant(1, getShiftAmountTy())); 1828 if (!DCI.isCalledByLegalizer()) 1829 DCI.AddToWorklist(SH.getNode()); 1830 return DAG.getSetCC(dl, VT, SH, N1.getOperand(0), Cond); 1831 } 1832 } 1833 } 1834 1835 // Simplify x&y == y to x&y != 0 if y has exactly one bit set. 1836 // Note that where y is variable and is known to have at most 1837 // one bit set (for example, if it is z&1) we cannot do this; 1838 // the expressions are not equivalent when y==0. 1839 if (N0.getOpcode() == ISD::AND) 1840 if (N0.getOperand(0) == N1 || N0.getOperand(1) == N1) { 1841 if (ValueHasExactlyOneBitSet(N1, DAG)) { 1842 Cond = ISD::getSetCCInverse(Cond, /*isInteger=*/true); 1843 SDValue Zero = DAG.getConstant(0, N1.getValueType()); 1844 return DAG.getSetCC(dl, VT, N0, Zero, Cond); 1845 } 1846 } 1847 if (N1.getOpcode() == ISD::AND) 1848 if (N1.getOperand(0) == N0 || N1.getOperand(1) == N0) { 1849 if (ValueHasExactlyOneBitSet(N0, DAG)) { 1850 Cond = ISD::getSetCCInverse(Cond, /*isInteger=*/true); 1851 SDValue Zero = DAG.getConstant(0, N0.getValueType()); 1852 return DAG.getSetCC(dl, VT, N1, Zero, Cond); 1853 } 1854 } 1855 } 1856 1857 // Fold away ALL boolean setcc's. 1858 SDValue Temp; 1859 if (N0.getValueType() == MVT::i1 && foldBooleans) { 1860 switch (Cond) { 1861 default: assert(0 && "Unknown integer setcc!"); 1862 case ISD::SETEQ: // X == Y -> ~(X^Y) 1863 Temp = DAG.getNode(ISD::XOR, dl, MVT::i1, N0, N1); 1864 N0 = DAG.getNOT(dl, Temp, MVT::i1); 1865 if (!DCI.isCalledByLegalizer()) 1866 DCI.AddToWorklist(Temp.getNode()); 1867 break; 1868 case ISD::SETNE: // X != Y --> (X^Y) 1869 N0 = DAG.getNode(ISD::XOR, dl, MVT::i1, N0, N1); 1870 break; 1871 case ISD::SETGT: // X >s Y --> X == 0 & Y == 1 --> ~X & Y 1872 case ISD::SETULT: // X <u Y --> X == 0 & Y == 1 --> ~X & Y 1873 Temp = DAG.getNOT(dl, N0, MVT::i1); 1874 N0 = DAG.getNode(ISD::AND, dl, MVT::i1, N1, Temp); 1875 if (!DCI.isCalledByLegalizer()) 1876 DCI.AddToWorklist(Temp.getNode()); 1877 break; 1878 case ISD::SETLT: // X <s Y --> X == 1 & Y == 0 --> ~Y & X 1879 case ISD::SETUGT: // X >u Y --> X == 1 & Y == 0 --> ~Y & X 1880 Temp = DAG.getNOT(dl, N1, MVT::i1); 1881 N0 = DAG.getNode(ISD::AND, dl, MVT::i1, N0, Temp); 1882 if (!DCI.isCalledByLegalizer()) 1883 DCI.AddToWorklist(Temp.getNode()); 1884 break; 1885 case ISD::SETULE: // X <=u Y --> X == 0 | Y == 1 --> ~X | Y 1886 case ISD::SETGE: // X >=s Y --> X == 0 | Y == 1 --> ~X | Y 1887 Temp = DAG.getNOT(dl, N0, MVT::i1); 1888 N0 = DAG.getNode(ISD::OR, dl, MVT::i1, N1, Temp); 1889 if (!DCI.isCalledByLegalizer()) 1890 DCI.AddToWorklist(Temp.getNode()); 1891 break; 1892 case ISD::SETUGE: // X >=u Y --> X == 1 | Y == 0 --> ~Y | X 1893 case ISD::SETLE: // X <=s Y --> X == 1 | Y == 0 --> ~Y | X 1894 Temp = DAG.getNOT(dl, N1, MVT::i1); 1895 N0 = DAG.getNode(ISD::OR, dl, MVT::i1, N0, Temp); 1896 break; 1897 } 1898 if (VT != MVT::i1) { 1899 if (!DCI.isCalledByLegalizer()) 1900 DCI.AddToWorklist(N0.getNode()); 1901 // FIXME: If running after legalize, we probably can't do this. 1902 N0 = DAG.getNode(ISD::ZERO_EXTEND, dl, VT, N0); 1903 } 1904 return N0; 1905 } 1906 1907 // Could not fold it. 1908 return SDValue(); 1909} 1910 1911/// isGAPlusOffset - Returns true (and the GlobalValue and the offset) if the 1912/// node is a GlobalAddress + offset. 