TargetLowering.cpp revision e853d2e2508e21b5c3156c7d8b6e6902a7d2604a
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> 34using namespace llvm; 35 36/// InitLibcallNames - Set default libcall names. 37/// 38static 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/// 302static 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. 310RTLIB::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. 321RTLIB::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. 341RTLIB::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. 384RTLIB::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. 427RTLIB::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. 461RTLIB::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/// 495static 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. 516TargetLowering::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 622TargetLowering::~TargetLowering() { 623 delete &TLOF; 624} 625 626MVT 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. 632bool 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 648static 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. 697bool 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". 708std::pair<const TargetRegisterClass*, uint8_t> 709TargetLowering::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. 736void 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 897const char *TargetLowering::getTargetNodeName(unsigned Opcode) const { 898 return NULL; 899} 900 901EVT TargetLowering::getSetCCResultType(EVT VT) const { 902 assert(!VT.isVector() && "No default SetCC type for vectors!"); 903 return PointerTy.SimpleTy; 904} 905 906MVT::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/// 919unsigned 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. 987void 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.hasSExtAttr()) 1000 ExtendKind = ISD::SIGN_EXTEND; 1001 else if (attr.hasZExtAttr()) 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.hasInRegAttr()) 1020 Flags.setInReg(); 1021 1022 // Propagate extension type if any 1023 if (attr.hasSExtAttr()) 1024 Flags.setSExt(); 1025 else if (attr.hasZExtAttr()) 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/// getByValTypeAlignment - Return the desired alignment for ByVal aggregate 1034/// function arguments in the caller parameter area. This is the actual 1035/// alignment, not its logarithm. 1036unsigned TargetLowering::getByValTypeAlignment(Type *Ty) const { 1037 return TD->getCallFrameTypeAlignment(Ty); 1038} 1039 1040/// getJumpTableEncoding - Return the entry encoding for a jump table in the 1041/// current function. The returned value is a member of the 1042/// MachineJumpTableInfo::JTEntryKind enum. 1043unsigned TargetLowering::getJumpTableEncoding() const { 1044 // In non-pic modes, just use the address of a block. 1045 if (getTargetMachine().getRelocationModel() != Reloc::PIC_) 1046 return MachineJumpTableInfo::EK_BlockAddress; 1047 1048 // In PIC mode, if the target supports a GPRel32 directive, use it. 1049 if (getTargetMachine().getMCAsmInfo()->getGPRel32Directive() != 0) 1050 return MachineJumpTableInfo::EK_GPRel32BlockAddress; 1051 1052 // Otherwise, use a label difference. 1053 return MachineJumpTableInfo::EK_LabelDifference32; 1054} 1055 1056SDValue TargetLowering::getPICJumpTableRelocBase(SDValue Table, 1057 SelectionDAG &DAG) const { 1058 // If our PIC model is GP relative, use the global offset table as the base. 1059 unsigned JTEncoding = getJumpTableEncoding(); 1060 1061 if ((JTEncoding == MachineJumpTableInfo::EK_GPRel64BlockAddress) || 1062 (JTEncoding == MachineJumpTableInfo::EK_GPRel32BlockAddress)) 1063 return DAG.getGLOBAL_OFFSET_TABLE(getPointerTy()); 1064 1065 return Table; 1066} 1067 1068/// getPICJumpTableRelocBaseExpr - This returns the relocation base for the 1069/// given PIC jumptable, the same as getPICJumpTableRelocBase, but as an 1070/// MCExpr. 