1 //===-- TargetLowering.cpp - Implement the TargetLowering class -----------===// 2 // 3 // The LLVM Compiler Infrastructure 4 // 5 // This file is distributed under the University of Illinois Open Source 6 // License. See LICENSE.TXT for details. 7 // 8 //===----------------------------------------------------------------------===// 9 // 10 // This implements the TargetLowering class. 11 // 12 //===----------------------------------------------------------------------===// 13 14 #include "llvm/Target/TargetLowering.h" 15 #include "llvm/MC/MCAsmInfo.h" 16 #include "llvm/MC/MCExpr.h" 17 #include "llvm/Target/TargetData.h" 18 #include "llvm/Target/TargetLoweringObjectFile.h" 19 #include "llvm/Target/TargetMachine.h" 20 #include "llvm/Target/TargetRegisterInfo.h" 21 #include "llvm/GlobalVariable.h" 22 #include "llvm/DerivedTypes.h" 23 #include "llvm/CodeGen/Analysis.h" 24 #include "llvm/CodeGen/MachineFrameInfo.h" 25 #include "llvm/CodeGen/MachineJumpTableInfo.h" 26 #include "llvm/CodeGen/MachineFunction.h" 27 #include "llvm/CodeGen/SelectionDAG.h" 28 #include "llvm/ADT/BitVector.h" 29 #include "llvm/ADT/STLExtras.h" 30 #include "llvm/Support/CommandLine.h" 31 #include "llvm/Support/ErrorHandling.h" 32 #include "llvm/Support/MathExtras.h" 33 #include <cctype> 34 using namespace llvm; 35 36 /// InitLibcallNames - Set default libcall names. 37 /// 38 static void InitLibcallNames(const char **Names) { 39 Names[RTLIB::SHL_I16] = "__ashlhi3"; 40 Names[RTLIB::SHL_I32] = "__ashlsi3"; 41 Names[RTLIB::SHL_I64] = "__ashldi3"; 42 Names[RTLIB::SHL_I128] = "__ashlti3"; 43 Names[RTLIB::SRL_I16] = "__lshrhi3"; 44 Names[RTLIB::SRL_I32] = "__lshrsi3"; 45 Names[RTLIB::SRL_I64] = "__lshrdi3"; 46 Names[RTLIB::SRL_I128] = "__lshrti3"; 47 Names[RTLIB::SRA_I16] = "__ashrhi3"; 48 Names[RTLIB::SRA_I32] = "__ashrsi3"; 49 Names[RTLIB::SRA_I64] = "__ashrdi3"; 50 Names[RTLIB::SRA_I128] = "__ashrti3"; 51 Names[RTLIB::MUL_I8] = "__mulqi3"; 52 Names[RTLIB::MUL_I16] = "__mulhi3"; 53 Names[RTLIB::MUL_I32] = "__mulsi3"; 54 Names[RTLIB::MUL_I64] = "__muldi3"; 55 Names[RTLIB::MUL_I128] = "__multi3"; 56 Names[RTLIB::MULO_I32] = "__mulosi4"; 57 Names[RTLIB::MULO_I64] = "__mulodi4"; 58 Names[RTLIB::MULO_I128] = "__muloti4"; 59 Names[RTLIB::SDIV_I8] = "__divqi3"; 60 Names[RTLIB::SDIV_I16] = "__divhi3"; 61 Names[RTLIB::SDIV_I32] = "__divsi3"; 62 Names[RTLIB::SDIV_I64] = "__divdi3"; 63 Names[RTLIB::SDIV_I128] = "__divti3"; 64 Names[RTLIB::UDIV_I8] = "__udivqi3"; 65 Names[RTLIB::UDIV_I16] = "__udivhi3"; 66 Names[RTLIB::UDIV_I32] = "__udivsi3"; 67 Names[RTLIB::UDIV_I64] = "__udivdi3"; 68 Names[RTLIB::UDIV_I128] = "__udivti3"; 69 Names[RTLIB::SREM_I8] = "__modqi3"; 70 Names[RTLIB::SREM_I16] = "__modhi3"; 71 Names[RTLIB::SREM_I32] = "__modsi3"; 72 Names[RTLIB::SREM_I64] = "__moddi3"; 73 Names[RTLIB::SREM_I128] = "__modti3"; 74 Names[RTLIB::UREM_I8] = "__umodqi3"; 75 Names[RTLIB::UREM_I16] = "__umodhi3"; 76 Names[RTLIB::UREM_I32] = "__umodsi3"; 77 Names[RTLIB::UREM_I64] = "__umoddi3"; 78 Names[RTLIB::UREM_I128] = "__umodti3"; 79 80 // These are generally not available. 81 Names[RTLIB::SDIVREM_I8] = 0; 82 Names[RTLIB::SDIVREM_I16] = 0; 83 Names[RTLIB::SDIVREM_I32] = 0; 84 Names[RTLIB::SDIVREM_I64] = 0; 85 Names[RTLIB::SDIVREM_I128] = 0; 86 Names[RTLIB::UDIVREM_I8] = 0; 87 Names[RTLIB::UDIVREM_I16] = 0; 88 Names[RTLIB::UDIVREM_I32] = 0; 89 Names[RTLIB::UDIVREM_I64] = 0; 90 Names[RTLIB::UDIVREM_I128] = 0; 91 92 Names[RTLIB::NEG_I32] = "__negsi2"; 93 Names[RTLIB::NEG_I64] = "__negdi2"; 94 Names[RTLIB::ADD_F32] = "__addsf3"; 95 Names[RTLIB::ADD_F64] = "__adddf3"; 96 Names[RTLIB::ADD_F80] = "__addxf3"; 97 Names[RTLIB::ADD_PPCF128] = "__gcc_qadd"; 98 Names[RTLIB::SUB_F32] = "__subsf3"; 99 Names[RTLIB::SUB_F64] = "__subdf3"; 100 Names[RTLIB::SUB_F80] = "__subxf3"; 101 Names[RTLIB::SUB_PPCF128] = "__gcc_qsub"; 102 Names[RTLIB::MUL_F32] = "__mulsf3"; 103 Names[RTLIB::MUL_F64] = "__muldf3"; 104 Names[RTLIB::MUL_F80] = "__mulxf3"; 105 Names[RTLIB::MUL_PPCF128] = "__gcc_qmul"; 106 Names[RTLIB::DIV_F32] = "__divsf3"; 107 Names[RTLIB::DIV_F64] = "__divdf3"; 108 Names[RTLIB::DIV_F80] = "__divxf3"; 109 Names[RTLIB::DIV_PPCF128] = "__gcc_qdiv"; 110 Names[RTLIB::REM_F32] = "fmodf"; 111 Names[RTLIB::REM_F64] = "fmod"; 112 Names[RTLIB::REM_F80] = "fmodl"; 113 Names[RTLIB::REM_PPCF128] = "fmodl"; 114 Names[RTLIB::FMA_F32] = "fmaf"; 115 Names[RTLIB::FMA_F64] = "fma"; 116 Names[RTLIB::FMA_F80] = "fmal"; 117 Names[RTLIB::FMA_PPCF128] = "fmal"; 118 Names[RTLIB::POWI_F32] = "__powisf2"; 119 Names[RTLIB::POWI_F64] = "__powidf2"; 120 Names[RTLIB::POWI_F80] = "__powixf2"; 121 Names[RTLIB::POWI_PPCF128] = "__powitf2"; 122 Names[RTLIB::SQRT_F32] = "sqrtf"; 123 Names[RTLIB::SQRT_F64] = "sqrt"; 124 Names[RTLIB::SQRT_F80] = "sqrtl"; 125 Names[RTLIB::SQRT_PPCF128] = "sqrtl"; 126 Names[RTLIB::LOG_F32] = "logf"; 127 Names[RTLIB::LOG_F64] = "log"; 128 Names[RTLIB::LOG_F80] = "logl"; 129 Names[RTLIB::LOG_PPCF128] = "logl"; 130 Names[RTLIB::LOG2_F32] = "log2f"; 131 Names[RTLIB::LOG2_F64] = "log2"; 132 Names[RTLIB::LOG2_F80] = "log2l"; 133 Names[RTLIB::LOG2_PPCF128] = "log2l"; 134 Names[RTLIB::LOG10_F32] = "log10f"; 135 Names[RTLIB::LOG10_F64] = "log10"; 136 Names[RTLIB::LOG10_F80] = "log10l"; 137 Names[RTLIB::LOG10_PPCF128] = "log10l"; 138 Names[RTLIB::EXP_F32] = "expf"; 139 Names[RTLIB::EXP_F64] = "exp"; 140 Names[RTLIB::EXP_F80] = "expl"; 141 Names[RTLIB::EXP_PPCF128] = "expl"; 142 Names[RTLIB::EXP2_F32] = "exp2f"; 143 Names[RTLIB::EXP2_F64] = "exp2"; 144 Names[RTLIB::EXP2_F80] = "exp2l"; 145 Names[RTLIB::EXP2_PPCF128] = "exp2l"; 146 Names[RTLIB::SIN_F32] = "sinf"; 147 Names[RTLIB::SIN_F64] = "sin"; 148 Names[RTLIB::SIN_F80] = "sinl"; 149 Names[RTLIB::SIN_PPCF128] = "sinl"; 150 Names[RTLIB::COS_F32] = "cosf"; 151 Names[RTLIB::COS_F64] = "cos"; 152 Names[RTLIB::COS_F80] = "cosl"; 153 Names[RTLIB::COS_PPCF128] = "cosl"; 154 Names[RTLIB::POW_F32] = "powf"; 155 Names[RTLIB::POW_F64] = "pow"; 156 Names[RTLIB::POW_F80] = "powl"; 157 Names[RTLIB::POW_PPCF128] = "powl"; 158 Names[RTLIB::CEIL_F32] = "ceilf"; 159 Names[RTLIB::CEIL_F64] = "ceil"; 160 Names[RTLIB::CEIL_F80] = "ceill"; 161 Names[RTLIB::CEIL_PPCF128] = "ceill"; 162 Names[RTLIB::TRUNC_F32] = "truncf"; 163 Names[RTLIB::TRUNC_F64] = "trunc"; 164 Names[RTLIB::TRUNC_F80] = "truncl"; 165 Names[RTLIB::TRUNC_PPCF128] = "truncl"; 166 Names[RTLIB::RINT_F32] = "rintf"; 167 Names[RTLIB::RINT_F64] = "rint"; 168 Names[RTLIB::RINT_F80] = "rintl"; 169 Names[RTLIB::RINT_PPCF128] = "rintl"; 170 Names[RTLIB::NEARBYINT_F32] = "nearbyintf"; 171 Names[RTLIB::NEARBYINT_F64] = "nearbyint"; 172 Names[RTLIB::NEARBYINT_F80] = "nearbyintl"; 173 Names[RTLIB::NEARBYINT_PPCF128] = "nearbyintl"; 174 Names[RTLIB::FLOOR_F32] = "floorf"; 175 Names[RTLIB::FLOOR_F64] = "floor"; 176 Names[RTLIB::FLOOR_F80] = "floorl"; 177 Names[RTLIB::FLOOR_PPCF128] = "floorl"; 178 Names[RTLIB::COPYSIGN_F32] = "copysignf"; 179 Names[RTLIB::COPYSIGN_F64] = "copysign"; 180 Names[RTLIB::COPYSIGN_F80] = "copysignl"; 181 Names[RTLIB::COPYSIGN_PPCF128] = "copysignl"; 182 Names[RTLIB::FPEXT_F32_F64] = "__extendsfdf2"; 183 Names[RTLIB::FPEXT_F16_F32] = "__gnu_h2f_ieee"; 184 Names[RTLIB::FPROUND_F32_F16] = "__gnu_f2h_ieee"; 185 Names[RTLIB::FPROUND_F64_F32] = "__truncdfsf2"; 186 Names[RTLIB::FPROUND_F80_F32] = "__truncxfsf2"; 187 Names[RTLIB::FPROUND_PPCF128_F32] = "__trunctfsf2"; 188 Names[RTLIB::FPROUND_F80_F64] = "__truncxfdf2"; 189 Names[RTLIB::FPROUND_PPCF128_F64] = "__trunctfdf2"; 190 Names[RTLIB::FPTOSINT_F32_I8] = "__fixsfqi"; 191 Names[RTLIB::FPTOSINT_F32_I16] = "__fixsfhi"; 192 Names[RTLIB::FPTOSINT_F32_I32] = "__fixsfsi"; 193 Names[RTLIB::FPTOSINT_F32_I64] = "__fixsfdi"; 194 Names[RTLIB::FPTOSINT_F32_I128] = "__fixsfti"; 195 Names[RTLIB::FPTOSINT_F64_I8] = "__fixdfqi"; 196 Names[RTLIB::FPTOSINT_F64_I16] = "__fixdfhi"; 197 Names[RTLIB::FPTOSINT_F64_I32] = "__fixdfsi"; 198 Names[RTLIB::FPTOSINT_F64_I64] = "__fixdfdi"; 199 Names[RTLIB::FPTOSINT_F64_I128] = "__fixdfti"; 200 Names[RTLIB::FPTOSINT_F80_I32] = "__fixxfsi"; 201 Names[RTLIB::FPTOSINT_F80_I64] = "__fixxfdi"; 202 Names[RTLIB::FPTOSINT_F80_I128] = "__fixxfti"; 203 Names[RTLIB::FPTOSINT_PPCF128_I32] = "__fixtfsi"; 204 Names[RTLIB::FPTOSINT_PPCF128_I64] = "__fixtfdi"; 205 Names[RTLIB::FPTOSINT_PPCF128_I128] = "__fixtfti"; 206 Names[RTLIB::FPTOUINT_F32_I8] = "__fixunssfqi"; 207 Names[RTLIB::FPTOUINT_F32_I16] = "__fixunssfhi"; 208 Names[RTLIB::FPTOUINT_F32_I32] = "__fixunssfsi"; 209 Names[RTLIB::FPTOUINT_F32_I64] = "__fixunssfdi"; 210 Names[RTLIB::FPTOUINT_F32_I128] = "__fixunssfti"; 211 Names[RTLIB::FPTOUINT_F64_I8] = "__fixunsdfqi"; 212 Names[RTLIB::FPTOUINT_F64_I16] = "__fixunsdfhi"; 213 Names[RTLIB::FPTOUINT_F64_I32] = "__fixunsdfsi"; 214 Names[RTLIB::FPTOUINT_F64_I64] = "__fixunsdfdi"; 215 Names[RTLIB::FPTOUINT_F64_I128] = "__fixunsdfti"; 216 Names[RTLIB::FPTOUINT_F80_I32] = "__fixunsxfsi"; 217 Names[RTLIB::FPTOUINT_F80_I64] = "__fixunsxfdi"; 218 Names[RTLIB::FPTOUINT_F80_I128] = "__fixunsxfti"; 219 Names[RTLIB::FPTOUINT_PPCF128_I32] = "__fixunstfsi"; 220 Names[RTLIB::FPTOUINT_PPCF128_I64] = "__fixunstfdi"; 221 Names[RTLIB::FPTOUINT_PPCF128_I128] = "__fixunstfti"; 222 Names[RTLIB::SINTTOFP_I32_F32] = "__floatsisf"; 223 Names[RTLIB::SINTTOFP_I32_F64] = "__floatsidf"; 224 Names[RTLIB::SINTTOFP_I32_F80] = "__floatsixf"; 225 Names[RTLIB::SINTTOFP_I32_PPCF128] = "__floatsitf"; 226 Names[RTLIB::SINTTOFP_I64_F32] = "__floatdisf"; 227 Names[RTLIB::SINTTOFP_I64_F64] = "__floatdidf"; 228 Names[RTLIB::SINTTOFP_I64_F80] = "__floatdixf"; 229 Names[RTLIB::SINTTOFP_I64_PPCF128] = "__floatditf"; 230 Names[RTLIB::SINTTOFP_I128_F32] = "__floattisf"; 231 Names[RTLIB::SINTTOFP_I128_F64] = "__floattidf"; 232 Names[RTLIB::SINTTOFP_I128_F80] = "__floattixf"; 233 Names[RTLIB::SINTTOFP_I128_PPCF128] = "__floattitf"; 234 Names[RTLIB::UINTTOFP_I32_F32] = "__floatunsisf"; 235 Names[RTLIB::UINTTOFP_I32_F64] = "__floatunsidf"; 236 Names[RTLIB::UINTTOFP_I32_F80] = "__floatunsixf"; 237 Names[RTLIB::UINTTOFP_I32_PPCF128] = "__floatunsitf"; 238 Names[RTLIB::UINTTOFP_I64_F32] = "__floatundisf"; 239 Names[RTLIB::UINTTOFP_I64_F64] = "__floatundidf"; 240 Names[RTLIB::UINTTOFP_I64_F80] = "__floatundixf"; 241 Names[RTLIB::UINTTOFP_I64_PPCF128] = "__floatunditf"; 242 Names[RTLIB::UINTTOFP_I128_F32] = "__floatuntisf"; 243 Names[RTLIB::UINTTOFP_I128_F64] = "__floatuntidf"; 244 Names[RTLIB::UINTTOFP_I128_F80] = "__floatuntixf"; 245 Names[RTLIB::UINTTOFP_I128_PPCF128] = "__floatuntitf"; 246 Names[RTLIB::OEQ_F32] = "__eqsf2"; 247 Names[RTLIB::OEQ_F64] = "__eqdf2"; 248 Names[RTLIB::UNE_F32] = "__nesf2"; 249 Names[RTLIB::UNE_F64] = "__nedf2"; 250 Names[RTLIB::OGE_F32] = "__gesf2"; 251 Names[RTLIB::OGE_F64] = "__gedf2"; 252 Names[RTLIB::OLT_F32] = "__ltsf2"; 253 Names[RTLIB::OLT_F64] = "__ltdf2"; 254 Names[RTLIB::OLE_F32] = "__lesf2"; 255 Names[RTLIB::OLE_F64] = "__ledf2"; 256 Names[RTLIB::OGT_F32] = "__gtsf2"; 257 Names[RTLIB::OGT_F64] = "__gtdf2"; 258 Names[RTLIB::UO_F32] = "__unordsf2"; 259 Names[RTLIB::UO_F64] = "__unorddf2"; 260 Names[RTLIB::O_F32] = "__unordsf2"; 261 Names[RTLIB::O_F64] = "__unorddf2"; 262 Names[RTLIB::MEMCPY] = "memcpy"; 263 Names[RTLIB::MEMMOVE] = "memmove"; 264 Names[RTLIB::MEMSET] = "memset"; 265 Names[RTLIB::UNWIND_RESUME] = "_Unwind_Resume"; 266 Names[RTLIB::SYNC_VAL_COMPARE_AND_SWAP_1] = "__sync_val_compare_and_swap_1"; 267 Names[RTLIB::SYNC_VAL_COMPARE_AND_SWAP_2] = "__sync_val_compare_and_swap_2"; 268 Names[RTLIB::SYNC_VAL_COMPARE_AND_SWAP_4] = "__sync_val_compare_and_swap_4"; 269 Names[RTLIB::SYNC_VAL_COMPARE_AND_SWAP_8] = "__sync_val_compare_and_swap_8"; 270 Names[RTLIB::SYNC_LOCK_TEST_AND_SET_1] = "__sync_lock_test_and_set_1"; 271 Names[RTLIB::SYNC_LOCK_TEST_AND_SET_2] = "__sync_lock_test_and_set_2"; 272 Names[RTLIB::SYNC_LOCK_TEST_AND_SET_4] = "__sync_lock_test_and_set_4"; 273 Names[RTLIB::SYNC_LOCK_TEST_AND_SET_8] = "__sync_lock_test_and_set_8"; 274 Names[RTLIB::SYNC_FETCH_AND_ADD_1] = "__sync_fetch_and_add_1"; 275 Names[RTLIB::SYNC_FETCH_AND_ADD_2] = "__sync_fetch_and_add_2"; 276 Names[RTLIB::SYNC_FETCH_AND_ADD_4] = "__sync_fetch_and_add_4"; 277 Names[RTLIB::SYNC_FETCH_AND_ADD_8] = "__sync_fetch_and_add_8"; 278 Names[RTLIB::SYNC_FETCH_AND_SUB_1] = "__sync_fetch_and_sub_1"; 279 Names[RTLIB::SYNC_FETCH_AND_SUB_2] = "__sync_fetch_and_sub_2"; 280 Names[RTLIB::SYNC_FETCH_AND_SUB_4] = "__sync_fetch_and_sub_4"; 281 Names[RTLIB::SYNC_FETCH_AND_SUB_8] = "__sync_fetch_and_sub_8"; 282 Names[RTLIB::SYNC_FETCH_AND_AND_1] = "__sync_fetch_and_and_1"; 283 Names[RTLIB::SYNC_FETCH_AND_AND_2] = "__sync_fetch_and_and_2"; 284 Names[RTLIB::SYNC_FETCH_AND_AND_4] = "__sync_fetch_and_and_4"; 285 Names[RTLIB::SYNC_FETCH_AND_AND_8] = "__sync_fetch_and_and_8"; 286 Names[RTLIB::SYNC_FETCH_AND_OR_1] = "__sync_fetch_and_or_1"; 287 Names[RTLIB::SYNC_FETCH_AND_OR_2] = "__sync_fetch_and_or_2"; 288 Names[RTLIB::SYNC_FETCH_AND_OR_4] = "__sync_fetch_and_or_4"; 289 Names[RTLIB::SYNC_FETCH_AND_OR_8] = "__sync_fetch_and_or_8"; 290 Names[RTLIB::SYNC_FETCH_AND_XOR_1] = "__sync_fetch_and_xor_1"; 291 Names[RTLIB::SYNC_FETCH_AND_XOR_2] = "__sync_fetch_and_xor_2"; 292 Names[RTLIB::SYNC_FETCH_AND_XOR_4] = "__sync_fetch_and_xor_4"; 293 Names[RTLIB::SYNC_FETCH_AND_XOR_8] = "__sync_fetch_and_xor_8"; 294 Names[RTLIB::SYNC_FETCH_AND_NAND_1] = "__sync_fetch_and_nand_1"; 295 Names[RTLIB::SYNC_FETCH_AND_NAND_2] = "__sync_fetch_and_nand_2"; 296 Names[RTLIB::SYNC_FETCH_AND_NAND_4] = "__sync_fetch_and_nand_4"; 297 Names[RTLIB::SYNC_FETCH_AND_NAND_8] = "__sync_fetch_and_nand_8"; 298 } 299 300 /// InitLibcallCallingConvs - Set default libcall CallingConvs. 301 /// 302 static void InitLibcallCallingConvs(CallingConv::ID *CCs) { 303 for (int i = 0; i < RTLIB::UNKNOWN_LIBCALL; ++i) { 304 CCs[i] = CallingConv::C; 305 } 306 } 307 308 /// getFPEXT - Return the FPEXT_*_* value for the given types, or 309 /// UNKNOWN_LIBCALL if there is none. 310 RTLIB::Libcall RTLIB::getFPEXT(EVT OpVT, EVT RetVT) { 311 if (OpVT == MVT::f32) { 312 if (RetVT == MVT::f64) 313 return FPEXT_F32_F64; 314 } 315 316 return UNKNOWN_LIBCALL; 317 } 318 319 /// getFPROUND - Return the FPROUND_*_* value for the given types, or 320 /// UNKNOWN_LIBCALL if there is none. 321 RTLIB::Libcall RTLIB::getFPROUND(EVT OpVT, EVT RetVT) { 322 if (RetVT == MVT::f32) { 323 if (OpVT == MVT::f64) 324 return FPROUND_F64_F32; 325 if (OpVT == MVT::f80) 326 return FPROUND_F80_F32; 327 if (OpVT == MVT::ppcf128) 328 return FPROUND_PPCF128_F32; 329 } else if (RetVT == MVT::f64) { 330 if (OpVT == MVT::f80) 331 return FPROUND_F80_F64; 332 if (OpVT == MVT::ppcf128) 333 return FPROUND_PPCF128_F64; 334 } 335 336 return UNKNOWN_LIBCALL; 337 } 338 339 /// getFPTOSINT - Return the FPTOSINT_*_* value for the given types, or 340 /// UNKNOWN_LIBCALL if there is none. 341 RTLIB::Libcall RTLIB::getFPTOSINT(EVT OpVT, EVT RetVT) { 342 if (OpVT == MVT::f32) { 343 if (RetVT == MVT::i8) 344 return FPTOSINT_F32_I8; 345 if (RetVT == MVT::i16) 346 return FPTOSINT_F32_I16; 347 if (RetVT == MVT::i32) 348 return FPTOSINT_F32_I32; 349 if (RetVT == MVT::i64) 350 return FPTOSINT_F32_I64; 351 if (RetVT == MVT::i128) 352 return FPTOSINT_F32_I128; 353 } else if (OpVT == MVT::f64) { 354 if (RetVT == MVT::i8) 355 return FPTOSINT_F64_I8; 356 if (RetVT == MVT::i16) 357 return FPTOSINT_F64_I16; 358 if (RetVT == MVT::i32) 359 return FPTOSINT_F64_I32; 360 if (RetVT == MVT::i64) 361 return FPTOSINT_F64_I64; 362 if (RetVT == MVT::i128) 363 return FPTOSINT_F64_I128; 364 } else if (OpVT == MVT::f80) { 365 if (RetVT == MVT::i32) 366 return FPTOSINT_F80_I32; 367 if (RetVT == MVT::i64) 368 return FPTOSINT_F80_I64; 369 if (RetVT == MVT::i128) 370 return FPTOSINT_F80_I128; 371 } else if (OpVT == MVT::ppcf128) { 372 if (RetVT == MVT::i32) 373 return FPTOSINT_PPCF128_I32; 374 if (RetVT == MVT::i64) 375 return FPTOSINT_PPCF128_I64; 376 if (RetVT == MVT::i128) 377 return FPTOSINT_PPCF128_I128; 378 } 379 return UNKNOWN_LIBCALL; 380 } 381 382 /// getFPTOUINT - Return the FPTOUINT_*_* value for the given types, or 383 /// UNKNOWN_LIBCALL if there is none. 384 RTLIB::Libcall RTLIB::getFPTOUINT(EVT OpVT, EVT RetVT) { 385 if (OpVT == MVT::f32) { 386 if (RetVT == MVT::i8) 387 return FPTOUINT_F32_I8; 388 if (RetVT == MVT::i16) 389 return FPTOUINT_F32_I16; 390 if (RetVT == MVT::i32) 391 return FPTOUINT_F32_I32; 392 if (RetVT == MVT::i64) 393 return FPTOUINT_F32_I64; 394 if (RetVT == MVT::i128) 395 return FPTOUINT_F32_I128; 396 } else if (OpVT == MVT::f64) { 397 if (RetVT == MVT::i8) 398 return FPTOUINT_F64_I8; 399 if (RetVT == MVT::i16) 400 return FPTOUINT_F64_I16; 401 if (RetVT == MVT::i32) 402 return FPTOUINT_F64_I32; 403 if (RetVT == MVT::i64) 404 return FPTOUINT_F64_I64; 405 if (RetVT == MVT::i128) 406 return FPTOUINT_F64_I128; 407 } else if (OpVT == MVT::f80) { 408 if (RetVT == MVT::i32) 409 return FPTOUINT_F80_I32; 410 if (RetVT == MVT::i64) 411 return FPTOUINT_F80_I64; 412 if (RetVT == MVT::i128) 413 return FPTOUINT_F80_I128; 414 } else if (OpVT == MVT::ppcf128) { 415 if (RetVT == MVT::i32) 416 return FPTOUINT_PPCF128_I32; 417 if (RetVT == MVT::i64) 418 return FPTOUINT_PPCF128_I64; 419 if (RetVT == MVT::i128) 420 return FPTOUINT_PPCF128_I128; 421 } 422 return UNKNOWN_LIBCALL; 423 } 424 425 /// getSINTTOFP - Return the SINTTOFP_*_* value for the given types, or 426 /// UNKNOWN_LIBCALL if there is none. 427 RTLIB::Libcall RTLIB::getSINTTOFP(EVT OpVT, EVT RetVT) { 428 if (OpVT == MVT::i32) { 429 if (RetVT == MVT::f32) 430 return SINTTOFP_I32_F32; 431 else if (RetVT == MVT::f64) 432 return SINTTOFP_I32_F64; 433 else if (RetVT == MVT::f80) 434 return SINTTOFP_I32_F80; 435 else if (RetVT == MVT::ppcf128) 436 return SINTTOFP_I32_PPCF128; 437 } else if (OpVT == MVT::i64) { 438 if (RetVT == MVT::f32) 439 return SINTTOFP_I64_F32; 440 else if (RetVT == MVT::f64) 441 return SINTTOFP_I64_F64; 442 else if (RetVT == MVT::f80) 443 return SINTTOFP_I64_F80; 444 else if (RetVT == MVT::ppcf128) 445 return SINTTOFP_I64_PPCF128; 446 } else if (OpVT == MVT::i128) { 447 if (RetVT == MVT::f32) 448 return SINTTOFP_I128_F32; 449 else if (RetVT == MVT::f64) 450 return SINTTOFP_I128_F64; 451 else if (RetVT == MVT::f80) 452 return SINTTOFP_I128_F80; 453 else if (RetVT == MVT::ppcf128) 454 return SINTTOFP_I128_PPCF128; 455 } 456 return UNKNOWN_LIBCALL; 457 } 458 459 /// getUINTTOFP - Return the UINTTOFP_*_* value for the given types, or 460 /// UNKNOWN_LIBCALL if there is none. 461 RTLIB::Libcall RTLIB::getUINTTOFP(EVT OpVT, EVT RetVT) { 462 if (OpVT == MVT::i32) { 463 if (RetVT == MVT::f32) 464 return UINTTOFP_I32_F32; 465 else if (RetVT == MVT::f64) 466 return UINTTOFP_I32_F64; 467 else if (RetVT == MVT::f80) 468 return UINTTOFP_I32_F80; 469 else if (RetVT == MVT::ppcf128) 470 return UINTTOFP_I32_PPCF128; 471 } else if (OpVT == MVT::i64) { 472 if (RetVT == MVT::f32) 473 return UINTTOFP_I64_F32; 474 else if (RetVT == MVT::f64) 475 return UINTTOFP_I64_F64; 476 else if (RetVT == MVT::f80) 477 return UINTTOFP_I64_F80; 478 else if (RetVT == MVT::ppcf128) 479 return UINTTOFP_I64_PPCF128; 480 } else if (OpVT == MVT::i128) { 481 if (RetVT == MVT::f32) 482 return UINTTOFP_I128_F32; 483 else if (RetVT == MVT::f64) 484 return UINTTOFP_I128_F64; 485 else if (RetVT == MVT::f80) 486 return UINTTOFP_I128_F80; 487 else if (RetVT == MVT::ppcf128) 488 return UINTTOFP_I128_PPCF128; 489 } 490 return UNKNOWN_LIBCALL; 491 } 492 493 /// InitCmpLibcallCCs - Set default comparison libcall CC. 494 /// 495 static void InitCmpLibcallCCs(ISD::CondCode *CCs) { 496 memset(CCs, ISD::SETCC_INVALID, sizeof(ISD::CondCode)*RTLIB::UNKNOWN_LIBCALL); 497 CCs[RTLIB::OEQ_F32] = ISD::SETEQ; 498 CCs[RTLIB::OEQ_F64] = ISD::SETEQ; 499 CCs[RTLIB::UNE_F32] = ISD::SETNE; 500 CCs[RTLIB::UNE_F64] = ISD::SETNE; 501 CCs[RTLIB::OGE_F32] = ISD::SETGE; 502 CCs[RTLIB::OGE_F64] = ISD::SETGE; 503 CCs[RTLIB::OLT_F32] = ISD::SETLT; 504 CCs[RTLIB::OLT_F64] = ISD::SETLT; 505 CCs[RTLIB::OLE_F32] = ISD::SETLE; 506 CCs[RTLIB::OLE_F64] = ISD::SETLE; 507 CCs[RTLIB::OGT_F32] = ISD::SETGT; 508 CCs[RTLIB::OGT_F64] = ISD::SETGT; 509 CCs[RTLIB::UO_F32] = ISD::SETNE; 510 CCs[RTLIB::UO_F64] = ISD::SETNE; 511 CCs[RTLIB::O_F32] = ISD::SETEQ; 512 CCs[RTLIB::O_F64] = ISD::SETEQ; 513 } 514 515 /// NOTE: The constructor takes ownership of TLOF. 516 TargetLowering::TargetLowering(const TargetMachine &tm, 517 const TargetLoweringObjectFile *tlof) 518 : TM(tm), TD(TM.getTargetData()), TLOF(*tlof) { 519 // All operations default to being supported. 520 memset(OpActions, 0, sizeof(OpActions)); 521 memset(LoadExtActions, 0, sizeof(LoadExtActions)); 522 memset(TruncStoreActions, 0, sizeof(TruncStoreActions)); 523 memset(IndexedModeActions, 0, sizeof(IndexedModeActions)); 524 memset(CondCodeActions, 0, sizeof(CondCodeActions)); 525 526 // Set default actions for various operations. 527 for (unsigned VT = 0; VT != (unsigned)MVT::LAST_VALUETYPE; ++VT) { 528 // Default all indexed load / store to expand. 529 for (unsigned IM = (unsigned)ISD::PRE_INC; 530 IM != (unsigned)ISD::LAST_INDEXED_MODE; ++IM) { 531 setIndexedLoadAction(IM, (MVT::SimpleValueType)VT, Expand); 532 setIndexedStoreAction(IM, (MVT::SimpleValueType)VT, Expand); 533 } 534 535 // These operations default to expand. 536 setOperationAction(ISD::FGETSIGN, (MVT::SimpleValueType)VT, Expand); 537 setOperationAction(ISD::CONCAT_VECTORS, (MVT::SimpleValueType)VT, Expand); 538 } 539 540 // Most targets ignore the @llvm.prefetch intrinsic. 541 setOperationAction(ISD::PREFETCH, MVT::Other, Expand); 542 543 // ConstantFP nodes default to expand. Targets can either change this to 544 // Legal, in which case all fp constants are legal, or use isFPImmLegal() 545 // to optimize expansions for certain constants. 546 setOperationAction(ISD::ConstantFP, MVT::f16, Expand); 547 setOperationAction(ISD::ConstantFP, MVT::f32, Expand); 548 setOperationAction(ISD::ConstantFP, MVT::f64, Expand); 549 setOperationAction(ISD::ConstantFP, MVT::f80, Expand); 550 551 // These library functions default to expand. 552 setOperationAction(ISD::FLOG , MVT::f16, Expand); 553 setOperationAction(ISD::FLOG2, MVT::f16, Expand); 554 setOperationAction(ISD::FLOG10, MVT::f16, Expand); 555 setOperationAction(ISD::FEXP , MVT::f16, Expand); 556 setOperationAction(ISD::FEXP2, MVT::f16, Expand); 557 setOperationAction(ISD::FFLOOR, MVT::f16, Expand); 558 setOperationAction(ISD::FNEARBYINT, MVT::f16, Expand); 559 setOperationAction(ISD::FCEIL, MVT::f16, Expand); 560 setOperationAction(ISD::FRINT, MVT::f16, Expand); 561 setOperationAction(ISD::FTRUNC, MVT::f16, Expand); 562 setOperationAction(ISD::FLOG , MVT::f32, Expand); 563 setOperationAction(ISD::FLOG2, MVT::f32, Expand); 564 setOperationAction(ISD::FLOG10, MVT::f32, Expand); 565 setOperationAction(ISD::FEXP , MVT::f32, Expand); 566 setOperationAction(ISD::FEXP2, MVT::f32, Expand); 567 setOperationAction(ISD::FFLOOR, MVT::f32, Expand); 568 setOperationAction(ISD::FNEARBYINT, MVT::f32, Expand); 569 setOperationAction(ISD::FCEIL, MVT::f32, Expand); 570 setOperationAction(ISD::FRINT, MVT::f32, Expand); 571 setOperationAction(ISD::FTRUNC, MVT::f32, Expand); 572 setOperationAction(ISD::FLOG , MVT::f64, Expand); 573 setOperationAction(ISD::FLOG2, MVT::f64, Expand); 574 setOperationAction(ISD::FLOG10, MVT::f64, Expand); 575 setOperationAction(ISD::FEXP , MVT::f64, Expand); 576 setOperationAction(ISD::FEXP2, MVT::f64, Expand); 577 setOperationAction(ISD::FFLOOR, MVT::f64, Expand); 578 setOperationAction(ISD::FNEARBYINT, MVT::f64, Expand); 579 setOperationAction(ISD::FCEIL, MVT::f64, Expand); 580 setOperationAction(ISD::FRINT, MVT::f64, Expand); 581 setOperationAction(ISD::FTRUNC, MVT::f64, Expand); 582 583 // Default ISD::TRAP to expand (which turns it into abort). 584 setOperationAction(ISD::TRAP, MVT::Other, Expand); 585 586 IsLittleEndian = TD->isLittleEndian(); 587 PointerTy = MVT::getIntegerVT(8*TD->getPointerSize()); 588 memset(RegClassForVT, 0,MVT::LAST_VALUETYPE*sizeof(TargetRegisterClass*)); 589 memset(TargetDAGCombineArray, 0, array_lengthof(TargetDAGCombineArray)); 590 maxStoresPerMemset = maxStoresPerMemcpy = maxStoresPerMemmove = 8; 591 maxStoresPerMemsetOptSize = maxStoresPerMemcpyOptSize 592 = maxStoresPerMemmoveOptSize = 4; 593 benefitFromCodePlacementOpt = false; 594 UseUnderscoreSetJmp = false; 595 UseUnderscoreLongJmp = false; 596 SelectIsExpensive = false; 597 IntDivIsCheap = false; 598 Pow2DivIsCheap = false; 599 JumpIsExpensive = false; 600 predictableSelectIsExpensive = false; 601 StackPointerRegisterToSaveRestore = 0; 602 ExceptionPointerRegister = 0; 603 ExceptionSelectorRegister = 0; 604 BooleanContents = UndefinedBooleanContent; 605 BooleanVectorContents = UndefinedBooleanContent; 606 SchedPreferenceInfo = Sched::ILP; 607 JumpBufSize = 0; 608 JumpBufAlignment = 0; 609 MinFunctionAlignment = 0; 610 PrefFunctionAlignment = 0; 611 PrefLoopAlignment = 0; 612 MinStackArgumentAlignment = 1; 613 ShouldFoldAtomicFences = false; 614 InsertFencesForAtomic = false; 615 SupportJumpTables = true; 616 617 InitLibcallNames(LibcallRoutineNames); 618 InitCmpLibcallCCs(CmpLibcallCCs); 619 InitLibcallCallingConvs(LibcallCallingConvs); 620 } 621 622 TargetLowering::~TargetLowering() { 623 delete &TLOF; 624 } 625 626 MVT TargetLowering::getShiftAmountTy(EVT LHSTy) const { 627 return MVT::getIntegerVT(8*TD->getPointerSize()); 628 } 629 630 /// canOpTrap - Returns true if the operation can trap for the value type. 631 /// VT must be a legal type. 632 bool TargetLowering::canOpTrap(unsigned Op, EVT VT) const { 633 assert(isTypeLegal(VT)); 634 switch (Op) { 635 default: 636 return false; 637 case ISD::FDIV: 638 case ISD::FREM: 639 case ISD::SDIV: 640 case ISD::UDIV: 641 case ISD::SREM: 642 case ISD::UREM: 643 return true; 644 } 645 } 646 647 648 static unsigned getVectorTypeBreakdownMVT(MVT VT, MVT &IntermediateVT, 649 unsigned &NumIntermediates, 650 EVT &RegisterVT, 651 TargetLowering *TLI) { 652 // Figure out the right, legal destination reg to copy into. 653 unsigned NumElts = VT.getVectorNumElements(); 654 MVT EltTy = VT.getVectorElementType(); 655 656 unsigned NumVectorRegs = 1; 657 658 // FIXME: We don't support non-power-of-2-sized vectors for now. Ideally we 659 // could break down into LHS/RHS like LegalizeDAG does. 660 if (!isPowerOf2_32(NumElts)) { 661 NumVectorRegs = NumElts; 662 NumElts = 1; 663 } 664 665 // Divide the input until we get to a supported size. This will always 666 // end with a scalar if the target doesn't support vectors. 667 while (NumElts > 1 && !TLI->isTypeLegal(MVT::getVectorVT(EltTy, NumElts))) { 668 NumElts >>= 1; 669 NumVectorRegs <<= 1; 670 } 671 672 NumIntermediates = NumVectorRegs; 673 674 MVT NewVT = MVT::getVectorVT(EltTy, NumElts); 675 if (!TLI->isTypeLegal(NewVT)) 676 NewVT = EltTy; 677 IntermediateVT = NewVT; 678 679 unsigned NewVTSize = NewVT.getSizeInBits(); 680 681 // Convert sizes such as i33 to i64. 682 if (!isPowerOf2_32(NewVTSize)) 683 NewVTSize = NextPowerOf2(NewVTSize); 684 685 EVT DestVT = TLI->getRegisterType(NewVT); 686 RegisterVT = DestVT; 687 if (EVT(DestVT).bitsLT(NewVT)) // Value is expanded, e.g. i64 -> i16. 688 return NumVectorRegs*(NewVTSize/DestVT.getSizeInBits()); 689 690 // Otherwise, promotion or legal types use the same number of registers as 691 // the vector decimated to the appropriate level. 692 return NumVectorRegs; 693 } 694 695 /// isLegalRC - Return true if the value types that can be represented by the 696 /// specified register class are all legal. 697 bool TargetLowering::isLegalRC(const TargetRegisterClass *RC) const { 698 for (TargetRegisterClass::vt_iterator I = RC->vt_begin(), E = RC->vt_end(); 699 I != E; ++I) { 700 if (isTypeLegal(*I)) 701 return true; 702 } 703 return false; 704 } 705 706 /// findRepresentativeClass - Return the largest legal super-reg register class 707 /// of the register class for the specified type and its associated "cost". 708 std::pair<const TargetRegisterClass*, uint8_t> 709 TargetLowering::findRepresentativeClass(EVT VT) const { 710 const TargetRegisterInfo *TRI = getTargetMachine().getRegisterInfo(); 711 const TargetRegisterClass *RC = RegClassForVT[VT.getSimpleVT().SimpleTy]; 712 if (!RC) 713 return std::make_pair(RC, 0); 714 715 // Compute the set of all super-register classes. 716 BitVector SuperRegRC(TRI->getNumRegClasses()); 717 for (SuperRegClassIterator RCI(RC, TRI); RCI.isValid(); ++RCI) 718 SuperRegRC.setBitsInMask(RCI.getMask()); 719 720 // Find the first legal register class with the largest spill size. 721 const TargetRegisterClass *BestRC = RC; 722 for (int i = SuperRegRC.find_first(); i >= 0; i = SuperRegRC.find_next(i)) { 723 const TargetRegisterClass *SuperRC = TRI->getRegClass(i); 724 // We want the largest possible spill size. 725 if (SuperRC->getSize() <= BestRC->getSize()) 726 continue; 727 if (!isLegalRC(SuperRC)) 728 continue; 729 BestRC = SuperRC; 730 } 731 return std::make_pair(BestRC, 1); 732 } 733 734 /// computeRegisterProperties - Once all of the register classes are added, 735 /// this allows us to compute derived properties we expose. 736 void TargetLowering::computeRegisterProperties() { 737 assert(MVT::LAST_VALUETYPE <= MVT::MAX_ALLOWED_VALUETYPE && 738 "Too many value types for ValueTypeActions to hold!"); 739 740 // Everything defaults to needing one register. 741 for (unsigned i = 0; i != MVT::LAST_VALUETYPE; ++i) { 742 NumRegistersForVT[i] = 1; 743 RegisterTypeForVT[i] = TransformToType[i] = (MVT::SimpleValueType)i; 744 } 745 // ...except isVoid, which doesn't need any registers. 746 NumRegistersForVT[MVT::isVoid] = 0; 747 748 // Find the largest integer register class. 749 unsigned LargestIntReg = MVT::LAST_INTEGER_VALUETYPE; 750 for (; RegClassForVT[LargestIntReg] == 0; --LargestIntReg) 751 assert(LargestIntReg != MVT::i1 && "No integer registers defined!"); 752 753 // Every integer value type larger than this largest register takes twice as 754 // many registers to represent as the previous ValueType. 755 for (unsigned ExpandedReg = LargestIntReg + 1; ; ++ExpandedReg) { 756 EVT ExpandedVT = (MVT::SimpleValueType)ExpandedReg; 757 if (!ExpandedVT.isInteger()) 758 break; 759 NumRegistersForVT[ExpandedReg] = 2*NumRegistersForVT[ExpandedReg-1]; 760 RegisterTypeForVT[ExpandedReg] = (MVT::SimpleValueType)LargestIntReg; 761 TransformToType[ExpandedReg] = (MVT::SimpleValueType)(ExpandedReg - 1); 762 ValueTypeActions.setTypeAction(ExpandedVT, TypeExpandInteger); 763 } 764 765 // Inspect all of the ValueType's smaller than the largest integer 766 // register to see which ones need promotion. 767 unsigned LegalIntReg = LargestIntReg; 768 for (unsigned IntReg = LargestIntReg - 1; 769 IntReg >= (unsigned)MVT::i1; --IntReg) { 770 EVT IVT = (MVT::SimpleValueType)IntReg; 771 if (isTypeLegal(IVT)) { 772 LegalIntReg = IntReg; 773 } else { 774 RegisterTypeForVT[IntReg] = TransformToType[IntReg] = 775 (const MVT::SimpleValueType)LegalIntReg; 776 ValueTypeActions.setTypeAction(IVT, TypePromoteInteger); 777 } 778 } 779 780 // ppcf128 type is really two f64's. 781 if (!isTypeLegal(MVT::ppcf128)) { 782 NumRegistersForVT[MVT::ppcf128] = 2*NumRegistersForVT[MVT::f64]; 783 RegisterTypeForVT[MVT::ppcf128] = MVT::f64; 784 TransformToType[MVT::ppcf128] = MVT::f64; 785 ValueTypeActions.setTypeAction(MVT::ppcf128, TypeExpandFloat); 786 } 787 788 // Decide how to handle f64. If the target does not have native f64 support, 789 // expand it to i64 and we will be generating soft float library calls. 790 if (!isTypeLegal(MVT::f64)) { 791 NumRegistersForVT[MVT::f64] = NumRegistersForVT[MVT::i64]; 792 RegisterTypeForVT[MVT::f64] = RegisterTypeForVT[MVT::i64]; 793 TransformToType[MVT::f64] = MVT::i64; 794 ValueTypeActions.setTypeAction(MVT::f64, TypeSoftenFloat); 795 } 796 797 // Decide how to handle f32. If the target does not have native support for 798 // f32, promote it to f64 if it is legal. Otherwise, expand it to i32. 799 if (!isTypeLegal(MVT::f32)) { 800 if (isTypeLegal(MVT::f64)) { 801 NumRegistersForVT[MVT::f32] = NumRegistersForVT[MVT::f64]; 802 RegisterTypeForVT[MVT::f32] = RegisterTypeForVT[MVT::f64]; 803 TransformToType[MVT::f32] = MVT::f64; 804 ValueTypeActions.setTypeAction(MVT::f32, TypePromoteInteger); 805 } else { 806 NumRegistersForVT[MVT::f32] = NumRegistersForVT[MVT::i32]; 807 RegisterTypeForVT[MVT::f32] = RegisterTypeForVT[MVT::i32]; 808 TransformToType[MVT::f32] = MVT::i32; 809 ValueTypeActions.setTypeAction(MVT::f32, TypeSoftenFloat); 810 } 811 } 812 813 // Loop over all of the vector value types to see which need transformations. 814 for (unsigned i = MVT::FIRST_VECTOR_VALUETYPE; 815 i <= (unsigned)MVT::LAST_VECTOR_VALUETYPE; ++i) { 816 MVT VT = (MVT::SimpleValueType)i; 817 if (isTypeLegal(VT)) continue; 818 819 // Determine if there is a legal wider type. If so, we should promote to 820 // that wider vector type. 821 EVT EltVT = VT.getVectorElementType(); 822 unsigned NElts = VT.getVectorNumElements(); 823 if (NElts != 1) { 824 bool IsLegalWiderType = false; 825 // First try to promote the elements of integer vectors. If no legal 826 // promotion was found, fallback to the widen-vector method. 827 for (unsigned nVT = i+1; nVT <= MVT::LAST_VECTOR_VALUETYPE; ++nVT) { 828 EVT SVT = (MVT::SimpleValueType)nVT; 829 // Promote vectors of integers to vectors with the same number 830 // of elements, with a wider element type. 831 if (SVT.getVectorElementType().getSizeInBits() > EltVT.getSizeInBits() 832 && SVT.getVectorNumElements() == NElts && 833 isTypeLegal(SVT) && SVT.getScalarType().isInteger()) { 834 TransformToType[i] = SVT; 835 RegisterTypeForVT[i] = SVT; 836 NumRegistersForVT[i] = 1; 837 ValueTypeActions.setTypeAction(VT, TypePromoteInteger); 838 IsLegalWiderType = true; 839 break; 840 } 841 } 842 843 if (IsLegalWiderType) continue; 844 845 // Try to widen the vector. 