1913bool TargetLowering::isGAPlusOffset(SDNode *N, GlobalValue* &GA, 1914 int64_t &Offset) const { 1915 if (isa<GlobalAddressSDNode>(N)) { 1916 GlobalAddressSDNode *GASD = cast<GlobalAddressSDNode>(N); 1917 GA = GASD->getGlobal(); 1918 Offset += GASD->getOffset(); 1919 return true; 1920 } 1921 1922 if (N->getOpcode() == ISD::ADD) { 1923 SDValue N1 = N->getOperand(0); 1924 SDValue N2 = N->getOperand(1); 1925 if (isGAPlusOffset(N1.getNode(), GA, Offset)) { 1926 ConstantSDNode *V = dyn_cast<ConstantSDNode>(N2); 1927 if (V) { 1928 Offset += V->getSExtValue(); 1929 return true; 1930 } 1931 } else if (isGAPlusOffset(N2.getNode(), GA, Offset)) { 1932 ConstantSDNode *V = dyn_cast<ConstantSDNode>(N1); 1933 if (V) { 1934 Offset += V->getSExtValue(); 1935 return true; 1936 } 1937 } 1938 } 1939 return false; 1940} 1941 1942 1943/// isConsecutiveLoad - Return true if LD (which must be a LoadSDNode) is 1944/// loading 'Bytes' bytes from a location that is 'Dist' units away from the 1945/// location that the 'Base' load is loading from. 1946bool TargetLowering::isConsecutiveLoad(SDNode *LD, SDNode *Base, 1947 unsigned Bytes, int Dist, 1948 const MachineFrameInfo *MFI) const { 1949 if (LD->getOperand(0).getNode() != Base->getOperand(0).getNode()) 1950 return false; 1951 MVT VT = LD->getValueType(0); 1952 if (VT.getSizeInBits() / 8 != Bytes) 1953 return false; 1954 1955 SDValue Loc = LD->getOperand(1); 1956 SDValue BaseLoc = Base->getOperand(1); 1957 if (Loc.getOpcode() == ISD::FrameIndex) { 1958 if (BaseLoc.getOpcode() != ISD::FrameIndex) 1959 return false; 1960 int FI = cast<FrameIndexSDNode>(Loc)->getIndex(); 1961 int BFI = cast<FrameIndexSDNode>(BaseLoc)->getIndex(); 1962 int FS = MFI->getObjectSize(FI); 1963 int BFS = MFI->getObjectSize(BFI); 1964 if (FS != BFS || FS != (int)Bytes) return false; 1965 return MFI->getObjectOffset(FI) == (MFI->getObjectOffset(BFI) + Dist*Bytes); 1966 } 1967 1968 GlobalValue *GV1 = NULL; 1969 GlobalValue *GV2 = NULL; 1970 int64_t Offset1 = 0; 1971 int64_t Offset2 = 0; 1972 bool isGA1 = isGAPlusOffset(Loc.getNode(), GV1, Offset1); 1973 bool isGA2 = isGAPlusOffset(BaseLoc.getNode(), GV2, Offset2); 1974 if (isGA1 && isGA2 && GV1 == GV2) 1975 return Offset1 == (Offset2 + Dist*Bytes); 1976 return false; 1977} 1978 1979 1980SDValue TargetLowering:: 1981PerformDAGCombine(SDNode *N, DAGCombinerInfo &DCI) const { 1982 // Default implementation: no optimization. 1983 return SDValue(); 1984} 1985 1986//===----------------------------------------------------------------------===// 1987// Inline Assembler Implementation Methods 1988//===----------------------------------------------------------------------===// 1989 1990 1991TargetLowering::ConstraintType 1992TargetLowering::getConstraintType(const std::string &Constraint) const { 1993 // FIXME: lots more standard ones to handle. 1994 if (Constraint.size() == 1) { 1995 switch (Constraint[0]) { 1996 default: break; 1997 case 'r': return C_RegisterClass; 1998 case 'm': // memory 1999 case 'o': // offsetable 2000 case 'V': // not offsetable 2001 return C_Memory; 2002 case 'i': // Simple Integer or Relocatable Constant 2003 case 'n': // Simple Integer 2004 case 's': // Relocatable Constant 2005 case 'X': // Allow ANY value. 