LLVM API Documentation
00001 //===-- TargetLowering.cpp - Implement the TargetLowering class -----------===// 00002 // 00003 // The LLVM Compiler Infrastructure 00004 // 00005 // This file is distributed under the University of Illinois Open Source 00006 // License. See LICENSE.TXT for details. 00007 // 00008 //===----------------------------------------------------------------------===// 00009 // 00010 // This implements the TargetLowering class. 00011 // 00012 //===----------------------------------------------------------------------===// 00013 00014 #include "llvm/Target/TargetLowering.h" 00015 #include "llvm/Target/TargetAsmInfo.h" 00016 #include "llvm/Target/TargetData.h" 00017 #include "llvm/Target/TargetLoweringObjectFile.h" 00018 #include "llvm/Target/TargetMachine.h" 00019 #include "llvm/Target/TargetRegisterInfo.h" 00020 #include "llvm/Target/TargetSubtarget.h" 00021 #include "llvm/GlobalVariable.h" 00022 #include "llvm/DerivedTypes.h" 00023 #include "llvm/CodeGen/MachineFrameInfo.h" 00024 #include "llvm/CodeGen/SelectionDAG.h" 00025 #include "llvm/ADT/StringExtras.h" 00026 #include "llvm/ADT/STLExtras.h" 00027 #include "llvm/Support/ErrorHandling.h" 00028 #include "llvm/Support/MathExtras.h" 00029 using namespace llvm; 00030 00031 namespace llvm { 00032 TLSModel::Model getTLSModel(const GlobalValue *GV, Reloc::Model reloc) { 00033 bool isLocal = GV->hasLocalLinkage(); 00034 bool isDeclaration = GV->isDeclaration(); 00035 // FIXME: what should we do for protected and internal visibility? 00036 // For variables, is internal different from hidden? 00037 bool isHidden = GV->hasHiddenVisibility(); 00038 00039 if (reloc == Reloc::PIC_) { 00040 if (isLocal || isHidden) 00041 return TLSModel::LocalDynamic; 00042 else 00043 return TLSModel::GeneralDynamic; 00044 } else { 00045 if (!isDeclaration || isHidden) 00046 return TLSModel::LocalExec; 00047 else 00048 return TLSModel::InitialExec; 00049 } 00050 } 00051 } 00052 00053 /// InitLibcallNames - Set default libcall names. 00054 /// 00055 static void InitLibcallNames(const char **Names) { 00056 Names[RTLIB::SHL_I16] = "__ashlhi3"; 00057 Names[RTLIB::SHL_I32] = "__ashlsi3"; 00058 Names[RTLIB::SHL_I64] = "__ashldi3"; 00059 Names[RTLIB::SHL_I128] = "__ashlti3"; 00060 Names[RTLIB::SRL_I16] = "__lshrhi3"; 00061 Names[RTLIB::SRL_I32] = "__lshrsi3"; 00062 Names[RTLIB::SRL_I64] = "__lshrdi3"; 00063 Names[RTLIB::SRL_I128] = "__lshrti3"; 00064 Names[RTLIB::SRA_I16] = "__ashrhi3"; 00065 Names[RTLIB::SRA_I32] = "__ashrsi3"; 00066 Names[RTLIB::SRA_I64] = "__ashrdi3"; 00067 Names[RTLIB::SRA_I128] = "__ashrti3"; 00068 Names[RTLIB::MUL_I16] = "__mulhi3"; 00069 Names[RTLIB::MUL_I32] = "__mulsi3"; 00070 Names[RTLIB::MUL_I64] = "__muldi3"; 00071 Names[RTLIB::MUL_I128] = "__multi3"; 00072 Names[RTLIB::SDIV_I16] = "__divhi3"; 00073 Names[RTLIB::SDIV_I32] = "__divsi3"; 00074 Names[RTLIB::SDIV_I64] = "__divdi3"; 00075 Names[RTLIB::SDIV_I128] = "__divti3"; 00076 Names[RTLIB::UDIV_I16] = "__udivhi3"; 00077 Names[RTLIB::UDIV_I32] = "__udivsi3"; 00078 Names[RTLIB::UDIV_I64] = "__udivdi3"; 00079 Names[RTLIB::UDIV_I128] = "__udivti3"; 00080 Names[RTLIB::SREM_I16] = "__modhi3"; 00081 Names[RTLIB::SREM_I32] = "__modsi3"; 00082 Names[RTLIB::SREM_I64] = "__moddi3"; 00083 Names[RTLIB::SREM_I128] = "__modti3"; 00084 Names[RTLIB::UREM_I16] = "__umodhi3"; 00085 Names[RTLIB::UREM_I32] = "__umodsi3"; 00086 Names[RTLIB::UREM_I64] = "__umoddi3"; 00087 Names[RTLIB::UREM_I128] = "__umodti3"; 00088 Names[RTLIB::NEG_I32] = "__negsi2"; 00089 Names[RTLIB::NEG_I64] = "__negdi2"; 00090 Names[RTLIB::ADD_F32] = "__addsf3"; 00091 Names[RTLIB::ADD_F64] = "__adddf3"; 00092 Names[RTLIB::ADD_F80] = "__addxf3"; 00093 Names[RTLIB::ADD_PPCF128] = "__gcc_qadd"; 00094 Names[RTLIB::SUB_F32] = "__subsf3"; 00095 Names[RTLIB::SUB_F64] = "__subdf3"; 00096 Names[RTLIB::SUB_F80] = "__subxf3"; 00097 Names[RTLIB::SUB_PPCF128] = "__gcc_qsub"; 00098 Names[RTLIB::MUL_F32] = "__mulsf3"; 00099 Names[RTLIB::MUL_F64] = "__muldf3"; 00100 Names[RTLIB::MUL_F80] = "__mulxf3"; 00101 Names[RTLIB::MUL_PPCF128] = "__gcc_qmul"; 00102 Names[RTLIB::DIV_F32] = "__divsf3"; 00103 Names[RTLIB::DIV_F64] = "__divdf3"; 00104 Names[RTLIB::DIV_F80] = "__divxf3"; 00105 Names[RTLIB::DIV_PPCF128] = "__gcc_qdiv"; 00106 Names[RTLIB::REM_F32] = "fmodf"; 00107 Names[RTLIB::REM_F64] = "fmod"; 00108 Names[RTLIB::REM_F80] = "fmodl"; 00109 Names[RTLIB::REM_PPCF128] = "fmodl"; 00110 Names[RTLIB::POWI_F32] = "__powisf2"; 00111 Names[RTLIB::POWI_F64] = "__powidf2"; 00112 Names[RTLIB::POWI_F80] = "__powixf2"; 00113 Names[RTLIB::POWI_PPCF128] = "__powitf2"; 00114 Names[RTLIB::SQRT_F32] = "sqrtf"; 00115 Names[RTLIB::SQRT_F64] = "sqrt"; 00116 Names[RTLIB::SQRT_F80] = "sqrtl"; 00117 Names[RTLIB::SQRT_PPCF128] = "sqrtl"; 00118 Names[RTLIB::LOG_F32] = "logf"; 00119 Names[RTLIB::LOG_F64] = "log"; 00120 Names[RTLIB::LOG_F80] = "logl"; 00121 Names[RTLIB::LOG_PPCF128] = "logl"; 00122 Names[RTLIB::LOG2_F32] = "log2f"; 00123 Names[RTLIB::LOG2_F64] = "log2"; 00124 Names[RTLIB::LOG2_F80] = "log2l"; 00125 Names[RTLIB::LOG2_PPCF128] = "log2l"; 00126 Names[RTLIB::LOG10_F32] = "log10f"; 00127 Names[RTLIB::LOG10_F64] = "log10"; 00128 Names[RTLIB::LOG10_F80] = "log10l"; 00129 Names[RTLIB::LOG10_PPCF128] = "log10l"; 00130 Names[RTLIB::EXP_F32] = "expf"; 00131 Names[RTLIB::EXP_F64] = "exp"; 00132 Names[RTLIB::EXP_F80] = "expl"; 00133 Names[RTLIB::EXP_PPCF128] = "expl"; 00134 Names[RTLIB::EXP2_F32] = "exp2f"; 00135 Names[RTLIB::EXP2_F64] = "exp2"; 00136 Names[RTLIB::EXP2_F80] = "exp2l"; 00137 Names[RTLIB::EXP2_PPCF128] = "exp2l"; 00138 Names[RTLIB::SIN_F32] = "sinf"; 00139 Names[RTLIB::SIN_F64] = "sin"; 00140 Names[RTLIB::SIN_F80] = "sinl"; 00141 Names[RTLIB::SIN_PPCF128] = "sinl"; 00142 Names[RTLIB::COS_F32] = "cosf"; 00143 Names[RTLIB::COS_F64] = "cos"; 00144 Names[RTLIB::COS_F80] = "cosl"; 00145 Names[RTLIB::COS_PPCF128] = "cosl"; 00146 Names[RTLIB::POW_F32] = "powf"; 00147 Names[RTLIB::POW_F64] = "pow"; 00148 Names[RTLIB::POW_F80] = "powl"; 00149 Names[RTLIB::POW_PPCF128] = "powl"; 00150 Names[RTLIB::CEIL_F32] = "ceilf"; 00151 Names[RTLIB::CEIL_F64] = "ceil"; 00152 Names[RTLIB::CEIL_F80] = "ceill"; 00153 Names[RTLIB::CEIL_PPCF128] = "ceill"; 00154 Names[RTLIB::TRUNC_F32] = "truncf"; 00155 Names[RTLIB::TRUNC_F64] = "trunc"; 00156 Names[RTLIB::TRUNC_F80] = "truncl"; 00157 Names[RTLIB::TRUNC_PPCF128] = "truncl"; 00158 Names[RTLIB::RINT_F32] = "rintf"; 00159 Names[RTLIB::RINT_F64] = "rint"; 00160 Names[RTLIB::RINT_F80] = "rintl"; 00161 Names[RTLIB::RINT_PPCF128] = "rintl"; 00162 Names[RTLIB::NEARBYINT_F32] = "nearbyintf"; 00163 Names[RTLIB::NEARBYINT_F64] = "nearbyint"; 00164 Names[RTLIB::NEARBYINT_F80] = "nearbyintl"; 00165 Names[RTLIB::NEARBYINT_PPCF128] = "nearbyintl"; 00166 Names[RTLIB::FLOOR_F32] = "floorf"; 00167 Names[RTLIB::FLOOR_F64] = "floor"; 00168 Names[RTLIB::FLOOR_F80] = "floorl"; 00169 Names[RTLIB::FLOOR_PPCF128] = "floorl"; 00170 Names[RTLIB::FPEXT_F32_F64] = "__extendsfdf2"; 00171 Names[RTLIB::FPROUND_F64_F32] = "__truncdfsf2"; 00172 Names[RTLIB::FPROUND_F80_F32] = "__truncxfsf2"; 00173 Names[RTLIB::FPROUND_PPCF128_F32] = "__trunctfsf2"; 00174 Names[RTLIB::FPROUND_F80_F64] = "__truncxfdf2"; 00175 Names[RTLIB::FPROUND_PPCF128_F64] = "__trunctfdf2"; 00176 Names[RTLIB::FPTOSINT_F32_I8] = "__fixsfi8"; 00177 Names[RTLIB::FPTOSINT_F32_I16] = "__fixsfi16"; 00178 Names[RTLIB::FPTOSINT_F32_I32] = "__fixsfsi"; 00179 Names[RTLIB::FPTOSINT_F32_I64] = "__fixsfdi"; 00180 Names[RTLIB::FPTOSINT_F32_I128] = "__fixsfti"; 00181 Names[RTLIB::FPTOSINT_F64_I32] = "__fixdfsi"; 00182 Names[RTLIB::FPTOSINT_F64_I64] = "__fixdfdi"; 00183 Names[RTLIB::FPTOSINT_F64_I128] = "__fixdfti"; 00184 Names[RTLIB::FPTOSINT_F80_I32] = "__fixxfsi"; 00185 Names[RTLIB::FPTOSINT_F80_I64] = "__fixxfdi"; 00186 Names[RTLIB::FPTOSINT_F80_I128] = "__fixxfti"; 00187 Names[RTLIB::FPTOSINT_PPCF128_I32] = "__fixtfsi"; 00188 Names[RTLIB::FPTOSINT_PPCF128_I64] = "__fixtfdi"; 00189 Names[RTLIB::FPTOSINT_PPCF128_I128] = "__fixtfti"; 00190 Names[RTLIB::FPTOUINT_F32_I8] = "__fixunssfi8"; 00191 Names[RTLIB::FPTOUINT_F32_I16] = "__fixunssfi16"; 00192 Names[RTLIB::FPTOUINT_F32_I32] = "__fixunssfsi"; 00193 Names[RTLIB::FPTOUINT_F32_I64] = "__fixunssfdi"; 00194 Names[RTLIB::FPTOUINT_F32_I128] = "__fixunssfti"; 00195 Names[RTLIB::FPTOUINT_F64_I32] = "__fixunsdfsi"; 00196 Names[RTLIB::FPTOUINT_F64_I64] = "__fixunsdfdi"; 00197 Names[RTLIB::FPTOUINT_F64_I128] = "__fixunsdfti"; 00198 Names[RTLIB::FPTOUINT_F80_I32] = "__fixunsxfsi"; 00199 Names[RTLIB::FPTOUINT_F80_I64] = "__fixunsxfdi"; 00200 Names[RTLIB::FPTOUINT_F80_I128] = "__fixunsxfti"; 00201 Names[RTLIB::FPTOUINT_PPCF128_I32] = "__fixunstfsi"; 00202 Names[RTLIB::FPTOUINT_PPCF128_I64] = "__fixunstfdi"; 00203 Names[RTLIB::FPTOUINT_PPCF128_I128] = "__fixunstfti"; 00204 Names[RTLIB::SINTTOFP_I32_F32] = "__floatsisf"; 00205 Names[RTLIB::SINTTOFP_I32_F64] = "__floatsidf"; 00206 Names[RTLIB::SINTTOFP_I32_F80] = "__floatsixf"; 00207 Names[RTLIB::SINTTOFP_I32_PPCF128] = "__floatsitf"; 00208 Names[RTLIB::SINTTOFP_I64_F32] = "__floatdisf"; 00209 Names[RTLIB::SINTTOFP_I64_F64] = "__floatdidf"; 00210 Names[RTLIB::SINTTOFP_I64_F80] = "__floatdixf"; 00211 Names[RTLIB::SINTTOFP_I64_PPCF128] = "__floatditf"; 00212 Names[RTLIB::SINTTOFP_I128_F32] = "__floattisf"; 00213 Names[RTLIB::SINTTOFP_I128_F64] = "__floattidf"; 00214 Names[RTLIB::SINTTOFP_I128_F80] = "__floattixf"; 00215 Names[RTLIB::SINTTOFP_I128_PPCF128] = "__floattitf"; 00216 Names[RTLIB::UINTTOFP_I32_F32] = "__floatunsisf"; 00217 Names[RTLIB::UINTTOFP_I32_F64] = "__floatunsidf"; 00218 Names[RTLIB::UINTTOFP_I32_F80] = "__floatunsixf"; 00219 Names[RTLIB::UINTTOFP_I32_PPCF128] = "__floatunsitf"; 00220 Names[RTLIB::UINTTOFP_I64_F32] = "__floatundisf"; 00221 Names[RTLIB::UINTTOFP_I64_F64] = "__floatundidf"; 00222 Names[RTLIB::UINTTOFP_I64_F80] = "__floatundixf"; 00223 Names[RTLIB::UINTTOFP_I64_PPCF128] = "__floatunditf"; 00224 Names[RTLIB::UINTTOFP_I128_F32] = "__floatuntisf"; 00225 Names[RTLIB::UINTTOFP_I128_F64] = "__floatuntidf"; 00226 Names[RTLIB::UINTTOFP_I128_F80] = "__floatuntixf"; 00227 Names[RTLIB::UINTTOFP_I128_PPCF128] = "__floatuntitf"; 00228 Names[RTLIB::OEQ_F32] = "__eqsf2"; 00229 Names[RTLIB::OEQ_F64] = "__eqdf2"; 00230 Names[RTLIB::UNE_F32] = "__nesf2"; 00231 Names[RTLIB::UNE_F64] = "__nedf2"; 00232 Names[RTLIB::OGE_F32] = "__gesf2"; 00233 Names[RTLIB::OGE_F64] = "__gedf2"; 00234 Names[RTLIB::OLT_F32] = "__ltsf2"; 00235 Names[RTLIB::OLT_F64] = "__ltdf2"; 00236 Names[RTLIB::OLE_F32] = "__lesf2"; 00237 Names[RTLIB::OLE_F64] = "__ledf2"; 00238 Names[RTLIB::OGT_F32] = "__gtsf2"; 00239 Names[RTLIB::OGT_F64] = "__gtdf2"; 00240 Names[RTLIB::UO_F32] = "__unordsf2"; 00241 Names[RTLIB::UO_F64] = "__unorddf2"; 00242 Names[RTLIB::O_F32] = "__unordsf2"; 00243 Names[RTLIB::O_F64] = "__unorddf2"; 00244 Names[RTLIB::MEMCPY] = "memcpy"; 00245 Names[RTLIB::MEMMOVE] = "memmove"; 00246 Names[RTLIB::MEMSET] = "memset"; 00247 Names[RTLIB::UNWIND_RESUME] = "_Unwind_Resume"; 00248 } 00249 00250 /// InitLibcallCallingConvs - Set default libcall CallingConvs. 00251 /// 00252 static void InitLibcallCallingConvs(CallingConv::ID *CCs) { 00253 for (int i = 0; i < RTLIB::UNKNOWN_LIBCALL; ++i) { 00254 CCs[i] = CallingConv::C; 00255 } 00256 } 00257 00258 /// getFPEXT - Return the FPEXT_*_* value for the given types, or 00259 /// UNKNOWN_LIBCALL if there is none. 00260 RTLIB::Libcall RTLIB::getFPEXT(EVT OpVT, EVT RetVT) { 00261 if (OpVT == MVT::f32) { 00262 if (RetVT == MVT::f64) 00263 return FPEXT_F32_F64; 00264 } 00265 return UNKNOWN_LIBCALL; 00266 } 00267 00268 /// getFPROUND - Return the FPROUND_*_* value for the given types, or 00269 /// UNKNOWN_LIBCALL if there is none. 00270 RTLIB::Libcall RTLIB::getFPROUND(EVT OpVT, EVT RetVT) { 00271 if (RetVT == MVT::f32) { 00272 if (OpVT == MVT::f64) 00273 return FPROUND_F64_F32; 00274 if (OpVT == MVT::f80) 00275 return FPROUND_F80_F32; 00276 if (OpVT == MVT::ppcf128) 00277 return FPROUND_PPCF128_F32; 00278 } else if (RetVT == MVT::f64) { 00279 if (OpVT == MVT::f80) 00280 return FPROUND_F80_F64; 00281 if (OpVT == MVT::ppcf128) 00282 return FPROUND_PPCF128_F64; 00283 } 00284 return UNKNOWN_LIBCALL; 00285 } 00286 00287 /// getFPTOSINT - Return the FPTOSINT_*_* value for the given types, or 00288 /// UNKNOWN_LIBCALL if there is none. 00289 RTLIB::Libcall RTLIB::getFPTOSINT(EVT OpVT, EVT RetVT) { 00290 if (OpVT == MVT::f32) { 00291 if (RetVT == MVT::i8) 00292 return FPTOSINT_F32_I8; 00293 if (RetVT == MVT::i16) 00294 return FPTOSINT_F32_I16; 00295 if (RetVT == MVT::i32) 00296 return FPTOSINT_F32_I32; 00297 if (RetVT == MVT::i64) 00298 return FPTOSINT_F32_I64; 00299 if (RetVT == MVT::i128) 00300 return FPTOSINT_F32_I128; 00301 } else if (OpVT == MVT::f64) { 00302 if (RetVT == MVT::i32) 00303 return FPTOSINT_F64_I32; 00304 if (RetVT == MVT::i64) 00305 return FPTOSINT_F64_I64; 00306 if (RetVT == MVT::i128) 00307 return FPTOSINT_F64_I128; 00308 } else if (OpVT == MVT::f80) { 00309 if (RetVT == MVT::i32) 00310 return FPTOSINT_F80_I32; 00311 if (RetVT == MVT::i64) 00312 return FPTOSINT_F80_I64; 00313 if (RetVT == MVT::i128) 00314 return FPTOSINT_F80_I128; 00315 } else if (OpVT == MVT::ppcf128) { 00316 if (RetVT == MVT::i32) 00317 return FPTOSINT_PPCF128_I32; 