846 for (unsigned nVT = i+1; nVT <= MVT::LAST_VECTOR_VALUETYPE; ++nVT) { 847 EVT SVT = (MVT::SimpleValueType)nVT; 848 if (SVT.getVectorElementType() == EltVT && 849 SVT.getVectorNumElements() > NElts && 850 isTypeLegal(SVT)) { 851 TransformToType[i] = SVT; 852 RegisterTypeForVT[i] = SVT; 853 NumRegistersForVT[i] = 1; 854 ValueTypeActions.setTypeAction(VT, TypeWidenVector); 855 IsLegalWiderType = true; 856 break; 857 } 858 } 859 if (IsLegalWiderType) continue; 860 } 861 862 MVT IntermediateVT; 863 EVT RegisterVT; 864 unsigned NumIntermediates; 865 NumRegistersForVT[i] = 866 getVectorTypeBreakdownMVT(VT, IntermediateVT, NumIntermediates, 867 RegisterVT, this); 868 RegisterTypeForVT[i] = RegisterVT; 869 870 EVT NVT = VT.getPow2VectorType(); 871 if (NVT == VT) { 872 // Type is already a power of 2. The default action is to split. 873 TransformToType[i] = MVT::Other; 874 unsigned NumElts = VT.getVectorNumElements(); 875 ValueTypeActions.setTypeAction(VT, 876 NumElts > 1 ? TypeSplitVector : TypeScalarizeVector); 877 } else { 878 TransformToType[i] = NVT; 879 ValueTypeActions.setTypeAction(VT, TypeWidenVector); 880 } 881 } 882 883 // Determine the 'representative' register class for each value type. 884 // An representative register class is the largest (meaning one which is 885 // not a sub-register class / subreg register class) legal register class for 886 // a group of value types. For example, on i386, i8, i16, and i32 887 // representative would be GR32; while on x86_64 it's GR64. 888 for (unsigned i = 0; i != MVT::LAST_VALUETYPE; ++i) { 889 const TargetRegisterClass* RRC; 890 uint8_t Cost; 891 tie(RRC, Cost) = findRepresentativeClass((MVT::SimpleValueType)i); 892 RepRegClassForVT[i] = RRC; 893 RepRegClassCostForVT[i] = Cost; 894 } 895 } 896 897 const char *TargetLowering::getTargetNodeName(unsigned Opcode) const { 898 return NULL; 899 } 900 901 EVT TargetLowering::getSetCCResultType(EVT VT) const { 902 assert(!VT.isVector() && "No default SetCC type for vectors!"); 903 return PointerTy.SimpleTy; 904 } 905 906 MVT::SimpleValueType TargetLowering::getCmpLibcallReturnType() const { 907 return MVT::i32; // return the default value 908 } 909 910 /// getVectorTypeBreakdown - Vector types are broken down into some number of 911 /// legal first class types. For example, MVT::v8f32 maps to 2 MVT::v4f32 912 /// with Altivec or SSE1, or 8 promoted MVT::f64 values with the X86 FP stack. 913 /// Similarly, MVT::v2i64 turns into 4 MVT::i32 values with both PPC and X86. 914 /// 915 /// This method returns the number of registers needed, and the VT for each 916 /// register. It also returns the VT and quantity of the intermediate values 917 /// before they are promoted/expanded. 918 /// 919 unsigned TargetLowering::getVectorTypeBreakdown(LLVMContext &Context, EVT VT, 920 EVT &IntermediateVT, 921 unsigned &NumIntermediates, 922 EVT &RegisterVT) const { 923 unsigned NumElts = VT.getVectorNumElements(); 924 925 // If there is a wider vector type with the same element type as this one, 926 // or a promoted vector type that has the same number of elements which 927 // are wider, then we should convert to that legal vector type. 928 // This handles things like <2 x float> -> <4 x float> and 929 // <4 x i1> -> <4 x i32>. 930 LegalizeTypeAction TA = getTypeAction(Context, VT); 931 if (NumElts != 1 && (TA == TypeWidenVector || TA == TypePromoteInteger)) { 932 RegisterVT = getTypeToTransformTo(Context, VT); 933 if (isTypeLegal(RegisterVT)) { 934 IntermediateVT = RegisterVT; 935 NumIntermediates = 1; 936 return 1; 937 } 938 } 939 940 // Figure out the right, legal destination reg to copy into. 941 EVT EltTy = VT.getVectorElementType(); 942 943 unsigned NumVectorRegs = 1; 944 945 // FIXME: We don't support non-power-of-2-sized vectors for now. Ideally we 946 // could break down into LHS/RHS like LegalizeDAG does. 947 if (!isPowerOf2_32(NumElts)) { 948 NumVectorRegs = NumElts; 949 NumElts = 1; 950 } 951 952 // Divide the input until we get to a supported size. This will always 953 // end with a scalar if the target doesn't support vectors. 954 while (NumElts > 1 && !isTypeLegal( 955 EVT::getVectorVT(Context, EltTy, NumElts))) { 956 NumElts >>= 1; 957 NumVectorRegs <<= 1; 958 } 959 960 NumIntermediates = NumVectorRegs; 961 962 EVT NewVT = EVT::getVectorVT(Context, EltTy, NumElts); 963 if (!isTypeLegal(NewVT)) 964 NewVT = EltTy; 965 IntermediateVT = NewVT; 966 967 EVT DestVT = getRegisterType(Context, NewVT); 968 RegisterVT = DestVT; 969 unsigned NewVTSize = NewVT.getSizeInBits(); 970 971 // Convert sizes such as i33 to i64. 972 if (!isPowerOf2_32(NewVTSize)) 973 NewVTSize = NextPowerOf2(NewVTSize); 974 975 if (DestVT.bitsLT(NewVT)) // Value is expanded, e.g. i64 -> i16. 976 return NumVectorRegs*(NewVTSize/DestVT.getSizeInBits()); 977 978 // Otherwise, promotion or legal types use the same number of registers as 979 // the vector decimated to the appropriate level. 980 return NumVectorRegs; 981 } 982 983 /// Get the EVTs and ArgFlags collections that represent the legalized return 984 /// type of the given function. This does not require a DAG or a return value, 985 /// and is suitable for use before any DAGs for the function are constructed. 986 /// TODO: Move this out of TargetLowering.cpp. 987 void llvm::GetReturnInfo(Type* ReturnType, Attributes attr, 988 SmallVectorImpl<ISD::OutputArg> &Outs, 989 const TargetLowering &TLI) { 990 SmallVector<EVT, 4> ValueVTs; 991 ComputeValueVTs(TLI, ReturnType, ValueVTs); 992 unsigned NumValues = ValueVTs.size(); 993 if (NumValues == 0) return; 994 995 for (unsigned j = 0, f = NumValues; j != f; ++j) { 996 EVT VT = ValueVTs[j]; 997 ISD::NodeType ExtendKind = ISD::ANY_EXTEND; 998 999 if (attr & Attribute::SExt) 1000 ExtendKind = ISD::SIGN_EXTEND; 1001 else if (attr & Attribute::ZExt) 1002 ExtendKind = ISD::ZERO_EXTEND; 1003 1004 // FIXME: C calling convention requires the return type to be promoted to 1005 // at least 32-bit. But this is not necessary for non-C calling 1006 // conventions. The frontend should mark functions whose return values 1007 // require promoting with signext or zeroext attributes. 1008 if (ExtendKind != ISD::ANY_EXTEND && VT.isInteger()) { 1009 EVT MinVT = TLI.getRegisterType(ReturnType->getContext(), MVT::i32); 1010 if (VT.bitsLT(MinVT)) 1011 VT = MinVT; 1012 } 1013 1014 unsigned NumParts = TLI.getNumRegisters(ReturnType->getContext(), VT); 1015 EVT PartVT = TLI.getRegisterType(ReturnType->getContext(), VT); 1016 1017 // 'inreg' on function refers to return value 1018 ISD::ArgFlagsTy Flags = ISD::ArgFlagsTy(); 1019 if (attr & Attribute::InReg) 1020 Flags.setInReg(); 1021 1022 // Propagate extension type if any 1023 if (attr & Attribute::SExt) 1024 Flags.setSExt(); 1025 else if (attr & Attribute::ZExt) 1026 Flags.setZExt(); 1027 1028 for (unsigned i = 0; i < NumParts; ++i) { 1029 Outs.push_back(ISD::OutputArg(Flags, PartVT, /*isFixed=*/true)); 1030 } 1031 } 1032 } 1033 1034 /// getByValTypeAlignment - Return the desired alignment for ByVal aggregate 1035 /// function arguments in the caller parameter area. This is the actual 1036 /// alignment, not its logarithm. 1037 unsigned TargetLowering::getByValTypeAlignment(Type *Ty) const { 1038 return TD->getCallFrameTypeAlignment(Ty); 1039 } 1040 1041 /// getJumpTableEncoding - Return the entry encoding for a jump table in the 1042 /// current function. The returned value is a member of the 1043 /// MachineJumpTableInfo::JTEntryKind enum. 1044 unsigned TargetLowering::getJumpTableEncoding() const { 1045 // In non-pic modes, just use the address of a block. 1046 if (getTargetMachine().getRelocationModel() != Reloc::PIC_) 1047 return MachineJumpTableInfo::EK_BlockAddress; 1048 1049 // In PIC mode, if the target supports a GPRel32 directive, use it. 1050 if (getTargetMachine().getMCAsmInfo()->getGPRel32Directive() != 0) 1051 return MachineJumpTableInfo::EK_GPRel32BlockAddress; 1052 1053 // Otherwise, use a label difference. 1054 return MachineJumpTableInfo::EK_LabelDifference32; 1055 } 1056 1057 SDValue TargetLowering::getPICJumpTableRelocBase(SDValue Table, 1058 SelectionDAG &DAG) const { 1059 // If our PIC model is GP relative, use the global offset table as the base. 1060 unsigned JTEncoding = getJumpTableEncoding(); 1061 1062 if ((JTEncoding == MachineJumpTableInfo::EK_GPRel64BlockAddress) || 1063 (JTEncoding == MachineJumpTableInfo::EK_GPRel32BlockAddress)) 1064 return DAG.getGLOBAL_OFFSET_TABLE(getPointerTy()); 1065 1066 return Table; 1067 } 1068 1069 /// getPICJumpTableRelocBaseExpr - This returns the relocation base for the 1070 /// given PIC jumptable, the same as getPICJumpTableRelocBase, but as an 1071 /// MCExpr. 1072 const MCExpr * 1073 TargetLowering::getPICJumpTableRelocBaseExpr(const MachineFunction *MF, 1074 unsigned JTI,MCContext &Ctx) const{ 1075 // The normal PIC reloc base is the label at the start of the jump table. 1076 return MCSymbolRefExpr::Create(MF->getJTISymbol(JTI, Ctx), Ctx); 1077 } 1078 1079 bool 1080 TargetLowering::isOffsetFoldingLegal(const GlobalAddressSDNode *GA) const { 1081 // Assume that everything is safe in static mode. 1082 if (getTargetMachine().getRelocationModel() == Reloc::Static) 1083 return true; 1084 1085 // In dynamic-no-pic mode, assume that known defined values are safe. 1086 if (getTargetMachine().getRelocationModel() == Reloc::DynamicNoPIC && 1087 GA && 1088 !GA->getGlobal()->isDeclaration() && 1089 !GA->getGlobal()->isWeakForLinker()) 1090 return true; 1091 1092 // Otherwise assume nothing is safe. 1093 return false; 1094 } 1095 1096 //===----------------------------------------------------------------------===// 1097 // Optimization Methods 1098 //===----------------------------------------------------------------------===// 1099 1100 /// ShrinkDemandedConstant - Check to see if the specified operand of the 1101 /// specified instruction is a constant integer. If so, check to see if there 1102 /// are any bits set in the constant that are not demanded. If so, shrink the 1103 /// constant and return true. 1104 bool TargetLowering::TargetLoweringOpt::ShrinkDemandedConstant(SDValue Op, 1105 const APInt &Demanded) { 1106 DebugLoc dl = Op.getDebugLoc(); 1107 1108 // FIXME: ISD::SELECT, ISD::SELECT_CC 1109 switch (Op.getOpcode()) { 1110 default: break; 1111 case ISD::XOR: 1112 case ISD::AND: 1113 case ISD::OR: { 1114 ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op.getOperand(1)); 1115 if (!C) return false; 1116 1117 if (Op.getOpcode() == ISD::XOR && 1118 (C->getAPIntValue() | (~Demanded)).isAllOnesValue()) 1119 return false; 1120 1121 // if we can expand it to have all bits set, do it 1122 if (C->getAPIntValue().intersects(~Demanded)) { 1123 EVT VT = Op.getValueType(); 1124 SDValue New = DAG.getNode(Op.getOpcode(), dl, VT, Op.getOperand(0), 1125 DAG.getConstant(Demanded & 1126 C->getAPIntValue(), 1127 VT)); 1128 return CombineTo(Op, New); 1129 } 1130 1131 break; 1132 } 1133 } 1134 1135 return false; 1136 } 1137 1138 /// ShrinkDemandedOp - Convert x+y to (VT)((SmallVT)x+(SmallVT)y) if the 1139 /// casts are free. This uses isZExtFree and ZERO_EXTEND for the widening 1140 /// cast, but it could be generalized for targets with other types of 1141 /// implicit widening casts. 1142 bool 1143 TargetLowering::TargetLoweringOpt::ShrinkDemandedOp(SDValue Op, 1144 unsigned BitWidth, 1145 const APInt &Demanded, 1146 DebugLoc dl) { 1147 assert(Op.getNumOperands() == 2 && 1148 "ShrinkDemandedOp only supports binary operators!"); 1149 assert(Op.getNode()->getNumValues() == 1 && 1150 "ShrinkDemandedOp only supports nodes with one result!"); 1151 1152 // Don't do this if the node has another user, which may require the 1153 // full value. 1154 if (!Op.getNode()->hasOneUse()) 1155 return false; 1156 1157 // Search for the smallest integer type with free casts to and from 1158 // Op's type. For expedience, just check power-of-2 integer types. 1159 const TargetLowering &TLI = DAG.getTargetLoweringInfo(); 1160 unsigned SmallVTBits = BitWidth - Demanded.countLeadingZeros(); 1161 if (!isPowerOf2_32(SmallVTBits)) 1162 SmallVTBits = NextPowerOf2(SmallVTBits); 1163 for (; SmallVTBits < BitWidth; SmallVTBits = NextPowerOf2(SmallVTBits)) { 1164 EVT SmallVT = EVT::getIntegerVT(*DAG.getContext(), SmallVTBits); 1165 if (TLI.isTruncateFree(Op.getValueType(), SmallVT) && 1166 TLI.isZExtFree(SmallVT, Op.getValueType())) { 1167 // We found a type with free casts. 1168 SDValue X = DAG.getNode(Op.getOpcode(), dl, SmallVT, 1169 DAG.getNode(ISD::TRUNCATE, dl, SmallVT, 1170 Op.getNode()->getOperand(0)), 1171 DAG.getNode(ISD::TRUNCATE, dl, SmallVT, 1172 Op.getNode()->getOperand(1))); 1173 SDValue Z = DAG.getNode(ISD::ZERO_EXTEND, dl, Op.getValueType(), X); 1174 return CombineTo(Op, Z); 1175 } 1176 } 1177 return false; 1178 } 1179 1180 /// SimplifyDemandedBits - Look at Op. At this point, we know that only the 1181 /// DemandedMask bits of the result of Op are ever used downstream. If we can 1182 /// use this information to simplify Op, create a new simplified DAG node and 1183 /// return true, returning the original and new nodes in Old and New. Otherwise, 1184 /// analyze the expression and return a mask of KnownOne and KnownZero bits for 1185 /// the expression (used to simplify the caller). The KnownZero/One bits may 1186 /// only be accurate for those bits in the DemandedMask. 1187 bool TargetLowering::SimplifyDemandedBits(SDValue Op, 1188 const APInt &DemandedMask, 1189 APInt &KnownZero, 1190 APInt &KnownOne, 1191 TargetLoweringOpt &TLO, 1192 unsigned Depth) const { 1193 unsigned BitWidth = DemandedMask.getBitWidth(); 1194 assert(Op.getValueType().getScalarType().getSizeInBits() == BitWidth && 1195 "Mask size mismatches value type size!"); 1196 APInt NewMask = DemandedMask; 1197 DebugLoc dl = Op.getDebugLoc(); 1198 1199 // Don't know anything. 1200 KnownZero = KnownOne = APInt(BitWidth, 0); 1201 1202 // Other users may use these bits. 1203 if (!Op.getNode()->hasOneUse()) { 1204 if (Depth != 0) { 1205 // If not at the root, Just compute the KnownZero/KnownOne bits to 1206 // simplify things downstream. 1207 TLO.DAG.ComputeMaskedBits(Op, KnownZero, KnownOne, Depth); 1208 return false; 1209 } 1210 // If this is the root being simplified, allow it to have multiple uses, 1211 // just set the NewMask to all bits. 1212 NewMask = APInt::getAllOnesValue(BitWidth); 1213 } else if (DemandedMask == 0) { 1214 // Not demanding any bits from Op. 1215 if (Op.getOpcode() != ISD::UNDEF) 1216 return TLO.CombineTo(Op, TLO.DAG.getUNDEF(Op.getValueType())); 1217 return false; 1218 } else if (Depth == 6) { // Limit search depth. 1219 return false; 1220 } 1221 1222 APInt KnownZero2, KnownOne2, KnownZeroOut, KnownOneOut; 1223 switch (Op.getOpcode()) { 1224 case ISD::Constant: 1225 // We know all of the bits for a constant! 1226 KnownOne = cast<ConstantSDNode>(Op)->getAPIntValue(); 1227 KnownZero = ~KnownOne; 1228 return false; // Don't fall through, will infinitely loop. 1229 case ISD::AND: 1230 // If the RHS is a constant, check to see if the LHS would be zero without 1231 // using the bits from the RHS. Below, we use knowledge about the RHS to 1232 // simplify the LHS, here we're using information from the LHS to simplify 1233 // the RHS. 1234 if (ConstantSDNode *RHSC = dyn_cast<ConstantSDNode>(Op.getOperand(1))) { 1235 APInt LHSZero, LHSOne; 1236 // Do not increment Depth here; that can cause an infinite loop. 1237 TLO.DAG.ComputeMaskedBits(Op.getOperand(0), LHSZero, LHSOne, Depth); 1238 // If the LHS already has zeros where RHSC does, this and is dead. 1239 if ((LHSZero & NewMask) == (~RHSC->getAPIntValue() & NewMask)) 1240 return TLO.CombineTo(Op, Op.getOperand(0)); 1241 // If any of the set bits in the RHS are known zero on the LHS, shrink 1242 // the constant. 1243 if (TLO.ShrinkDemandedConstant(Op, ~LHSZero & NewMask)) 1244 return true; 1245 } 1246 1247 if (SimplifyDemandedBits(Op.getOperand(1), NewMask, KnownZero, 1248 KnownOne, TLO, Depth+1)) 1249 return true; 1250 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?"); 1251 if (SimplifyDemandedBits(Op.getOperand(0), ~KnownZero & NewMask, 1252 KnownZero2, KnownOne2, TLO, Depth+1)) 1253 return true; 1254 assert((KnownZero2 & KnownOne2) == 0 && "Bits known to be one AND zero?"); 1255 1256 // If all of the demanded bits are known one on one side, return the other. 1257 // These bits cannot contribute to the result of the 'and'. 1258 if ((NewMask & ~KnownZero2 & KnownOne) == (~KnownZero2 & NewMask)) 1259 return TLO.CombineTo(Op, Op.getOperand(0)); 1260 if ((NewMask & ~KnownZero & KnownOne2) == (~KnownZero & NewMask)) 1261 return TLO.CombineTo(Op, Op.getOperand(1)); 1262 // If all of the demanded bits in the inputs are known zeros, return zero. 1263 if ((NewMask & (KnownZero|KnownZero2)) == NewMask) 1264 return TLO.CombineTo(Op, TLO.DAG.getConstant(0, Op.getValueType())); 1265 // If the RHS is a constant, see if we can simplify it. 1266 if (TLO.ShrinkDemandedConstant(Op, ~KnownZero2 & NewMask)) 1267 return true; 1268 // If the operation can be done in a smaller type, do so. 1269 if (TLO.ShrinkDemandedOp(Op, BitWidth, NewMask, dl)) 1270 return true; 1271 1272 // Output known-1 bits are only known if set in both the LHS & RHS. 1273 KnownOne &= KnownOne2; 1274 // Output known-0 are known to be clear if zero in either the LHS | RHS. 1275 KnownZero |= KnownZero2; 1276 break; 1277 case ISD::OR: 1278 if (SimplifyDemandedBits(Op.getOperand(1), NewMask, KnownZero, 1279 KnownOne, TLO, Depth+1)) 1280 return true; 1281 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?"); 1282 if (SimplifyDemandedBits(Op.getOperand(0), ~KnownOne & NewMask, 1283 KnownZero2, KnownOne2, TLO, Depth+1)) 1284 return true; 1285 assert((KnownZero2 & KnownOne2) == 0 && "Bits known to be one AND zero?"); 1286 1287 // If all of the demanded bits are known zero on one side, return the other. 