2006 case 'I': // Target registers. 2007 case 'J': 2008 case 'K': 2009 case 'L': 2010 case 'M': 2011 case 'N': 2012 case 'O': 2013 case 'P': 2014 return C_Other; 2015 } 2016 } 2017 2018 if (Constraint.size() > 1 && Constraint[0] == '{' && 2019 Constraint[Constraint.size()-1] == '}') 2020 return C_Register; 2021 return C_Unknown; 2022} 2023 2024/// LowerXConstraint - try to replace an X constraint, which matches anything, 2025/// with another that has more specific requirements based on the type of the 2026/// corresponding operand. 2027const char *TargetLowering::LowerXConstraint(MVT ConstraintVT) const{ 2028 if (ConstraintVT.isInteger()) 2029 return "r"; 2030 if (ConstraintVT.isFloatingPoint()) 2031 return "f"; // works for many targets 2032 return 0; 2033} 2034 2035/// LowerAsmOperandForConstraint - Lower the specified operand into the Ops 2036/// vector. If it is invalid, don't add anything to Ops. 2037void TargetLowering::LowerAsmOperandForConstraint(SDValue Op, 2038 char ConstraintLetter, 2039 bool hasMemory, 2040 std::vector<SDValue> &Ops, 2041 SelectionDAG &DAG) const { 2042 switch (ConstraintLetter) { 2043 default: break; 2044 case 'X': // Allows any operand; labels (basic block) use this. 2045 if (Op.getOpcode() == ISD::BasicBlock) { 2046 Ops.push_back(Op); 2047 return; 2048 } 2049 // fall through 2050 case 'i': // Simple Integer or Relocatable Constant 2051 case 'n': // Simple Integer 2052 case 's': { // Relocatable Constant 2053 // These operands are interested in values of the form (GV+C), where C may 2054 // be folded in as an offset of GV, or it may be explicitly added. Also, it 2055 // is possible and fine if either GV or C are missing. 2056 ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op); 2057 GlobalAddressSDNode *GA = dyn_cast<GlobalAddressSDNode>(Op); 2058 2059 // If we have "(add GV, C)", pull out GV/C 2060 if (Op.getOpcode() == ISD::ADD) { 2061 C = dyn_cast<ConstantSDNode>(Op.getOperand(1)); 2062 GA = dyn_cast<GlobalAddressSDNode>(Op.getOperand(0)); 2063 if (C == 0 || GA == 0) { 2064 C = dyn_cast<ConstantSDNode>(Op.getOperand(0)); 2065 GA = dyn_cast<GlobalAddressSDNode>(Op.getOperand(1)); 2066 } 2067 if (C == 0 || GA == 0) 2068 C = 0, GA = 0; 2069 } 2070 2071 // If we find a valid operand, map to the TargetXXX version so that the 2072 // value itself doesn't get selected. 2073 if (GA) { // Either &GV or &GV+C 2074 if (ConstraintLetter != 'n') { 2075 int64_t Offs = GA->getOffset(); 2076 if (C) Offs += C->getZExtValue(); 2077 Ops.push_back(DAG.getTargetGlobalAddress(GA->getGlobal(), 2078 Op.getValueType(), Offs)); 2079 return; 2080 } 2081 } 2082 if (C) { // just C, no GV. 2083 // Simple constants are not allowed for 's'. 2084 if (ConstraintLetter != 's') { 2085 // gcc prints these as sign extended. Sign extend value to 64 bits 2086 // now; without this it would get ZExt'd later in 2087 // ScheduleDAGSDNodes::EmitNode, which is very generic. 