00318 if (RetVT == MVT::i64) 00319 return FPTOSINT_PPCF128_I64; 00320 if (RetVT == MVT::i128) 00321 return FPTOSINT_PPCF128_I128; 00322 } 00323 return UNKNOWN_LIBCALL; 00324 } 00325 00326 /// getFPTOUINT - Return the FPTOUINT_*_* value for the given types, or 00327 /// UNKNOWN_LIBCALL if there is none. 00328 RTLIB::Libcall RTLIB::getFPTOUINT(EVT OpVT, EVT RetVT) { 00329 if (OpVT == MVT::f32) { 00330 if (RetVT == MVT::i8) 00331 return FPTOUINT_F32_I8; 00332 if (RetVT == MVT::i16) 00333 return FPTOUINT_F32_I16; 00334 if (RetVT == MVT::i32) 00335 return FPTOUINT_F32_I32; 00336 if (RetVT == MVT::i64) 00337 return FPTOUINT_F32_I64; 00338 if (RetVT == MVT::i128) 00339 return FPTOUINT_F32_I128; 00340 } else if (OpVT == MVT::f64) { 00341 if (RetVT == MVT::i32) 00342 return FPTOUINT_F64_I32; 00343 if (RetVT == MVT::i64) 00344 return FPTOUINT_F64_I64; 00345 if (RetVT == MVT::i128) 00346 return FPTOUINT_F64_I128; 00347 } else if (OpVT == MVT::f80) { 00348 if (RetVT == MVT::i32) 00349 return FPTOUINT_F80_I32; 00350 if (RetVT == MVT::i64) 00351 return FPTOUINT_F80_I64; 00352 if (RetVT == MVT::i128) 00353 return FPTOUINT_F80_I128; 00354 } else if (OpVT == MVT::ppcf128) { 00355 if (RetVT == MVT::i32) 00356 return FPTOUINT_PPCF128_I32; 00357 if (RetVT == MVT::i64) 00358 return FPTOUINT_PPCF128_I64; 00359 if (RetVT == MVT::i128) 00360 return FPTOUINT_PPCF128_I128; 00361 } 00362 return UNKNOWN_LIBCALL; 00363 } 00364 00365 /// getSINTTOFP - Return the SINTTOFP_*_* value for the given types, or 00366 /// UNKNOWN_LIBCALL if there is none. 00367 RTLIB::Libcall RTLIB::getSINTTOFP(EVT OpVT, EVT RetVT) { 00368 if (OpVT == MVT::i32) { 00369 if (RetVT == MVT::f32) 00370 return SINTTOFP_I32_F32; 00371 else if (RetVT == MVT::f64) 00372 return SINTTOFP_I32_F64; 00373 else if (RetVT == MVT::f80) 00374 return SINTTOFP_I32_F80; 00375 else if (RetVT == MVT::ppcf128) 00376 return SINTTOFP_I32_PPCF128; 00377 } else if (OpVT == MVT::i64) { 00378 if (RetVT == MVT::f32) 00379 return SINTTOFP_I64_F32; 00380 else if (RetVT == MVT::f64) 00381 return SINTTOFP_I64_F64; 00382 else if (RetVT == MVT::f80) 00383 return SINTTOFP_I64_F80; 00384 else if (RetVT == MVT::ppcf128) 00385 return SINTTOFP_I64_PPCF128; 00386 } else if (OpVT == MVT::i128) { 00387 if (RetVT == MVT::f32) 00388 return SINTTOFP_I128_F32; 00389 else if (RetVT == MVT::f64) 00390 return SINTTOFP_I128_F64; 00391 else if (RetVT == MVT::f80) 00392 return SINTTOFP_I128_F80; 00393 else if (RetVT == MVT::ppcf128) 00394 return SINTTOFP_I128_PPCF128; 00395 } 00396 return UNKNOWN_LIBCALL; 00397 } 00398 00399 /// getUINTTOFP - Return the UINTTOFP_*_* value for the given types, or 00400 /// UNKNOWN_LIBCALL if there is none. 00401 RTLIB::Libcall RTLIB::getUINTTOFP(EVT OpVT, EVT RetVT) { 00402 if (OpVT == MVT::i32) { 00403 if (RetVT == MVT::f32) 00404 return UINTTOFP_I32_F32; 00405 else if (RetVT == MVT::f64) 00406 return UINTTOFP_I32_F64; 00407 else if (RetVT == MVT::f80) 00408 return UINTTOFP_I32_F80; 00409 else if (RetVT == MVT::ppcf128) 00410 return UINTTOFP_I32_PPCF128; 00411 } else if (OpVT == MVT::i64) { 00412 if (RetVT == MVT::f32) 00413 return UINTTOFP_I64_F32; 00414 else if (RetVT == MVT::f64) 00415 return UINTTOFP_I64_F64; 00416 else if (RetVT == MVT::f80) 00417 return UINTTOFP_I64_F80; 00418 else if (RetVT == MVT::ppcf128) 00419 return UINTTOFP_I64_PPCF128; 00420 } else if (OpVT == MVT::i128) { 00421 if (RetVT == MVT::f32) 00422 return UINTTOFP_I128_F32; 00423 else if (RetVT == MVT::f64) 00424 return UINTTOFP_I128_F64; 00425 else if (RetVT == MVT::f80) 00426 return UINTTOFP_I128_F80; 00427 else if (RetVT == MVT::ppcf128) 00428 return UINTTOFP_I128_PPCF128; 00429 } 00430 return UNKNOWN_LIBCALL; 00431 } 00432 00433 /// InitCmpLibcallCCs - Set default comparison libcall CC. 00434 /// 00435 static void InitCmpLibcallCCs(ISD::CondCode *CCs) { 00436 memset(CCs, ISD::SETCC_INVALID, sizeof(ISD::CondCode)*RTLIB::UNKNOWN_LIBCALL); 00437 CCs[RTLIB::OEQ_F32] = ISD::SETEQ; 00438 CCs[RTLIB::OEQ_F64] = ISD::SETEQ; 00439 CCs[RTLIB::UNE_F32] = ISD::SETNE; 00440 CCs[RTLIB::UNE_F64] = ISD::SETNE; 00441 CCs[RTLIB::OGE_F32] = ISD::SETGE; 00442 CCs[RTLIB::OGE_F64] = ISD::SETGE; 00443 CCs[RTLIB::OLT_F32] = ISD::SETLT; 00444 CCs[RTLIB::OLT_F64] = ISD::SETLT; 00445 CCs[RTLIB::OLE_F32] = ISD::SETLE; 00446 CCs[RTLIB::OLE_F64] = ISD::SETLE; 00447 CCs[RTLIB::OGT_F32] = ISD::SETGT; 00448 CCs[RTLIB::OGT_F64] = ISD::SETGT; 00449 CCs[RTLIB::UO_F32] = ISD::SETNE; 00450 CCs[RTLIB::UO_F64] = ISD::SETNE; 00451 CCs[RTLIB::O_F32] = ISD::SETEQ; 00452 CCs[RTLIB::O_F64] = ISD::SETEQ; 00453 } 00454 00455 /// NOTE: The constructor takes ownership of TLOF. 00456 TargetLowering::TargetLowering(TargetMachine &tm,TargetLoweringObjectFile *tlof) 00457 : TM(tm), TD(TM.getTargetData()), TLOF(*tlof) { 00458 // All operations default to being supported. 00459 memset(OpActions, 0, sizeof(OpActions)); 00460 memset(LoadExtActions, 0, sizeof(LoadExtActions)); 00461 memset(TruncStoreActions, 0, sizeof(TruncStoreActions)); 00462 memset(IndexedModeActions, 0, sizeof(IndexedModeActions)); 00463 memset(ConvertActions, 0, sizeof(ConvertActions)); 00464 memset(CondCodeActions, 0, sizeof(CondCodeActions)); 00465 00466 // Set default actions for various operations. 00467 for (unsigned VT = 0; VT != (unsigned)MVT::LAST_VALUETYPE; ++VT) { 00468 // Default all indexed load / store to expand. 00469 for (unsigned IM = (unsigned)ISD::PRE_INC; 00470 IM != (unsigned)ISD::LAST_INDEXED_MODE; ++IM) { 00471 setIndexedLoadAction(IM, (MVT::SimpleValueType)VT, Expand); 00472 setIndexedStoreAction(IM, (MVT::SimpleValueType)VT, Expand); 00473 } 00474 00475 // These operations default to expand. 00476 setOperationAction(ISD::FGETSIGN, (MVT::SimpleValueType)VT, Expand); 00477 setOperationAction(ISD::CONCAT_VECTORS, (MVT::SimpleValueType)VT, Expand); 00478 } 00479 00480 // Most targets ignore the @llvm.prefetch intrinsic. 00481 setOperationAction(ISD::PREFETCH, MVT::Other, Expand); 00482 00483 // ConstantFP nodes default to expand. Targets can either change this to 00484 // Legal, in which case all fp constants are legal, or use addLegalFPImmediate 00485 // to optimize expansions for certain constants. 00486 setOperationAction(ISD::ConstantFP, MVT::f32, Expand); 00487 setOperationAction(ISD::ConstantFP, MVT::f64, Expand); 00488 setOperationAction(ISD::ConstantFP, MVT::f80, Expand); 00489 00490 // These library functions default to expand. 00491 setOperationAction(ISD::FLOG , MVT::f64, Expand); 00492 setOperationAction(ISD::FLOG2, MVT::f64, Expand); 00493 setOperationAction(ISD::FLOG10,MVT::f64, Expand); 00494 setOperationAction(ISD::FEXP , MVT::f64, Expand); 00495 setOperationAction(ISD::FEXP2, MVT::f64, Expand); 00496 setOperationAction(ISD::FLOG , MVT::f32, Expand); 00497 setOperationAction(ISD::FLOG2, MVT::f32, Expand); 00498 setOperationAction(ISD::FLOG10,MVT::f32, Expand); 00499 setOperationAction(ISD::FEXP , MVT::f32, Expand); 00500 setOperationAction(ISD::FEXP2, MVT::f32, Expand); 00501 00502 // Default ISD::TRAP to expand (which turns it into abort). 00503 setOperationAction(ISD::TRAP, MVT::Other, Expand); 00504 00505 IsLittleEndian = TD->isLittleEndian(); 00506 UsesGlobalOffsetTable = false; 00507 ShiftAmountTy = PointerTy = MVT::getIntegerVT(8*TD->getPointerSize()); 00508 memset(RegClassForVT, 0,MVT::LAST_VALUETYPE*sizeof(TargetRegisterClass*)); 00509 memset(TargetDAGCombineArray, 0, array_lengthof(TargetDAGCombineArray)); 00510 maxStoresPerMemset = maxStoresPerMemcpy = maxStoresPerMemmove = 8; 00511 benefitFromCodePlacementOpt = false; 00512 UseUnderscoreSetJmp = false; 00513 UseUnderscoreLongJmp = false; 00514 SelectIsExpensive = false; 00515 IntDivIsCheap = false; 00516 Pow2DivIsCheap = false; 00517 StackPointerRegisterToSaveRestore = 0; 00518 ExceptionPointerRegister = 0; 00519 ExceptionSelectorRegister = 0; 00520 BooleanContents = UndefinedBooleanContent; 00521 SchedPreferenceInfo = SchedulingForLatency; 00522 JumpBufSize = 0; 00523 JumpBufAlignment = 0; 00524 IfCvtBlockSizeLimit = 2; 00525 IfCvtDupBlockSizeLimit = 0; 00526 PrefLoopAlignment = 0; 00527 00528 InitLibcallNames(LibcallRoutineNames); 00529 InitCmpLibcallCCs(CmpLibcallCCs); 00530 InitLibcallCallingConvs(LibcallCallingConvs); 00531 00532 // Tell Legalize whether the assembler supports DEBUG_LOC. 00533 const TargetAsmInfo *TASM = TM.getTargetAsmInfo(); 00534 if (!TASM || !TASM->hasDotLocAndDotFile()) 00535 setOperationAction(ISD::DEBUG_LOC, MVT::Other, Expand); 00536 } 00537 00538 TargetLowering::~TargetLowering() { 00539 delete &TLOF; 00540 } 00541 00542 static unsigned getVectorTypeBreakdownMVT(MVT VT, MVT &IntermediateVT, 00543 unsigned &NumIntermediates, 00544 EVT &RegisterVT, 00545 TargetLowering* TLI) { 00546 // Figure out the right, legal destination reg to copy into. 00547 unsigned NumElts = VT.getVectorNumElements(); 00548 MVT EltTy = VT.getVectorElementType(); 00549 00550 unsigned NumVectorRegs = 1; 00551 00552 // FIXME: We don't support non-power-of-2-sized vectors for now. Ideally we 00553 // could break down into LHS/RHS like LegalizeDAG does. 00554 if (!isPowerOf2_32(NumElts)) { 00555 NumVectorRegs = NumElts; 00556 NumElts = 1; 00557 } 00558 00559 // Divide the input until we get to a supported size. This will always 00560 // end with a scalar if the target doesn't support vectors. 00561 while (NumElts > 1 && !TLI->isTypeLegal(MVT::getVectorVT(EltTy, NumElts))) { 00562 NumElts >>= 1; 00563 NumVectorRegs <<= 1; 00564 } 00565 00566 NumIntermediates = NumVectorRegs; 00567 00568 MVT NewVT = MVT::getVectorVT(EltTy, NumElts); 00569 if (!TLI->isTypeLegal(NewVT)) 00570 NewVT = EltTy; 00571 IntermediateVT = NewVT; 00572 00573 EVT DestVT = TLI->getRegisterType(NewVT); 00574 RegisterVT = DestVT; 00575 if (EVT(DestVT).bitsLT(NewVT)) { 00576 // Value is expanded, e.g. i64 -> i16. 00577 return NumVectorRegs*(NewVT.getSizeInBits()/DestVT.getSizeInBits()); 00578 } else { 00579 // Otherwise, promotion or legal types use the same number of registers as 00580 // the vector decimated to the appropriate level. 00581 return NumVectorRegs; 00582 } 00583 00584 return 1; 00585 } 00586 00587 /// computeRegisterProperties - Once all of the register classes are added, 00588 /// this allows us to compute derived properties we expose. 00589 void TargetLowering::computeRegisterProperties() { 00590 assert(MVT::LAST_VALUETYPE <= MVT::MAX_ALLOWED_VALUETYPE && 00591 "Too many value types for ValueTypeActions to hold!"); 00592 00593 // Everything defaults to needing one register. 00594 for (unsigned i = 0; i != MVT::LAST_VALUETYPE; ++i) { 00595 NumRegistersForVT[i] = 1; 00596 RegisterTypeForVT[i] = TransformToType[i] = (MVT::SimpleValueType)i; 00597 } 00598 // ...except isVoid, which doesn't need any registers. 00599 NumRegistersForVT[MVT::isVoid] = 0; 00600 00601 // Find the largest integer register class. 00602 unsigned LargestIntReg = MVT::LAST_INTEGER_VALUETYPE; 00603 for (; RegClassForVT[LargestIntReg] == 0; --LargestIntReg) 00604 assert(LargestIntReg != MVT::i1 && "No integer registers defined!"); 00605 00606 // Every integer value type larger than this largest register takes twice as 00607 // many registers to represent as the previous ValueType. 00608 for (unsigned ExpandedReg = LargestIntReg + 1; ; ++ExpandedReg) { 00609 EVT EVT = (MVT::SimpleValueType)ExpandedReg; 00610 if (!EVT.isInteger()) 00611 break; 00612 NumRegistersForVT[ExpandedReg] = 2*NumRegistersForVT[ExpandedReg-1]; 00613 RegisterTypeForVT[ExpandedReg] = (MVT::SimpleValueType)LargestIntReg; 00614 TransformToType[ExpandedReg] = (MVT::SimpleValueType)(ExpandedReg - 1); 00615 ValueTypeActions.setTypeAction(EVT, Expand); 00616 } 00617 00618 // Inspect all of the ValueType's smaller than the largest integer 00619 // register to see which ones need promotion. 00620 unsigned LegalIntReg = LargestIntReg; 00621 for (unsigned IntReg = LargestIntReg - 1; 00622 IntReg >= (unsigned)MVT::i1; --IntReg) { 00623 EVT IVT = (MVT::SimpleValueType)IntReg; 00624 if (isTypeLegal(IVT)) { 00625 LegalIntReg = IntReg; 00626 } else { 00627 RegisterTypeForVT[IntReg] = TransformToType[IntReg] = 00628 (MVT::SimpleValueType)LegalIntReg; 00629 ValueTypeActions.setTypeAction(IVT, Promote); 00630 } 00631 } 00632 00633 // ppcf128 type is really two f64's. 00634 if (!isTypeLegal(MVT::ppcf128)) { 00635 NumRegistersForVT[MVT::ppcf128] = 2*NumRegistersForVT[MVT::f64]; 00636 RegisterTypeForVT[MVT::ppcf128] = MVT::f64; 00637 TransformToType[MVT::ppcf128] = MVT::f64; 00638 ValueTypeActions.setTypeAction(MVT::ppcf128, Expand); 00639 } 00640 00641 // Decide how to handle f64. If the target does not have native f64 support, 00642 // expand it to i64 and we will be generating soft float library calls. 00643 if (!isTypeLegal(MVT::f64)) { 00644 NumRegistersForVT[MVT::f64] = NumRegistersForVT[MVT::i64]; 00645 RegisterTypeForVT[MVT::f64] = RegisterTypeForVT[MVT::i64]; 00646 TransformToType[MVT::f64] = MVT::i64; 00647 ValueTypeActions.setTypeAction(MVT::f64, Expand); 00648 } 00649 00650 // Decide how to handle f32. If the target does not have native support for 00651 // f32, promote it to f64 if it is legal. Otherwise, expand it to i32. 