1288 // These bits cannot contribute to the result of the 'or'. 1289 if ((NewMask & ~KnownOne2 & KnownZero) == (~KnownOne2 & NewMask)) 1290 return TLO.CombineTo(Op, Op.getOperand(0)); 1291 if ((NewMask & ~KnownOne & KnownZero2) == (~KnownOne & NewMask)) 1292 return TLO.CombineTo(Op, Op.getOperand(1)); 1293 // If all of the potentially set bits on one side are known to be set on 1294 // the other side, just use the 'other' side. 1295 if ((NewMask & ~KnownZero & KnownOne2) == (~KnownZero & NewMask)) 1296 return TLO.CombineTo(Op, Op.getOperand(0)); 1297 if ((NewMask & ~KnownZero2 & KnownOne) == (~KnownZero2 & NewMask)) 1298 return TLO.CombineTo(Op, Op.getOperand(1)); 1299 // If the RHS is a constant, see if we can simplify it. 1300 if (TLO.ShrinkDemandedConstant(Op, NewMask)) 1301 return true; 1302 // If the operation can be done in a smaller type, do so. 1303 if (TLO.ShrinkDemandedOp(Op, BitWidth, NewMask, dl)) 1304 return true; 1305 1306 // Output known-0 bits are only known if clear in both the LHS & RHS. 1307 KnownZero &= KnownZero2; 1308 // Output known-1 are known to be set if set in either the LHS | RHS. 1309 KnownOne |= KnownOne2; 1310 break; 1311 case ISD::XOR: 1312 if (SimplifyDemandedBits(Op.getOperand(1), NewMask, KnownZero, 1313 KnownOne, TLO, Depth+1)) 1314 return true; 1315 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?"); 1316 if (SimplifyDemandedBits(Op.getOperand(0), NewMask, KnownZero2, 1317 KnownOne2, TLO, Depth+1)) 1318 return true; 1319 assert((KnownZero2 & KnownOne2) == 0 && "Bits known to be one AND zero?"); 1320 1321 // If all of the demanded bits are known zero on one side, return the other. 1322 // These bits cannot contribute to the result of the 'xor'. 1323 if ((KnownZero & NewMask) == NewMask) 1324 return TLO.CombineTo(Op, Op.getOperand(0)); 1325 if ((KnownZero2 & NewMask) == NewMask) 1326 return TLO.CombineTo(Op, Op.getOperand(1)); 1327 // If the operation can be done in a smaller type, do so. 1328 if (TLO.ShrinkDemandedOp(Op, BitWidth, NewMask, dl)) 1329 return true; 1330 1331 // If all of the unknown bits are known to be zero on one side or the other 1332 // (but not both) turn this into an *inclusive* or. 1333 // e.g. (A & C1)^(B & C2) -> (A & C1)|(B & C2) iff C1&C2 == 0 1334 if ((NewMask & ~KnownZero & ~KnownZero2) == 0) 1335 return TLO.CombineTo(Op, TLO.DAG.getNode(ISD::OR, dl, Op.getValueType(), 1336 Op.getOperand(0), 1337 Op.getOperand(1))); 1338 1339 // Output known-0 bits are known if clear or set in both the LHS & RHS. 1340 KnownZeroOut = (KnownZero & KnownZero2) | (KnownOne & KnownOne2); 1341 // Output known-1 are known to be set if set in only one of the LHS, RHS. 1342 KnownOneOut = (KnownZero & KnownOne2) | (KnownOne & KnownZero2); 1343 1344 // If all of the demanded bits on one side are known, and all of the set 1345 // bits on that side are also known to be set on the other side, turn this 1346 // into an AND, as we know the bits will be cleared. 1347 // e.g. (X | C1) ^ C2 --> (X | C1) & ~C2 iff (C1&C2) == C2 1348 // NB: it is okay if more bits are known than are requested 1349 if ((NewMask & (KnownZero|KnownOne)) == NewMask) { // all known on one side 1350 if (KnownOne == KnownOne2) { // set bits are the same on both sides 1351 EVT VT = Op.getValueType(); 1352 SDValue ANDC = TLO.DAG.getConstant(~KnownOne & NewMask, VT); 1353 return TLO.CombineTo(Op, TLO.DAG.getNode(ISD::AND, dl, VT, 1354 Op.getOperand(0), ANDC)); 1355 } 1356 } 1357 1358 // If the RHS is a constant, see if we can simplify it. 1359 // for XOR, we prefer to force bits to 1 if they will make a -1. 1360 // if we can't force bits, try to shrink constant 1361 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op.getOperand(1))) { 1362 APInt Expanded = C->getAPIntValue() | (~NewMask); 1363 // if we can expand it to have all bits set, do it 1364 if (Expanded.isAllOnesValue()) { 1365 if (Expanded != C->getAPIntValue()) { 1366 EVT VT = Op.getValueType(); 1367 SDValue New = TLO.DAG.getNode(Op.getOpcode(), dl,VT, Op.getOperand(0), 1368 TLO.DAG.getConstant(Expanded, VT)); 1369 return TLO.CombineTo(Op, New); 1370 } 1371 // if it already has all the bits set, nothing to change 1372 // but don't shrink either! 1373 } else if (TLO.ShrinkDemandedConstant(Op, NewMask)) { 1374 return true; 1375 } 1376 } 1377 1378 KnownZero = KnownZeroOut; 1379 KnownOne = KnownOneOut; 1380 break; 1381 case ISD::SELECT: 1382 if (SimplifyDemandedBits(Op.getOperand(2), NewMask, KnownZero, 1383 KnownOne, TLO, Depth+1)) 1384 return true; 1385 if (SimplifyDemandedBits(Op.getOperand(1), NewMask, KnownZero2, 1386 KnownOne2, TLO, Depth+1)) 1387 return true; 1388 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?"); 1389 assert((KnownZero2 & KnownOne2) == 0 && "Bits known to be one AND zero?"); 1390 1391 // If the operands are constants, see if we can simplify them. 1392 if (TLO.ShrinkDemandedConstant(Op, NewMask)) 1393 return true; 1394 1395 // Only known if known in both the LHS and RHS. 1396 KnownOne &= KnownOne2; 1397 KnownZero &= KnownZero2; 1398 break; 1399 case ISD::SELECT_CC: 1400 if (SimplifyDemandedBits(Op.getOperand(3), NewMask, KnownZero, 1401 KnownOne, TLO, Depth+1)) 1402 return true; 1403 if (SimplifyDemandedBits(Op.getOperand(2), NewMask, KnownZero2, 1404 KnownOne2, TLO, Depth+1)) 1405 return true; 1406 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?"); 1407 assert((KnownZero2 & KnownOne2) == 0 && "Bits known to be one AND zero?"); 1408 1409 // If the operands are constants, see if we can simplify them. 1410 if (TLO.ShrinkDemandedConstant(Op, NewMask)) 1411 return true; 1412 1413 // Only known if known in both the LHS and RHS. 1414 KnownOne &= KnownOne2; 1415 KnownZero &= KnownZero2; 1416 break; 1417 case ISD::SHL: 1418 if (ConstantSDNode *SA = dyn_cast<ConstantSDNode>(Op.getOperand(1))) { 1419 unsigned ShAmt = SA->getZExtValue(); 1420 SDValue InOp = Op.getOperand(0); 1421 1422 // If the shift count is an invalid immediate, don't do anything. 1423 if (ShAmt >= BitWidth) 1424 break; 1425 1426 // If this is ((X >>u C1) << ShAmt), see if we can simplify this into a 1427 // single shift. We can do this if the bottom bits (which are shifted 1428 // out) are never demanded. 1429 if (InOp.getOpcode() == ISD::SRL && 1430 isa<ConstantSDNode>(InOp.getOperand(1))) { 1431 if (ShAmt && (NewMask & APInt::getLowBitsSet(BitWidth, ShAmt)) == 0) { 1432 unsigned C1= cast<ConstantSDNode>(InOp.getOperand(1))->getZExtValue(); 1433 unsigned Opc = ISD::SHL; 1434 int Diff = ShAmt-C1; 1435 if (Diff < 0) { 1436 Diff = -Diff; 1437 Opc = ISD::SRL; 1438 } 1439 1440 SDValue NewSA = 1441 TLO.DAG.getConstant(Diff, Op.getOperand(1).getValueType()); 1442 EVT VT = Op.getValueType(); 1443 return TLO.CombineTo(Op, TLO.DAG.getNode(Opc, dl, VT, 1444 InOp.getOperand(0), NewSA)); 1445 } 1446 } 1447 1448 if (SimplifyDemandedBits(InOp, NewMask.lshr(ShAmt), 1449 KnownZero, KnownOne, TLO, Depth+1)) 1450 return true; 1451 1452 // Convert (shl (anyext x, c)) to (anyext (shl x, c)) if the high bits 1453 // are not demanded. This will likely allow the anyext to be folded away. 1454 if (InOp.getNode()->getOpcode() == ISD::ANY_EXTEND) { 1455 SDValue InnerOp = InOp.getNode()->getOperand(0); 1456 EVT InnerVT = InnerOp.getValueType(); 1457 unsigned InnerBits = InnerVT.getSizeInBits(); 1458 if (ShAmt < InnerBits && NewMask.lshr(InnerBits) == 0 && 1459 isTypeDesirableForOp(ISD::SHL, InnerVT)) { 1460 EVT ShTy = getShiftAmountTy(InnerVT); 1461 if (!APInt(BitWidth, ShAmt).isIntN(ShTy.getSizeInBits())) 1462 ShTy = InnerVT; 1463 SDValue NarrowShl = 1464 TLO.DAG.getNode(ISD::SHL, dl, InnerVT, InnerOp, 1465 TLO.DAG.getConstant(ShAmt, ShTy)); 1466 return 1467 TLO.CombineTo(Op, 1468 TLO.DAG.getNode(ISD::ANY_EXTEND, dl, Op.getValueType(), 1469 NarrowShl)); 1470 } 1471 } 1472 1473 KnownZero <<= SA->getZExtValue(); 1474 KnownOne <<= SA->getZExtValue(); 1475 // low bits known zero. 1476 KnownZero |= APInt::getLowBitsSet(BitWidth, SA->getZExtValue()); 1477 } 1478 break; 1479 case ISD::SRL: 1480 if (ConstantSDNode *SA = dyn_cast<ConstantSDNode>(Op.getOperand(1))) { 1481 EVT VT = Op.getValueType(); 1482 unsigned ShAmt = SA->getZExtValue(); 1483 unsigned VTSize = VT.getSizeInBits(); 1484 SDValue InOp = Op.getOperand(0); 1485 1486 // If the shift count is an invalid immediate, don't do anything. 1487 if (ShAmt >= BitWidth) 1488 break; 1489 1490 // If this is ((X << C1) >>u ShAmt), see if we can simplify this into a 1491 // single shift. We can do this if the top bits (which are shifted out) 1492 // are never demanded. 1493 if (InOp.getOpcode() == ISD::SHL && 1494 isa<ConstantSDNode>(InOp.getOperand(1))) { 1495 if (ShAmt && (NewMask & APInt::getHighBitsSet(VTSize, ShAmt)) == 0) { 1496 unsigned C1= cast<ConstantSDNode>(InOp.getOperand(1))->getZExtValue(); 1497 unsigned Opc = ISD::SRL; 1498 int Diff = ShAmt-C1; 1499 if (Diff < 0) { 1500 Diff = -Diff; 1501 Opc = ISD::SHL; 1502 } 1503 1504 SDValue NewSA = 1505 TLO.DAG.getConstant(Diff, Op.getOperand(1).getValueType()); 1506 return TLO.CombineTo(Op, TLO.DAG.getNode(Opc, dl, VT, 1507 InOp.getOperand(0), NewSA)); 1508 } 1509 } 1510 1511 // Compute the new bits that are at the top now. 1512 if (SimplifyDemandedBits(InOp, (NewMask << ShAmt), 1513 KnownZero, KnownOne, TLO, Depth+1)) 1514 return true; 1515 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?"); 1516 KnownZero = KnownZero.lshr(ShAmt); 1517 KnownOne = KnownOne.lshr(ShAmt); 1518 1519 APInt HighBits = APInt::getHighBitsSet(BitWidth, ShAmt); 1520 KnownZero |= HighBits; // High bits known zero. 1521 } 1522 break; 1523 case ISD::SRA: 1524 // If this is an arithmetic shift right and only the low-bit is set, we can 1525 // always convert this into a logical shr, even if the shift amount is 1526 // variable. The low bit of the shift cannot be an input sign bit unless 1527 // the shift amount is >= the size of the datatype, which is undefined. 1528 if (NewMask == 1) 1529 return TLO.CombineTo(Op, 1530 TLO.DAG.getNode(ISD::SRL, dl, Op.getValueType(), 1531 Op.getOperand(0), Op.getOperand(1))); 1532 1533 if (ConstantSDNode *SA = dyn_cast<ConstantSDNode>(Op.getOperand(1))) { 1534 EVT VT = Op.getValueType(); 1535 unsigned ShAmt = SA->getZExtValue(); 1536 1537 // If the shift count is an invalid immediate, don't do anything. 1538 if (ShAmt >= BitWidth) 1539 break; 1540 1541 APInt InDemandedMask = (NewMask << ShAmt); 1542 1543 // If any of the demanded bits are produced by the sign extension, we also 1544 // demand the input sign bit. 1545 APInt HighBits = APInt::getHighBitsSet(BitWidth, ShAmt); 1546 if (HighBits.intersects(NewMask)) 1547 InDemandedMask |= APInt::getSignBit(VT.getScalarType().getSizeInBits()); 1548 1549 if (SimplifyDemandedBits(Op.getOperand(0), InDemandedMask, 1550 KnownZero, KnownOne, TLO, Depth+1)) 1551 return true; 1552 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?"); 1553 KnownZero = KnownZero.lshr(ShAmt); 1554 KnownOne = KnownOne.lshr(ShAmt); 1555 1556 // Handle the sign bit, adjusted to where it is now in the mask. 1557 APInt SignBit = APInt::getSignBit(BitWidth).lshr(ShAmt); 1558 1559 // If the input sign bit is known to be zero, or if none of the top bits 1560 // are demanded, turn this into an unsigned shift right. 1561 if (KnownZero.intersects(SignBit) || (HighBits & ~NewMask) == HighBits) { 1562 return TLO.CombineTo(Op, TLO.DAG.getNode(ISD::SRL, dl, VT, 1563 Op.getOperand(0), 1564 Op.getOperand(1))); 1565 } else if (KnownOne.intersects(SignBit)) { // New bits are known one. 1566 KnownOne |= HighBits; 1567 } 1568 } 1569 break; 1570 case ISD::SIGN_EXTEND_INREG: { 1571 EVT ExVT = cast<VTSDNode>(Op.getOperand(1))->getVT(); 1572 1573 APInt MsbMask = APInt::getHighBitsSet(BitWidth, 1); 1574 // If we only care about the highest bit, don't bother shifting right. 1575 if (MsbMask == DemandedMask) { 1576 unsigned ShAmt = ExVT.getScalarType().getSizeInBits(); 1577 SDValue InOp = Op.getOperand(0); 1578 1579 // Compute the correct shift amount type, which must be getShiftAmountTy 1580 // for scalar types after legalization. 1581 EVT ShiftAmtTy = Op.getValueType(); 1582 if (TLO.LegalTypes() && !ShiftAmtTy.isVector()) 1583 ShiftAmtTy = getShiftAmountTy(ShiftAmtTy); 1584 1585 SDValue ShiftAmt = TLO.DAG.getConstant(BitWidth - ShAmt, ShiftAmtTy); 1586 return TLO.CombineTo(Op, TLO.DAG.getNode(ISD::SHL, dl, 1587 Op.getValueType(), InOp, ShiftAmt)); 1588 } 1589 1590 // Sign extension. Compute the demanded bits in the result that are not 1591 // present in the input. 1592 APInt NewBits = 1593 APInt::getHighBitsSet(BitWidth, 1594 BitWidth - ExVT.getScalarType().getSizeInBits()); 1595 1596 // If none of the extended bits are demanded, eliminate the sextinreg. 1597 if ((NewBits & NewMask) == 0) 1598 return TLO.CombineTo(Op, Op.getOperand(0)); 1599 1600 APInt InSignBit = 1601 APInt::getSignBit(ExVT.getScalarType().getSizeInBits()).zext(BitWidth); 1602 APInt InputDemandedBits = 1603 APInt::getLowBitsSet(BitWidth, 1604 ExVT.getScalarType().getSizeInBits()) & 1605 NewMask; 1606 1607 // Since the sign extended bits are demanded, we know that the sign 1608 // bit is demanded. 1609 InputDemandedBits |= InSignBit; 1610 1611 if (SimplifyDemandedBits(Op.getOperand(0), InputDemandedBits, 1612 KnownZero, KnownOne, TLO, Depth+1)) 1613 return true; 1614 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?"); 1615 1616 // If the sign bit of the input is known set or clear, then we know the 1617 // top bits of the result. 1618 1619 // If the input sign bit is known zero, convert this into a zero extension. 1620 if (KnownZero.intersects(InSignBit)) 1621 return TLO.CombineTo(Op, 1622 TLO.DAG.getZeroExtendInReg(Op.getOperand(0),dl,ExVT)); 1623 1624 if (KnownOne.intersects(InSignBit)) { // Input sign bit known set 1625 KnownOne |= NewBits; 1626 KnownZero &= ~NewBits; 1627 } else { // Input sign bit unknown 1628 KnownZero &= ~NewBits; 1629 KnownOne &= ~NewBits; 1630 } 1631 break; 1632 } 1633 case ISD::ZERO_EXTEND: { 1634 unsigned OperandBitWidth = 1635 Op.getOperand(0).getValueType().getScalarType().getSizeInBits(); 1636 APInt InMask = NewMask.trunc(OperandBitWidth); 1637 1638 // If none of the top bits are demanded, convert this into an any_extend. 1639 APInt NewBits = 1640 APInt::getHighBitsSet(BitWidth, BitWidth - OperandBitWidth) & NewMask; 1641 if (!NewBits.intersects(NewMask)) 1642 return TLO.CombineTo(Op, TLO.DAG.getNode(ISD::ANY_EXTEND, dl, 1643 Op.getValueType(), 1644 Op.getOperand(0))); 1645 1646 if (SimplifyDemandedBits(Op.getOperand(0), InMask, 1647 KnownZero, KnownOne, TLO, Depth+1)) 1648 return true; 1649 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?"); 1650 KnownZero = KnownZero.zext(BitWidth); 1651 KnownOne = KnownOne.zext(BitWidth); 1652 KnownZero |= NewBits; 1653 break; 1654 } 1655 case ISD::SIGN_EXTEND: { 1656 EVT InVT = Op.getOperand(0).getValueType(); 1657 unsigned InBits = InVT.getScalarType().getSizeInBits(); 1658 APInt InMask = APInt::getLowBitsSet(BitWidth, InBits); 1659 APInt InSignBit = APInt::getBitsSet(BitWidth, InBits - 1, InBits); 1660 APInt NewBits = ~InMask & NewMask; 1661 1662 // If none of the top bits are demanded, convert this into an any_extend. 1663 if (NewBits == 0) 1664 return TLO.CombineTo(Op,TLO.DAG.getNode(ISD::ANY_EXTEND, dl, 1665 Op.getValueType(), 1666 Op.getOperand(0))); 1667 1668 // Since some of the sign extended bits are demanded, we know that the sign 1669 // bit is demanded. 1670 APInt InDemandedBits = InMask & NewMask; 1671 InDemandedBits |= InSignBit; 1672 InDemandedBits = InDemandedBits.trunc(InBits); 1673 1674 if (SimplifyDemandedBits(Op.getOperand(0), InDemandedBits, KnownZero, 1675 KnownOne, TLO, Depth+1)) 1676 return true; 1677 KnownZero = KnownZero.zext(BitWidth); 1678 KnownOne = KnownOne.zext(BitWidth); 1679 1680 // If the sign bit is known zero, convert this to a zero extend. 1681 if (KnownZero.intersects(InSignBit)) 1682 return TLO.CombineTo(Op, TLO.DAG.getNode(ISD::ZERO_EXTEND, dl, 1683 Op.getValueType(), 1684 Op.getOperand(0))); 1685 1686 // If the sign bit is known one, the top bits match. 1687 if (KnownOne.intersects(InSignBit)) { 1688 KnownOne |= NewBits; 1689 assert((KnownZero & NewBits) == 0); 1690 } else { // Otherwise, top bits aren't known. 