1071const MCExpr * 1072TargetLowering::getPICJumpTableRelocBaseExpr(const MachineFunction *MF, 1073 unsigned JTI,MCContext &Ctx) const{ 1074 // The normal PIC reloc base is the label at the start of the jump table. 1075 return MCSymbolRefExpr::Create(MF->getJTISymbol(JTI, Ctx), Ctx); 1076} 1077 1078bool 1079TargetLowering::isOffsetFoldingLegal(const GlobalAddressSDNode *GA) const { 1080 // Assume that everything is safe in static mode. 1081 if (getTargetMachine().getRelocationModel() == Reloc::Static) 1082 return true; 1083 1084 // In dynamic-no-pic mode, assume that known defined values are safe. 1085 if (getTargetMachine().getRelocationModel() == Reloc::DynamicNoPIC && 1086 GA && 1087 !GA->getGlobal()->isDeclaration() && 1088 !GA->getGlobal()->isWeakForLinker()) 1089 return true; 1090 1091 // Otherwise assume nothing is safe. 1092 return false; 1093} 1094 1095//===----------------------------------------------------------------------===// 1096// Optimization Methods 1097//===----------------------------------------------------------------------===// 1098 1099/// ShrinkDemandedConstant - Check to see if the specified operand of the 1100/// specified instruction is a constant integer. If so, check to see if there 1101/// are any bits set in the constant that are not demanded. If so, shrink the 1102/// constant and return true. 1103bool TargetLowering::TargetLoweringOpt::ShrinkDemandedConstant(SDValue Op, 1104 const APInt &Demanded) { 1105 DebugLoc dl = Op.getDebugLoc(); 1106 1107 // FIXME: ISD::SELECT, ISD::SELECT_CC 1108 switch (Op.getOpcode()) { 1109 default: break; 1110 case ISD::XOR: 1111 case ISD::AND: 1112 case ISD::OR: { 1113 ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op.getOperand(1)); 1114 if (!C) return false; 1115 1116 if (Op.getOpcode() == ISD::XOR && 1117 (C->getAPIntValue() | (~Demanded)).isAllOnesValue()) 1118 return false; 1119 1120 // if we can expand it to have all bits set, do it 1121 if (C->getAPIntValue().intersects(~Demanded)) { 1122 EVT VT = Op.getValueType(); 1123 SDValue New = DAG.getNode(Op.getOpcode(), dl, VT, Op.getOperand(0), 1124 DAG.getConstant(Demanded & 1125 C->getAPIntValue(), 1126 VT)); 1127 return CombineTo(Op, New); 1128 } 1129 1130 break; 1131 } 1132 } 1133 1134 return false; 1135} 1136 1137/// ShrinkDemandedOp - Convert x+y to (VT)((SmallVT)x+(SmallVT)y) if the 1138/// casts are free. This uses isZExtFree and ZERO_EXTEND for the widening 1139/// cast, but it could be generalized for targets with other types of 1140/// implicit widening casts. 1141bool 1142TargetLowering::TargetLoweringOpt::ShrinkDemandedOp(SDValue Op, 1143 unsigned BitWidth, 1144 const APInt &Demanded, 1145 DebugLoc dl) { 1146 assert(Op.getNumOperands() == 2 && 1147 "ShrinkDemandedOp only supports binary operators!"); 1148 assert(Op.getNode()->getNumValues() == 1 && 1149 "ShrinkDemandedOp only supports nodes with one result!"); 1150 1151 // Don't do this if the node has another user, which may require the 1152 // full value. 1153 if (!Op.getNode()->hasOneUse()) 1154 return false; 1155 1156 // Search for the smallest integer type with free casts to and from 1157 // Op's type. For expedience, just check power-of-2 integer types. 