00652 if (!isTypeLegal(MVT::f32)) { 00653 if (isTypeLegal(MVT::f64)) { 00654 NumRegistersForVT[MVT::f32] = NumRegistersForVT[MVT::f64]; 00655 RegisterTypeForVT[MVT::f32] = RegisterTypeForVT[MVT::f64]; 00656 TransformToType[MVT::f32] = MVT::f64; 00657 ValueTypeActions.setTypeAction(MVT::f32, Promote); 00658 } else { 00659 NumRegistersForVT[MVT::f32] = NumRegistersForVT[MVT::i32]; 00660 RegisterTypeForVT[MVT::f32] = RegisterTypeForVT[MVT::i32]; 00661 TransformToType[MVT::f32] = MVT::i32; 00662 ValueTypeActions.setTypeAction(MVT::f32, Expand); 00663 } 00664 } 00665 00666 // Loop over all of the vector value types to see which need transformations. 00667 for (unsigned i = MVT::FIRST_VECTOR_VALUETYPE; 00668 i <= (unsigned)MVT::LAST_VECTOR_VALUETYPE; ++i) { 00669 MVT VT = (MVT::SimpleValueType)i; 00670 if (!isTypeLegal(VT)) { 00671 MVT IntermediateVT; 00672 EVT RegisterVT; 00673 unsigned NumIntermediates; 00674 NumRegistersForVT[i] = 00675 getVectorTypeBreakdownMVT(VT, IntermediateVT, NumIntermediates, 00676 RegisterVT, this); 00677 RegisterTypeForVT[i] = RegisterVT; 00678 00679 // Determine if there is a legal wider type. 00680 bool IsLegalWiderType = false; 00681 EVT EltVT = VT.getVectorElementType(); 00682 unsigned NElts = VT.getVectorNumElements(); 00683 for (unsigned nVT = i+1; nVT <= MVT::LAST_VECTOR_VALUETYPE; ++nVT) { 00684 EVT SVT = (MVT::SimpleValueType)nVT; 00685 if (isTypeLegal(SVT) && SVT.getVectorElementType() == EltVT && 00686 SVT.getVectorNumElements() > NElts) { 00687 TransformToType[i] = SVT; 00688 ValueTypeActions.setTypeAction(VT, Promote); 00689 IsLegalWiderType = true; 00690 break; 00691 } 00692 } 00693 if (!IsLegalWiderType) { 00694 EVT NVT = VT.getPow2VectorType(); 00695 if (NVT == VT) { 00696 // Type is already a power of 2. The default action is to split. 00697 TransformToType[i] = MVT::Other; 00698 ValueTypeActions.setTypeAction(VT, Expand); 00699 } else { 00700 TransformToType[i] = NVT; 00701 ValueTypeActions.setTypeAction(VT, Promote); 00702 } 00703 } 00704 } 00705 } 00706 } 00707 00708 const char *TargetLowering::getTargetNodeName(unsigned Opcode) const { 00709 return NULL; 00710 } 00711 00712 00713 MVT::SimpleValueType TargetLowering::getSetCCResultType(EVT VT) const { 00714 return PointerTy.SimpleTy; 00715 } 00716 00717 /// getVectorTypeBreakdown - Vector types are broken down into some number of 00718 /// legal first class types. For example, MVT::v8f32 maps to 2 MVT::v4f32 00719 /// with Altivec or SSE1, or 8 promoted MVT::f64 values with the X86 FP stack. 00720 /// Similarly, MVT::v2i64 turns into 4 MVT::i32 values with both PPC and X86. 00721 /// 00722 /// This method returns the number of registers needed, and the VT for each 00723 /// register. It also returns the VT and quantity of the intermediate values 00724 /// before they are promoted/expanded. 00725 /// 00726 unsigned TargetLowering::getVectorTypeBreakdown(LLVMContext &Context, EVT VT, 00727 EVT &IntermediateVT, 00728 unsigned &NumIntermediates, 00729 EVT &RegisterVT) const { 00730 // Figure out the right, legal destination reg to copy into. 00731 unsigned NumElts = VT.getVectorNumElements(); 00732 EVT EltTy = VT.getVectorElementType(); 00733 00734 unsigned NumVectorRegs = 1; 00735 00736 // FIXME: We don't support non-power-of-2-sized vectors for now. Ideally we 00737 // could break down into LHS/RHS like LegalizeDAG does. 00738 if (!isPowerOf2_32(NumElts)) { 00739 NumVectorRegs = NumElts; 00740 NumElts = 1; 00741 } 00742 00743 // Divide the input until we get to a supported size. This will always 00744 // end with a scalar if the target doesn't support vectors. 00745 while (NumElts > 1 && !isTypeLegal( 00746 EVT::getVectorVT(Context, EltTy, NumElts))) { 00747 NumElts >>= 1; 00748 NumVectorRegs <<= 1; 00749 } 00750 00751 NumIntermediates = NumVectorRegs; 00752 00753 EVT NewVT = EVT::getVectorVT(Context, EltTy, NumElts); 00754 if (!isTypeLegal(NewVT)) 00755 NewVT = EltTy; 00756 IntermediateVT = NewVT; 00757 00758 EVT DestVT = getRegisterType(Context, NewVT); 00759 RegisterVT = DestVT; 00760 if (DestVT.bitsLT(NewVT)) { 00761 // Value is expanded, e.g. i64 -> i16. 00762 return NumVectorRegs*(NewVT.getSizeInBits()/DestVT.getSizeInBits()); 00763 } else { 00764 // Otherwise, promotion or legal types use the same number of registers as 00765 // the vector decimated to the appropriate level. 00766 return NumVectorRegs; 00767 } 00768 00769 return 1; 00770 } 00771 00772 /// getWidenVectorType: given a vector type, returns the type to widen to 00773 /// (e.g., v7i8 to v8i8). If the vector type is legal, it returns itself. 00774 /// If there is no vector type that we want to widen to, returns MVT::Other 00775 /// When and where to widen is target dependent based on the cost of 00776 /// scalarizing vs using the wider vector type. 00777 EVT TargetLowering::getWidenVectorType(EVT VT) const { 00778 assert(VT.isVector()); 00779 if (isTypeLegal(VT)) 00780 return VT; 00781 00782 // Default is not to widen until moved to LegalizeTypes 00783 return MVT::Other; 00784 } 00785 00786 /// getByValTypeAlignment - Return the desired alignment for ByVal aggregate 00787 /// function arguments in the caller parameter area. This is the actual 00788 /// alignment, not its logarithm. 00789 unsigned TargetLowering::getByValTypeAlignment(const Type *Ty) const { 00790 return TD->getCallFrameTypeAlignment(Ty); 00791 } 00792 00793 SDValue TargetLowering::getPICJumpTableRelocBase(SDValue Table, 00794 SelectionDAG &DAG) const { 00795 if (usesGlobalOffsetTable()) 00796 return DAG.getGLOBAL_OFFSET_TABLE(getPointerTy()); 00797 return Table; 00798 } 00799 00800 bool 00801 TargetLowering::isOffsetFoldingLegal(const GlobalAddressSDNode *GA) const { 00802 // Assume that everything is safe in static mode. 00803 if (getTargetMachine().getRelocationModel() == Reloc::Static) 00804 return true; 00805 00806 // In dynamic-no-pic mode, assume that known defined values are safe. 00807 if (getTargetMachine().getRelocationModel() == Reloc::DynamicNoPIC && 00808 GA && 00809 !GA->getGlobal()->isDeclaration() && 00810 !GA->getGlobal()->isWeakForLinker()) 00811 return true; 00812 00813 // Otherwise assume nothing is safe. 00814 return false; 00815 } 00816 00817 //===----------------------------------------------------------------------===// 00818 // Optimization Methods 00819 //===----------------------------------------------------------------------===// 00820 00821 /// ShrinkDemandedConstant - Check to see if the specified operand of the 00822 /// specified instruction is a constant integer. If so, check to see if there 00823 /// are any bits set in the constant that are not demanded. If so, shrink the 00824 /// constant and return true. 00825 bool TargetLowering::TargetLoweringOpt::ShrinkDemandedConstant(SDValue Op, 00826 const APInt &Demanded) { 00827 DebugLoc dl = Op.getDebugLoc(); 00828 00829 // FIXME: ISD::SELECT, ISD::SELECT_CC 00830 switch (Op.getOpcode()) { 00831 default: break; 00832 case ISD::XOR: 00833 case ISD::AND: 00834 case ISD::OR: { 00835 ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op.getOperand(1)); 00836 if (!C) return false; 00837 00838 if (Op.getOpcode() == ISD::XOR && 00839 (C->getAPIntValue() | (~Demanded)).isAllOnesValue()) 00840 return false; 00841 00842 // if we can expand it to have all bits set, do it 00843 if (C->getAPIntValue().intersects(~Demanded)) { 00844 EVT VT = Op.getValueType(); 00845 SDValue New = DAG.getNode(Op.getOpcode(), dl, VT, Op.getOperand(0), 00846 DAG.getConstant(Demanded & 00847 C->getAPIntValue(), 00848 VT)); 00849 return CombineTo(Op, New); 00850 } 00851 00852 break; 00853 } 00854 } 00855 00856 return false; 00857 } 00858 00859 /// ShrinkDemandedOp - Convert x+y to (VT)((SmallVT)x+(SmallVT)y) if the 00860 /// casts are free. This uses isZExtFree and ZERO_EXTEND for the widening 00861 /// cast, but it could be generalized for targets with other types of 00862 /// implicit widening casts. 00863 bool 00864 TargetLowering::TargetLoweringOpt::ShrinkDemandedOp(SDValue Op, 00865 unsigned BitWidth, 00866 const APInt &Demanded, 00867 DebugLoc dl) { 00868 assert(Op.getNumOperands() == 2 && 00869 "ShrinkDemandedOp only supports binary operators!"); 00870 assert(Op.getNode()->getNumValues() == 1 && 00871 "ShrinkDemandedOp only supports nodes with one result!"); 00872 00873 // Don't do this if the node has another user, which may require the 00874 // full value. 00875 if (!Op.getNode()->hasOneUse()) 00876 return false; 00877 00878 // Search for the smallest integer type with free casts to and from 00879 // Op's type. For expedience, just check power-of-2 integer types. 00880 const TargetLowering &TLI = DAG.getTargetLoweringInfo(); 00881 unsigned SmallVTBits = BitWidth - Demanded.countLeadingZeros(); 00882 if (!isPowerOf2_32(SmallVTBits)) 00883 SmallVTBits = NextPowerOf2(SmallVTBits); 00884 for (; SmallVTBits < BitWidth; SmallVTBits = NextPowerOf2(SmallVTBits)) { 00885 EVT SmallVT = EVT::getIntegerVT(*DAG.getContext(), SmallVTBits); 00886 if (TLI.isTruncateFree(Op.getValueType(), SmallVT) && 00887 TLI.isZExtFree(SmallVT, Op.getValueType())) { 00888 // We found a type with free casts. 00889 SDValue X = DAG.getNode(Op.getOpcode(), dl, SmallVT, 00890 DAG.getNode(ISD::TRUNCATE, dl, SmallVT, 00891 Op.getNode()->getOperand(0)), 00892 DAG.getNode(ISD::TRUNCATE, dl, SmallVT, 00893 Op.getNode()->getOperand(1))); 00894 SDValue Z = DAG.getNode(ISD::ZERO_EXTEND, dl, Op.getValueType(), X); 00895 return CombineTo(Op, Z); 00896 } 00897 } 00898 return false; 00899 } 00900 00901 /// SimplifyDemandedBits - Look at Op. At this point, we know that only the 00902 /// DemandedMask bits of the result of Op are ever used downstream. If we can 00903 /// use this information to simplify Op, create a new simplified DAG node and 00904 /// return true, returning the original and new nodes in Old and New. Otherwise, 00905 /// analyze the expression and return a mask of KnownOne and KnownZero bits for 00906 /// the expression (used to simplify the caller). The KnownZero/One bits may 00907 /// only be accurate for those bits in the DemandedMask. 00908 bool TargetLowering::SimplifyDemandedBits(SDValue Op, 00909 const APInt &DemandedMask, 00910 APInt &KnownZero, 00911 APInt &KnownOne, 00912 TargetLoweringOpt &TLO, 00913 unsigned Depth) const { 00914 unsigned BitWidth = DemandedMask.getBitWidth(); 00915 assert(Op.getValueSizeInBits() == BitWidth && 00916 "Mask size mismatches value type size!"); 00917 APInt NewMask = DemandedMask; 00918 DebugLoc dl = Op.getDebugLoc(); 00919 00920 // Don't know anything. 00921 KnownZero = KnownOne = APInt(BitWidth, 0); 00922 00923 // Other users may use these bits. 00924 if (!Op.getNode()->hasOneUse()) { 00925 if (Depth != 0) { 00926 // If not at the root, Just compute the KnownZero/KnownOne bits to 00927 // simplify things downstream. 00928 TLO.DAG.ComputeMaskedBits(Op, DemandedMask, KnownZero, KnownOne, Depth); 00929 return false; 00930 } 00931 // If this is the root being simplified, allow it to have multiple uses, 00932 // just set the NewMask to all bits. 00933 NewMask = APInt::getAllOnesValue(BitWidth); 00934 } else if (DemandedMask == 0) { 00935 // Not demanding any bits from Op. 00936 if (Op.getOpcode() != ISD::UNDEF) 00937 return TLO.CombineTo(Op, TLO.DAG.getUNDEF(Op.getValueType())); 00938 return false; 00939 } else if (Depth == 6) { // Limit search depth. 00940 return false; 00941 } 00942 00943 APInt KnownZero2, KnownOne2, KnownZeroOut, KnownOneOut; 00944 switch (Op.getOpcode()) { 00945 case ISD::Constant: 00946 // We know all of the bits for a constant! 00947 KnownOne = cast<ConstantSDNode>(Op)->getAPIntValue() & NewMask; 00948 KnownZero = ~KnownOne & NewMask; 00949 return false; // Don't fall through, will infinitely loop. 00950 case ISD::AND: 00951 // If the RHS is a constant, check to see if the LHS would be zero without 00952 // using the bits from the RHS. Below, we use knowledge about the RHS to 00953 // simplify the LHS, here we're using information from the LHS to simplify 00954 // the RHS. 00955 if (ConstantSDNode *RHSC = dyn_cast<ConstantSDNode>(Op.getOperand(1))) { 00956 APInt LHSZero, LHSOne; 00957 TLO.DAG.ComputeMaskedBits(Op.getOperand(0), NewMask, 00958 LHSZero, LHSOne, Depth+1); 00959 // If the LHS already has zeros where RHSC does, this and is dead. 00960 if ((LHSZero & NewMask) == (~RHSC->getAPIntValue() & NewMask)) 00961 return TLO.CombineTo(Op, Op.getOperand(0)); 00962 // If any of the set bits in the RHS are known zero on the LHS, shrink 00963 // the constant. 00964 if (TLO.ShrinkDemandedConstant(Op, ~LHSZero & NewMask)) 00965 return true; 00966 } 00967 00968 if (SimplifyDemandedBits(Op.getOperand(1), NewMask, KnownZero, 00969 KnownOne, TLO, Depth+1)) 00970 return true; 00971 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?"); 00972 if (SimplifyDemandedBits(Op.getOperand(0), ~KnownZero & NewMask, 00973 KnownZero2, KnownOne2, TLO, Depth+1)) 00974 return true; 00975 assert((KnownZero2 & KnownOne2) == 0 && "Bits known to be one AND zero?"); 00976 00977 // If all of the demanded bits are known one on one side, return the other. 00978 // These bits cannot contribute to the result of the 'and'. 