2088 Ops.push_back(DAG.getTargetConstant(C->getAPIntValue().getSExtValue(), 2089 MVT::i64)); 2090 return; 2091 } 2092 } 2093 break; 2094 } 2095 } 2096} 2097 2098std::vector<unsigned> TargetLowering:: 2099getRegClassForInlineAsmConstraint(const std::string &Constraint, 2100 MVT VT) const { 2101 return std::vector<unsigned>(); 2102} 2103 2104 2105std::pair<unsigned, const TargetRegisterClass*> TargetLowering:: 2106getRegForInlineAsmConstraint(const std::string &Constraint, 2107 MVT VT) const { 2108 if (Constraint[0] != '{') 2109 return std::pair<unsigned, const TargetRegisterClass*>(0, 0); 2110 assert(*(Constraint.end()-1) == '}' && "Not a brace enclosed constraint?"); 2111 2112 // Remove the braces from around the name. 2113 std::string RegName(Constraint.begin()+1, Constraint.end()-1); 2114 2115 // Figure out which register class contains this reg. 2116 const TargetRegisterInfo *RI = TM.getRegisterInfo(); 2117 for (TargetRegisterInfo::regclass_iterator RCI = RI->regclass_begin(), 2118 E = RI->regclass_end(); RCI != E; ++RCI) { 2119 const TargetRegisterClass *RC = *RCI; 2120 2121 // If none of the the value types for this register class are valid, we 2122 // can't use it. For example, 64-bit reg classes on 32-bit targets. 2123 bool isLegal = false; 2124 for (TargetRegisterClass::vt_iterator I = RC->vt_begin(), E = RC->vt_end(); 2125 I != E; ++I) { 2126 if (isTypeLegal(*I)) { 2127 isLegal = true; 2128 break; 2129 } 2130 } 2131 2132 if (!isLegal) continue; 2133 2134 for (TargetRegisterClass::iterator I = RC->begin(), E = RC->end(); 2135 I != E; ++I) { 2136 if (StringsEqualNoCase(RegName, RI->get(*I).AsmName)) 2137 return std::make_pair(*I, RC); 2138 } 2139 } 2140 2141 return std::pair<unsigned, const TargetRegisterClass*>(0, 0); 2142} 2143 2144//===----------------------------------------------------------------------===// 2145// Constraint Selection. 2146 2147/// isMatchingInputConstraint - Return true of this is an input operand that is 2148/// a matching constraint like "4". 2149bool TargetLowering::AsmOperandInfo::isMatchingInputConstraint() const { 2150 assert(!ConstraintCode.empty() && "No known constraint!"); 2151 return isdigit(ConstraintCode[0]); 2152} 2153 2154/// getMatchedOperand - If this is an input matching constraint, this method 2155/// returns the output operand it matches. 2156unsigned TargetLowering::AsmOperandInfo::getMatchedOperand() const { 2157 assert(!ConstraintCode.empty() && "No known constraint!"); 2158 return atoi(ConstraintCode.c_str()); 2159} 2160 2161 2162/// getConstraintGenerality - Return an integer indicating how general CT 2163/// is. 2164static unsigned getConstraintGenerality(TargetLowering::ConstraintType CT) { 2165 switch (CT) { 2166 default: assert(0 && "Unknown constraint type!"); 2167 case TargetLowering::C_Other: 2168 case TargetLowering::C_Unknown: 2169 return 0; 2170 case TargetLowering::C_Register: 2171 return 1; 2172 case TargetLowering::C_RegisterClass: 2173 return 2; 2174 case TargetLowering::C_Memory: 2175 return 3; 2176 } 2177} 2178 2179/// ChooseConstraint - If there are multiple different constraints that we 2180/// could pick for this operand (e.g. "imr") try to pick the 'best' one. 2181/// This is somewhat tricky: constraints fall into four classes: 2182/// Other -> immediates and magic values 2183/// Register -> one specific register 2184/// RegisterClass -> a group of regs 2185/// Memory -> memory 2186/// Ideally, we