1158 const TargetLowering &TLI = DAG.getTargetLoweringInfo(); 1159 unsigned SmallVTBits = BitWidth - Demanded.countLeadingZeros(); 1160 if (!isPowerOf2_32(SmallVTBits)) 1161 SmallVTBits = NextPowerOf2(SmallVTBits); 1162 for (; SmallVTBits < BitWidth; SmallVTBits = NextPowerOf2(SmallVTBits)) { 1163 EVT SmallVT = EVT::getIntegerVT(*DAG.getContext(), SmallVTBits); 1164 if (TLI.isTruncateFree(Op.getValueType(), SmallVT) && 1165 TLI.isZExtFree(SmallVT, Op.getValueType())) { 1166 // We found a type with free casts. 1167 SDValue X = DAG.getNode(Op.getOpcode(), dl, SmallVT, 1168 DAG.getNode(ISD::TRUNCATE, dl, SmallVT, 1169 Op.getNode()->getOperand(0)), 1170 DAG.getNode(ISD::TRUNCATE, dl, SmallVT, 1171 Op.getNode()->getOperand(1))); 1172 SDValue Z = DAG.getNode(ISD::ZERO_EXTEND, dl, Op.getValueType(), X); 1173 return CombineTo(Op, Z); 1174 } 1175 } 1176 return false; 1177} 1178 1179/// SimplifyDemandedBits - Look at Op. At this point, we know that only the 1180/// DemandedMask bits of the result of Op are ever used downstream. If we can 1181/// use this information to simplify Op, create a new simplified DAG node and 1182/// return true, returning the original and new nodes in Old and New. Otherwise, 1183/// analyze the expression and return a mask of KnownOne and KnownZero bits for 1184/// the expression (used to simplify the caller). The KnownZero/One bits may 1185/// only be accurate for those bits in the DemandedMask. 1186bool TargetLowering::SimplifyDemandedBits(SDValue Op, 1187 const APInt &DemandedMask, 1188 APInt &KnownZero, 1189 APInt &KnownOne, 1190 TargetLoweringOpt &TLO, 1191 unsigned Depth) const { 1192 unsigned BitWidth = DemandedMask.getBitWidth(); 1193 assert(Op.getValueType().getScalarType().getSizeInBits() == BitWidth && 1194 "Mask size mismatches value type size!"); 1195 APInt NewMask = DemandedMask; 1196 DebugLoc dl = Op.getDebugLoc(); 1197 1198 // Don't know anything. 1199 KnownZero = KnownOne = APInt(BitWidth, 0); 1200 1201 // Other users may use these bits. 1202 if (!Op.getNode()->hasOneUse()) { 1203 if (Depth != 0) { 1204 // If not at the root, Just compute the KnownZero/KnownOne bits to 1205 // simplify things downstream. 1206 TLO.DAG.ComputeMaskedBits(Op, KnownZero, KnownOne, Depth); 1207 return false; 1208 } 1209 // If this is the root being simplified, allow it to have multiple uses, 1210 // just set the NewMask to all bits. 1211 NewMask = APInt::getAllOnesValue(BitWidth); 1212 } else if (DemandedMask == 0) { 1213 // Not demanding any bits from Op. 1214 if (Op.getOpcode() != ISD::UNDEF) 1215 return TLO.CombineTo(Op, TLO.DAG.getUNDEF(Op.getValueType())); 1216 return false; 1217 } else if (Depth == 6) { // Limit search depth. 1218 return false; 1219 } 1220 1221 APInt KnownZero2, KnownOne2, KnownZeroOut, KnownOneOut; 1222 switch (Op.getOpcode()) { 1223 case ISD::Constant: 1224 // We know all of the bits for a constant! 1225 KnownOne = cast<ConstantSDNode>(Op)->getAPIntValue(); 1226 KnownZero = ~KnownOne; 1227 return false; // Don't fall through, will infinitely loop. 1228 case ISD::AND: 1229 // If the RHS is a constant, check to see if the LHS would be zero without 1230 // using the bits from the RHS. Below, we use knowledge about the RHS to 1231 // simplify the LHS, here we're using information from the LHS to simplify 1232 // the RHS. 1233 if (ConstantSDNode *RHSC = dyn_cast<ConstantSDNode>(Op.getOperand(1))) { 1234 APInt LHSZero, LHSOne; 1235 // Do not increment Depth here; that can cause an infinite loop. 1236 TLO.DAG.ComputeMaskedBits(Op.getOperand(0), LHSZero, LHSOne, Depth); 1237 // If the LHS already has zeros where RHSC does, this and is dead. 1238 if ((LHSZero & NewMask) == (~RHSC->getAPIntValue() & NewMask)) 1239 return TLO.CombineTo(Op, Op.getOperand(0)); 1240 // If any of the set bits in the RHS are known zero on the LHS, shrink 1241 // the constant. 