00979 if ((NewMask & ~KnownZero2 & KnownOne) == (~KnownZero2 & NewMask)) 00980 return TLO.CombineTo(Op, Op.getOperand(0)); 00981 if ((NewMask & ~KnownZero & KnownOne2) == (~KnownZero & NewMask)) 00982 return TLO.CombineTo(Op, Op.getOperand(1)); 00983 // If all of the demanded bits in the inputs are known zeros, return zero. 00984 if ((NewMask & (KnownZero|KnownZero2)) == NewMask) 00985 return TLO.CombineTo(Op, TLO.DAG.getConstant(0, Op.getValueType())); 00986 // If the RHS is a constant, see if we can simplify it. 00987 if (TLO.ShrinkDemandedConstant(Op, ~KnownZero2 & NewMask)) 00988 return true; 00989 // If the operation can be done in a smaller type, do so. 00990 if (TLO.ShrinkDemandedOp(Op, BitWidth, NewMask, dl)) 00991 return true; 00992 00993 // Output known-1 bits are only known if set in both the LHS & RHS. 00994 KnownOne &= KnownOne2; 00995 // Output known-0 are known to be clear if zero in either the LHS | RHS. 00996 KnownZero |= KnownZero2; 00997 break; 00998 case ISD::OR: 00999 if (SimplifyDemandedBits(Op.getOperand(1), NewMask, KnownZero, 01000 KnownOne, TLO, Depth+1)) 01001 return true; 01002 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?"); 01003 if (SimplifyDemandedBits(Op.getOperand(0), ~KnownOne & NewMask, 01004 KnownZero2, KnownOne2, TLO, Depth+1)) 01005 return true; 01006 assert((KnownZero2 & KnownOne2) == 0 && "Bits known to be one AND zero?"); 01007 01008 // If all of the demanded bits are known zero on one side, return the other. 01009 // These bits cannot contribute to the result of the 'or'. 01010 if ((NewMask & ~KnownOne2 & KnownZero) == (~KnownOne2 & NewMask)) 01011 return TLO.CombineTo(Op, Op.getOperand(0)); 01012 if ((NewMask & ~KnownOne & KnownZero2) == (~KnownOne & NewMask)) 01013 return TLO.CombineTo(Op, Op.getOperand(1)); 01014 // If all of the potentially set bits on one side are known to be set on 01015 // the other side, just use the 'other' side. 01016 if ((NewMask & ~KnownZero & KnownOne2) == (~KnownZero & NewMask)) 01017 return TLO.CombineTo(Op, Op.getOperand(0)); 01018 if ((NewMask & ~KnownZero2 & KnownOne) == (~KnownZero2 & NewMask)) 01019 return TLO.CombineTo(Op, Op.getOperand(1)); 01020 // If the RHS is a constant, see if we can simplify it. 01021 if (TLO.ShrinkDemandedConstant(Op, NewMask)) 01022 return true; 01023 // If the operation can be done in a smaller type, do so. 01024 if (TLO.ShrinkDemandedOp(Op, BitWidth, NewMask, dl)) 01025 return true; 01026 01027 // Output known-0 bits are only known if clear in both the LHS & RHS. 01028 KnownZero &= KnownZero2; 01029 // Output known-1 are known to be set if set in either the LHS | RHS. 01030 KnownOne |= KnownOne2; 01031 break; 01032 case ISD::XOR: 01033 if (SimplifyDemandedBits(Op.getOperand(1), NewMask, KnownZero, 01034 KnownOne, TLO, Depth+1)) 01035 return true; 01036 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?"); 01037 if (SimplifyDemandedBits(Op.getOperand(0), NewMask, KnownZero2, 01038 KnownOne2, TLO, Depth+1)) 01039 return true; 01040 assert((KnownZero2 & KnownOne2) == 0 && "Bits known to be one AND zero?"); 01041 01042 // If all of the demanded bits are known zero on one side, return the other. 01043 // These bits cannot contribute to the result of the 'xor'. 01044 if ((KnownZero & NewMask) == NewMask) 01045 return TLO.CombineTo(Op, Op.getOperand(0)); 01046 if ((KnownZero2 & NewMask) == NewMask) 01047 return TLO.CombineTo(Op, Op.getOperand(1)); 01048 // If the operation can be done in a smaller type, do so. 01049 if (TLO.ShrinkDemandedOp(Op, BitWidth, NewMask, dl)) 01050 return true; 01051 01052 // If all of the unknown bits are known to be zero on one side or the other 01053 // (but not both) turn this into an *inclusive* or. 01054 // e.g. (A & C1)^(B & C2) -> (A & C1)|(B & C2) iff C1&C2 == 0 01055 if ((NewMask & ~KnownZero & ~KnownZero2) == 0) 01056 return TLO.CombineTo(Op, TLO.DAG.getNode(ISD::OR, dl, Op.getValueType(), 01057 Op.getOperand(0), 01058 Op.getOperand(1))); 01059 01060 // Output known-0 bits are known if clear or set in both the LHS & RHS. 01061 KnownZeroOut = (KnownZero & KnownZero2) | (KnownOne & KnownOne2); 01062 // Output known-1 are known to be set if set in only one of the LHS, RHS. 01063 KnownOneOut = (KnownZero & KnownOne2) | (KnownOne & KnownZero2); 01064 01065 // If all of the demanded bits on one side are known, and all of the set 01066 // bits on that side are also known to be set on the other side, turn this 01067 // into an AND, as we know the bits will be cleared. 01068 // e.g. (X | C1) ^ C2 --> (X | C1) & ~C2 iff (C1&C2) == C2 01069 if ((NewMask & (KnownZero|KnownOne)) == NewMask) { // all known 01070 if ((KnownOne & KnownOne2) == KnownOne) { 01071 EVT VT = Op.getValueType(); 01072 SDValue ANDC = TLO.DAG.getConstant(~KnownOne & NewMask, VT); 01073 return TLO.CombineTo(Op, TLO.DAG.getNode(ISD::AND, dl, VT, 01074 Op.getOperand(0), ANDC)); 01075 } 01076 } 01077 01078 // If the RHS is a constant, see if we can simplify it. 01079 // for XOR, we prefer to force bits to 1 if they will make a -1. 01080 // if we can't force bits, try to shrink constant 01081 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op.getOperand(1))) { 01082 APInt Expanded = C->getAPIntValue() | (~NewMask); 01083 // if we can expand it to have all bits set, do it 01084 if (Expanded.isAllOnesValue()) { 01085 if (Expanded != C->getAPIntValue()) { 01086 EVT VT = Op.getValueType(); 01087 SDValue New = TLO.DAG.getNode(Op.getOpcode(), dl,VT, Op.getOperand(0), 01088 TLO.DAG.getConstant(Expanded, VT)); 01089 return TLO.CombineTo(Op, New); 01090 } 01091 // if it already has all the bits set, nothing to change 01092 // but don't shrink either! 01093 } else if (TLO.ShrinkDemandedConstant(Op, NewMask)) { 01094 return true; 01095 } 01096 } 01097 01098 KnownZero = KnownZeroOut; 01099 KnownOne = KnownOneOut; 01100 break; 01101 case ISD::SELECT: 01102 if (SimplifyDemandedBits(Op.getOperand(2), NewMask, KnownZero, 01103 KnownOne, TLO, Depth+1)) 01104 return true; 01105 if (SimplifyDemandedBits(Op.getOperand(1), NewMask, KnownZero2, 01106 KnownOne2, TLO, Depth+1)) 01107 return true; 01108 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?"); 01109 assert((KnownZero2 & KnownOne2) == 0 && "Bits known to be one AND zero?"); 01110 01111 // If the operands are constants, see if we can simplify them. 01112 if (TLO.ShrinkDemandedConstant(Op, NewMask)) 01113 return true; 01114 01115 // Only known if known in both the LHS and RHS. 01116 KnownOne &= KnownOne2; 01117 KnownZero &= KnownZero2; 01118 break; 01119 case ISD::SELECT_CC: 01120 if (SimplifyDemandedBits(Op.getOperand(3), NewMask, KnownZero, 01121 KnownOne, TLO, Depth+1)) 01122 return true; 01123 if (SimplifyDemandedBits(Op.getOperand(2), NewMask, KnownZero2, 01124 KnownOne2, TLO, Depth+1)) 01125 return true; 01126 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?"); 01127 assert((KnownZero2 & KnownOne2) == 0 && "Bits known to be one AND zero?"); 01128 01129 // If the operands are constants, see if we can simplify them. 01130 if (TLO.ShrinkDemandedConstant(Op, NewMask)) 01131 return true; 01132 01133 // Only known if known in both the LHS and RHS. 01134 KnownOne &= KnownOne2; 01135 KnownZero &= KnownZero2; 01136 break; 01137 case ISD::SHL: 01138 if (ConstantSDNode *SA = dyn_cast<ConstantSDNode>(Op.getOperand(1))) { 01139 unsigned ShAmt = SA->getZExtValue(); 01140 SDValue InOp = Op.getOperand(0); 01141 01142 // If the shift count is an invalid immediate, don't do anything. 01143 if (ShAmt >= BitWidth) 01144 break; 01145 01146 // If this is ((X >>u C1) << ShAmt), see if we can simplify this into a 01147 // single shift. We can do this if the bottom bits (which are shifted 01148 // out) are never demanded. 01149 if (InOp.getOpcode() == ISD::SRL && 01150 isa<ConstantSDNode>(InOp.getOperand(1))) { 01151 if (ShAmt && (NewMask & APInt::getLowBitsSet(BitWidth, ShAmt)) == 0) { 01152 unsigned C1= cast<ConstantSDNode>(InOp.getOperand(1))->getZExtValue(); 01153 unsigned Opc = ISD::SHL; 01154 int Diff = ShAmt-C1; 01155 if (Diff < 0) { 01156 Diff = -Diff; 01157 Opc = ISD::SRL; 01158 } 01159 01160 SDValue NewSA = 01161 TLO.DAG.getConstant(Diff, Op.getOperand(1).getValueType()); 01162 EVT VT = Op.getValueType(); 01163 return TLO.CombineTo(Op, TLO.DAG.getNode(Opc, dl, VT, 01164 InOp.getOperand(0), NewSA)); 01165 } 01166 } 01167 01168 if (SimplifyDemandedBits(Op.getOperand(0), NewMask.lshr(ShAmt), 01169 KnownZero, KnownOne, TLO, Depth+1)) 01170 return true; 01171 KnownZero <<= SA->getZExtValue(); 01172 KnownOne <<= SA->getZExtValue(); 01173 // low bits known zero. 01174 KnownZero |= APInt::getLowBitsSet(BitWidth, SA->getZExtValue()); 01175 } 01176 break; 01177 case ISD::SRL: 01178 if (ConstantSDNode *SA = dyn_cast<ConstantSDNode>(Op.getOperand(1))) { 01179 EVT VT = Op.getValueType(); 01180 unsigned ShAmt = SA->getZExtValue(); 01181 unsigned VTSize = VT.getSizeInBits(); 01182 SDValue InOp = Op.getOperand(0); 01183 01184 // If the shift count is an invalid immediate, don't do anything. 01185 if (ShAmt >= BitWidth) 01186 break; 01187 01188 // If this is ((X << C1) >>u ShAmt), see if we can simplify this into a 01189 // single shift. We can do this if the top bits (which are shifted out) 01190 // are never demanded. 01191 if (InOp.getOpcode() == ISD::SHL && 01192 isa<ConstantSDNode>(InOp.getOperand(1))) { 01193 if (ShAmt && (NewMask & APInt::getHighBitsSet(VTSize, ShAmt)) == 0) { 01194 unsigned C1= cast<ConstantSDNode>(InOp.getOperand(1))->getZExtValue(); 01195 unsigned Opc = ISD::SRL; 01196 int Diff = ShAmt-C1; 01197 if (Diff < 0) { 01198 Diff = -Diff; 01199 Opc = ISD::SHL; 01200 } 01201 01202 SDValue NewSA = 01203 TLO.DAG.getConstant(Diff, Op.getOperand(1).getValueType()); 01204 return TLO.CombineTo(Op, TLO.DAG.getNode(Opc, dl, VT, 01205 InOp.getOperand(0), NewSA)); 01206 } 01207 } 01208 01209 // Compute the new bits that are at the top now. 01210 if (SimplifyDemandedBits(InOp, (NewMask << ShAmt), 01211 KnownZero, KnownOne, TLO, Depth+1)) 01212 return true; 01213 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?"); 01214 KnownZero = KnownZero.lshr(ShAmt); 01215 KnownOne = KnownOne.lshr(ShAmt); 01216 01217 APInt HighBits = APInt::getHighBitsSet(BitWidth, ShAmt); 01218 KnownZero |= HighBits; // High bits known zero. 01219 } 01220 break; 01221 case ISD::SRA: 01222 // If this is an arithmetic shift right and only the low-bit is set, we can 01223 // always convert this into a logical shr, even if the shift amount is 01224 // variable. The low bit of the shift cannot be an input sign bit unless 01225 // the shift amount is >= the size of the datatype, which is undefined. 01226 if (DemandedMask == 1) 01227 return TLO.CombineTo(Op, TLO.DAG.getNode(ISD::SRL, dl, Op.getValueType(), 01228 Op.getOperand(0), Op.getOperand(1))); 01229 01230 if (ConstantSDNode *SA = dyn_cast<ConstantSDNode>(Op.getOperand(1))) { 01231 EVT VT = Op.getValueType(); 01232 unsigned ShAmt = SA->getZExtValue(); 01233 01234 // If the shift count is an invalid immediate, don't do anything. 01235 if (ShAmt >= BitWidth) 01236 break; 01237 01238 APInt InDemandedMask = (NewMask << ShAmt); 01239 01240 // If any of the demanded bits are produced by the sign extension, we also 01241 // demand the input sign bit. 01242 APInt HighBits = APInt::getHighBitsSet(BitWidth, ShAmt); 01243 if (HighBits.intersects(NewMask)) 01244 InDemandedMask |= APInt::getSignBit(VT.getSizeInBits()); 01245 01246 if (SimplifyDemandedBits(Op.getOperand(0), InDemandedMask, 01247 KnownZero, KnownOne, TLO, Depth+1)) 01248 return true; 01249 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?"); 01250 KnownZero = KnownZero.lshr(ShAmt); 01251 KnownOne = KnownOne.lshr(ShAmt); 01252 01253 // Handle the sign bit, adjusted to where it is now in the mask. 01254 APInt SignBit = APInt::getSignBit(BitWidth).lshr(ShAmt); 01255 01256 // If the input sign bit is known to be zero, or if none of the top bits 01257 // are demanded, turn this into an unsigned shift right. 01258 if (KnownZero.intersects(SignBit) || (HighBits & ~NewMask) == HighBits) { 01259 return TLO.CombineTo(Op, TLO.DAG.getNode(ISD::SRL, dl, VT, 01260 Op.getOperand(0), 01261 Op.getOperand(1))); 01262 } else if (KnownOne.intersects(SignBit)) { // New bits are known one. 01263 KnownOne |= HighBits; 01264 } 01265 } 01266 break; 01267 case ISD::SIGN_EXTEND_INREG: { 01268 EVT EVT = cast<VTSDNode>(Op.getOperand(1))->getVT(); 01269 01270 // Sign extension. Compute the demanded bits in the result that are not 01271 // present in the input. 01272 APInt NewBits = APInt::getHighBitsSet(BitWidth, 01273 BitWidth - EVT.getSizeInBits()) & 01274 NewMask; 01275 01276 // If none of the extended bits are demanded, eliminate the sextinreg. 01277 if (NewBits == 0) 01278 return TLO.CombineTo(Op, Op.getOperand(0)); 01279 01280 APInt InSignBit = APInt::getSignBit(EVT.getSizeInBits()); 01281 InSignBit.zext(BitWidth); 01282 APInt InputDemandedBits = APInt::getLowBitsSet(BitWidth, 01283 EVT.getSizeInBits()) & 01284 NewMask; 01285 01286 // Since the sign extended bits are demanded, we know that the sign 01287 // bit is demanded. 01288 InputDemandedBits |= InSignBit; 01289 01290 if (SimplifyDemandedBits(Op.getOperand(0), InputDemandedBits, 01291 KnownZero, KnownOne, TLO, Depth+1)) 01292 return true; 01293 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?"); 01294 01295 // If the sign bit of the input is known set or clear, then we know the 01296 // top bits of the result. 