would pick the most specific constraint possible: if we have 2187/// something that fits into a register, we would pick it. The problem here 2188/// is that if we have something that could either be in a register or in 2189/// memory that use of the register could cause selection of *other* 2190/// operands to fail: they might only succeed if we pick memory. Because of 2191/// this the heuristic we use is: 2192/// 2193/// 1) If there is an 'other' constraint, and if the operand is valid for 2194/// that constraint, use it. This makes us take advantage of 'i' 2195/// constraints when available. 2196/// 2) Otherwise, pick the most general constraint present. This prefers 2197/// 'm' over 'r', for example. 2198/// 2199static void ChooseConstraint(TargetLowering::AsmOperandInfo &OpInfo, 2200 bool hasMemory, const TargetLowering &TLI, 2201 SDValue Op, SelectionDAG *DAG) { 2202 assert(OpInfo.Codes.size() > 1 && "Doesn't have multiple constraint options"); 2203 unsigned BestIdx = 0; 2204 TargetLowering::ConstraintType BestType = TargetLowering::C_Unknown; 2205 int BestGenerality = -1; 2206 2207 // Loop over the options, keeping track of the most general one. 2208 for (unsigned i = 0, e = OpInfo.Codes.size(); i != e; ++i) { 2209 TargetLowering::ConstraintType CType = 2210 TLI.getConstraintType(OpInfo.Codes[i]); 2211 2212 // If this is an 'other' constraint, see if the operand is valid for it. 2213 // For example, on X86 we might have an 'rI' constraint. If the operand 2214 // is an integer in the range [0..31] we want to use I (saving a load 2215 // of a register), otherwise we must use 'r'. 2216 if (CType == TargetLowering::C_Other && Op.getNode()) { 2217 assert(OpInfo.Codes[i].size() == 1 && 2218 "Unhandled multi-letter 'other' constraint"); 2219 std::vector<SDValue> ResultOps; 2220 TLI.LowerAsmOperandForConstraint(Op, OpInfo.Codes[i][0], hasMemory, 2221 ResultOps, *DAG); 2222 if (!ResultOps.empty()) { 2223 BestType = CType; 2224 BestIdx = i; 2225 break; 2226 } 2227 } 2228 2229 // This constraint letter is more general than the previous one, use it. 2230 int Generality = getConstraintGenerality(CType); 2231 if (Generality > BestGenerality) { 2232 BestType = CType; 2233 BestIdx = i; 2234 BestGenerality = Generality; 2235 } 2236 } 2237 2238 OpInfo.ConstraintCode = OpInfo.Codes[BestIdx]; 2239 OpInfo.ConstraintType = BestType; 2240} 2241 2242/// ComputeConstraintToUse - Determines the constraint code and constraint 2243/// type to use for the specific AsmOperandInfo, setting 2244/// OpInfo.ConstraintCode and OpInfo.ConstraintType. 2245void TargetLowering::ComputeConstraintToUse(AsmOperandInfo &OpInfo, 2246 SDValue Op, 2247 bool hasMemory, 2248 SelectionDAG *DAG) const { 2249 assert(!OpInfo.Codes.empty() && "Must have at least one constraint"); 2250 2251 // Single-letter constraints ('r') are very common. 2252 if (OpInfo.Codes.size() == 1) { 2253 OpInfo.ConstraintCode = OpInfo.Codes[0]; 2254 OpInfo.ConstraintType = getConstraintType(OpInfo.ConstraintCode); 2255 } else { 2256 ChooseConstraint(OpInfo, hasMemory, *this, Op, DAG); 2257 } 2258 2259 // 'X' matches anything. 2260 if (OpInfo.ConstraintCode == "X" && OpInfo.CallOperandVal) { 2261 // Labels and constants are handled elsewhere ('X' is the only thing 2262 // that matches labels). 