1242 if (TLO.ShrinkDemandedConstant(Op, ~LHSZero & NewMask)) 1243 return true; 1244 } 1245 1246 if (SimplifyDemandedBits(Op.getOperand(1), NewMask, KnownZero, 1247 KnownOne, TLO, Depth+1)) 1248 return true; 1249 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?"); 1250 if (SimplifyDemandedBits(Op.getOperand(0), ~KnownZero & NewMask, 1251 KnownZero2, KnownOne2, TLO, Depth+1)) 1252 return true; 1253 assert((KnownZero2 & KnownOne2) == 0 && "Bits known to be one AND zero?"); 1254 1255 // If all of the demanded bits are known one on one side, return the other. 1256 // These bits cannot contribute to the result of the 'and'. 1257 if ((NewMask & ~KnownZero2 & KnownOne) == (~KnownZero2 & NewMask)) 1258 return TLO.CombineTo(Op, Op.getOperand(0)); 1259 if ((NewMask & ~KnownZero & KnownOne2) == (~KnownZero & NewMask)) 1260 return TLO.CombineTo(Op, Op.getOperand(1)); 1261 // If all of the demanded bits in the inputs are known zeros, return zero. 1262 if ((NewMask & (KnownZero|KnownZero2)) == NewMask) 1263 return TLO.CombineTo(Op, TLO.DAG.getConstant(0, Op.getValueType())); 1264 // If the RHS is a constant, see if we can simplify it. 1265 if (TLO.ShrinkDemandedConstant(Op, ~KnownZero2 & NewMask)) 1266 return true; 1267 // If the operation can be done in a smaller type, do so. 1268 if (TLO.ShrinkDemandedOp(Op, BitWidth, NewMask, dl)) 1269 return true; 1270 1271 // Output known-1 bits are only known if set in both the LHS & RHS. 1272 KnownOne &= KnownOne2; 1273 // Output known-0 are known to be clear if zero in either the LHS | RHS. 1274 KnownZero |= KnownZero2; 1275 break; 1276 case ISD::OR: 1277 if (SimplifyDemandedBits(Op.getOperand(1), NewMask, KnownZero, 1278 KnownOne, TLO, Depth+1)) 1279 return true; 1280 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?"); 1281 if (SimplifyDemandedBits(Op.getOperand(0), ~KnownOne & NewMask, 1282 KnownZero2, KnownOne2, TLO, Depth+1)) 1283 return true; 1284 assert((KnownZero2 & KnownOne2) == 0 && "Bits known to be one AND zero?"); 1285 1286 // If all of the demanded bits are known zero on one side, return the other. 1287 // These bits cannot contribute to the result of the 'or'. 1288 if ((NewMask & ~KnownOne2 & KnownZero) == (~KnownOne2 & NewMask)) 1289 return TLO.CombineTo(Op, Op.getOperand(0)); 1290 if ((NewMask & ~KnownOne & KnownZero2) == (~KnownOne & NewMask)) 1291 return TLO.CombineTo(Op, Op.getOperand(1)); 1292 // If all of the potentially set bits on one side are known to be set on 1293 // the other side, just use the 'other' side. 1294 if ((NewMask & ~KnownZero & KnownOne2) == (~KnownZero & NewMask)) 1295 return TLO.CombineTo(Op, Op.getOperand(0)); 1296 if ((NewMask & ~KnownZero2 & KnownOne) == (~KnownZero2 & NewMask)) 1297 return TLO.CombineTo(Op, Op.getOperand(1)); 1298 // If the RHS is a constant, see if we can simplify it. 1299 if (TLO.ShrinkDemandedConstant(Op, NewMask)) 1300 return true; 1301 // If the operation can be done in a smaller type, do so. 1302 if (TLO.ShrinkDemandedOp(Op, BitWidth, NewMask, dl)) 1303 return true; 1304 1305 // Output known-0 bits are only known if clear in both the LHS & RHS. 1306 KnownZero &= KnownZero2; 1307 // Output known-1 are known to be set if set in either the LHS | RHS. 1308 KnownOne |= KnownOne2; 1309 break; 1310 case ISD::XOR: 1311 if (SimplifyDemandedBits(Op.getOperand(1), NewMask, KnownZero, 1312 KnownOne, TLO, Depth+1)) 1313 return true; 1314 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?"); 1315 if (SimplifyDemandedBits(Op.getOperand(0), NewMask, KnownZero2, 1316 KnownOne2, TLO, Depth+1)) 1317 return true; 1318 assert((KnownZero2 & KnownOne2) == 0 && "Bits known to be one AND zero?"); 1319 1320 // If all of the demanded bits are known zero on one side, return the other. 