01297 01298 // If the input sign bit is known zero, convert this into a zero extension. 01299 if (KnownZero.intersects(InSignBit)) 01300 return TLO.CombineTo(Op, 01301 TLO.DAG.getZeroExtendInReg(Op.getOperand(0),dl,EVT)); 01302 01303 if (KnownOne.intersects(InSignBit)) { // Input sign bit known set 01304 KnownOne |= NewBits; 01305 KnownZero &= ~NewBits; 01306 } else { // Input sign bit unknown 01307 KnownZero &= ~NewBits; 01308 KnownOne &= ~NewBits; 01309 } 01310 break; 01311 } 01312 case ISD::ZERO_EXTEND: { 01313 unsigned OperandBitWidth = Op.getOperand(0).getValueSizeInBits(); 01314 APInt InMask = NewMask; 01315 InMask.trunc(OperandBitWidth); 01316 01317 // If none of the top bits are demanded, convert this into an any_extend. 01318 APInt NewBits = 01319 APInt::getHighBitsSet(BitWidth, BitWidth - OperandBitWidth) & NewMask; 01320 if (!NewBits.intersects(NewMask)) 01321 return TLO.CombineTo(Op, TLO.DAG.getNode(ISD::ANY_EXTEND, dl, 01322 Op.getValueType(), 01323 Op.getOperand(0))); 01324 01325 if (SimplifyDemandedBits(Op.getOperand(0), InMask, 01326 KnownZero, KnownOne, TLO, Depth+1)) 01327 return true; 01328 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?"); 01329 KnownZero.zext(BitWidth); 01330 KnownOne.zext(BitWidth); 01331 KnownZero |= NewBits; 01332 break; 01333 } 01334 case ISD::SIGN_EXTEND: { 01335 EVT InVT = Op.getOperand(0).getValueType(); 01336 unsigned InBits = InVT.getSizeInBits(); 01337 APInt InMask = APInt::getLowBitsSet(BitWidth, InBits); 01338 APInt InSignBit = APInt::getBitsSet(BitWidth, InBits - 1, InBits); 01339 APInt NewBits = ~InMask & NewMask; 01340 01341 // If none of the top bits are demanded, convert this into an any_extend. 01342 if (NewBits == 0) 01343 return TLO.CombineTo(Op,TLO.DAG.getNode(ISD::ANY_EXTEND, dl, 01344 Op.getValueType(), 01345 Op.getOperand(0))); 01346 01347 // Since some of the sign extended bits are demanded, we know that the sign 01348 // bit is demanded. 01349 APInt InDemandedBits = InMask & NewMask; 01350 InDemandedBits |= InSignBit; 01351 InDemandedBits.trunc(InBits); 01352 01353 if (SimplifyDemandedBits(Op.getOperand(0), InDemandedBits, KnownZero, 01354 KnownOne, TLO, Depth+1)) 01355 return true; 01356 KnownZero.zext(BitWidth); 01357 KnownOne.zext(BitWidth); 01358 01359 // If the sign bit is known zero, convert this to a zero extend. 01360 if (KnownZero.intersects(InSignBit)) 01361 return TLO.CombineTo(Op, TLO.DAG.getNode(ISD::ZERO_EXTEND, dl, 01362 Op.getValueType(), 01363 Op.getOperand(0))); 01364 01365 // If the sign bit is known one, the top bits match. 01366 if (KnownOne.intersects(InSignBit)) { 01367 KnownOne |= NewBits; 01368 KnownZero &= ~NewBits; 01369 } else { // Otherwise, top bits aren't known. 01370 KnownOne &= ~NewBits; 01371 KnownZero &= ~NewBits; 01372 } 01373 break; 01374 } 01375 case ISD::ANY_EXTEND: { 01376 unsigned OperandBitWidth = Op.getOperand(0).getValueSizeInBits(); 01377 APInt InMask = NewMask; 01378 InMask.trunc(OperandBitWidth); 01379 if (SimplifyDemandedBits(Op.getOperand(0), InMask, 01380 KnownZero, KnownOne, TLO, Depth+1)) 01381 return true; 01382 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?"); 01383 KnownZero.zext(BitWidth); 01384 KnownOne.zext(BitWidth); 01385 break; 01386 } 01387 case ISD::TRUNCATE: { 01388 // Simplify the input, using demanded bit information, and compute the known 01389 // zero/one bits live out. 01390 APInt TruncMask = NewMask; 01391 TruncMask.zext(Op.getOperand(0).getValueSizeInBits()); 01392 if (SimplifyDemandedBits(Op.getOperand(0), TruncMask, 01393 KnownZero, KnownOne, TLO, Depth+1)) 01394 return true; 01395 KnownZero.trunc(BitWidth); 01396 KnownOne.trunc(BitWidth); 01397 01398 // If the input is only used by this truncate, see if we can shrink it based 01399 // on the known demanded bits. 01400 if (Op.getOperand(0).getNode()->hasOneUse()) { 01401 SDValue In = Op.getOperand(0); 01402 unsigned InBitWidth = In.getValueSizeInBits(); 01403 switch (In.getOpcode()) { 01404 default: break; 01405 case ISD::SRL: 01406 // Shrink SRL by a constant if none of the high bits shifted in are 01407 // demanded. 01408 if (ConstantSDNode *ShAmt = dyn_cast<ConstantSDNode>(In.getOperand(1))){ 01409 APInt HighBits = APInt::getHighBitsSet(InBitWidth, 01410 InBitWidth - BitWidth); 01411 HighBits = HighBits.lshr(ShAmt->getZExtValue()); 01412 HighBits.trunc(BitWidth); 01413 01414 if (ShAmt->getZExtValue() < BitWidth && !(HighBits & NewMask)) { 01415 // None of the shifted in bits are needed. Add a truncate of the 01416 // shift input, then shift it. 01417 SDValue NewTrunc = TLO.DAG.getNode(ISD::TRUNCATE, dl, 01418 Op.getValueType(), 01419 In.getOperand(0)); 01420 return TLO.CombineTo(Op, TLO.DAG.getNode(ISD::SRL, dl, 01421 Op.getValueType(), 01422 NewTrunc, 01423 In.getOperand(1))); 01424 } 01425 } 01426 break; 01427 } 01428 } 01429 01430 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?"); 01431 break; 01432 } 01433 case ISD::AssertZext: { 01434 EVT VT = cast<VTSDNode>(Op.getOperand(1))->getVT(); 01435 APInt InMask = APInt::getLowBitsSet(BitWidth, 01436 VT.getSizeInBits()); 01437 if (SimplifyDemandedBits(Op.getOperand(0), InMask & NewMask, 01438 KnownZero, KnownOne, TLO, Depth+1)) 01439 return true; 01440 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?"); 01441 KnownZero |= ~InMask & NewMask; 01442 break; 01443 } 01444 case ISD::BIT_CONVERT: 01445 #if 0 01446 // If this is an FP->Int bitcast and if the sign bit is the only thing that 01447 // is demanded, turn this into a FGETSIGN. 01448 if (NewMask == EVT::getIntegerVTSignBit(Op.getValueType()) && 01449 MVT::isFloatingPoint(Op.getOperand(0).getValueType()) && 01450 !MVT::isVector(Op.getOperand(0).getValueType())) { 01451 // Only do this xform if FGETSIGN is valid or if before legalize. 01452 if (!TLO.AfterLegalize || 01453 isOperationLegal(ISD::FGETSIGN, Op.getValueType())) { 01454 // Make a FGETSIGN + SHL to move the sign bit into the appropriate 01455 // place. We expect the SHL to be eliminated by other optimizations. 01456 SDValue Sign = TLO.DAG.getNode(ISD::FGETSIGN, Op.getValueType(), 01457 Op.getOperand(0)); 01458 unsigned ShVal = Op.getValueType().getSizeInBits()-1; 01459 SDValue ShAmt = TLO.DAG.getConstant(ShVal, getShiftAmountTy()); 01460 return TLO.CombineTo(Op, TLO.DAG.getNode(ISD::SHL, Op.getValueType(), 01461 Sign, ShAmt)); 01462 } 01463 } 01464 #endif 01465 break; 01466 case ISD::ADD: 01467 case ISD::MUL: 01468 case ISD::SUB: { 01469 // Add, Sub, and Mul don't demand any bits in positions beyond that 01470 // of the highest bit demanded of them. 01471 APInt LoMask = APInt::getLowBitsSet(BitWidth, 01472 BitWidth - NewMask.countLeadingZeros()); 01473 if (SimplifyDemandedBits(Op.getOperand(0), LoMask, KnownZero2, 01474 KnownOne2, TLO, Depth+1)) 01475 return true; 01476 if (SimplifyDemandedBits(Op.getOperand(1), LoMask, KnownZero2, 01477 KnownOne2, TLO, Depth+1)) 01478 return true; 01479 // See if the operation should be performed at a smaller bit width. 01480 if (TLO.ShrinkDemandedOp(Op, BitWidth, NewMask, dl)) 01481 return true; 01482 } 01483 // FALL THROUGH 01484 default: 01485 // Just use ComputeMaskedBits to compute output bits. 01486 TLO.DAG.ComputeMaskedBits(Op, NewMask, KnownZero, KnownOne, Depth); 01487 break; 01488 } 01489 01490 // If we know the value of all of the demanded bits, return this as a 01491 // constant. 01492 if ((NewMask & (KnownZero|KnownOne)) == NewMask) 01493 return TLO.CombineTo(Op, TLO.DAG.getConstant(KnownOne, Op.getValueType())); 01494 01495 return false; 01496 } 01497 01498 /// computeMaskedBitsForTargetNode - Determine which of the bits specified 01499 /// in Mask are known to be either zero or one and return them in the 01500 /// KnownZero/KnownOne bitsets. 01501 void TargetLowering::computeMaskedBitsForTargetNode(const SDValue Op, 01502 const APInt &Mask, 01503 APInt &KnownZero, 01504 APInt &KnownOne, 01505 const SelectionDAG &DAG, 01506 unsigned Depth) const { 01507 assert((Op.getOpcode() >= ISD::BUILTIN_OP_END || 01508 Op.getOpcode() == ISD::INTRINSIC_WO_CHAIN || 01509 Op.getOpcode() == ISD::INTRINSIC_W_CHAIN || 01510 Op.getOpcode() == ISD::INTRINSIC_VOID) && 01511 "Should use MaskedValueIsZero if you don't know whether Op" 01512 " is a target node!"); 01513 KnownZero = KnownOne = APInt(Mask.getBitWidth(), 0); 01514 } 01515 01516 /// ComputeNumSignBitsForTargetNode - This method can be implemented by 01517 /// targets that want to expose additional information about sign bits to the 01518 /// DAG Combiner. 01519 unsigned TargetLowering::ComputeNumSignBitsForTargetNode(SDValue Op, 01520 unsigned Depth) const { 01521 assert((Op.getOpcode() >= ISD::BUILTIN_OP_END || 01522 Op.getOpcode() == ISD::INTRINSIC_WO_CHAIN || 01523 Op.getOpcode() == ISD::INTRINSIC_W_CHAIN || 01524 Op.getOpcode() == ISD::INTRINSIC_VOID) && 01525 "Should use ComputeNumSignBits if you don't know whether Op" 01526 " is a target node!"); 01527 return 1; 01528 } 01529 01530 /// ValueHasExactlyOneBitSet - Test if the given value is known to have exactly 01531 /// one bit set. This differs from ComputeMaskedBits in that it doesn't need to 01532 /// determine which bit is set. 01533 /// 01534 static bool ValueHasExactlyOneBitSet(SDValue Val, const SelectionDAG &DAG) { 01535 // A left-shift of a constant one will have exactly one bit set, because 01536 // shifting the bit off the end is undefined. 01537 if (Val.getOpcode() == ISD::SHL) 01538 if (ConstantSDNode *C = 01539 dyn_cast<ConstantSDNode>(Val.getNode()->getOperand(0))) 01540 if (C->getAPIntValue() == 1) 01541 return true; 01542 01543 // Similarly, a right-shift of a constant sign-bit will have exactly 01544 // one bit set. 01545 if (Val.getOpcode() == ISD::SRL) 01546 if (ConstantSDNode *C = 01547 dyn_cast<ConstantSDNode>(Val.getNode()->getOperand(0))) 01548 if (C->getAPIntValue().isSignBit()) 01549 return true; 01550 01551 // More could be done here, though the above checks are enough 01552 // to handle some common cases. 01553 01554 // Fall back to ComputeMaskedBits to catch other known cases. 01555 EVT OpVT = Val.getValueType(); 01556 unsigned BitWidth = OpVT.getSizeInBits(); 01557 APInt Mask = APInt::getAllOnesValue(BitWidth); 01558 APInt KnownZero, KnownOne; 01559 DAG.ComputeMaskedBits(Val, Mask, KnownZero, KnownOne); 01560 return (KnownZero.countPopulation() == BitWidth - 1) && 01561 (KnownOne.countPopulation() == 1); 01562 } 01563 01564 /// SimplifySetCC - Try to simplify a setcc built with the specified operands 01565 /// and cc. If it is unable to simplify it, return a null SDValue. 01566 SDValue 01567 TargetLowering::SimplifySetCC(EVT VT, SDValue N0, SDValue N1, 01568 ISD::CondCode Cond, bool foldBooleans, 01569 DAGCombinerInfo &DCI, DebugLoc dl) const { 01570 SelectionDAG &DAG = DCI.DAG; 01571 LLVMContext &Context = *DAG.getContext(); 01572 01573 // These setcc operations always fold. 01574 switch (Cond) { 01575 default: break; 01576 case ISD::SETFALSE: 01577 case ISD::SETFALSE2: return DAG.getConstant(0, VT); 01578 case ISD::SETTRUE: 01579 case ISD::SETTRUE2: return DAG.getConstant(1, VT); 01580 } 01581 01582 if (isa<ConstantSDNode>(N0.getNode())) { 01583 // Ensure that the constant occurs on the RHS, and fold constant 01584 // comparisons. 01585 return DAG.getSetCC(dl, VT, N1, N0, ISD::getSetCCSwappedOperands(Cond)); 01586 } 01587 01588 if (ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1.getNode())) { 01589 const APInt &C1 = N1C->getAPIntValue(); 01590 01591 // If the LHS is '(srl (ctlz x), 5)', the RHS is 0/1, and this is an 01592 // equality comparison, then we're just comparing whether X itself is 01593 // zero. 01594 if (N0.getOpcode() == ISD::SRL && (C1 == 0 || C1 == 1) && 01595 N0.getOperand(0).getOpcode() == ISD::CTLZ && 01596 N0.getOperand(1).getOpcode() == ISD::Constant) { 01597 unsigned ShAmt = cast<ConstantSDNode>(N0.getOperand(1))->getZExtValue(); 01598 if ((Cond == ISD::SETEQ || Cond == ISD::SETNE) && 01599 ShAmt == Log2_32(N0.getValueType().getSizeInBits())) { 01600 if ((C1 == 0) == (Cond == ISD::SETEQ)) { 01601 // (srl (ctlz x), 5) == 0 -> X != 0 01602 // (srl (ctlz x), 5) != 1 -> X != 0 01603 Cond = ISD::SETNE; 01604 } else { 01605 // (srl (ctlz x), 5) != 0 -> X == 0 01606 // (srl (ctlz x), 5) == 1 -> X == 0 01607 Cond = ISD::SETEQ; 01608 } 01609 SDValue Zero = DAG.getConstant(0, N0.getValueType()); 01610 return DAG.getSetCC(dl, VT, N0.getOperand(0).getOperand(0), 01611 Zero, Cond); 01612 } 01613 } 01614 01615 // If the LHS is '(and load, const)', the RHS is 0, 01616 // the test is for equality or unsigned, and all 1 bits of the const are 01617 // in the same partial word, see if we can shorten the load. 01618 if (DCI.isBeforeLegalize() && 01619 N0.getOpcode() == ISD::AND && C1 == 0 && 01620 N0.getNode()->hasOneUse() && 01621 isa<LoadSDNode>(N0.getOperand(0)) && 01622 N0.getOperand(0).getNode()->hasOneUse() && 01623 isa<ConstantSDNode>(N0.getOperand(1))) { 01624 LoadSDNode *Lod = cast<LoadSDNode>(N0.getOperand(0)); 01625 uint64_t bestMask = 0; 01626 unsigned bestWidth = 0, bestOffset = 0; 01627 if (!Lod->isVolatile() && Lod->isUnindexed() && 01628 // FIXME: This uses getZExtValue() below so it only works on i64 and 01629 // below. 