2263 if (isa<BasicBlock>(OpInfo.CallOperandVal) || 2264 isa<ConstantInt>(OpInfo.CallOperandVal)) 2265 return; 2266 2267 // Otherwise, try to resolve it to something we know about by looking at 2268 // the actual operand type. 2269 if (const char *Repl = LowerXConstraint(OpInfo.ConstraintVT)) { 2270 OpInfo.ConstraintCode = Repl; 2271 OpInfo.ConstraintType = getConstraintType(OpInfo.ConstraintCode); 2272 } 2273 } 2274} 2275 2276//===----------------------------------------------------------------------===// 2277// Loop Strength Reduction hooks 2278//===----------------------------------------------------------------------===// 2279 2280/// isLegalAddressingMode - Return true if the addressing mode represented 2281/// by AM is legal for this target, for a load/store of the specified type. 2282bool TargetLowering::isLegalAddressingMode(const AddrMode &AM, 2283 const Type *Ty) const { 2284 // The default implementation of this implements a conservative RISCy, r+r and 2285 // r+i addr mode. 2286 2287 // Allows a sign-extended 16-bit immediate field. 2288 if (AM.BaseOffs <= -(1LL << 16) || AM.BaseOffs >= (1LL << 16)-1) 2289 return false; 2290 2291 // No global is ever allowed as a base. 2292 if (AM.BaseGV) 2293 return false; 2294 2295 // Only support r+r, 2296 switch (AM.Scale) { 2297 case 0: // "r+i" or just "i", depending on HasBaseReg. 2298 break; 2299 case 1: 2300 if (AM.HasBaseReg && AM.BaseOffs) // "r+r+i" is not allowed. 2301 return false; 2302 // Otherwise we have r+r or r+i. 2303 break; 2304 case 2: 2305 if (AM.HasBaseReg || AM.BaseOffs) // 2*r+r or 2*r+i is not allowed. 2306 return false; 2307 // Allow 2*r as r+r. 2308 break; 2309 } 2310 2311 return true; 2312} 2313 2314struct mu { 2315 APInt m; // magic number 2316 bool a; // add indicator 2317 unsigned s; // shift amount 2318}; 2319 2320/// magicu - calculate the magic numbers required to codegen an integer udiv as 2321/// a sequence of multiply, add and shifts. Requires that the divisor not be 0. 2322static mu magicu(const APInt& d) { 2323 unsigned p; 2324 APInt nc, delta, q1, r1, q2, r2; 2325 struct mu magu; 2326 magu.a = 0; // initialize "add" indicator 2327 APInt allOnes = APInt::getAllOnesValue(d.getBitWidth()); 2328 APInt signedMin = APInt::getSignedMinValue(d.getBitWidth()); 2329 APInt signedMax = APInt::getSignedMaxValue(d.getBitWidth()); 2330 2331 nc = allOnes - (-d).urem(d); 2332 p = d.getBitWidth() - 1; // initialize p 2333 q1 = signedMin.udiv(nc); // initialize q1 = 2p/nc 2334 r1 = signedMin - q1*nc; // initialize r1 = rem(2p,nc) 2335 q2 = signedMax.udiv(d); // initialize q2 = (2p-1)/d 2336 r2 = signedMax - q2*d; // initialize r2 = rem((2p-1),d) 2337 do { 2338 p = p + 1; 2339 if (r1.uge(nc - r1)) { 2340 q1 = q1 + q1 + 1; // update q1 2341 r1 = r1 + r1 - nc; // update r1 2342 } 2343 else { 2344 q1 = q1+q1; // update q1 2345 r1 = r1+r1; // update r1 2346 } 2347 if ((r2 + 1).uge(d - r2)) { 2348 if (q2.uge(signedMax)) magu.a = 1; 2349 q2 = q2+q2 + 1; // update q2 2350 r2 = r2+r2 + 1 - d; // update r2 2351 } 2352 else { 2353 if (q2.uge(signedMin)) magu.a = 1; 2354 q2 = q2+q2; // update q2 2355 r2 = r2+r2 + 