1691 assert((KnownOne & NewBits) == 0); 1692 assert((KnownZero & NewBits) == 0); 1693 } 1694 break; 1695 } 1696 case ISD::ANY_EXTEND: { 1697 unsigned OperandBitWidth = 1698 Op.getOperand(0).getValueType().getScalarType().getSizeInBits(); 1699 APInt InMask = NewMask.trunc(OperandBitWidth); 1700 if (SimplifyDemandedBits(Op.getOperand(0), InMask, 1701 KnownZero, KnownOne, TLO, Depth+1)) 1702 return true; 1703 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?"); 1704 KnownZero = KnownZero.zext(BitWidth); 1705 KnownOne = KnownOne.zext(BitWidth); 1706 break; 1707 } 1708 case ISD::TRUNCATE: { 1709 // Simplify the input, using demanded bit information, and compute the known 1710 // zero/one bits live out. 1711 unsigned OperandBitWidth = 1712 Op.getOperand(0).getValueType().getScalarType().getSizeInBits(); 1713 APInt TruncMask = NewMask.zext(OperandBitWidth); 1714 if (SimplifyDemandedBits(Op.getOperand(0), TruncMask, 1715 KnownZero, KnownOne, TLO, Depth+1)) 1716 return true; 1717 KnownZero = KnownZero.trunc(BitWidth); 1718 KnownOne = KnownOne.trunc(BitWidth); 1719 1720 // If the input is only used by this truncate, see if we can shrink it based 1721 // on the known demanded bits. 1722 if (Op.getOperand(0).getNode()->hasOneUse()) { 1723 SDValue In = Op.getOperand(0); 1724 switch (In.getOpcode()) { 1725 default: break; 1726 case ISD::SRL: 1727 // Shrink SRL by a constant if none of the high bits shifted in are 1728 // demanded. 1729 if (TLO.LegalTypes() && 1730 !isTypeDesirableForOp(ISD::SRL, Op.getValueType())) 1731 // Do not turn (vt1 truncate (vt2 srl)) into (vt1 srl) if vt1 is 1732 // undesirable. 1733 break; 1734 ConstantSDNode *ShAmt = dyn_cast<ConstantSDNode>(In.getOperand(1)); 1735 if (!ShAmt) 1736 break; 1737 SDValue Shift = In.getOperand(1); 1738 if (TLO.LegalTypes()) { 1739 uint64_t ShVal = ShAmt->getZExtValue(); 1740 Shift = 1741 TLO.DAG.getConstant(ShVal, getShiftAmountTy(Op.getValueType())); 1742 } 1743 1744 APInt HighBits = APInt::getHighBitsSet(OperandBitWidth, 1745 OperandBitWidth - BitWidth); 1746 HighBits = HighBits.lshr(ShAmt->getZExtValue()).trunc(BitWidth); 1747 1748 if (ShAmt->getZExtValue() < BitWidth && !(HighBits & NewMask)) { 1749 // None of the shifted in bits are needed. Add a truncate of the 1750 // shift input, then shift it. 1751 SDValue NewTrunc = TLO.DAG.getNode(ISD::TRUNCATE, dl, 1752 Op.getValueType(), 1753 In.getOperand(0)); 1754 return TLO.CombineTo(Op, TLO.DAG.getNode(ISD::SRL, dl, 1755 Op.getValueType(), 1756 NewTrunc, 1757 Shift)); 1758 } 1759 break; 1760 } 1761 } 1762 1763 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?"); 1764 break; 1765 } 1766 case ISD::AssertZext: { 1767 // AssertZext demands all of the high bits, plus any of the low bits 1768 // demanded by its users. 1769 EVT VT = cast<VTSDNode>(Op.getOperand(1))->getVT(); 1770 APInt InMask = APInt::getLowBitsSet(BitWidth, 1771 VT.getSizeInBits()); 1772 if (SimplifyDemandedBits(Op.getOperand(0), ~InMask | NewMask, 1773 KnownZero, KnownOne, TLO, Depth+1)) 1774 return true; 1775 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?"); 1776 1777 KnownZero |= ~InMask & NewMask; 1778 break; 1779 } 1780 case ISD::BITCAST: 1781 // If this is an FP->Int bitcast and if the sign bit is the only 1782 // thing demanded, turn this into a FGETSIGN. 1783 if (!TLO.LegalOperations() && 1784 !Op.getValueType().isVector() && 1785 !Op.getOperand(0).getValueType().isVector() && 1786 NewMask == APInt::getSignBit(Op.getValueType().getSizeInBits()) && 1787 Op.getOperand(0).getValueType().isFloatingPoint()) { 1788 bool OpVTLegal = isOperationLegalOrCustom(ISD::FGETSIGN, Op.getValueType()); 1789 bool i32Legal = isOperationLegalOrCustom(ISD::FGETSIGN, MVT::i32); 1790 if ((OpVTLegal || i32Legal) && Op.getValueType().isSimple()) { 1791 EVT Ty = OpVTLegal ? Op.getValueType() : MVT::i32; 1792 // Make a FGETSIGN + SHL to move the sign bit into the appropriate 1793 // place. We expect the SHL to be eliminated by other optimizations. 1794 SDValue Sign = TLO.DAG.getNode(ISD::FGETSIGN, dl, Ty, Op.getOperand(0)); 1795 unsigned OpVTSizeInBits = Op.getValueType().getSizeInBits(); 1796 if (!OpVTLegal && OpVTSizeInBits > 32) 1797 Sign = TLO.DAG.getNode(ISD::ZERO_EXTEND, dl, Op.getValueType(), Sign); 1798 unsigned ShVal = Op.getValueType().getSizeInBits()-1; 1799 SDValue ShAmt = TLO.DAG.getConstant(ShVal, Op.getValueType()); 1800 return TLO.CombineTo(Op, TLO.DAG.getNode(ISD::SHL, dl, 1801 Op.getValueType(), 1802 Sign, ShAmt)); 1803 } 1804 } 1805 break; 1806 case ISD::ADD: 1807 case ISD::MUL: 1808 case ISD::SUB: { 1809 // Add, Sub, and Mul don't demand any bits in positions beyond that 1810 // of the highest bit demanded of them. 1811 APInt LoMask = APInt::getLowBitsSet(BitWidth, 1812 BitWidth - NewMask.countLeadingZeros()); 1813 if (SimplifyDemandedBits(Op.getOperand(0), LoMask, KnownZero2, 1814 KnownOne2, TLO, Depth+1)) 1815 return true; 1816 if (SimplifyDemandedBits(Op.getOperand(1), LoMask, KnownZero2, 1817 KnownOne2, TLO, Depth+1)) 1818 return true; 1819 // See if the operation should be performed at a smaller bit width. 1820 if (TLO.ShrinkDemandedOp(Op, BitWidth, NewMask, dl)) 1821 return true; 1822 } 1823 // FALL THROUGH 1824 default: 1825 // Just use ComputeMaskedBits to compute output bits. 1826 TLO.DAG.ComputeMaskedBits(Op, KnownZero, KnownOne, Depth); 1827 break; 1828 } 1829 1830 // If we know the value of all of the demanded bits, return this as a 1831 // constant. 1832 if ((NewMask & (KnownZero|KnownOne)) == NewMask) 1833 return TLO.CombineTo(Op, TLO.DAG.getConstant(KnownOne, Op.getValueType())); 1834 1835 return false; 1836 } 1837 1838 /// computeMaskedBitsForTargetNode - Determine which of the bits specified 1839 /// in Mask are known to be either zero or one and return them in the 1840 /// KnownZero/KnownOne bitsets. 1841 void TargetLowering::computeMaskedBitsForTargetNode(const SDValue Op, 1842 APInt &KnownZero, 1843 APInt &KnownOne, 1844 const SelectionDAG &DAG, 1845 unsigned Depth) const { 1846 assert((Op.getOpcode() >= ISD::BUILTIN_OP_END || 1847 Op.getOpcode() == ISD::INTRINSIC_WO_CHAIN || 1848 Op.getOpcode() == ISD::INTRINSIC_W_CHAIN || 1849 Op.getOpcode() == ISD::INTRINSIC_VOID) && 1850 "Should use MaskedValueIsZero if you don't know whether Op" 1851 " is a target node!"); 1852 KnownZero = KnownOne = APInt(KnownOne.getBitWidth(), 0); 1853 } 1854 1855 /// ComputeNumSignBitsForTargetNode - This method can be implemented by 1856 /// targets that want to expose additional information about sign bits to the 1857 /// DAG Combiner. 1858 unsigned TargetLowering::ComputeNumSignBitsForTargetNode(SDValue Op, 1859 unsigned Depth) const { 1860 assert((Op.getOpcode() >= ISD::BUILTIN_OP_END || 1861 Op.getOpcode() == ISD::INTRINSIC_WO_CHAIN || 1862 Op.getOpcode() == ISD::INTRINSIC_W_CHAIN || 1863 Op.getOpcode() == ISD::INTRINSIC_VOID) && 1864 "Should use ComputeNumSignBits if you don't know whether Op" 1865 " is a target node!"); 1866 return 1; 1867 } 1868 1869 /// ValueHasExactlyOneBitSet - Test if the given value is known to have exactly 1870 /// one bit set. This differs from ComputeMaskedBits in that it doesn't need to 1871 /// determine which bit is set. 1872 /// 1873 static bool ValueHasExactlyOneBitSet(SDValue Val, const SelectionDAG &DAG) { 1874 // A left-shift of a constant one will have exactly one bit set, because 1875 // shifting the bit off the end is undefined. 1876 if (Val.getOpcode() == ISD::SHL) 1877 if (ConstantSDNode *C = 1878 dyn_cast<ConstantSDNode>(Val.getNode()->getOperand(0))) 1879 if (C->getAPIntValue() == 1) 1880 return true; 1881 1882 // Similarly, a right-shift of a constant sign-bit will have exactly 1883 // one bit set. 1884 if (Val.getOpcode() == ISD::SRL) 1885 if (ConstantSDNode *C = 1886 dyn_cast<ConstantSDNode>(Val.getNode()->getOperand(0))) 1887 if (C->getAPIntValue().isSignBit()) 1888 return true; 1889 1890 // More could be done here, though the above checks are enough 1891 // to handle some common cases. 1892 1893 // Fall back to ComputeMaskedBits to catch other known cases. 1894 EVT OpVT = Val.getValueType(); 1895 unsigned BitWidth = OpVT.getScalarType().getSizeInBits(); 1896 APInt KnownZero, KnownOne; 1897 DAG.ComputeMaskedBits(Val, KnownZero, KnownOne); 1898 return (KnownZero.countPopulation() == BitWidth - 1) && 1899 (KnownOne.countPopulation() == 1); 1900 } 1901 1902 /// SimplifySetCC - Try to simplify a setcc built with the specified operands 1903 /// and cc. If it is unable to simplify it, return a null SDValue. 1904 SDValue 1905 TargetLowering::SimplifySetCC(EVT VT, SDValue N0, SDValue N1, 1906 ISD::CondCode Cond, bool foldBooleans, 1907 DAGCombinerInfo &DCI, DebugLoc dl) const { 1908 SelectionDAG &DAG = DCI.DAG; 1909 1910 // These setcc operations always fold. 1911 switch (Cond) { 1912 default: break; 1913 case ISD::SETFALSE: 1914 case ISD::SETFALSE2: return DAG.getConstant(0, VT); 1915 case ISD::SETTRUE: 1916 case ISD::SETTRUE2: return DAG.getConstant(1, VT); 1917 } 1918 1919 // Ensure that the constant occurs on the RHS, and fold constant 1920 // comparisons. 1921 if (isa<ConstantSDNode>(N0.getNode())) 1922 return DAG.getSetCC(dl, VT, N1, N0, ISD::getSetCCSwappedOperands(Cond)); 1923 1924 if (ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1.getNode())) { 1925 const APInt &C1 = N1C->getAPIntValue(); 1926 1927 // If the LHS is '(srl (ctlz x), 5)', the RHS is 0/1, and this is an 1928 // equality comparison, then we're just comparing whether X itself is 1929 // zero. 1930 if (N0.getOpcode() == ISD::SRL && (C1 == 0 || C1 == 1) && 1931 N0.getOperand(0).getOpcode() == ISD::CTLZ && 1932 N0.getOperand(1).getOpcode() == ISD::Constant) { 1933 const APInt &ShAmt 1934 = cast<ConstantSDNode>(N0.getOperand(1))->getAPIntValue(); 1935 if ((Cond == ISD::SETEQ || Cond == ISD::SETNE) && 1936 ShAmt == Log2_32(N0.getValueType().getSizeInBits())) { 1937 if ((C1 == 0) == (Cond == ISD::SETEQ)) { 1938 // (srl (ctlz x), 5) == 0 -> X != 0 1939 // (srl (ctlz x), 5) != 1 -> X != 0 1940 Cond = ISD::SETNE; 1941 } else { 1942 // (srl (ctlz x), 5) != 0 -> X == 0 1943 // (srl (ctlz x), 5) == 1 -> X == 0 1944 Cond = ISD::SETEQ; 1945 } 1946 SDValue Zero = DAG.getConstant(0, N0.getValueType()); 1947 return DAG.getSetCC(dl, VT, N0.getOperand(0).getOperand(0), 1948 Zero, Cond); 1949 } 1950 } 1951 1952 SDValue CTPOP = N0; 1953 // Look through truncs that don't change the value of a ctpop. 1954 if (N0.hasOneUse() && N0.getOpcode() == ISD::TRUNCATE) 1955 CTPOP = N0.getOperand(0); 1956 1957 if (CTPOP.hasOneUse() && CTPOP.getOpcode() == ISD::CTPOP && 1958 (N0 == CTPOP || N0.getValueType().getSizeInBits() > 1959 Log2_32_Ceil(CTPOP.getValueType().getSizeInBits()))) { 1960 EVT CTVT = CTPOP.getValueType(); 1961 SDValue CTOp = CTPOP.getOperand(0); 1962 1963 // (ctpop x) u< 2 -> (x & x-1) == 0 1964 // (ctpop x) u> 1 -> (x & x-1) != 0 1965 if ((Cond == ISD::SETULT && C1 == 2) || (Cond == ISD::SETUGT && C1 == 1)){ 1966 SDValue Sub = DAG.getNode(ISD::SUB, dl, CTVT, CTOp, 1967 DAG.getConstant(1, CTVT)); 1968 SDValue And = DAG.getNode(ISD::AND, dl, CTVT, CTOp, Sub); 1969 ISD::CondCode CC = Cond == ISD::SETULT ? ISD::SETEQ : ISD::SETNE; 1970 return DAG.getSetCC(dl, VT, And, DAG.getConstant(0, CTVT), CC); 1971 } 1972 1973 // TODO: (ctpop x) == 1 -> x && (x & x-1) == 0 iff ctpop is illegal. 1974 } 1975 1976 // (zext x) == C --> x == (trunc C) 1977 if (DCI.isBeforeLegalize() && N0->hasOneUse() && 1978 (Cond == ISD::SETEQ || Cond == ISD::SETNE)) { 1979 unsigned MinBits = N0.getValueSizeInBits(); 1980 SDValue PreZExt; 1981 if (N0->getOpcode() == ISD::ZERO_EXTEND) { 1982 // ZExt 1983 MinBits = N0->getOperand(0).getValueSizeInBits(); 1984 PreZExt = N0->getOperand(0); 1985 } else if (N0->getOpcode() == ISD::AND) { 1986 // DAGCombine turns costly ZExts into ANDs 1987 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(N0->getOperand(1))) 1988 if ((C->getAPIntValue()+1).isPowerOf2()) { 1989 MinBits = C->getAPIntValue().countTrailingOnes(); 1990 PreZExt = N0->getOperand(0); 1991 } 1992 } else if (LoadSDNode *LN0 = dyn_cast<LoadSDNode>(N0)) { 1993 // ZEXTLOAD 1994 if (LN0->getExtensionType() == ISD::ZEXTLOAD) { 1995 MinBits = LN0->getMemoryVT().getSizeInBits(); 1996 PreZExt = N0; 1997 } 1998 } 1999 2000 // Make sure we're not losing bits from the constant. 2001 if (MinBits < C1.getBitWidth() && MinBits > C1.getActiveBits()) { 2002 EVT MinVT = EVT::getIntegerVT(*DAG.getContext(), MinBits); 2003 if (isTypeDesirableForOp(ISD::SETCC, MinVT)) { 2004 // Will get folded away. 2005 SDValue Trunc = DAG.getNode(ISD::TRUNCATE, dl, MinVT, PreZExt); 2006 SDValue C = DAG.getConstant(C1.trunc(MinBits), MinVT); 2007 return DAG.getSetCC(dl, VT, Trunc, C, Cond); 2008 } 2009 } 2010 } 2011 2012 // If the LHS is '(and load, const)', the RHS is 0, 2013 // the test is for equality or unsigned, and all 1 bits of the const are 2014 // in the same partial word, see if we can shorten the load. 2015 if (DCI.isBeforeLegalize() && 2016 N0.getOpcode() == ISD::AND && C1 == 0 && 2017 N0.getNode()->hasOneUse() && 2018 isa<LoadSDNode>(N0.getOperand(0)) && 2019 N0.getOperand(0).getNode()->hasOneUse() && 2020 isa<ConstantSDNode>(N0.getOperand(1))) { 2021 LoadSDNode *Lod = cast<LoadSDNode>(N0.getOperand(0)); 2022 APInt bestMask; 2023 unsigned bestWidth = 0, bestOffset = 0; 2024 if (!Lod->isVolatile() && Lod->isUnindexed()) { 2025 unsigned origWidth = N0.getValueType().getSizeInBits(); 2026 unsigned maskWidth = origWidth; 2027 // We can narrow (e.g.) 16-bit extending loads on 32-bit target to 2028 // 8 bits, but have to be careful... 2029 if (Lod->getExtensionType() != ISD::NON_EXTLOAD) 2030 origWidth = Lod->getMemoryVT().getSizeInBits(); 2031 const APInt &Mask = 2032 cast<ConstantSDNode>(N0.getOperand(1))->getAPIntValue(); 2033 for (unsigned width = origWidth / 2; width>=8; width /= 2) { 2034 APInt newMask = APInt::getLowBitsSet(maskWidth, width); 2035 for (unsigned offset=0; offset<origWidth/width; offset++) { 2036 if ((newMask & Mask) == Mask) { 2037 if (!TD->isLittleEndian()) 2038 bestOffset = (origWidth/width - offset - 1) * (width/8); 2039 else 2040 bestOffset = (uint64_t)offset * (width/8); 2041 bestMask = Mask.lshr(offset * (width/8) * 8); 2042 bestWidth = width; 2043 break; 2044 } 2045 newMask = newMask << width; 2046 } 2047 } 2048 } 2049 if (bestWidth) { 2050 EVT newVT = EVT::getIntegerVT(*DAG.getContext(), bestWidth); 2051 if (newVT.isRound()) { 2052 EVT PtrType = Lod->getOperand(1).getValueType(); 2053 SDValue Ptr = Lod->getBasePtr(); 2054 if (bestOffset != 0) 2055 Ptr = DAG.getNode(ISD::ADD, dl, PtrType, Lod->getBasePtr(), 2056 DAG.getConstant(bestOffset, PtrType)); 2057 unsigned NewAlign = MinAlign(Lod->getAlignment(), bestOffset); 2058 SDValue NewLoad = DAG.getLoad(newVT, dl, Lod->getChain(), Ptr, 2059 Lod->getPointerInfo().getWithOffset(bestOffset), 2060 false, false, false, NewAlign); 2061 return DAG.getSetCC(dl, VT, 2062 DAG.getNode(ISD::AND, dl, newVT, NewLoad, 2063 DAG.getConstant(bestMask.trunc(bestWidth), 2064 newVT)), 2065 DAG.getConstant(0LL, newVT), Cond); 2066 } 2067 } 2068 } 2069 2070 // If the LHS is a ZERO_EXTEND, perform the comparison on the input. 2071 if (N0.getOpcode() == ISD::ZERO_EXTEND) { 2072 unsigned InSize = N0.getOperand(0).getValueType().getSizeInBits(); 2073 2074 // If the comparison constant has bits in the upper part, the 2075 // zero-extended value could never match. 2076 if (C1.intersects(APInt::getHighBitsSet(C1.getBitWidth(), 2077 C1.getBitWidth() - InSize))) { 2078 switch (Cond) { 2079 case ISD::SETUGT: 2080 case ISD::SETUGE: 2081 case ISD::SETEQ: return DAG.getConstant(0, VT); 2082 case ISD::SETULT: 2083 case ISD::SETULE: 2084 case ISD::SETNE: return DAG.getConstant(1, VT); 2085 case ISD::SETGT: 2086 case ISD::SETGE: 2087 // True if the sign bit of C1 is set. 2088 return DAG.getConstant(C1.isNegative(), VT); 2089 case ISD::SETLT: 2090 case ISD::SETLE: 2091 // True if the sign bit of C1 isn't set. 2092 return DAG.getConstant(C1.isNonNegative(), VT); 2093 default: 2094 break; 2095 } 2096 } 2097 2098 // Otherwise, we can perform the comparison with the low bits. 2099 switch (Cond) { 2100 case ISD::SETEQ: 2101 case ISD::SETNE: 2102 case ISD::SETUGT: 2103 case ISD::SETUGE: 2104 case ISD::SETULT: 2105 case ISD::SETULE: { 2106 EVT newVT = N0.getOperand(0).getValueType(); 2107 if (DCI.isBeforeLegalizeOps() || 2108 (isOperationLegal(ISD::SETCC, newVT) && 2109 getCondCodeAction(Cond, newVT)==Legal)) 2110 return DAG.getSetCC(dl, VT, N0.getOperand(0), 2111 DAG.getConstant(C1.trunc(InSize), newVT), 2112 Cond); 2113 break; 2114 } 2115 default: 2116 break; // todo, be more careful with signed comparisons 2117 } 2118 } else if (N0.getOpcode() == ISD::SIGN_EXTEND_INREG && 2119 (Cond == ISD::SETEQ || Cond == ISD::SETNE)) { 2120 EVT ExtSrcTy = cast<VTSDNode>(N0.getOperand(1))->getVT(); 2121 unsigned ExtSrcTyBits = ExtSrcTy.getSizeInBits(); 2122 EVT ExtDstTy = N0.getValueType(); 2123 unsigned ExtDstTyBits = ExtDstTy.getSizeInBits(); 2124 2125 // If the constant doesn't fit into the number of bits for the source of 2126 // the sign extension, it is impossible for both sides to be equal. 