01630 N0.getValueType().getSizeInBits() <= 64) { 01631 unsigned origWidth = N0.getValueType().getSizeInBits(); 01632 // We can narrow (e.g.) 16-bit extending loads on 32-bit target to 01633 // 8 bits, but have to be careful... 01634 if (Lod->getExtensionType() != ISD::NON_EXTLOAD) 01635 origWidth = Lod->getMemoryVT().getSizeInBits(); 01636 uint64_t Mask =cast<ConstantSDNode>(N0.getOperand(1))->getZExtValue(); 01637 for (unsigned width = origWidth / 2; width>=8; width /= 2) { 01638 uint64_t newMask = (1ULL << width) - 1; 01639 for (unsigned offset=0; offset<origWidth/width; offset++) { 01640 if ((newMask & Mask) == Mask) { 01641 if (!TD->isLittleEndian()) 01642 bestOffset = (origWidth/width - offset - 1) * (width/8); 01643 else 01644 bestOffset = (uint64_t)offset * (width/8); 01645 bestMask = Mask >> (offset * (width/8) * 8); 01646 bestWidth = width; 01647 break; 01648 } 01649 newMask = newMask << width; 01650 } 01651 } 01652 } 01653 if (bestWidth) { 01654 EVT newVT = EVT::getIntegerVT(Context, bestWidth); 01655 if (newVT.isRound()) { 01656 EVT PtrType = Lod->getOperand(1).getValueType(); 01657 SDValue Ptr = Lod->getBasePtr(); 01658 if (bestOffset != 0) 01659 Ptr = DAG.getNode(ISD::ADD, dl, PtrType, Lod->getBasePtr(), 01660 DAG.getConstant(bestOffset, PtrType)); 01661 unsigned NewAlign = MinAlign(Lod->getAlignment(), bestOffset); 01662 SDValue NewLoad = DAG.getLoad(newVT, dl, Lod->getChain(), Ptr, 01663 Lod->getSrcValue(), 01664 Lod->getSrcValueOffset() + bestOffset, 01665 false, NewAlign); 01666 return DAG.getSetCC(dl, VT, 01667 DAG.getNode(ISD::AND, dl, newVT, NewLoad, 01668 DAG.getConstant(bestMask, newVT)), 01669 DAG.getConstant(0LL, newVT), Cond); 01670 } 01671 } 01672 } 01673 01674 // If the LHS is a ZERO_EXTEND, perform the comparison on the input. 01675 if (N0.getOpcode() == ISD::ZERO_EXTEND) { 01676 unsigned InSize = N0.getOperand(0).getValueType().getSizeInBits(); 01677 01678 // If the comparison constant has bits in the upper part, the 01679 // zero-extended value could never match. 01680 if (C1.intersects(APInt::getHighBitsSet(C1.getBitWidth(), 01681 C1.getBitWidth() - InSize))) { 01682 switch (Cond) { 01683 case ISD::SETUGT: 01684 case ISD::SETUGE: 01685 case ISD::SETEQ: return DAG.getConstant(0, VT); 01686 case ISD::SETULT: 01687 case ISD::SETULE: 01688 case ISD::SETNE: return DAG.getConstant(1, VT); 01689 case ISD::SETGT: 01690 case ISD::SETGE: 01691 // True if the sign bit of C1 is set. 01692 return DAG.getConstant(C1.isNegative(), VT); 01693 case ISD::SETLT: 01694 case ISD::SETLE: 01695 // True if the sign bit of C1 isn't set. 01696 return DAG.getConstant(C1.isNonNegative(), VT); 01697 default: 01698 break; 01699 } 01700 } 01701 01702 // Otherwise, we can perform the comparison with the low bits. 01703 switch (Cond) { 01704 case ISD::SETEQ: 01705 case ISD::SETNE: 01706 case ISD::SETUGT: 01707 case ISD::SETUGE: 01708 case ISD::SETULT: 01709 case ISD::SETULE: { 01710 EVT newVT = N0.getOperand(0).getValueType(); 01711 if (DCI.isBeforeLegalizeOps() || 01712 (isOperationLegal(ISD::SETCC, newVT) && 01713 getCondCodeAction(Cond, newVT)==Legal)) 01714 return DAG.getSetCC(dl, VT, N0.getOperand(0), 01715 DAG.getConstant(APInt(C1).trunc(InSize), newVT), 01716 Cond); 01717 break; 01718 } 01719 default: 01720 break; // todo, be more careful with signed comparisons 01721 } 01722 } else if (N0.getOpcode() == ISD::SIGN_EXTEND_INREG && 01723 (Cond == ISD::SETEQ || Cond == ISD::SETNE)) { 01724 EVT ExtSrcTy = cast<VTSDNode>(N0.getOperand(1))->getVT(); 01725 unsigned ExtSrcTyBits = ExtSrcTy.getSizeInBits(); 01726 EVT ExtDstTy = N0.getValueType(); 01727 unsigned ExtDstTyBits = ExtDstTy.getSizeInBits(); 01728 01729 // If the extended part has any inconsistent bits, it cannot ever 01730 // compare equal. In other words, they have to be all ones or all 01731 // zeros. 01732 APInt ExtBits = 01733 APInt::getHighBitsSet(ExtDstTyBits, ExtDstTyBits - ExtSrcTyBits); 01734 if ((C1 & ExtBits) != 0 && (C1 & ExtBits) != ExtBits) 01735 return DAG.getConstant(Cond == ISD::SETNE, VT); 01736 01737 SDValue ZextOp; 01738 EVT Op0Ty = N0.getOperand(0).getValueType(); 01739 if (Op0Ty == ExtSrcTy) { 01740 ZextOp = N0.getOperand(0); 01741 } else { 01742 APInt Imm = APInt::getLowBitsSet(ExtDstTyBits, ExtSrcTyBits); 01743 ZextOp = DAG.getNode(ISD::AND, dl, Op0Ty, N0.getOperand(0), 01744 DAG.getConstant(Imm, Op0Ty)); 01745 } 01746 if (!DCI.isCalledByLegalizer()) 01747 DCI.AddToWorklist(ZextOp.getNode()); 01748 // Otherwise, make this a use of a zext. 01749 return DAG.getSetCC(dl, VT, ZextOp, 01750 DAG.getConstant(C1 & APInt::getLowBitsSet( 01751 ExtDstTyBits, 01752 ExtSrcTyBits), 01753 ExtDstTy), 01754 Cond); 01755 } else if ((N1C->isNullValue() || N1C->getAPIntValue() == 1) && 01756 (Cond == ISD::SETEQ || Cond == ISD::SETNE)) { 01757 01758 // SETCC (SETCC), [0|1], [EQ|NE] -> SETCC 01759 if (N0.getOpcode() == ISD::SETCC) { 01760 bool TrueWhenTrue = (Cond == ISD::SETEQ) ^ (N1C->getZExtValue() != 1); 01761 if (TrueWhenTrue) 01762 return N0; 01763 01764 // Invert the condition. 01765 ISD::CondCode CC = cast<CondCodeSDNode>(N0.getOperand(2))->get(); 01766 CC = ISD::getSetCCInverse(CC, 01767 N0.getOperand(0).getValueType().isInteger()); 01768 return DAG.getSetCC(dl, VT, N0.getOperand(0), N0.getOperand(1), CC); 01769 } 01770 01771 if ((N0.getOpcode() == ISD::XOR || 01772 (N0.getOpcode() == ISD::AND && 01773 N0.getOperand(0).getOpcode() == ISD::XOR && 01774 N0.getOperand(1) == N0.getOperand(0).getOperand(1))) && 01775 isa<ConstantSDNode>(N0.getOperand(1)) && 01776 cast<ConstantSDNode>(N0.getOperand(1))->getAPIntValue() == 1) { 01777 // If this is (X^1) == 0/1, swap the RHS and eliminate the xor. We 01778 // can only do this if the top bits are known zero. 01779 unsigned BitWidth = N0.getValueSizeInBits(); 01780 if (DAG.MaskedValueIsZero(N0, 01781 APInt::getHighBitsSet(BitWidth, 01782 BitWidth-1))) { 01783 // Okay, get the un-inverted input value. 01784 SDValue Val; 01785 if (N0.getOpcode() == ISD::XOR) 01786 Val = N0.getOperand(0); 01787 else { 01788 assert(N0.getOpcode() == ISD::AND && 01789 N0.getOperand(0).getOpcode() == ISD::XOR); 01790 // ((X^1)&1)^1 -> X & 1 01791 Val = DAG.getNode(ISD::AND, dl, N0.getValueType(), 01792 N0.getOperand(0).getOperand(0), 01793 N0.getOperand(1)); 01794 } 01795 return DAG.getSetCC(dl, VT, Val, N1, 01796 Cond == ISD::SETEQ ? ISD::SETNE : ISD::SETEQ); 01797 } 01798 } 01799 } 01800 01801 APInt MinVal, MaxVal; 01802 unsigned OperandBitSize = N1C->getValueType(0).getSizeInBits(); 01803 if (ISD::isSignedIntSetCC(Cond)) { 01804 MinVal = APInt::getSignedMinValue(OperandBitSize); 01805 MaxVal = APInt::getSignedMaxValue(OperandBitSize); 01806 } else { 01807 MinVal = APInt::getMinValue(OperandBitSize); 01808 MaxVal = APInt::getMaxValue(OperandBitSize); 01809 } 01810 01811 // Canonicalize GE/LE comparisons to use GT/LT comparisons. 01812 if (Cond == ISD::SETGE || Cond == ISD::SETUGE) { 01813 if (C1 == MinVal) return DAG.getConstant(1, VT); // X >= MIN --> true 01814 // X >= C0 --> X > (C0-1) 01815 return DAG.getSetCC(dl, VT, N0, 01816 DAG.getConstant(C1-1, N1.getValueType()), 01817 (Cond == ISD::SETGE) ? ISD::SETGT : ISD::SETUGT); 01818 } 01819 01820 if (Cond == ISD::SETLE || Cond == ISD::SETULE) { 01821 if (C1 == MaxVal) return DAG.getConstant(1, VT); // X <= MAX --> true 01822 // X <= C0 --> X < (C0+1) 01823 return DAG.getSetCC(dl, VT, N0, 01824 DAG.getConstant(C1+1, N1.getValueType()), 01825 (Cond == ISD::SETLE) ? ISD::SETLT : ISD::SETULT); 01826 } 01827 01828 if ((Cond == ISD::SETLT || Cond == ISD::SETULT) && C1 == MinVal) 01829 return DAG.getConstant(0, VT); // X < MIN --> false 01830 if ((Cond == ISD::SETGE || Cond == ISD::SETUGE) && C1 == MinVal) 01831 return DAG.getConstant(1, VT); // X >= MIN --> true 01832 if ((Cond == ISD::SETGT || Cond == ISD::SETUGT) && C1 == MaxVal) 01833 return DAG.getConstant(0, VT); // X > MAX --> false 01834 if ((Cond == ISD::SETLE || Cond == ISD::SETULE) && C1 == MaxVal) 01835 return DAG.getConstant(1, VT); // X <= MAX --> true 01836 01837 // Canonicalize setgt X, Min --> setne X, Min 01838 if ((Cond == ISD::SETGT || Cond == ISD::SETUGT) && C1 == MinVal) 01839 return DAG.getSetCC(dl, VT, N0, N1, ISD::SETNE); 01840 // Canonicalize setlt X, Max --> setne X, Max 01841 if ((Cond == ISD::SETLT || Cond == ISD::SETULT) && C1 == MaxVal) 01842 return DAG.getSetCC(dl, VT, N0, N1, ISD::SETNE); 01843 01844 // If we have setult X, 1, turn it into seteq X, 0 01845 if ((Cond == ISD::SETLT || Cond == ISD::SETULT) && C1 == MinVal+1) 01846 return DAG.getSetCC(dl, VT, N0, 01847 DAG.getConstant(MinVal, N0.getValueType()), 01848 ISD::SETEQ); 01849 // If we have setugt X, Max-1, turn it into seteq X, Max 01850 else if ((Cond == ISD::SETGT || Cond == ISD::SETUGT) && C1 == MaxVal-1) 01851 return DAG.getSetCC(dl, VT, N0, 01852 DAG.getConstant(MaxVal, N0.getValueType()), 01853 ISD::SETEQ); 01854 01855 // If we have "setcc X, C0", check to see if we can shrink the immediate 01856 // by changing cc. 01857 01858 // SETUGT X, SINTMAX -> SETLT X, 0 01859 if (Cond == ISD::SETUGT && 01860 C1 == APInt::getSignedMaxValue(OperandBitSize)) 01861 return DAG.getSetCC(dl, VT, N0, 01862 DAG.getConstant(0, N1.getValueType()), 01863 ISD::SETLT); 01864 01865 // SETULT X, SINTMIN -> SETGT X, -1 01866 if (Cond == ISD::SETULT && 01867 C1 == APInt::getSignedMinValue(OperandBitSize)) { 01868 SDValue ConstMinusOne = 01869 DAG.getConstant(APInt::getAllOnesValue(OperandBitSize), 01870 N1.getValueType()); 01871 return DAG.getSetCC(dl, VT, N0, ConstMinusOne, ISD::SETGT); 01872 } 01873 01874 // Fold bit comparisons when we can. 01875 if ((Cond == ISD::SETEQ || Cond == ISD::SETNE) && 01876 VT == N0.getValueType() && N0.getOpcode() == ISD::AND) 01877 if (ConstantSDNode *AndRHS = 01878 dyn_cast<ConstantSDNode>(N0.getOperand(1))) { 01879 EVT ShiftTy = DCI.isBeforeLegalize() ? 01880 getPointerTy() : getShiftAmountTy(); 01881 if (Cond == ISD::SETNE && C1 == 0) {// (X & 8) != 0 --> (X & 8) >> 3 01882 // Perform the xform if the AND RHS is a single bit. 01883 if (isPowerOf2_64(AndRHS->getZExtValue())) { 01884 return DAG.getNode(ISD::SRL, dl, VT, N0, 01885 DAG.getConstant(Log2_64(AndRHS->getZExtValue()), 01886 ShiftTy)); 01887 } 01888 } else if (Cond == ISD::SETEQ && C1 == AndRHS->getZExtValue()) { 01889 // (X & 8) == 8 --> (X & 8) >> 3 01890 // Perform the xform if C1 is a single bit. 01891 if (C1.isPowerOf2()) { 01892 return DAG.getNode(ISD::SRL, dl, VT, N0, 01893 DAG.getConstant(C1.logBase2(), ShiftTy)); 01894 } 01895 } 01896 } 01897 } 01898 01899 if (isa<ConstantFPSDNode>(N0.getNode())) { 01900 // Constant fold or commute setcc. 01901 SDValue O = DAG.FoldSetCC(VT, N0, N1, Cond, dl); 01902 if (O.getNode()) return O; 01903 } else if (ConstantFPSDNode *CFP = dyn_cast<ConstantFPSDNode>(N1.getNode())) { 01904 // If the RHS of an FP comparison is a constant, simplify it away in 01905 // some cases. 01906 if (CFP->getValueAPF().isNaN()) { 01907 // If an operand is known to be a nan, we can fold it. 01908 switch (ISD::getUnorderedFlavor(Cond)) { 01909 default: llvm_unreachable("Unknown flavor!"); 01910 case 0: // Known false. 01911 return DAG.getConstant(0, VT); 01912 case 1: // Known true. 01913 return DAG.getConstant(1, VT); 01914 case 2: // Undefined. 01915 return DAG.getUNDEF(VT); 01916 } 01917 } 01918 01919 // Otherwise, we know the RHS is not a NaN. Simplify the node to drop the 01920 // constant if knowing that the operand is non-nan is enough. We prefer to 01921 // have SETO(x,x) instead of SETO(x, 0.0) because this avoids having to 01922 // materialize 0.0. 01923 if (Cond == ISD::SETO || Cond == ISD::SETUO) 01924 return DAG.getSetCC(dl, VT, N0, N0, Cond); 01925 } 01926 01927 if (N0 == N1) { 01928 // We can always fold X == X for integer setcc's. 01929 if (N0.getValueType().isInteger()) 01930 return DAG.getConstant(ISD::isTrueWhenEqual(Cond), VT); 01931 unsigned UOF = ISD::getUnorderedFlavor(Cond); 01932 if (UOF == 2) // FP operators that are undefined on NaNs. 01933 return DAG.getConstant(ISD::isTrueWhenEqual(Cond), VT); 01934 if (UOF == unsigned(ISD::isTrueWhenEqual(Cond))) 01935 return DAG.getConstant(UOF, VT); 01936 // Otherwise, we can't fold it. However, we can simplify it to SETUO/SETO 01937 // if it is not already. 01938 ISD::CondCode NewCond = UOF == 0 ? ISD::SETO : ISD::SETUO; 01939 if (NewCond != Cond) 01940 return DAG.getSetCC(dl, VT, N0, N1, NewCond); 01941 } 01942 01943 if ((Cond == ISD::SETEQ || Cond == ISD::SETNE) && 01944 N0.getValueType().isInteger()) { 01945 if (N0.getOpcode() == ISD::ADD || N0.getOpcode() == ISD::SUB || 01946 N0.getOpcode() == ISD::XOR) { 01947 // Simplify (X+Y) == (X+Z) --> Y == Z 01948 if (N0.getOpcode() == N1.getOpcode()) { 01949 if (N0.getOperand(0) == N1.getOperand(0)) 01950 return DAG.getSetCC(dl, VT, N0.getOperand(1), N1.getOperand(1), Cond); 01951 if (N0.getOperand(1) == N1.getOperand(1)) 01952 return DAG.getSetCC(dl, VT, N0.getOperand(0), N1.getOperand(0), Cond); 01953 if (DAG.isCommutativeBinOp(N0.getOpcode())) { 01954 // If X op Y == Y op X, try other combinations. 