1; // update r2 2356 } 2357 delta = d - 1 - r2; 2358 } while (p < d.getBitWidth()*2 && 2359 (q1.ult(delta) || (q1 == delta && r1 == 0))); 2360 magu.m = q2 + 1; // resulting magic number 2361 magu.s = p - d.getBitWidth(); // resulting shift 2362 return magu; 2363} 2364 2365// Magic for divide replacement 2366struct ms { 2367 APInt m; // magic number 2368 unsigned s; // shift amount 2369}; 2370 2371/// magic - calculate the magic numbers required to codegen an integer sdiv as 2372/// a sequence of multiply and shifts. Requires that the divisor not be 0, 1, 2373/// or -1. 2374static ms magic(const APInt& d) { 2375 unsigned p; 2376 APInt ad, anc, delta, q1, r1, q2, r2, t; 2377 APInt allOnes = APInt::getAllOnesValue(d.getBitWidth()); 2378 APInt signedMin = APInt::getSignedMinValue(d.getBitWidth()); 2379 APInt signedMax = APInt::getSignedMaxValue(d.getBitWidth()); 2380 struct ms mag; 2381 2382 ad = d.abs(); 2383 t = signedMin + (d.lshr(d.getBitWidth() - 1)); 2384 anc = t - 1 - t.urem(ad); // absolute value of nc 2385 p = d.getBitWidth() - 1; // initialize p 2386 q1 = signedMin.udiv(anc); // initialize q1 = 2p/abs(nc) 2387 r1 = signedMin - q1*anc; // initialize r1 = rem(2p,abs(nc)) 2388 q2 = signedMin.udiv(ad); // initialize q2 = 2p/abs(d) 2389 r2 = signedMin - q2*ad; // initialize r2 = rem(2p,abs(d)) 2390 do { 2391 p = p + 1; 2392 q1 = q1<<1; // update q1 = 2p/abs(nc) 2393 r1 = r1<<1; // update r1 = rem(2p/abs(nc)) 2394 if (r1.uge(anc)) { // must be unsigned comparison 2395 q1 = q1 + 1; 2396 r1 = r1 - anc; 2397 } 2398 q2 = q2<<1; // update q2 = 2p/abs(d) 2399 r2 = r2<<1; // update r2 = rem(2p/abs(d)) 2400 if (r2.uge(ad)) { // must be unsigned comparison 2401 q2 = q2 + 1; 2402 r2 = r2 - ad; 2403 } 2404 delta = ad - r2; 2405 } while (q1.ule(delta) || (q1 == delta && r1 == 0)); 2406 2407 mag.m = q2 + 1; 2408 if (d.isNegative()) mag.m = -mag.m; // resulting magic number 2409 mag.s = p - d.getBitWidth(); // resulting shift 2410 return mag; 2411} 2412 2413/// BuildSDIVSequence - Given an ISD::SDIV node expressing a divide by constant, 2414/// return a DAG expression to select that will generate the same value by 2415/// multiplying by a magic number. See: 2416/// <http://the.wall.riscom.net/books/proc/ppc/cwg/code2.html> 2417SDValue TargetLowering::BuildSDIV(SDNode *N, SelectionDAG &DAG, 2418 std::vector<SDNode*>* Created) const { 2419 MVT VT = N->getValueType(0); 2420 DebugLoc dl= N->getDebugLoc(); 2421 2422 // Check to see if we can do this. 2423 // FIXME: We should be more aggressive here. 2424 if (!isTypeLegal(VT)) 2425 return SDValue(); 2426 2427 APInt d = cast<ConstantSDNode>(N->getOperand(1))->getAPIntValue(); 2428 ms magics = magic(d); 2429 2430 // Multiply the numerator (operand 0) by the magic value 2431 // FIXME: We should support doing a MUL in a wider type 2432 SDValue Q; 2433 if (isOperationLegalOrCustom(ISD::MULHS, VT)) 2434 Q = DAG.getNode(ISD::MULHS, dl, VT, N->getOperand(0), 2435 DAG.getConstant(magics.m, VT)); 2436 else if (isOperationLegalOrCustom(ISD::SMUL_LOHI, VT)) 2437 Q = SDValue(DAG.getNode(ISD::SMUL_LOHI, dl, DAG.getVTList(VT, VT), 2438 N->getOperand(0), 2439 DAG.getConstant(magics.m, VT)).getNode(), 1); 2440 else 2441 return SDValue(); // No mulhs or equvialent 2442 // If d > 0 and m < 0, add the numerator 2443 if (d.isStrictlyPositive() && magics.m.isNegative()) { 2444 Q = DAG.getNode(ISD::ADD, dl, VT, Q, N->getOperand(0)); 2445 if (Created) 2446 Created->push_back(Q.getNode()); 2447 } 2448 // If d < 0 and m > 0, subtract the numerator. 