2127 if (C1.getMinSignedBits() > ExtSrcTyBits) 2128 return DAG.getConstant(Cond == ISD::SETNE, VT); 2129 2130 SDValue ZextOp; 2131 EVT Op0Ty = N0.getOperand(0).getValueType(); 2132 if (Op0Ty == ExtSrcTy) { 2133 ZextOp = N0.getOperand(0); 2134 } else { 2135 APInt Imm = APInt::getLowBitsSet(ExtDstTyBits, ExtSrcTyBits); 2136 ZextOp = DAG.getNode(ISD::AND, dl, Op0Ty, N0.getOperand(0), 2137 DAG.getConstant(Imm, Op0Ty)); 2138 } 2139 if (!DCI.isCalledByLegalizer()) 2140 DCI.AddToWorklist(ZextOp.getNode()); 2141 // Otherwise, make this a use of a zext. 2142 return DAG.getSetCC(dl, VT, ZextOp, 2143 DAG.getConstant(C1 & APInt::getLowBitsSet( 2144 ExtDstTyBits, 2145 ExtSrcTyBits), 2146 ExtDstTy), 2147 Cond); 2148 } else if ((N1C->isNullValue() || N1C->getAPIntValue() == 1) && 2149 (Cond == ISD::SETEQ || Cond == ISD::SETNE)) { 2150 // SETCC (SETCC), [0|1], [EQ|NE] -> SETCC 2151 if (N0.getOpcode() == ISD::SETCC && 2152 isTypeLegal(VT) && VT.bitsLE(N0.getValueType())) { 2153 bool TrueWhenTrue = (Cond == ISD::SETEQ) ^ (N1C->getAPIntValue() != 1); 2154 if (TrueWhenTrue) 2155 return DAG.getNode(ISD::TRUNCATE, dl, VT, N0); 2156 // Invert the condition. 2157 ISD::CondCode CC = cast<CondCodeSDNode>(N0.getOperand(2))->get(); 2158 CC = ISD::getSetCCInverse(CC, 2159 N0.getOperand(0).getValueType().isInteger()); 2160 return DAG.getSetCC(dl, VT, N0.getOperand(0), N0.getOperand(1), CC); 2161 } 2162 2163 if ((N0.getOpcode() == ISD::XOR || 2164 (N0.getOpcode() == ISD::AND && 2165 N0.getOperand(0).getOpcode() == ISD::XOR && 2166 N0.getOperand(1) == N0.getOperand(0).getOperand(1))) && 2167 isa<ConstantSDNode>(N0.getOperand(1)) && 2168 cast<ConstantSDNode>(N0.getOperand(1))->getAPIntValue() == 1) { 2169 // If this is (X^1) == 0/1, swap the RHS and eliminate the xor. We 2170 // can only do this if the top bits are known zero. 2171 unsigned BitWidth = N0.getValueSizeInBits(); 2172 if (DAG.MaskedValueIsZero(N0, 2173 APInt::getHighBitsSet(BitWidth, 2174 BitWidth-1))) { 2175 // Okay, get the un-inverted input value. 2176 SDValue Val; 2177 if (N0.getOpcode() == ISD::XOR) 2178 Val = N0.getOperand(0); 2179 else { 2180 assert(N0.getOpcode() == ISD::AND && 2181 N0.getOperand(0).getOpcode() == ISD::XOR); 2182 // ((X^1)&1)^1 -> X & 1 2183 Val = DAG.getNode(ISD::AND, dl, N0.getValueType(), 2184 N0.getOperand(0).getOperand(0), 2185 N0.getOperand(1)); 2186 } 2187 2188 return DAG.getSetCC(dl, VT, Val, N1, 2189 Cond == ISD::SETEQ ? ISD::SETNE : ISD::SETEQ); 2190 } 2191 } else if (N1C->getAPIntValue() == 1 && 2192 (VT == MVT::i1 || 2193 getBooleanContents(false) == ZeroOrOneBooleanContent)) { 2194 SDValue Op0 = N0; 2195 if (Op0.getOpcode() == ISD::TRUNCATE) 2196 Op0 = Op0.getOperand(0); 2197 2198 if ((Op0.getOpcode() == ISD::XOR) && 2199 Op0.getOperand(0).getOpcode() == ISD::SETCC && 2200 Op0.getOperand(1).getOpcode() == ISD::SETCC) { 2201 // (xor (setcc), (setcc)) == / != 1 -> (setcc) != / == (setcc) 2202 Cond = (Cond == ISD::SETEQ) ? ISD::SETNE : ISD::SETEQ; 2203 return DAG.getSetCC(dl, VT, Op0.getOperand(0), Op0.getOperand(1), 2204 Cond); 2205 } else if (Op0.getOpcode() == ISD::AND && 2206 isa<ConstantSDNode>(Op0.getOperand(1)) && 2207 cast<ConstantSDNode>(Op0.getOperand(1))->getAPIntValue() == 1) { 2208 // If this is (X&1) == / != 1, normalize it to (X&1) != / == 0. 2209 if (Op0.getValueType().bitsGT(VT)) 2210 Op0 = DAG.getNode(ISD::AND, dl, VT, 2211 DAG.getNode(ISD::TRUNCATE, dl, VT, Op0.getOperand(0)), 2212 DAG.getConstant(1, VT)); 2213 else if (Op0.getValueType().bitsLT(VT)) 2214 Op0 = DAG.getNode(ISD::AND, dl, VT, 2215 DAG.getNode(ISD::ANY_EXTEND, dl, VT, Op0.getOperand(0)), 2216 DAG.getConstant(1, VT)); 2217 2218 return DAG.getSetCC(dl, VT, Op0, 2219 DAG.getConstant(0, Op0.getValueType()), 2220 Cond == ISD::SETEQ ? ISD::SETNE : ISD::SETEQ); 2221 } 2222 } 2223 } 2224 2225 APInt MinVal, MaxVal; 2226 unsigned OperandBitSize = N1C->getValueType(0).getSizeInBits(); 2227 if (ISD::isSignedIntSetCC(Cond)) { 2228 MinVal = APInt::getSignedMinValue(OperandBitSize); 2229 MaxVal = APInt::getSignedMaxValue(OperandBitSize); 2230 } else { 2231 MinVal = APInt::getMinValue(OperandBitSize); 2232 MaxVal = APInt::getMaxValue(OperandBitSize); 2233 } 2234 2235 // Canonicalize GE/LE comparisons to use GT/LT comparisons. 2236 if (Cond == ISD::SETGE || Cond == ISD::SETUGE) { 2237 if (C1 == MinVal) return DAG.getConstant(1, VT); // X >= MIN --> true 2238 // X >= C0 --> X > (C0-1) 2239 return DAG.getSetCC(dl, VT, N0, 2240 DAG.getConstant(C1-1, N1.getValueType()), 2241 (Cond == ISD::SETGE) ? ISD::SETGT : ISD::SETUGT); 2242 } 2243 2244 if (Cond == ISD::SETLE || Cond == ISD::SETULE) { 2245 if (C1 == MaxVal) return DAG.getConstant(1, VT); // X <= MAX --> true 2246 // X <= C0 --> X < (C0+1) 2247 return DAG.getSetCC(dl, VT, N0, 2248 DAG.getConstant(C1+1, N1.getValueType()), 2249 (Cond == ISD::SETLE) ? ISD::SETLT : ISD::SETULT); 2250 } 2251 2252 if ((Cond == ISD::SETLT || Cond == ISD::SETULT) && C1 == MinVal) 2253 return DAG.getConstant(0, VT); // X < MIN --> false 2254 if ((Cond == ISD::SETGE || Cond == ISD::SETUGE) && C1 == MinVal) 2255 return DAG.getConstant(1, VT); // X >= MIN --> true 2256 if ((Cond == ISD::SETGT || Cond == ISD::SETUGT) && C1 == MaxVal) 2257 return DAG.getConstant(0, VT); // X > MAX --> false 2258 if ((Cond == ISD::SETLE || Cond == ISD::SETULE) && C1 == MaxVal) 2259 return DAG.getConstant(1, VT); // X <= MAX --> true 2260 2261 // Canonicalize setgt X, Min --> setne X, Min 2262 if ((Cond == ISD::SETGT || Cond == ISD::SETUGT) && C1 == MinVal) 2263 return DAG.getSetCC(dl, VT, N0, N1, ISD::SETNE); 2264 // Canonicalize setlt X, Max --> setne X, Max 2265 if ((Cond == ISD::SETLT || Cond == ISD::SETULT) && C1 == MaxVal) 2266 return DAG.getSetCC(dl, VT, N0, N1, ISD::SETNE); 2267 2268 // If we have setult X, 1, turn it into seteq X, 0 2269 if ((Cond == ISD::SETLT || Cond == ISD::SETULT) && C1 == MinVal+1) 2270 return DAG.getSetCC(dl, VT, N0, 2271 DAG.getConstant(MinVal, N0.getValueType()), 2272 ISD::SETEQ); 2273 // If we have setugt X, Max-1, turn it into seteq X, Max 2274 else if ((Cond == ISD::SETGT || Cond == ISD::SETUGT) && C1 == MaxVal-1) 2275 return DAG.getSetCC(dl, VT, N0, 2276 DAG.getConstant(MaxVal, N0.getValueType()), 2277 ISD::SETEQ); 2278 2279 // If we have "setcc X, C0", check to see if we can shrink the immediate 2280 // by changing cc. 2281 2282 // SETUGT X, SINTMAX -> SETLT X, 0 2283 if (Cond == ISD::SETUGT && 2284 C1 == APInt::getSignedMaxValue(OperandBitSize)) 2285 return DAG.getSetCC(dl, VT, N0, 2286 DAG.getConstant(0, N1.getValueType()), 2287 ISD::SETLT); 2288 2289 // SETULT X, SINTMIN -> SETGT X, -1 2290 if (Cond == ISD::SETULT && 2291 C1 == APInt::getSignedMinValue(OperandBitSize)) { 2292 SDValue ConstMinusOne = 2293 DAG.getConstant(APInt::getAllOnesValue(OperandBitSize), 2294 N1.getValueType()); 2295 return DAG.getSetCC(dl, VT, N0, ConstMinusOne, ISD::SETGT); 2296 } 2297 2298 // Fold bit comparisons when we can. 2299 if ((Cond == ISD::SETEQ || Cond == ISD::SETNE) && 2300 (VT == N0.getValueType() || 2301 (isTypeLegal(VT) && VT.bitsLE(N0.getValueType()))) && 2302 N0.getOpcode() == ISD::AND) 2303 if (ConstantSDNode *AndRHS = 2304 dyn_cast<ConstantSDNode>(N0.getOperand(1))) { 2305 EVT ShiftTy = DCI.isBeforeLegalizeOps() ? 2306 getPointerTy() : getShiftAmountTy(N0.getValueType()); 2307 if (Cond == ISD::SETNE && C1 == 0) {// (X & 8) != 0 --> (X & 8) >> 3 2308 // Perform the xform if the AND RHS is a single bit. 2309 if (AndRHS->getAPIntValue().isPowerOf2()) { 2310 return DAG.getNode(ISD::TRUNCATE, dl, VT, 2311 DAG.getNode(ISD::SRL, dl, N0.getValueType(), N0, 2312 DAG.getConstant(AndRHS->getAPIntValue().logBase2(), ShiftTy))); 2313 } 2314 } else if (Cond == ISD::SETEQ && C1 == AndRHS->getAPIntValue()) { 2315 // (X & 8) == 8 --> (X & 8) >> 3 2316 // Perform the xform if C1 is a single bit. 2317 if (C1.isPowerOf2()) { 2318 return DAG.getNode(ISD::TRUNCATE, dl, VT, 2319 DAG.getNode(ISD::SRL, dl, N0.getValueType(), N0, 2320 DAG.getConstant(C1.logBase2(), ShiftTy))); 2321 } 2322 } 2323 } 2324 2325 if (C1.getMinSignedBits() <= 64 && 2326 !isLegalICmpImmediate(C1.getSExtValue())) { 2327 // (X & -256) == 256 -> (X >> 8) == 1 2328 if ((Cond == ISD::SETEQ || Cond == ISD::SETNE) && 2329 N0.getOpcode() == ISD::AND && N0.hasOneUse()) { 2330 if (ConstantSDNode *AndRHS = 2331 dyn_cast<ConstantSDNode>(N0.getOperand(1))) { 2332 const APInt &AndRHSC = AndRHS->getAPIntValue(); 2333 if ((-AndRHSC).isPowerOf2() && (AndRHSC & C1) == C1) { 2334 unsigned ShiftBits = AndRHSC.countTrailingZeros(); 2335 EVT ShiftTy = DCI.isBeforeLegalizeOps() ? 2336 getPointerTy() : getShiftAmountTy(N0.getValueType()); 2337 EVT CmpTy = N0.getValueType(); 2338 SDValue Shift = DAG.getNode(ISD::SRL, dl, CmpTy, N0.getOperand(0), 2339 DAG.getConstant(ShiftBits, ShiftTy)); 2340 SDValue CmpRHS = DAG.getConstant(C1.lshr(ShiftBits), CmpTy); 2341 return DAG.getSetCC(dl, VT, Shift, CmpRHS, Cond); 2342 } 2343 } 2344 } else if (Cond == ISD::SETULT || Cond == ISD::SETUGE || 2345 Cond == ISD::SETULE || Cond == ISD::SETUGT) { 2346 bool AdjOne = (Cond == ISD::SETULE || Cond == ISD::SETUGT); 2347 // X < 0x100000000 -> (X >> 32) < 1 2348 // X >= 0x100000000 -> (X >> 32) >= 1 2349 // X <= 0x0ffffffff -> (X >> 32) < 1 2350 // X > 0x0ffffffff -> (X >> 32) >= 1 2351 unsigned ShiftBits; 2352 APInt NewC = C1; 2353 ISD::CondCode NewCond = Cond; 2354 if (AdjOne) { 2355 ShiftBits = C1.countTrailingOnes(); 2356 NewC = NewC + 1; 2357 NewCond = (Cond == ISD::SETULE) ? ISD::SETULT : ISD::SETUGE; 2358 } else { 2359 ShiftBits = C1.countTrailingZeros(); 2360 } 2361 NewC = NewC.lshr(ShiftBits); 2362 if (ShiftBits && isLegalICmpImmediate(NewC.getSExtValue())) { 2363 EVT ShiftTy = DCI.isBeforeLegalizeOps() ? 2364 getPointerTy() : getShiftAmountTy(N0.getValueType()); 2365 EVT CmpTy = N0.getValueType(); 2366 SDValue Shift = DAG.getNode(ISD::SRL, dl, CmpTy, N0, 2367 DAG.getConstant(ShiftBits, ShiftTy)); 2368 SDValue CmpRHS = DAG.getConstant(NewC, CmpTy); 2369 return DAG.getSetCC(dl, VT, Shift, CmpRHS, NewCond); 2370 } 2371 } 2372 } 2373 } 2374 2375 if (isa<ConstantFPSDNode>(N0.getNode())) { 2376 // Constant fold or commute setcc. 2377 SDValue O = DAG.FoldSetCC(VT, N0, N1, Cond, dl); 2378 if (O.getNode()) return O; 2379 } else if (ConstantFPSDNode *CFP = dyn_cast<ConstantFPSDNode>(N1.getNode())) { 2380 // If the RHS of an FP comparison is a constant, simplify it away in 2381 // some cases. 2382 if (CFP->getValueAPF().isNaN()) { 2383 // If an operand is known to be a nan, we can fold it. 2384 switch (ISD::getUnorderedFlavor(Cond)) { 2385 default: llvm_unreachable("Unknown flavor!"); 2386 case 0: // Known false. 2387 return DAG.getConstant(0, VT); 2388 case 1: // Known true. 2389 return DAG.getConstant(1, VT); 2390 case 2: // Undefined. 2391 return DAG.getUNDEF(VT); 2392 } 2393 } 2394 2395 // Otherwise, we know the RHS is not a NaN. Simplify the node to drop the 2396 // constant if knowing that the operand is non-nan is enough. We prefer to 2397 // have SETO(x,x) instead of SETO(x, 0.0) because this avoids having to 2398 // materialize 0.0. 2399 if (Cond == ISD::SETO || Cond == ISD::SETUO) 2400 return DAG.getSetCC(dl, VT, N0, N0, Cond); 2401 2402 // If the condition is not legal, see if we can find an equivalent one 2403 // which is legal. 2404 if (!isCondCodeLegal(Cond, N0.getValueType())) { 2405 // If the comparison was an awkward floating-point == or != and one of 2406 // the comparison operands is infinity or negative infinity, convert the 2407 // condition to a less-awkward <= or >=. 2408 if (CFP->getValueAPF().isInfinity()) { 2409 if (CFP->getValueAPF().isNegative()) { 2410 if (Cond == ISD::SETOEQ && 2411 isCondCodeLegal(ISD::SETOLE, N0.getValueType())) 2412 return DAG.getSetCC(dl, VT, N0, N1, ISD::SETOLE); 2413 if (Cond == ISD::SETUEQ && 2414 isCondCodeLegal(ISD::SETOLE, N0.getValueType())) 2415 return DAG.getSetCC(dl, VT, N0, N1, ISD::SETULE); 2416 if (Cond == ISD::SETUNE && 2417 isCondCodeLegal(ISD::SETUGT, N0.getValueType())) 2418 return DAG.getSetCC(dl, VT, N0, N1, ISD::SETUGT); 2419 if (Cond == ISD::SETONE && 2420 isCondCodeLegal(ISD::SETUGT, N0.getValueType())) 2421 return DAG.getSetCC(dl, VT, N0, N1, ISD::SETOGT); 2422 } else { 2423 if (Cond == ISD::SETOEQ && 2424 isCondCodeLegal(ISD::SETOGE, N0.getValueType())) 2425 return DAG.getSetCC(dl, VT, N0, N1, ISD::SETOGE); 2426 if (Cond == ISD::SETUEQ && 2427 isCondCodeLegal(ISD::SETOGE, N0.getValueType())) 2428 return DAG.getSetCC(dl, VT, N0, N1, ISD::SETUGE); 2429 if (Cond == ISD::SETUNE && 2430 isCondCodeLegal(ISD::SETULT, N0.getValueType())) 2431 return DAG.getSetCC(dl, VT, N0, N1, ISD::SETULT); 2432 if (Cond == ISD::SETONE && 2433 isCondCodeLegal(ISD::SETULT, N0.getValueType())) 2434 return DAG.getSetCC(dl, VT, N0, N1, ISD::SETOLT); 2435 } 2436 } 2437 } 2438 } 2439 2440 if (N0 == N1) { 2441 // The sext(setcc()) => setcc() optimization relies on the appropriate 2442 // constant being emitted. 2443 uint64_t EqVal = 0; 2444 switch (getBooleanContents(N0.getValueType().isVector())) { 2445 case UndefinedBooleanContent: 2446 case ZeroOrOneBooleanContent: 2447 EqVal = ISD::isTrueWhenEqual(Cond); 2448 break; 2449 case ZeroOrNegativeOneBooleanContent: 2450 EqVal = ISD::isTrueWhenEqual(Cond) ? -1 : 0; 2451 break; 2452 } 2453 2454 // We can always fold X == X for integer setcc's. 2455 if (N0.getValueType().isInteger()) { 2456 return DAG.getConstant(EqVal, VT); 2457 } 2458 unsigned UOF = ISD::getUnorderedFlavor(Cond); 2459 if (UOF == 2) // FP operators that are undefined on NaNs. 2460 return DAG.getConstant(EqVal, VT); 2461 if (UOF == unsigned(ISD::isTrueWhenEqual(Cond))) 2462 return DAG.getConstant(EqVal, VT); 2463 // Otherwise, we can't fold it. However, we can simplify it to SETUO/SETO 2464 // if it is not already. 2465 ISD::CondCode NewCond = UOF == 0 ? ISD::SETO : ISD::SETUO; 2466 if (NewCond != Cond && (DCI.isBeforeLegalizeOps() || 2467 getCondCodeAction(NewCond, N0.getValueType()) == Legal)) 2468 return DAG.getSetCC(dl, VT, N0, N1, NewCond); 2469 } 2470 2471 if ((Cond == ISD::SETEQ || Cond == ISD::SETNE) && 2472 N0.getValueType().isInteger()) { 2473 if (N0.getOpcode() == ISD::ADD || N0.getOpcode() == ISD::SUB || 2474 N0.getOpcode() == ISD::XOR) { 2475 // Simplify (X+Y) == (X+Z) --> Y == Z 2476 if (N0.getOpcode() == N1.getOpcode()) { 2477 if (N0.getOperand(0) == N1.getOperand(0)) 2478 return DAG.getSetCC(dl, VT, N0.getOperand(1), N1.getOperand(1), Cond); 2479 if (N0.getOperand(1) == N1.getOperand(1)) 2480 return DAG.getSetCC(dl, VT, N0.getOperand(0), N1.getOperand(0), Cond); 2481 if (DAG.isCommutativeBinOp(N0.getOpcode())) { 2482 // If X op Y == Y op X, try other combinations. 2483 if (N0.getOperand(0) == N1.getOperand(1)) 2484 return DAG.getSetCC(dl, VT, N0.getOperand(1), N1.getOperand(0), 2485 Cond); 2486 if (N0.getOperand(1) == N1.getOperand(0)) 2487 return DAG.getSetCC(dl, VT, N0.getOperand(0), N1.getOperand(1), 2488 Cond); 2489 } 2490 } 2491 2492 // If RHS is a legal immediate value for a compare instruction, we need 2493 // to be careful about increasing register pressure needlessly. 2494 bool LegalRHSImm = false; 2495 2496 if (ConstantSDNode *RHSC = dyn_cast<ConstantSDNode>(N1)) { 2497 if (ConstantSDNode *LHSR = dyn_cast<ConstantSDNode>(N0.getOperand(1))) { 2498 // Turn (X+C1) == C2 --> X == C2-C1 2499 if (N0.getOpcode() == ISD::ADD && N0.getNode()->hasOneUse()) { 2500 return DAG.getSetCC(dl, VT, N0.getOperand(0), 2501 DAG.getConstant(RHSC->getAPIntValue()- 2502 LHSR->getAPIntValue(), 2503 N0.getValueType()), Cond); 2504 } 2505 2506 // Turn (X^C1) == C2 into X == C1^C2 iff X&~C1 = 0. 2507 if (N0.getOpcode() == ISD::XOR) 2508 // If we know that all of the inverted bits are zero, don't bother 2509 // performing the inversion. 2510 if (DAG.MaskedValueIsZero(N0.getOperand(0), ~LHSR->getAPIntValue())) 2511 return 2512 DAG.getSetCC(dl, VT, N0.getOperand(0), 2513 DAG.getConstant(LHSR->getAPIntValue() ^ 2514 RHSC->getAPIntValue(), 2515 N0.getValueType()), 2516 Cond); 2517 } 2518 2519 // Turn (C1-X) == C2 --> X == C1-C2 2520 if (ConstantSDNode *SUBC = dyn_cast<ConstantSDNode>(N0.getOperand(0))) { 2521 if (N0.getOpcode() == ISD::SUB && N0.getNode()->hasOneUse()) { 2522 return 2523 DAG.getSetCC(dl, VT, N0.getOperand(1), 2524 DAG.getConstant(SUBC->getAPIntValue() - 2525 RHSC->getAPIntValue(), 2526 N0.getValueType()), 2527 Cond); 2528 } 2529 } 2530 2531 // Could RHSC fold directly into a compare? 2532 if (RHSC->getValueType(0).getSizeInBits() <= 64) 2533 LegalRHSImm = isLegalICmpImmediate(RHSC->getSExtValue()); 2534 } 2535 2536 // Simplify (X+Z) == X --> Z == 0 2537 // Don't do this if X is an immediate that can fold into a cmp 2538 // instruction and X+Z has other uses. It could be an induction variable 2539 // chain, and the transform would increase register pressure. 