01955 if (N0.getOperand(0) == N1.getOperand(1)) 01956 return DAG.getSetCC(dl, VT, N0.getOperand(1), N1.getOperand(0), 01957 Cond); 01958 if (N0.getOperand(1) == N1.getOperand(0)) 01959 return DAG.getSetCC(dl, VT, N0.getOperand(0), N1.getOperand(1), 01960 Cond); 01961 } 01962 } 01963 01964 if (ConstantSDNode *RHSC = dyn_cast<ConstantSDNode>(N1)) { 01965 if (ConstantSDNode *LHSR = dyn_cast<ConstantSDNode>(N0.getOperand(1))) { 01966 // Turn (X+C1) == C2 --> X == C2-C1 01967 if (N0.getOpcode() == ISD::ADD && N0.getNode()->hasOneUse()) { 01968 return DAG.getSetCC(dl, VT, N0.getOperand(0), 01969 DAG.getConstant(RHSC->getAPIntValue()- 01970 LHSR->getAPIntValue(), 01971 N0.getValueType()), Cond); 01972 } 01973 01974 // Turn (X^C1) == C2 into X == C1^C2 iff X&~C1 = 0. 01975 if (N0.getOpcode() == ISD::XOR) 01976 // If we know that all of the inverted bits are zero, don't bother 01977 // performing the inversion. 01978 if (DAG.MaskedValueIsZero(N0.getOperand(0), ~LHSR->getAPIntValue())) 01979 return 01980 DAG.getSetCC(dl, VT, N0.getOperand(0), 01981 DAG.getConstant(LHSR->getAPIntValue() ^ 01982 RHSC->getAPIntValue(), 01983 N0.getValueType()), 01984 Cond); 01985 } 01986 01987 // Turn (C1-X) == C2 --> X == C1-C2 01988 if (ConstantSDNode *SUBC = dyn_cast<ConstantSDNode>(N0.getOperand(0))) { 01989 if (N0.getOpcode() == ISD::SUB && N0.getNode()->hasOneUse()) { 01990 return 01991 DAG.getSetCC(dl, VT, N0.getOperand(1), 01992 DAG.getConstant(SUBC->getAPIntValue() - 01993 RHSC->getAPIntValue(), 01994 N0.getValueType()), 01995 Cond); 01996 } 01997 } 01998 } 01999 02000 // Simplify (X+Z) == X --> Z == 0 02001 if (N0.getOperand(0) == N1) 02002 return DAG.getSetCC(dl, VT, N0.getOperand(1), 02003 DAG.getConstant(0, N0.getValueType()), Cond); 02004 if (N0.getOperand(1) == N1) { 02005 if (DAG.isCommutativeBinOp(N0.getOpcode())) 02006 return DAG.getSetCC(dl, VT, N0.getOperand(0), 02007 DAG.getConstant(0, N0.getValueType()), Cond); 02008 else if (N0.getNode()->hasOneUse()) { 02009 assert(N0.getOpcode() == ISD::SUB && "Unexpected operation!"); 02010 // (Z-X) == X --> Z == X<<1 02011 SDValue SH = DAG.getNode(ISD::SHL, dl, N1.getValueType(), 02012 N1, 02013 DAG.getConstant(1, getShiftAmountTy())); 02014 if (!DCI.isCalledByLegalizer()) 02015 DCI.AddToWorklist(SH.getNode()); 02016 return DAG.getSetCC(dl, VT, N0.getOperand(0), SH, Cond); 02017 } 02018 } 02019 } 02020 02021 if (N1.getOpcode() == ISD::ADD || N1.getOpcode() == ISD::SUB || 02022 N1.getOpcode() == ISD::XOR) { 02023 // Simplify X == (X+Z) --> Z == 0 02024 if (N1.getOperand(0) == N0) { 02025 return DAG.getSetCC(dl, VT, N1.getOperand(1), 02026 DAG.getConstant(0, N1.getValueType()), Cond); 02027 } else if (N1.getOperand(1) == N0) { 02028 if (DAG.isCommutativeBinOp(N1.getOpcode())) { 02029 return DAG.getSetCC(dl, VT, N1.getOperand(0), 02030 DAG.getConstant(0, N1.getValueType()), Cond); 02031 } else if (N1.getNode()->hasOneUse()) { 02032 assert(N1.getOpcode() == ISD::SUB && "Unexpected operation!"); 02033 // X == (Z-X) --> X<<1 == Z 02034 SDValue SH = DAG.getNode(ISD::SHL, dl, N1.getValueType(), N0, 02035 DAG.getConstant(1, getShiftAmountTy())); 02036 if (!DCI.isCalledByLegalizer()) 02037 DCI.AddToWorklist(SH.getNode()); 02038 return DAG.getSetCC(dl, VT, SH, N1.getOperand(0), Cond); 02039 } 02040 } 02041 } 02042 02043 // Simplify x&y == y to x&y != 0 if y has exactly one bit set. 02044 // Note that where y is variable and is known to have at most 02045 // one bit set (for example, if it is z&1) we cannot do this; 02046 // the expressions are not equivalent when y==0. 02047 if (N0.getOpcode() == ISD::AND) 02048 if (N0.getOperand(0) == N1 || N0.getOperand(1) == N1) { 02049 if (ValueHasExactlyOneBitSet(N1, DAG)) { 02050 Cond = ISD::getSetCCInverse(Cond, /*isInteger=*/true); 02051 SDValue Zero = DAG.getConstant(0, N1.getValueType()); 02052 return DAG.getSetCC(dl, VT, N0, Zero, Cond); 02053 } 02054 } 02055 if (N1.getOpcode() == ISD::AND) 02056 if (N1.getOperand(0) == N0 || N1.getOperand(1) == N0) { 02057 if (ValueHasExactlyOneBitSet(N0, DAG)) { 02058 Cond = ISD::getSetCCInverse(Cond, /*isInteger=*/true); 02059 SDValue Zero = DAG.getConstant(0, N0.getValueType()); 02060 return DAG.getSetCC(dl, VT, N1, Zero, Cond); 02061 } 02062 } 02063 } 02064 02065 // Fold away ALL boolean setcc's. 02066 SDValue Temp; 02067 if (N0.getValueType() == MVT::i1 && foldBooleans) { 02068 switch (Cond) { 02069 default: llvm_unreachable("Unknown integer setcc!"); 02070 case ISD::SETEQ: // X == Y -> ~(X^Y) 02071 Temp = DAG.getNode(ISD::XOR, dl, MVT::i1, N0, N1); 02072 N0 = DAG.getNOT(dl, Temp, MVT::i1); 02073 if (!DCI.isCalledByLegalizer()) 02074 DCI.AddToWorklist(Temp.getNode()); 02075 break; 02076 case ISD::SETNE: // X != Y --> (X^Y) 02077 N0 = DAG.getNode(ISD::XOR, dl, MVT::i1, N0, N1); 02078 break; 02079 case ISD::SETGT: // X >s Y --> X == 0 & Y == 1 --> ~X & Y 02080 case ISD::SETULT: // X <u Y --> X == 0 & Y == 1 --> ~X & Y 02081 Temp = DAG.getNOT(dl, N0, MVT::i1); 02082 N0 = DAG.getNode(ISD::AND, dl, MVT::i1, N1, Temp); 02083 if (!DCI.isCalledByLegalizer()) 02084 DCI.AddToWorklist(Temp.getNode()); 02085 break; 02086 case ISD::SETLT: // X <s Y --> X == 1 & Y == 0 --> ~Y & X 02087 case ISD::SETUGT: // X >u Y --> X == 1 & Y == 0 --> ~Y & X 02088 Temp = DAG.getNOT(dl, N1, MVT::i1); 02089 N0 = DAG.getNode(ISD::AND, dl, MVT::i1, N0, Temp); 02090 if (!DCI.isCalledByLegalizer()) 02091 DCI.AddToWorklist(Temp.getNode()); 02092 break; 02093 case ISD::SETULE: // X <=u Y --> X == 0 | Y == 1 --> ~X | Y 02094 case ISD::SETGE: // X >=s Y --> X == 0 | Y == 1 --> ~X | Y 02095 Temp = DAG.getNOT(dl, N0, MVT::i1); 02096 N0 = DAG.getNode(ISD::OR, dl, MVT::i1, N1, Temp); 02097 if (!DCI.isCalledByLegalizer()) 02098 DCI.AddToWorklist(Temp.getNode()); 02099 break; 02100 case ISD::SETUGE: // X >=u Y --> X == 1 | Y == 0 --> ~Y | X 02101 case ISD::SETLE: // X <=s Y --> X == 1 | Y == 0 --> ~Y | X 02102 Temp = DAG.getNOT(dl, N1, MVT::i1); 02103 N0 = DAG.getNode(ISD::OR, dl, MVT::i1, N0, Temp); 02104 break; 02105 } 02106 if (VT != MVT::i1) { 02107 if (!DCI.isCalledByLegalizer()) 02108 DCI.AddToWorklist(N0.getNode()); 02109 // FIXME: If running after legalize, we probably can't do this. 02110 N0 = DAG.getNode(ISD::ZERO_EXTEND, dl, VT, N0); 02111 } 02112 return N0; 02113 } 02114 02115 // Could not fold it. 02116 return SDValue(); 02117 } 02118 02119 /// isGAPlusOffset - Returns true (and the GlobalValue and the offset) if the 02120 /// node is a GlobalAddress + offset. 02121 bool TargetLowering::isGAPlusOffset(SDNode *N, GlobalValue* &GA, 02122 int64_t &Offset) const { 02123 if (isa<GlobalAddressSDNode>(N)) { 02124 GlobalAddressSDNode *GASD = cast<GlobalAddressSDNode>(N); 02125 GA = GASD->getGlobal(); 02126 Offset += GASD->getOffset(); 02127 return true; 02128 } 02129 02130 if (N->getOpcode() == ISD::ADD) { 02131 SDValue N1 = N->getOperand(0); 02132 SDValue N2 = N->getOperand(1); 02133 if (isGAPlusOffset(N1.getNode(), GA, Offset)) { 02134 ConstantSDNode *V = dyn_cast<ConstantSDNode>(N2); 02135 if (V) { 02136 Offset += V->getSExtValue(); 02137 return true; 02138 } 02139 } else if (isGAPlusOffset(N2.getNode(), GA, Offset)) { 02140 ConstantSDNode *V = dyn_cast<ConstantSDNode>(N1); 02141 if (V) { 02142 Offset += V->getSExtValue(); 02143 return true; 02144 } 02145 } 02146 } 02147 return false; 02148 } 02149 02150 02151 /// isConsecutiveLoad - Return true if LD is loading 'Bytes' bytes from a 02152 /// location that is 'Dist' units away from the location that the 'Base' load 02153 /// is loading from. 02154 bool TargetLowering::isConsecutiveLoad(LoadSDNode *LD, LoadSDNode *Base, 02155 unsigned Bytes, int Dist, 02156 const MachineFrameInfo *MFI) const { 02157 if (LD->getChain() != Base->getChain()) 02158 return false; 02159 EVT VT = LD->getValueType(0); 02160 if (VT.getSizeInBits() / 8 != Bytes) 02161 return false; 02162 02163 SDValue Loc = LD->getOperand(1); 02164 SDValue BaseLoc = Base->getOperand(1); 02165 if (Loc.getOpcode() == ISD::FrameIndex) { 02166 if (BaseLoc.getOpcode() != ISD::FrameIndex) 02167 return false; 02168 int FI = cast<FrameIndexSDNode>(Loc)->getIndex(); 02169 int BFI = cast<FrameIndexSDNode>(BaseLoc)->getIndex(); 02170 int FS = MFI->getObjectSize(FI); 02171 int BFS = MFI->getObjectSize(BFI); 02172 if (FS != BFS || FS != (int)Bytes) return false; 02173 return MFI->getObjectOffset(FI) == (MFI->getObjectOffset(BFI) + Dist*Bytes); 02174 } 02175 if (Loc.getOpcode() == ISD::ADD && Loc.getOperand(0) == BaseLoc) { 02176 ConstantSDNode *V = dyn_cast<ConstantSDNode>(Loc.getOperand(1)); 02177 if (V && (V->getSExtValue() == Dist*Bytes)) 02178 return true; 02179 } 02180 02181 GlobalValue *GV1 = NULL; 02182 GlobalValue *GV2 = NULL; 02183 int64_t Offset1 = 0; 02184 int64_t Offset2 = 0; 02185 bool isGA1 = isGAPlusOffset(Loc.getNode(), GV1, Offset1); 02186 bool isGA2 = isGAPlusOffset(BaseLoc.getNode(), GV2, Offset2); 02187 if (isGA1 && isGA2 && GV1 == GV2) 02188 return Offset1 == (Offset2 + Dist*Bytes); 02189 return false; 02190 } 02191 02192 02193 SDValue TargetLowering:: 02194 PerformDAGCombine(SDNode *N, DAGCombinerInfo &DCI) const { 02195 // Default implementation: no optimization. 02196 return SDValue(); 02197 } 02198 02199 //===----------------------------------------------------------------------===// 02200 // Inline Assembler Implementation Methods 02201 //===----------------------------------------------------------------------===// 02202 02203 02204 TargetLowering::ConstraintType 02205 TargetLowering::getConstraintType(const std::string &Constraint) const { 02206 // FIXME: lots more standard ones to handle. 02207 if (Constraint.size() == 1) { 02208 switch (Constraint[0]) { 02209 default: break; 02210 case 'r': return C_RegisterClass; 02211 case 'm': // memory 02212 case 'o': // offsetable 02213 case 'V': // not offsetable 02214 return C_Memory; 02215 case 'i': // Simple Integer or Relocatable Constant 02216 case 'n': // Simple Integer 02217 case 's': // Relocatable Constant 02218 case 'X': // Allow ANY value. 02219 case 'I': // Target registers. 02220 case 'J': 02221 case 'K': 02222 case 'L': 02223 case 'M': 02224 case 'N': 02225 case 'O': 02226 case 'P': 02227 return C_Other; 02228 } 02229 } 02230 02231 if (Constraint.size() > 1 && Constraint[0] == '{' && 02232 Constraint[Constraint.size()-1] == '}') 02233 return C_Register; 02234 return C_Unknown; 02235 } 02236 02237 /// LowerXConstraint - try to replace an X constraint, which matches anything, 02238 /// with another that has more specific requirements based on the type of the 02239 /// corresponding operand. 02240 const char *TargetLowering::LowerXConstraint(EVT ConstraintVT) const{ 02241 if (ConstraintVT.isInteger()) 02242 return "r"; 02243 if (ConstraintVT.isFloatingPoint()) 02244 return "f"; // works for many targets 02245 return 0; 02246 } 02247 02248 /// LowerAsmOperandForConstraint - Lower the specified operand into the Ops 02249 /// vector. If it is invalid, don't add anything to Ops. 02250 void TargetLowering::LowerAsmOperandForConstraint(SDValue Op, 02251 char ConstraintLetter, 02252 bool hasMemory, 02253 std::vector<SDValue> &Ops, 02254 SelectionDAG &DAG) const { 02255 switch (ConstraintLetter) { 02256 default: break; 02257 case 'X': // Allows any operand; labels (basic block) use this. 02258 if (Op.getOpcode() == ISD::BasicBlock) { 02259 Ops.push_back(Op); 02260 return; 02261 } 02262 // fall through 02263 case 'i': // Simple Integer or Relocatable Constant 02264 case 'n': // Simple Integer 02265 case 's': { // Relocatable Constant 02266 // These operands are interested in values of the form (GV+C), where C may 02267 // be folded in as an offset of GV, or it may be explicitly added. Also, it 02268 // is possible and fine if either GV or C are missing. 02269 ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op); 02270 GlobalAddressSDNode *GA = dyn_cast<GlobalAddressSDNode>(Op); 02271 02272 // If we have "(add GV, C)", pull out GV/C 02273 if (Op.getOpcode() == ISD::ADD) { 02274 C = dyn_cast<ConstantSDNode>(Op.getOperand(1)); 02275 GA = dyn_cast<GlobalAddressSDNode>(Op.getOperand(0)); 02276 if (C == 0 || GA == 0) { 02277 C = dyn_cast<ConstantSDNode>(Op.getOperand(0)); 02278 GA = dyn_cast<GlobalAddressSDNode>(Op.getOperand(1)); 02279 } 02280 if (C == 0 || GA == 0) 02281 C = 0, GA = 0; 02282 } 02283 02284 // If we find a valid operand, map to the TargetXXX version so that the 02285 // value itself doesn't get selected. 02286 if (GA) { // Either &GV or &GV+C 02287 if (ConstraintLetter != 'n') { 02288 int64_t Offs = GA->getOffset(); 02289 if (C) Offs += C->getZExtValue(); 02290 Ops.push_back(DAG.getTargetGlobalAddress(GA->getGlobal(), 02291 Op.getValueType(), Offs)); 02292 return; 02293 } 02294 } 02295 if (C) { // just C, no GV. 02296 // Simple constants are not allowed for 's'. 02297 if (ConstraintLetter != 's') { 02298 // gcc prints these as sign extended. Sign extend value to 64 bits 02299 // now; without this it would get ZExt'd later in 02300 // ScheduleDAGSDNodes::EmitNode, which is very generic. 