2449 if (d.isNegative() && magics.m.isStrictlyPositive()) { 2450 Q = DAG.getNode(ISD::SUB, dl, VT, Q, N->getOperand(0)); 2451 if (Created) 2452 Created->push_back(Q.getNode()); 2453 } 2454 // Shift right algebraic if shift value is nonzero 2455 if (magics.s > 0) { 2456 Q = DAG.getNode(ISD::SRA, dl, VT, Q, 2457 DAG.getConstant(magics.s, getShiftAmountTy())); 2458 if (Created) 2459 Created->push_back(Q.getNode()); 2460 } 2461 // Extract the sign bit and add it to the quotient 2462 SDValue T = 2463 DAG.getNode(ISD::SRL, dl, VT, Q, DAG.getConstant(VT.getSizeInBits()-1, 2464 getShiftAmountTy())); 2465 if (Created) 2466 Created->push_back(T.getNode()); 2467 return DAG.getNode(ISD::ADD, dl, VT, Q, T); 2468} 2469 2470/// BuildUDIVSequence - Given an ISD::UDIV node expressing a divide by constant, 2471/// return a DAG expression to select that will generate the same value by 2472/// multiplying by a magic number. See: 2473/// <http://the.wall.riscom.net/books/proc/ppc/cwg/code2.html> 2474SDValue TargetLowering::BuildUDIV(SDNode *N, SelectionDAG &DAG, 2475 std::vector<SDNode*>* Created) const { 2476 MVT VT = N->getValueType(0); 2477 DebugLoc dl = N->getDebugLoc(); 2478 2479 // Check to see if we can do this. 2480 // FIXME: We should be more aggressive here. 2481 if (!isTypeLegal(VT)) 2482 return SDValue(); 2483 2484 // FIXME: We should use a narrower constant when the upper 2485 // bits are known to be zero. 2486 ConstantSDNode *N1C = cast<ConstantSDNode>(N->getOperand(1)); 2487 mu magics = magicu(N1C->getAPIntValue()); 2488 2489 // Multiply the numerator (operand 0) by the magic value 2490 // FIXME: We should support doing a MUL in a wider type 2491 SDValue Q; 2492 if (isOperationLegalOrCustom(ISD::MULHU, VT)) 2493 Q = DAG.getNode(ISD::MULHU, dl, VT, N->getOperand(0), 2494 DAG.getConstant(magics.m, VT)); 2495 else if (isOperationLegalOrCustom(ISD::UMUL_LOHI, VT)) 2496 Q = SDValue(DAG.getNode(ISD::UMUL_LOHI, dl, DAG.getVTList(VT, VT), 2497 N->getOperand(0), 2498 DAG.getConstant(magics.m, VT)).getNode(), 1); 2499 else 2500 return SDValue(); // No mulhu or equvialent 2501 if (Created) 2502 Created->push_back(Q.getNode()); 2503 2504 if (magics.a == 0) { 2505 assert(magics.s < N1C->getAPIntValue().getBitWidth() && 2506 "We shouldn't generate an undefined shift!"); 2507 return DAG.getNode(ISD::SRL, dl, VT, Q, 2508 DAG.getConstant(magics.s, getShiftAmountTy())); 2509 } else { 2510 SDValue NPQ = DAG.getNode(ISD::SUB, dl, VT, N->getOperand(0), Q); 2511 if (Created) 2512 Created->push_back(NPQ.getNode()); 2513 NPQ = DAG.getNode(ISD::SRL, dl, VT, NPQ, 2514 DAG.getConstant(1, getShiftAmountTy())); 2515 if (Created) 2516 Created->push_back(NPQ.getNode()); 2517 NPQ = DAG.getNode(ISD::ADD, dl, VT, NPQ, Q); 2518 if (Created) 2519 Created->push_back(NPQ.getNode()); 2520 return DAG.getNode(ISD::SRL, dl, VT, NPQ, 2521 DAG.getConstant(magics.s-1, getShiftAmountTy())); 2522 } 2523} 2524