2540 if (!LegalRHSImm || N0.getNode()->hasOneUse()) { 2541 if (N0.getOperand(0) == N1) 2542 return DAG.getSetCC(dl, VT, N0.getOperand(1), 2543 DAG.getConstant(0, N0.getValueType()), Cond); 2544 if (N0.getOperand(1) == N1) { 2545 if (DAG.isCommutativeBinOp(N0.getOpcode())) 2546 return DAG.getSetCC(dl, VT, N0.getOperand(0), 2547 DAG.getConstant(0, N0.getValueType()), Cond); 2548 else if (N0.getNode()->hasOneUse()) { 2549 assert(N0.getOpcode() == ISD::SUB && "Unexpected operation!"); 2550 // (Z-X) == X --> Z == X<<1 2551 SDValue SH = DAG.getNode(ISD::SHL, dl, N1.getValueType(), N1, 2552 DAG.getConstant(1, getShiftAmountTy(N1.getValueType()))); 2553 if (!DCI.isCalledByLegalizer()) 2554 DCI.AddToWorklist(SH.getNode()); 2555 return DAG.getSetCC(dl, VT, N0.getOperand(0), SH, Cond); 2556 } 2557 } 2558 } 2559 } 2560 2561 if (N1.getOpcode() == ISD::ADD || N1.getOpcode() == ISD::SUB || 2562 N1.getOpcode() == ISD::XOR) { 2563 // Simplify X == (X+Z) --> Z == 0 2564 if (N1.getOperand(0) == N0) { 2565 return DAG.getSetCC(dl, VT, N1.getOperand(1), 2566 DAG.getConstant(0, N1.getValueType()), Cond); 2567 } else if (N1.getOperand(1) == N0) { 2568 if (DAG.isCommutativeBinOp(N1.getOpcode())) { 2569 return DAG.getSetCC(dl, VT, N1.getOperand(0), 2570 DAG.getConstant(0, N1.getValueType()), Cond); 2571 } else if (N1.getNode()->hasOneUse()) { 2572 assert(N1.getOpcode() == ISD::SUB && "Unexpected operation!"); 2573 // X == (Z-X) --> X<<1 == Z 2574 SDValue SH = DAG.getNode(ISD::SHL, dl, N1.getValueType(), N0, 2575 DAG.getConstant(1, getShiftAmountTy(N0.getValueType()))); 2576 if (!DCI.isCalledByLegalizer()) 2577 DCI.AddToWorklist(SH.getNode()); 2578 return DAG.getSetCC(dl, VT, SH, N1.getOperand(0), Cond); 2579 } 2580 } 2581 } 2582 2583 // Simplify x&y == y to x&y != 0 if y has exactly one bit set. 2584 // Note that where y is variable and is known to have at most 2585 // one bit set (for example, if it is z&1) we cannot do this; 2586 // the expressions are not equivalent when y==0. 2587 if (N0.getOpcode() == ISD::AND) 2588 if (N0.getOperand(0) == N1 || N0.getOperand(1) == N1) { 2589 if (ValueHasExactlyOneBitSet(N1, DAG)) { 2590 Cond = ISD::getSetCCInverse(Cond, /*isInteger=*/true); 2591 SDValue Zero = DAG.getConstant(0, N1.getValueType()); 2592 return DAG.getSetCC(dl, VT, N0, Zero, Cond); 2593 } 2594 } 2595 if (N1.getOpcode() == ISD::AND) 2596 if (N1.getOperand(0) == N0 || N1.getOperand(1) == N0) { 2597 if (ValueHasExactlyOneBitSet(N0, DAG)) { 2598 Cond = ISD::getSetCCInverse(Cond, /*isInteger=*/true); 2599 SDValue Zero = DAG.getConstant(0, N0.getValueType()); 2600 return DAG.getSetCC(dl, VT, N1, Zero, Cond); 2601 } 2602 } 2603 } 2604 2605 // Fold away ALL boolean setcc's. 2606 SDValue Temp; 2607 if (N0.getValueType() == MVT::i1 && foldBooleans) { 2608 switch (Cond) { 2609 default: llvm_unreachable("Unknown integer setcc!"); 2610 case ISD::SETEQ: // X == Y -> ~(X^Y) 2611 Temp = DAG.getNode(ISD::XOR, dl, MVT::i1, N0, N1); 2612 N0 = DAG.getNOT(dl, Temp, MVT::i1); 2613 if (!DCI.isCalledByLegalizer()) 2614 DCI.AddToWorklist(Temp.getNode()); 2615 break; 2616 case ISD::SETNE: // X != Y --> (X^Y) 2617 N0 = DAG.getNode(ISD::XOR, dl, MVT::i1, N0, N1); 2618 break; 2619 case ISD::SETGT: // X >s Y --> X == 0 & Y == 1 --> ~X & Y 2620 case ISD::SETULT: // X <u Y --> X == 0 & Y == 1 --> ~X & Y 2621 Temp = DAG.getNOT(dl, N0, MVT::i1); 2622 N0 = DAG.getNode(ISD::AND, dl, MVT::i1, N1, Temp); 2623 if (!DCI.isCalledByLegalizer()) 2624 DCI.AddToWorklist(Temp.getNode()); 2625 break; 2626 case ISD::SETLT: // X <s Y --> X == 1 & Y == 0 --> ~Y & X 2627 case ISD::SETUGT: // X >u Y --> X == 1 & Y == 0 --> ~Y & X 2628 Temp = DAG.getNOT(dl, N1, MVT::i1); 2629 N0 = DAG.getNode(ISD::AND, dl, MVT::i1, N0, Temp); 2630 if (!DCI.isCalledByLegalizer()) 2631 DCI.AddToWorklist(Temp.getNode()); 2632 break; 2633 case ISD::SETULE: // X <=u Y --> X == 0 | Y == 1 --> ~X | Y 2634 case ISD::SETGE: // X >=s Y --> X == 0 | Y == 1 --> ~X | Y 2635 Temp = DAG.getNOT(dl, N0, MVT::i1); 2636 N0 = DAG.getNode(ISD::OR, dl, MVT::i1, N1, Temp); 2637 if (!DCI.isCalledByLegalizer()) 2638 DCI.AddToWorklist(Temp.getNode()); 2639 break; 2640 case ISD::SETUGE: // X >=u Y --> X == 1 | Y == 0 --> ~Y | X 2641 case ISD::SETLE: // X <=s Y --> X == 1 | Y == 0 --> ~Y | X 2642 Temp = DAG.getNOT(dl, N1, MVT::i1); 2643 N0 = DAG.getNode(ISD::OR, dl, MVT::i1, N0, Temp); 2644 break; 2645 } 2646 if (VT != MVT::i1) { 2647 if (!DCI.isCalledByLegalizer()) 2648 DCI.AddToWorklist(N0.getNode()); 2649 // FIXME: If running after legalize, we probably can't do this. 2650 N0 = DAG.getNode(ISD::ZERO_EXTEND, dl, VT, N0); 2651 } 2652 return N0; 2653 } 2654 2655 // Could not fold it. 2656 return SDValue(); 2657 } 2658 2659 /// isGAPlusOffset - Returns true (and the GlobalValue and the offset) if the 2660 /// node is a GlobalAddress + offset. 2661 bool TargetLowering::isGAPlusOffset(SDNode *N, const GlobalValue *&GA, 2662 int64_t &Offset) const { 2663 if (isa<GlobalAddressSDNode>(N)) { 2664 GlobalAddressSDNode *GASD = cast<GlobalAddressSDNode>(N); 2665 GA = GASD->getGlobal(); 2666 Offset += GASD->getOffset(); 2667 return true; 2668 } 2669 2670 if (N->getOpcode() == ISD::ADD) { 2671 SDValue N1 = N->getOperand(0); 2672 SDValue N2 = N->getOperand(1); 2673 if (isGAPlusOffset(N1.getNode(), GA, Offset)) { 2674 ConstantSDNode *V = dyn_cast<ConstantSDNode>(N2); 2675 if (V) { 2676 Offset += V->getSExtValue(); 2677 return true; 2678 } 2679 } else if (isGAPlusOffset(N2.getNode(), GA, Offset)) { 2680 ConstantSDNode *V = dyn_cast<ConstantSDNode>(N1); 2681 if (V) { 2682 Offset += V->getSExtValue(); 2683 return true; 2684 } 2685 } 2686 } 2687 2688 return false; 2689 } 2690 2691 2692 SDValue TargetLowering:: 2693 PerformDAGCombine(SDNode *N, DAGCombinerInfo &DCI) const { 2694 // Default implementation: no optimization. 2695 return SDValue(); 2696 } 2697 2698 //===----------------------------------------------------------------------===// 2699 // Inline Assembler Implementation Methods 2700 //===----------------------------------------------------------------------===// 2701 2702 2703 TargetLowering::ConstraintType 2704 TargetLowering::getConstraintType(const std::string &Constraint) const { 2705 if (Constraint.size() == 1) { 2706 switch (Constraint[0]) { 2707 default: break; 2708 case 'r': return C_RegisterClass; 2709 case 'm': // memory 2710 case 'o': // offsetable 2711 case 'V': // not offsetable 2712 return C_Memory; 2713 case 'i': // Simple Integer or Relocatable Constant 2714 case 'n': // Simple Integer 2715 case 'E': // Floating Point Constant 2716 case 'F': // Floating Point Constant 2717 case 's': // Relocatable Constant 2718 case 'p': // Address. 2719 case 'X': // Allow ANY value. 2720 case 'I': // Target registers. 2721 case 'J': 2722 case 'K': 2723 case 'L': 2724 case 'M': 2725 case 'N': 2726 case 'O': 2727 case 'P': 2728 case '<': 2729 case '>': 2730 return C_Other; 2731 } 2732 } 2733 2734 if (Constraint.size() > 1 && Constraint[0] == '{' && 2735 Constraint[Constraint.size()-1] == '}') 2736 return C_Register; 2737 return C_Unknown; 2738 } 2739 2740 /// LowerXConstraint - try to replace an X constraint, which matches anything, 2741 /// with another that has more specific requirements based on the type of the 2742 /// corresponding operand. 2743 const char *TargetLowering::LowerXConstraint(EVT ConstraintVT) const{ 2744 if (ConstraintVT.isInteger()) 2745 return "r"; 2746 if (ConstraintVT.isFloatingPoint()) 2747 return "f"; // works for many targets 2748 return 0; 2749 } 2750 2751 /// LowerAsmOperandForConstraint - Lower the specified operand into the Ops 2752 /// vector. If it is invalid, don't add anything to Ops. 2753 void TargetLowering::LowerAsmOperandForConstraint(SDValue Op, 2754 std::string &Constraint, 2755 std::vector<SDValue> &Ops, 2756 SelectionDAG &DAG) const { 2757 2758 if (Constraint.length() > 1) return; 2759 2760 char ConstraintLetter = Constraint[0]; 2761 switch (ConstraintLetter) { 2762 default: break; 2763 case 'X': // Allows any operand; labels (basic block) use this. 2764 if (Op.getOpcode() == ISD::BasicBlock) { 2765 Ops.push_back(Op); 2766 return; 2767 } 2768 // fall through 2769 case 'i': // Simple Integer or Relocatable Constant 2770 case 'n': // Simple Integer 2771 case 's': { // Relocatable Constant 2772 // These operands are interested in values of the form (GV+C), where C may 2773 // be folded in as an offset of GV, or it may be explicitly added. Also, it 2774 // is possible and fine if either GV or C are missing. 2775 ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op); 2776 GlobalAddressSDNode *GA = dyn_cast<GlobalAddressSDNode>(Op); 2777 2778 // If we have "(add GV, C)", pull out GV/C 2779 if (Op.getOpcode() == ISD::ADD) { 2780 C = dyn_cast<ConstantSDNode>(Op.getOperand(1)); 2781 GA = dyn_cast<GlobalAddressSDNode>(Op.getOperand(0)); 2782 if (C == 0 || GA == 0) { 2783 C = dyn_cast<ConstantSDNode>(Op.getOperand(0)); 2784 GA = dyn_cast<GlobalAddressSDNode>(Op.getOperand(1)); 2785 } 2786 if (C == 0 || GA == 0) 2787 C = 0, GA = 0; 2788 } 2789 2790 // If we find a valid operand, map to the TargetXXX version so that the 2791 // value itself doesn't get selected. 2792 if (GA) { // Either &GV or &GV+C 2793 if (ConstraintLetter != 'n') { 2794 int64_t Offs = GA->getOffset(); 2795 if (C) Offs += C->getZExtValue(); 2796 Ops.push_back(DAG.getTargetGlobalAddress(GA->getGlobal(), 2797 C ? C->getDebugLoc() : DebugLoc(), 2798 Op.getValueType(), Offs)); 2799 return; 2800 } 2801 } 2802 if (C) { // just C, no GV. 2803 // Simple constants are not allowed for 's'. 2804 if (ConstraintLetter != 's') { 2805 // gcc prints these as sign extended. Sign extend value to 64 bits 2806 // now; without this it would get ZExt'd later in 2807 // ScheduleDAGSDNodes::EmitNode, which is very generic. 2808 Ops.push_back(DAG.getTargetConstant(C->getAPIntValue().getSExtValue(), 2809 MVT::i64)); 2810 return; 2811 } 2812 } 2813 break; 2814 } 2815 } 2816 } 2817 2818 std::pair<unsigned, const TargetRegisterClass*> TargetLowering:: 2819 getRegForInlineAsmConstraint(const std::string &Constraint, 2820 EVT VT) const { 2821 if (Constraint[0] != '{') 2822 return std::make_pair(0u, static_cast<TargetRegisterClass*>(0)); 2823 assert(*(Constraint.end()-1) == '}' && "Not a brace enclosed constraint?"); 2824 2825 // Remove the braces from around the name. 2826 StringRef RegName(Constraint.data()+1, Constraint.size()-2); 2827 2828 // Figure out which register class contains this reg. 2829 const TargetRegisterInfo *RI = TM.getRegisterInfo(); 2830 for (TargetRegisterInfo::regclass_iterator RCI = RI->regclass_begin(), 2831 E = RI->regclass_end(); RCI != E; ++RCI) { 2832 const TargetRegisterClass *RC = *RCI; 2833 2834 // If none of the value types for this register class are valid, we 2835 // can't use it. For example, 64-bit reg classes on 32-bit targets. 2836 if (!isLegalRC(RC)) 2837 continue; 2838 2839 for (TargetRegisterClass::iterator I = RC->begin(), E = RC->end(); 2840 I != E; ++I) { 2841 if (RegName.equals_lower(RI->getName(*I))) 2842 return std::make_pair(*I, RC); 2843 } 2844 } 2845 2846 return std::make_pair(0u, static_cast<const TargetRegisterClass*>(0)); 2847 } 2848 2849 //===----------------------------------------------------------------------===// 2850 // Constraint Selection. 2851 2852 /// isMatchingInputConstraint - Return true of this is an input operand that is 2853 /// a matching constraint like "4". 2854 bool TargetLowering::AsmOperandInfo::isMatchingInputConstraint() const { 2855 assert(!ConstraintCode.empty() && "No known constraint!"); 2856 return isdigit(ConstraintCode[0]); 2857 } 2858 2859 /// getMatchedOperand - If this is an input matching constraint, this method 2860 /// returns the output operand it matches. 2861 unsigned TargetLowering::AsmOperandInfo::getMatchedOperand() const { 2862 assert(!ConstraintCode.empty() && "No known constraint!"); 2863 return atoi(ConstraintCode.c_str()); 2864 } 2865 2866 2867 /// ParseConstraints - Split up the constraint string from the inline 2868 /// assembly value into the specific constraints and their prefixes, 2869 /// and also tie in the associated operand values. 2870 /// If this returns an empty vector, and if the constraint string itself 2871 /// isn't empty, there was an error parsing. 2872 TargetLowering::AsmOperandInfoVector TargetLowering::ParseConstraints( 2873 ImmutableCallSite CS) const { 2874 /// ConstraintOperands - Information about all of the constraints. 2875 AsmOperandInfoVector ConstraintOperands; 2876 const InlineAsm *IA = cast<InlineAsm>(CS.getCalledValue()); 2877 unsigned maCount = 0; // Largest number of multiple alternative constraints. 2878 2879 // Do a prepass over the constraints, canonicalizing them, and building up the 2880 // ConstraintOperands list. 2881 InlineAsm::ConstraintInfoVector 2882 ConstraintInfos = IA->ParseConstraints(); 2883 2884 unsigned ArgNo = 0; // ArgNo - The argument of the CallInst. 2885 unsigned ResNo = 0; // ResNo - The result number of the next output. 2886 2887 for (unsigned i = 0, e = ConstraintInfos.size(); i != e; ++i) { 2888 ConstraintOperands.push_back(AsmOperandInfo(ConstraintInfos[i])); 2889 AsmOperandInfo &OpInfo = ConstraintOperands.back(); 2890 2891 // Update multiple alternative constraint count. 2892 if (OpInfo.multipleAlternatives.size() > maCount) 2893 maCount = OpInfo.multipleAlternatives.size(); 2894 2895 OpInfo.ConstraintVT = MVT::Other; 2896 2897 // Compute the value type for each operand. 2898 switch (OpInfo.Type) { 2899 case InlineAsm::isOutput: 2900 // Indirect outputs just consume an argument. 2901 if (OpInfo.isIndirect) { 2902 OpInfo.CallOperandVal = const_cast<Value *>(CS.getArgument(ArgNo++)); 2903 break; 2904 } 2905 2906 // The return value of the call is this value. As such, there is no 2907 // corresponding argument. 2908 assert(!CS.getType()->isVoidTy() && 2909 "Bad inline asm!"); 2910 if (StructType *STy = dyn_cast<StructType>(CS.getType())) { 2911 OpInfo.ConstraintVT = getValueType(STy->getElementType(ResNo)); 2912 } else { 2913 assert(ResNo == 0 && "Asm only has one result!"); 2914 OpInfo.ConstraintVT = getValueType(CS.getType()); 2915 } 2916 ++ResNo; 2917 break; 2918 case InlineAsm::isInput: 2919 OpInfo.CallOperandVal = const_cast<Value *>(CS.getArgument(ArgNo++)); 2920 break; 2921 case InlineAsm::isClobber: 2922 // Nothing to do. 2923 break; 2924 } 2925 2926 if (OpInfo.CallOperandVal) { 2927 llvm::Type *OpTy = OpInfo.CallOperandVal->getType(); 2928 if (OpInfo.isIndirect) { 2929 llvm::PointerType *PtrTy = dyn_cast<PointerType>(OpTy); 2930 if (!PtrTy) 2931 report_fatal_error("Indirect operand for inline asm not a pointer!"); 2932 OpTy = PtrTy->getElementType(); 2933 } 2934 2935 // Look for vector wrapped in a struct. e.g. { <16 x i8> }. 2936 if (StructType *STy = dyn_cast<StructType>(OpTy)) 2937 if (STy->getNumElements() == 1) 2938 OpTy = STy->getElementType(0); 2939 2940 // If OpTy is not a single value, it may be a struct/union that we 2941 // can tile with integers. 2942 if (!OpTy->isSingleValueType() && OpTy->isSized()) { 2943 unsigned BitSize = TD->getTypeSizeInBits(OpTy); 2944 switch (BitSize) { 2945 default: break; 2946 case 1: 2947 case 8: 2948 case 16: 2949 case 32: 2950 case 64: 2951 case 128: 2952 OpInfo.ConstraintVT = 2953 EVT::getEVT(IntegerType::get(OpTy->getContext(), BitSize), true); 2954 break; 2955 } 2956 } else if (dyn_cast<PointerType>(OpTy)) { 2957 OpInfo.ConstraintVT = MVT::getIntegerVT(8*TD->getPointerSize()); 2958 } else { 2959 OpInfo.ConstraintVT = EVT::getEVT(OpTy, true); 2960 } 2961 } 2962 } 2963 2964 // If we have multiple alternative constraints, select the best alternative. 2965 if (ConstraintInfos.size()) { 2966 if (maCount) { 2967 unsigned bestMAIndex = 0; 2968 int bestWeight = -1; 2969 // weight: -1 = invalid match, and 0 = so-so match to 5 = good match. 2970 int weight = -1; 2971 unsigned maIndex; 2972 // Compute the sums of the weights for each alternative, keeping track 2973 // of the best (highest weight) one so far. 2974 for (maIndex = 0; maIndex < maCount; ++maIndex) { 2975 int weightSum = 0; 2976 for (unsigned cIndex = 0, eIndex = ConstraintOperands.size(); 2977 cIndex != eIndex; ++cIndex) { 2978 AsmOperandInfo& OpInfo = ConstraintOperands[cIndex]; 2979 if (OpInfo.Type == InlineAsm::isClobber) 2980 continue; 2981 2982 // If this is an output operand with a matching input operand, 2983 // look up the matching input. If their types mismatch, e.g. one 2984 // is an integer, the other is floating point, or their sizes are 2985 // different, flag it as an maCantMatch. 2986 if (OpInfo.hasMatchingInput()) { 2987 AsmOperandInfo &Input = ConstraintOperands[OpInfo.MatchingInput]; 2988 if (OpInfo.ConstraintVT != Input.ConstraintVT) { 2989 if ((OpInfo.ConstraintVT.isInteger() != 2990 Input.ConstraintVT.isInteger()) || 2991 (OpInfo.ConstraintVT.getSizeInBits() != 2992 Input.ConstraintVT.getSizeInBits())) { 2993 weightSum = -1; // Can't match. 2994 break; 2995 } 2996 } 2997 } 2998 weight = getMultipleConstraintMatchWeight(OpInfo, maIndex); 2999 if (weight == -1) { 3000 weightSum = -1; 3001 break; 3002 } 3003 weightSum += weight; 3004 } 3005 // Update best. 3006 if (weightSum > bestWeight) { 3007 bestWeight = weightSum; 3008 bestMAIndex = maIndex; 3009 } 3010 } 3011 3012 // Now select chosen alternative in each constraint. 