02301 Ops.push_back(DAG.getTargetConstant(C->getAPIntValue().getSExtValue(), 02302 MVT::i64)); 02303 return; 02304 } 02305 } 02306 break; 02307 } 02308 } 02309 } 02310 02311 std::vector<unsigned> TargetLowering:: 02312 getRegClassForInlineAsmConstraint(const std::string &Constraint, 02313 EVT VT) const { 02314 return std::vector<unsigned>(); 02315 } 02316 02317 02318 std::pair<unsigned, const TargetRegisterClass*> TargetLowering:: 02319 getRegForInlineAsmConstraint(const std::string &Constraint, 02320 EVT VT) const { 02321 if (Constraint[0] != '{') 02322 return std::pair<unsigned, const TargetRegisterClass*>(0, 0); 02323 assert(*(Constraint.end()-1) == '}' && "Not a brace enclosed constraint?"); 02324 02325 // Remove the braces from around the name. 02326 std::string RegName(Constraint.begin()+1, Constraint.end()-1); 02327 02328 // Figure out which register class contains this reg. 02329 const TargetRegisterInfo *RI = TM.getRegisterInfo(); 02330 for (TargetRegisterInfo::regclass_iterator RCI = RI->regclass_begin(), 02331 E = RI->regclass_end(); RCI != E; ++RCI) { 02332 const TargetRegisterClass *RC = *RCI; 02333 02334 // If none of the the value types for this register class are valid, we 02335 // can't use it. For example, 64-bit reg classes on 32-bit targets. 02336 bool isLegal = false; 02337 for (TargetRegisterClass::vt_iterator I = RC->vt_begin(), E = RC->vt_end(); 02338 I != E; ++I) { 02339 if (isTypeLegal(*I)) { 02340 isLegal = true; 02341 break; 02342 } 02343 } 02344 02345 if (!isLegal) continue; 02346 02347 for (TargetRegisterClass::iterator I = RC->begin(), E = RC->end(); 02348 I != E; ++I) { 02349 if (StringsEqualNoCase(RegName, RI->get(*I).AsmName)) 02350 return std::make_pair(*I, RC); 02351 } 02352 } 02353 02354 return std::pair<unsigned, const TargetRegisterClass*>(0, 0); 02355 } 02356 02357 //===----------------------------------------------------------------------===// 02358 // Constraint Selection. 02359 02360 /// isMatchingInputConstraint - Return true of this is an input operand that is 02361 /// a matching constraint like "4". 02362 bool TargetLowering::AsmOperandInfo::isMatchingInputConstraint() const { 02363 assert(!ConstraintCode.empty() && "No known constraint!"); 02364 return isdigit(ConstraintCode[0]); 02365 } 02366 02367 /// getMatchedOperand - If this is an input matching constraint, this method 02368 /// returns the output operand it matches. 02369 unsigned TargetLowering::AsmOperandInfo::getMatchedOperand() const { 02370 assert(!ConstraintCode.empty() && "No known constraint!"); 02371 return atoi(ConstraintCode.c_str()); 02372 } 02373 02374 02375 /// getConstraintGenerality - Return an integer indicating how general CT 02376 /// is. 02377 static unsigned getConstraintGenerality(TargetLowering::ConstraintType CT) { 02378 switch (CT) { 02379 default: llvm_unreachable("Unknown constraint type!"); 02380 case TargetLowering::C_Other: 02381 case TargetLowering::C_Unknown: 02382 return 0; 02383 case TargetLowering::C_Register: 02384 return 1; 02385 case TargetLowering::C_RegisterClass: 02386 return 2; 02387 case TargetLowering::C_Memory: 02388 return 3; 02389 } 02390 } 02391 02392 /// ChooseConstraint - If there are multiple different constraints that we 02393 /// could pick for this operand (e.g. "imr") try to pick the 'best' one. 02394 /// This is somewhat tricky: constraints fall into four classes: 02395 /// Other -> immediates and magic values 02396 /// Register -> one specific register 02397 /// RegisterClass -> a group of regs 02398 /// Memory -> memory 02399 /// Ideally, we would pick the most specific constraint possible: if we have 02400 /// something that fits into a register, we would pick it. The problem here 02401 /// is that if we have something that could either be in a register or in 02402 /// memory that use of the register could cause selection of *other* 02403 /// operands to fail: they might only succeed if we pick memory. Because of 02404 /// this the heuristic we use is: 02405 /// 02406 /// 1) If there is an 'other' constraint, and if the operand is valid for 02407 /// that constraint, use it. This makes us take advantage of 'i' 02408 /// constraints when available. 02409 /// 2) Otherwise, pick the most general constraint present. This prefers 02410 /// 'm' over 'r', for example. 02411 /// 02412 static void ChooseConstraint(TargetLowering::AsmOperandInfo &OpInfo, 02413 bool hasMemory, const TargetLowering &TLI, 02414 SDValue Op, SelectionDAG *DAG) { 02415 assert(OpInfo.Codes.size() > 1 && "Doesn't have multiple constraint options"); 02416 unsigned BestIdx = 0; 02417 TargetLowering::ConstraintType BestType = TargetLowering::C_Unknown; 02418 int BestGenerality = -1; 02419 02420 // Loop over the options, keeping track of the most general one. 02421 for (unsigned i = 0, e = OpInfo.Codes.size(); i != e; ++i) { 02422 TargetLowering::ConstraintType CType = 02423 TLI.getConstraintType(OpInfo.Codes[i]); 02424 02425 // If this is an 'other' constraint, see if the operand is valid for it. 02426 // For example, on X86 we might have an 'rI' constraint. If the operand 02427 // is an integer in the range [0..31] we want to use I (saving a load 02428 // of a register), otherwise we must use 'r'. 02429 if (CType == TargetLowering::C_Other && Op.getNode()) { 02430 assert(OpInfo.Codes[i].size() == 1 && 02431 "Unhandled multi-letter 'other' constraint"); 02432 std::vector<SDValue> ResultOps; 02433 TLI.LowerAsmOperandForConstraint(Op, OpInfo.Codes[i][0], hasMemory, 02434 ResultOps, *DAG); 02435 if (!ResultOps.empty()) { 02436 BestType = CType; 02437 BestIdx = i; 02438 break; 02439 } 02440 } 02441 02442 // This constraint letter is more general than the previous one, use it. 02443 int Generality = getConstraintGenerality(CType); 02444 if (Generality > BestGenerality) { 02445 BestType = CType; 02446 BestIdx = i; 02447 BestGenerality = Generality; 02448 } 02449 } 02450 02451 OpInfo.ConstraintCode = OpInfo.Codes[BestIdx]; 02452 OpInfo.ConstraintType = BestType; 02453 } 02454 02455 /// ComputeConstraintToUse - Determines the constraint code and constraint 02456 /// type to use for the specific AsmOperandInfo, setting 02457 /// OpInfo.ConstraintCode and OpInfo.ConstraintType. 02458 void TargetLowering::ComputeConstraintToUse(AsmOperandInfo &OpInfo, 02459 SDValue Op, 02460 bool hasMemory, 02461 SelectionDAG *DAG) const { 02462 assert(!OpInfo.Codes.empty() && "Must have at least one constraint"); 02463 02464 // Single-letter constraints ('r') are very common. 02465 if (OpInfo.Codes.size() == 1) { 02466 OpInfo.ConstraintCode = OpInfo.Codes[0]; 02467 OpInfo.ConstraintType = getConstraintType(OpInfo.ConstraintCode); 02468 } else { 02469 ChooseConstraint(OpInfo, hasMemory, *this, Op, DAG); 02470 } 02471 02472 // 'X' matches anything. 02473 if (OpInfo.ConstraintCode == "X" && OpInfo.CallOperandVal) { 02474 // Labels and constants are handled elsewhere ('X' is the only thing 02475 // that matches labels). For Functions, the type here is the type of 02476 // the result, which is not what we want to look at; leave them alone. 02477 Value *v = OpInfo.CallOperandVal; 02478 if (isa<BasicBlock>(v) || isa<ConstantInt>(v) || isa<Function>(v)) { 02479 OpInfo.CallOperandVal = v; 02480 return; 02481 } 02482 02483 // Otherwise, try to resolve it to something we know about by looking at 02484 // the actual operand type. 02485 if (const char *Repl = LowerXConstraint(OpInfo.ConstraintVT)) { 02486 OpInfo.ConstraintCode = Repl; 02487 OpInfo.ConstraintType = getConstraintType(OpInfo.ConstraintCode); 02488 } 02489 } 02490 } 02491 02492 //===----------------------------------------------------------------------===// 02493 // Loop Strength Reduction hooks 02494 //===----------------------------------------------------------------------===// 02495 02496 /// isLegalAddressingMode - Return true if the addressing mode represented 02497 /// by AM is legal for this target, for a load/store of the specified type. 02498 bool TargetLowering::isLegalAddressingMode(const AddrMode &AM, 02499 const Type *Ty) const { 02500 // The default implementation of this implements a conservative RISCy, r+r and 02501 // r+i addr mode. 02502 02503 // Allows a sign-extended 16-bit immediate field. 02504 if (AM.BaseOffs <= -(1LL << 16) || AM.BaseOffs >= (1LL << 16)-1) 02505 return false; 02506 02507 // No global is ever allowed as a base. 02508 if (AM.BaseGV) 02509 return false; 02510 02511 // Only support r+r, 02512 switch (AM.Scale) { 02513 case 0: // "r+i" or just "i", depending on HasBaseReg. 02514 break; 02515 case 1: 02516 if (AM.HasBaseReg && AM.BaseOffs) // "r+r+i" is not allowed. 02517 return false; 02518 // Otherwise we have r+r or r+i. 02519 break; 02520 case 2: 02521 if (AM.HasBaseReg || AM.BaseOffs) // 2*r+r or 2*r+i is not allowed. 02522 return false; 02523 // Allow 2*r as r+r. 02524 break; 02525 } 02526 02527 return true; 02528 } 02529 02530 /// BuildSDIVSequence - Given an ISD::SDIV node expressing a divide by constant, 02531 /// return a DAG expression to select that will generate the same value by 02532 /// multiplying by a magic number. See: 02533 /// <http://the.wall.riscom.net/books/proc/ppc/cwg/code2.html> 02534 SDValue TargetLowering::BuildSDIV(SDNode *N, SelectionDAG &DAG, 02535 std::vector<SDNode*>* Created) const { 02536 EVT VT = N->getValueType(0); 02537 DebugLoc dl= N->getDebugLoc(); 02538 02539 // Check to see if we can do this. 02540 // FIXME: We should be more aggressive here. 02541 if (!isTypeLegal(VT)) 02542 return SDValue(); 02543 02544 APInt d = cast<ConstantSDNode>(N->getOperand(1))->getAPIntValue(); 02545 APInt::ms magics = d.magic(); 02546 02547 // Multiply the numerator (operand 0) by the magic value 02548 // FIXME: We should support doing a MUL in a wider type 02549 SDValue Q; 02550 if (isOperationLegalOrCustom(ISD::MULHS, VT)) 02551 Q = DAG.getNode(ISD::MULHS, dl, VT, N->getOperand(0), 02552 DAG.getConstant(magics.m, VT)); 02553 else if (isOperationLegalOrCustom(ISD::SMUL_LOHI, VT)) 02554 Q = SDValue(DAG.getNode(ISD::SMUL_LOHI, dl, DAG.getVTList(VT, VT), 02555 N->getOperand(0), 02556 DAG.getConstant(magics.m, VT)).getNode(), 1); 02557 else 02558 return SDValue(); // No mulhs or equvialent 02559 // If d > 0 and m < 0, add the numerator 02560 if (d.isStrictlyPositive() && magics.m.isNegative()) { 02561 Q = DAG.getNode(ISD::ADD, dl, VT, Q, N->getOperand(0)); 02562 if (Created) 02563 Created->push_back(Q.getNode()); 02564 } 02565 // If d < 0 and m > 0, subtract the numerator. 02566 if (d.isNegative() && magics.m.isStrictlyPositive()) { 02567 Q = DAG.getNode(ISD::SUB, dl, VT, Q, N->getOperand(0)); 02568 if (Created) 02569 Created->push_back(Q.getNode()); 02570 } 02571 // Shift right algebraic if shift value is nonzero 02572 if (magics.s > 0) { 02573 Q = DAG.getNode(ISD::SRA, dl, VT, Q, 02574 DAG.getConstant(magics.s, getShiftAmountTy())); 02575 if (Created) 02576 Created->push_back(Q.getNode()); 02577 } 02578 // Extract the sign bit and add it to the quotient 02579 SDValue T = 02580 DAG.getNode(ISD::SRL, dl, VT, Q, DAG.getConstant(VT.getSizeInBits()-1, 02581 getShiftAmountTy())); 02582 if (Created) 02583 Created->push_back(T.getNode()); 02584 return DAG.getNode(ISD::ADD, dl, VT, Q, T); 02585 } 02586 02587 /// BuildUDIVSequence - Given an ISD::UDIV node expressing a divide by constant, 02588 /// return a DAG expression to select that will generate the same value by 02589 /// multiplying by a magic number. See: 02590 /// <http://the.wall.riscom.net/books/proc/ppc/cwg/code2.html> 02591 SDValue TargetLowering::BuildUDIV(SDNode *N, SelectionDAG &DAG, 02592 std::vector<SDNode*>* Created) const { 02593 EVT VT = N->getValueType(0); 02594 DebugLoc dl = N->getDebugLoc(); 02595 02596 // Check to see if we can do this. 02597 // FIXME: We should be more aggressive here. 02598 if (!isTypeLegal(VT)) 02599 return SDValue(); 02600 02601 // FIXME: We should use a narrower constant when the upper 02602 // bits are known to be zero. 02603 ConstantSDNode *N1C = cast<ConstantSDNode>(N->getOperand(1)); 02604 APInt::mu magics = N1C->getAPIntValue().magicu(); 02605 02606 // Multiply the numerator (operand 0) by the magic value 02607 // FIXME: We should support doing a MUL in a wider type 02608 SDValue Q; 02609 if (isOperationLegalOrCustom(ISD::MULHU, VT)) 02610 Q = DAG.getNode(ISD::MULHU, dl, VT, N->getOperand(0), 02611 DAG.getConstant(magics.m, VT)); 02612 else if (isOperationLegalOrCustom(ISD::UMUL_LOHI, VT)) 02613 Q = SDValue(DAG.getNode(ISD::UMUL_LOHI, dl, DAG.getVTList(VT, VT), 02614 N->getOperand(0), 02615 DAG.getConstant(magics.m, VT)).getNode(), 1); 02616 else 02617 return SDValue(); // No mulhu or equvialent 02618 if (Created) 02619 Created->push_back(Q.getNode()); 02620 02621 if (magics.a == 0) { 02622 assert(magics.s < N1C->getAPIntValue().getBitWidth() && 02623 "We shouldn't generate an undefined shift!"); 02624 return DAG.getNode(ISD::SRL, dl, VT, Q, 02625 DAG.getConstant(magics.s, getShiftAmountTy())); 02626 } else { 02627 SDValue NPQ = DAG.getNode(ISD::SUB, dl, VT, N->getOperand(0), Q); 02628 if (Created) 02629 Created->push_back(NPQ.getNode()); 02630 NPQ = DAG.getNode(ISD::SRL, dl, VT, NPQ, 02631 DAG.getConstant(1, getShiftAmountTy())); 02632 if (Created) 02633 Created->push_back(NPQ.getNode()); 02634 NPQ = DAG.getNode(ISD::ADD, dl, VT, NPQ, Q); 02635 if (Created) 02636 Created->push_back(NPQ.getNode()); 02637 return DAG.getNode(ISD::SRL, dl, VT, NPQ, 02638 DAG.getConstant(magics.s-1, getShiftAmountTy())); 02639 } 02640 }