3013 for (unsigned cIndex = 0, eIndex = ConstraintOperands.size(); 3014 cIndex != eIndex; ++cIndex) { 3015 AsmOperandInfo& cInfo = ConstraintOperands[cIndex]; 3016 if (cInfo.Type == InlineAsm::isClobber) 3017 continue; 3018 cInfo.selectAlternative(bestMAIndex); 3019 } 3020 } 3021 } 3022 3023 // Check and hook up tied operands, choose constraint code to use. 3024 for (unsigned cIndex = 0, eIndex = ConstraintOperands.size(); 3025 cIndex != eIndex; ++cIndex) { 3026 AsmOperandInfo& OpInfo = ConstraintOperands[cIndex]; 3027 3028 // If this is an output operand with a matching input operand, look up the 3029 // matching input. If their types mismatch, e.g. one is an integer, the 3030 // other is floating point, or their sizes are different, flag it as an 3031 // error. 3032 if (OpInfo.hasMatchingInput()) { 3033 AsmOperandInfo &Input = ConstraintOperands[OpInfo.MatchingInput]; 3034 3035 if (OpInfo.ConstraintVT != Input.ConstraintVT) { 3036 std::pair<unsigned, const TargetRegisterClass*> MatchRC = 3037 getRegForInlineAsmConstraint(OpInfo.ConstraintCode, 3038 OpInfo.ConstraintVT); 3039 std::pair<unsigned, const TargetRegisterClass*> InputRC = 3040 getRegForInlineAsmConstraint(Input.ConstraintCode, 3041 Input.ConstraintVT); 3042 if ((OpInfo.ConstraintVT.isInteger() != 3043 Input.ConstraintVT.isInteger()) || 3044 (MatchRC.second != InputRC.second)) { 3045 report_fatal_error("Unsupported asm: input constraint" 3046 " with a matching output constraint of" 3047 " incompatible type!"); 3048 } 3049 } 3050 3051 } 3052 } 3053 3054 return ConstraintOperands; 3055 } 3056 3057 3058 /// getConstraintGenerality - Return an integer indicating how general CT 3059 /// is. 3060 static unsigned getConstraintGenerality(TargetLowering::ConstraintType CT) { 3061 switch (CT) { 3062 case TargetLowering::C_Other: 3063 case TargetLowering::C_Unknown: 3064 return 0; 3065 case TargetLowering::C_Register: 3066 return 1; 3067 case TargetLowering::C_RegisterClass: 3068 return 2; 3069 case TargetLowering::C_Memory: 3070 return 3; 3071 } 3072 llvm_unreachable("Invalid constraint type"); 3073 } 3074 3075 /// Examine constraint type and operand type and determine a weight value. 3076 /// This object must already have been set up with the operand type 3077 /// and the current alternative constraint selected. 3078 TargetLowering::ConstraintWeight 3079 TargetLowering::getMultipleConstraintMatchWeight( 3080 AsmOperandInfo &info, int maIndex) const { 3081 InlineAsm::ConstraintCodeVector *rCodes; 3082 if (maIndex >= (int)info.multipleAlternatives.size()) 3083 rCodes = &info.Codes; 3084 else 3085 rCodes = &info.multipleAlternatives[maIndex].Codes; 3086 ConstraintWeight BestWeight = CW_Invalid; 3087 3088 // Loop over the options, keeping track of the most general one. 3089 for (unsigned i = 0, e = rCodes->size(); i != e; ++i) { 3090 ConstraintWeight weight = 3091 getSingleConstraintMatchWeight(info, (*rCodes)[i].c_str()); 3092 if (weight > BestWeight) 3093 BestWeight = weight; 3094 } 3095 3096 return BestWeight; 3097 } 3098 3099 /// Examine constraint type and operand type and determine a weight value. 3100 /// This object must already have been set up with the operand type 3101 /// and the current alternative constraint selected. 3102 TargetLowering::ConstraintWeight 3103 TargetLowering::getSingleConstraintMatchWeight( 3104 AsmOperandInfo &info, const char *constraint) const { 3105 ConstraintWeight weight = CW_Invalid; 3106 Value *CallOperandVal = info.CallOperandVal; 3107 // If we don't have a value, we can't do a match, 3108 // but allow it at the lowest weight. 3109 if (CallOperandVal == NULL) 3110 return CW_Default; 3111 // Look at the constraint type. 3112 switch (*constraint) { 3113 case 'i': // immediate integer. 3114 case 'n': // immediate integer with a known value. 3115 if (isa<ConstantInt>(CallOperandVal)) 3116 weight = CW_Constant; 3117 break; 3118 case 's': // non-explicit intregal immediate. 3119 if (isa<GlobalValue>(CallOperandVal)) 3120 weight = CW_Constant; 3121 break; 3122 case 'E': // immediate float if host format. 3123 case 'F': // immediate float. 3124 if (isa<ConstantFP>(CallOperandVal)) 3125 weight = CW_Constant; 3126 break; 3127 case '<': // memory operand with autodecrement. 3128 case '>': // memory operand with autoincrement. 3129 case 'm': // memory operand. 3130 case 'o': // offsettable memory operand 3131 case 'V': // non-offsettable memory operand 3132 weight = CW_Memory; 3133 break; 3134 case 'r': // general register. 3135 case 'g': // general register, memory operand or immediate integer. 3136 // note: Clang converts "g" to "imr". 3137 if (CallOperandVal->getType()->isIntegerTy()) 3138 weight = CW_Register; 3139 break; 3140 case 'X': // any operand. 3141 default: 3142 weight = CW_Default; 3143 break; 3144 } 3145 return weight; 3146 } 3147 3148 /// ChooseConstraint - If there are multiple different constraints that we 3149 /// could pick for this operand (e.g. "imr") try to pick the 'best' one. 3150 /// This is somewhat tricky: constraints fall into four classes: 3151 /// Other -> immediates and magic values 3152 /// Register -> one specific register 3153 /// RegisterClass -> a group of regs 3154 /// Memory -> memory 3155 /// Ideally, we would pick the most specific constraint possible: if we have 3156 /// something that fits into a register, we would pick it. The problem here 3157 /// is that if we have something that could either be in a register or in 3158 /// memory that use of the register could cause selection of *other* 3159 /// operands to fail: they might only succeed if we pick memory. Because of 3160 /// this the heuristic we use is: 3161 /// 3162 /// 1) If there is an 'other' constraint, and if the operand is valid for 3163 /// that constraint, use it. This makes us take advantage of 'i' 3164 /// constraints when available. 3165 /// 2) Otherwise, pick the most general constraint present. This prefers 3166 /// 'm' over 'r', for example. 3167 /// 3168 static void ChooseConstraint(TargetLowering::AsmOperandInfo &OpInfo, 3169 const TargetLowering &TLI, 3170 SDValue Op, SelectionDAG *DAG) { 3171 assert(OpInfo.Codes.size() > 1 && "Doesn't have multiple constraint options"); 3172 unsigned BestIdx = 0; 3173 TargetLowering::ConstraintType BestType = TargetLowering::C_Unknown; 3174 int BestGenerality = -1; 3175 3176 // Loop over the options, keeping track of the most general one. 3177 for (unsigned i = 0, e = OpInfo.Codes.size(); i != e; ++i) { 3178 TargetLowering::ConstraintType CType = 3179 TLI.getConstraintType(OpInfo.Codes[i]); 3180 3181 // If this is an 'other' constraint, see if the operand is valid for it. 3182 // For example, on X86 we might have an 'rI' constraint. If the operand 3183 // is an integer in the range [0..31] we want to use I (saving a load 3184 // of a register), otherwise we must use 'r'. 3185 if (CType == TargetLowering::C_Other && Op.getNode()) { 3186 assert(OpInfo.Codes[i].size() == 1 && 3187 "Unhandled multi-letter 'other' constraint"); 3188 std::vector<SDValue> ResultOps; 3189 TLI.LowerAsmOperandForConstraint(Op, OpInfo.Codes[i], 3190 ResultOps, *DAG); 3191 if (!ResultOps.empty()) { 3192 BestType = CType; 3193 BestIdx = i; 3194 break; 3195 } 3196 } 3197 3198 // Things with matching constraints can only be registers, per gcc 3199 // documentation. This mainly affects "g" constraints. 3200 if (CType == TargetLowering::C_Memory && OpInfo.hasMatchingInput()) 3201 continue; 3202 3203 // This constraint letter is more general than the previous one, use it. 3204 int Generality = getConstraintGenerality(CType); 3205 if (Generality > BestGenerality) { 3206 BestType = CType; 3207 BestIdx = i; 3208 BestGenerality = Generality; 3209 } 3210 } 3211 3212 OpInfo.ConstraintCode = OpInfo.Codes[BestIdx]; 3213 OpInfo.ConstraintType = BestType; 3214 } 3215 3216 /// ComputeConstraintToUse - Determines the constraint code and constraint 3217 /// type to use for the specific AsmOperandInfo, setting 3218 /// OpInfo.ConstraintCode and OpInfo.ConstraintType. 3219 void TargetLowering::ComputeConstraintToUse(AsmOperandInfo &OpInfo, 3220 SDValue Op, 3221 SelectionDAG *DAG) const { 3222 assert(!OpInfo.Codes.empty() && "Must have at least one constraint"); 3223 3224 // Single-letter constraints ('r') are very common. 3225 if (OpInfo.Codes.size() == 1) { 3226 OpInfo.ConstraintCode = OpInfo.Codes[0]; 3227 OpInfo.ConstraintType = getConstraintType(OpInfo.ConstraintCode); 3228 } else { 3229 ChooseConstraint(OpInfo, *this, Op, DAG); 3230 } 3231 3232 // 'X' matches anything. 3233 if (OpInfo.ConstraintCode == "X" && OpInfo.CallOperandVal) { 3234 // Labels and constants are handled elsewhere ('X' is the only thing 3235 // that matches labels). For Functions, the type here is the type of 3236 // the result, which is not what we want to look at; leave them alone. 3237 Value *v = OpInfo.CallOperandVal; 3238 if (isa<BasicBlock>(v) || isa<ConstantInt>(v) || isa<Function>(v)) { 3239 OpInfo.CallOperandVal = v; 3240 return; 3241 } 3242 3243 // Otherwise, try to resolve it to something we know about by looking at 3244 // the actual operand type. 3245 if (const char *Repl = LowerXConstraint(OpInfo.ConstraintVT)) { 3246 OpInfo.ConstraintCode = Repl; 3247 OpInfo.ConstraintType = getConstraintType(OpInfo.ConstraintCode); 3248 } 3249 } 3250 } 3251 3252 //===----------------------------------------------------------------------===// 3253 // Loop Strength Reduction hooks 3254 //===----------------------------------------------------------------------===// 3255 3256 /// isLegalAddressingMode - Return true if the addressing mode represented 3257 /// by AM is legal for this target, for a load/store of the specified type. 3258 bool TargetLowering::isLegalAddressingMode(const AddrMode &AM, 3259 Type *Ty) const { 3260 // The default implementation of this implements a conservative RISCy, r+r and 3261 // r+i addr mode. 3262 3263 // Allows a sign-extended 16-bit immediate field. 3264 if (AM.BaseOffs <= -(1LL << 16) || AM.BaseOffs >= (1LL << 16)-1) 3265 return false; 3266 3267 // No global is ever allowed as a base. 3268 if (AM.BaseGV) 3269 return false; 3270 3271 // Only support r+r, 3272 switch (AM.Scale) { 3273 case 0: // "r+i" or just "i", depending on HasBaseReg. 3274 break; 3275 case 1: 3276 if (AM.HasBaseReg && AM.BaseOffs) // "r+r+i" is not allowed. 3277 return false; 3278 // Otherwise we have r+r or r+i. 3279 break; 3280 case 2: 3281 if (AM.HasBaseReg || AM.BaseOffs) // 2*r+r or 2*r+i is not allowed. 3282 return false; 3283 // Allow 2*r as r+r. 3284 break; 3285 } 3286 3287 return true; 3288 } 3289 3290 /// BuildExactDiv - Given an exact SDIV by a constant, create a multiplication 3291 /// with the multiplicative inverse of the constant. 3292 SDValue TargetLowering::BuildExactSDIV(SDValue Op1, SDValue Op2, DebugLoc dl, 3293 SelectionDAG &DAG) const { 3294 ConstantSDNode *C = cast<ConstantSDNode>(Op2); 3295 APInt d = C->getAPIntValue(); 3296 assert(d != 0 && "Division by zero!"); 3297 3298 // Shift the value upfront if it is even, so the LSB is one. 3299 unsigned ShAmt = d.countTrailingZeros(); 3300 if (ShAmt) { 3301 // TODO: For UDIV use SRL instead of SRA. 3302 SDValue Amt = DAG.getConstant(ShAmt, getShiftAmountTy(Op1.getValueType())); 3303 Op1 = DAG.getNode(ISD::SRA, dl, Op1.getValueType(), Op1, Amt); 3304 d = d.ashr(ShAmt); 3305 } 3306 3307 // Calculate the multiplicative inverse, using Newton's method. 3308 APInt t, xn = d; 3309 while ((t = d*xn) != 1) 3310 xn *= APInt(d.getBitWidth(), 2) - t; 3311 3312 Op2 = DAG.getConstant(xn, Op1.getValueType()); 3313 return DAG.getNode(ISD::MUL, dl, Op1.getValueType(), Op1, Op2); 3314 } 3315 3316 /// BuildSDIVSequence - Given an ISD::SDIV node expressing a divide by constant, 3317 /// return a DAG expression to select that will generate the same value by 3318 /// multiplying by a magic number. See: 3319 /// <http://the.wall.riscom.net/books/proc/ppc/cwg/code2.html> 3320 SDValue TargetLowering:: 3321 BuildSDIV(SDNode *N, SelectionDAG &DAG, bool IsAfterLegalization, 3322 std::vector<SDNode*>* Created) const { 3323 EVT VT = N->getValueType(0); 3324 DebugLoc dl= N->getDebugLoc(); 3325 3326 // Check to see if we can do this. 3327 // FIXME: We should be more aggressive here. 3328 if (!isTypeLegal(VT)) 3329 return SDValue(); 3330 3331 APInt d = cast<ConstantSDNode>(N->getOperand(1))->getAPIntValue(); 3332 APInt::ms magics = d.magic(); 3333 3334 // Multiply the numerator (operand 0) by the magic value 3335 // FIXME: We should support doing a MUL in a wider type 3336 SDValue Q; 3337 if (IsAfterLegalization ? isOperationLegal(ISD::MULHS, VT) : 3338 isOperationLegalOrCustom(ISD::MULHS, VT)) 3339 Q = DAG.getNode(ISD::MULHS, dl, VT, N->getOperand(0), 3340 DAG.getConstant(magics.m, VT)); 3341 else if (IsAfterLegalization ? isOperationLegal(ISD::SMUL_LOHI, VT) : 3342 isOperationLegalOrCustom(ISD::SMUL_LOHI, VT)) 3343 Q = SDValue(DAG.getNode(ISD::SMUL_LOHI, dl, DAG.getVTList(VT, VT), 3344 N->getOperand(0), 3345 DAG.getConstant(magics.m, VT)).getNode(), 1); 3346 else 3347 return SDValue(); // No mulhs or equvialent 3348 // If d > 0 and m < 0, add the numerator 3349 if (d.isStrictlyPositive() && magics.m.isNegative()) { 3350 Q = DAG.getNode(ISD::ADD, dl, VT, Q, N->getOperand(0)); 3351 if (Created) 3352 Created->push_back(Q.getNode()); 3353 } 3354 // If d < 0 and m > 0, subtract the numerator. 3355 if (d.isNegative() && magics.m.isStrictlyPositive()) { 3356 Q = DAG.getNode(ISD::SUB, dl, VT, Q, N->getOperand(0)); 3357 if (Created) 3358 Created->push_back(Q.getNode()); 3359 } 3360 // Shift right algebraic if shift value is nonzero 3361 if (magics.s > 0) { 3362 Q = DAG.getNode(ISD::SRA, dl, VT, Q, 3363 DAG.getConstant(magics.s, getShiftAmountTy(Q.getValueType()))); 3364 if (Created) 3365 Created->push_back(Q.getNode()); 3366 } 3367 // Extract the sign bit and add it to the quotient 3368 SDValue T = 3369 DAG.getNode(ISD::SRL, dl, VT, Q, DAG.getConstant(VT.getSizeInBits()-1, 3370 getShiftAmountTy(Q.getValueType()))); 3371 if (Created) 3372 Created->push_back(T.getNode()); 3373 return DAG.getNode(ISD::ADD, dl, VT, Q, T); 3374 } 3375 3376 /// BuildUDIVSequence - Given an ISD::UDIV node expressing a divide by constant, 3377 /// return a DAG expression to select that will generate the same value by 3378 /// multiplying by a magic number. See: 3379 /// <http://the.wall.riscom.net/books/proc/ppc/cwg/code2.html> 3380 SDValue TargetLowering:: 3381 BuildUDIV(SDNode *N, SelectionDAG &DAG, bool IsAfterLegalization, 3382 std::vector<SDNode*>* Created) const { 3383 EVT VT = N->getValueType(0); 3384 DebugLoc dl = N->getDebugLoc(); 3385 3386 // Check to see if we can do this. 3387 // FIXME: We should be more aggressive here. 3388 if (!isTypeLegal(VT)) 3389 return SDValue(); 3390 3391 // FIXME: We should use a narrower constant when the upper 3392 // bits are known to be zero. 3393 const APInt &N1C = cast<ConstantSDNode>(N->getOperand(1))->getAPIntValue(); 3394 APInt::mu magics = N1C.magicu(); 3395 3396 SDValue Q = N->getOperand(0); 3397 3398 // If the divisor is even, we can avoid using the expensive fixup by shifting 3399 // the divided value upfront. 3400 if (magics.a != 0 && !N1C[0]) { 3401 unsigned Shift = N1C.countTrailingZeros(); 3402 Q = DAG.getNode(ISD::SRL, dl, VT, Q, 3403 DAG.getConstant(Shift, getShiftAmountTy(Q.getValueType()))); 3404 if (Created) 3405 Created->push_back(Q.getNode()); 3406 3407 // Get magic number for the shifted divisor. 3408 magics = N1C.lshr(Shift).magicu(Shift); 3409 assert(magics.a == 0 && "Should use cheap fixup now"); 3410 } 3411 3412 // Multiply the numerator (operand 0) by the magic value 3413 // FIXME: We should support doing a MUL in a wider type 3414 if (IsAfterLegalization ? isOperationLegal(ISD::MULHU, VT) : 3415 isOperationLegalOrCustom(ISD::MULHU, VT)) 3416 Q = DAG.getNode(ISD::MULHU, dl, VT, Q, DAG.getConstant(magics.m, VT)); 3417 else if (IsAfterLegalization ? isOperationLegal(ISD::UMUL_LOHI, VT) : 3418 isOperationLegalOrCustom(ISD::UMUL_LOHI, VT)) 3419 Q = SDValue(DAG.getNode(ISD::UMUL_LOHI, dl, DAG.getVTList(VT, VT), Q, 3420 DAG.getConstant(magics.m, VT)).getNode(), 1); 3421 else 3422 return SDValue(); // No mulhu or equvialent 3423 if (Created) 3424 Created->push_back(Q.getNode()); 3425 3426 if (magics.a == 0) { 3427 assert(magics.s < N1C.getBitWidth() && 3428 "We shouldn't generate an undefined shift!"); 3429 return DAG.getNode(ISD::SRL, dl, VT, Q, 3430 DAG.getConstant(magics.s, getShiftAmountTy(Q.getValueType()))); 3431 } else { 3432 SDValue NPQ = DAG.getNode(ISD::SUB, dl, VT, N->getOperand(0), Q); 3433 if (Created) 3434 Created->push_back(NPQ.getNode()); 3435 NPQ = DAG.getNode(ISD::SRL, dl, VT, NPQ, 3436 DAG.getConstant(1, getShiftAmountTy(NPQ.getValueType()))); 3437 if (Created) 3438 Created->push_back(NPQ.getNode()); 3439 NPQ = DAG.getNode(ISD::ADD, dl, VT, NPQ, Q); 3440 if (Created) 3441 Created->push_back(NPQ.getNode()); 3442 return DAG.getNode(ISD::SRL, dl, VT, NPQ, 3443 DAG.getConstant(magics.s-1, getShiftAmountTy(NPQ.getValueType()))); 3444 } 3445 } 3446