1321 // These bits cannot contribute to the result of the 'xor'. 1322 if ((KnownZero & NewMask) == NewMask) 1323 return TLO.CombineTo(Op, Op.getOperand(0)); 1324 if ((KnownZero2 & NewMask) == NewMask) 1325 return TLO.CombineTo(Op, Op.getOperand(1)); 1326 // If the operation can be done in a smaller type, do so. 1327 if (TLO.ShrinkDemandedOp(Op, BitWidth, NewMask, dl)) 1328 return true; 1329 1330 // If all of the unknown bits are known to be zero on one side or the other 1331 // (but not both) turn this into an *inclusive* or. 1332 // e.g. (A & C1)^(B & C2) -> (A & C1)|(B & C2) iff C1&C2 == 0 1333 if ((NewMask & ~KnownZero & ~KnownZero2) == 0) 1334 return TLO.CombineTo(Op, TLO.DAG.getNode(ISD::OR, dl, Op.getValueType(), 1335 Op.getOperand(0), 1336 Op.getOperand(1))); 1337 1338 // Output known-0 bits are known if clear or set in both the LHS & RHS. 1339 KnownZeroOut = (KnownZero & KnownZero2) | (KnownOne & KnownOne2); 1340 // Output known-1 are known to be set if set in only one of the LHS, RHS. 1341 KnownOneOut = (KnownZero & KnownOne2) | (KnownOne & KnownZero2); 1342 1343 // If all of the demanded bits on one side are known, and all of the set 1344 // bits on that side are also known to be set on the other side, turn this 1345 // into an AND, as we know the bits will be cleared. 1346 // e.g. (X | C1) ^ C2 --> (X | C1) & ~C2 iff (C1&C2) == C2 1347 // NB: it is okay if more bits are known than are requested 1348 if ((NewMask & (KnownZero|KnownOne)) == NewMask) { // all known on one side 1349 if (KnownOne == KnownOne2) { // set bits are the same on both sides 1350 EVT VT = Op.getValueType(); 1351 SDValue ANDC = TLO.DAG.getConstant(~KnownOne & NewMask, VT); 1352 return TLO.CombineTo(Op, TLO.DAG.getNode(ISD::AND, dl, VT, 1353 Op.getOperand(0), ANDC)); 1354 } 1355 } 1356 1357 // If the RHS is a constant, see if we can simplify it. 1358 // for XOR, we prefer to force bits to 1 if they will make a -1. 1359 // if we can't force bits, try to shrink constant 1360 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op.getOperand(1))) { 1361 APInt Expanded = C->getAPIntValue() | (~NewMask); 1362 // if we can expand it to have all bits set, do it 1363 if (Expanded.isAllOnesValue()) { 1364 if (Expanded != C->getAPIntValue()) { 1365 EVT VT = Op.getValueType(); 1366 SDValue New = TLO.DAG.getNode(Op.getOpcode(), dl,VT, Op.getOperand(0), 1367 TLO.DAG.getConstant(Expanded, VT)); 1368 return TLO.CombineTo(Op, New); 1369 } 1370 // if it already has all the bits set, nothing to change 1371 // but don't shrink either! 1372 } else if (TLO.ShrinkDemandedConstant(Op, NewMask)) { 1373 return true; 1374 } 1375 } 1376 1377 KnownZero = KnownZeroOut; 1378 KnownOne = KnownOneOut; 1379 break; 1380 case ISD::SELECT: 1381 if (SimplifyDemandedBits(Op.getOperand(2), NewMask, KnownZero, 1382 KnownOne, TLO, Depth+1)) 1383 return true; 1384 if (SimplifyDemandedBits(Op.getOperand(1), NewMask, KnownZero2, 1385 KnownOne2, TLO, Depth+1)) 1386 return true; 1387 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?"); 1388 assert((KnownZero2 & KnownOne2) == 0 && "Bits known to be one AND zero?"); 1389 1390 // If the operands are constants, see if we can simplify them. 1391 if (TLO.ShrinkDemandedConstant(Op, NewMask)) 1392 return true; 1393 1394 // Only known if known in both the LHS and RHS. 1395 KnownOne &= KnownOne2; 1396 KnownZero &= KnownZero2; 1397 break; 1398 case ISD::SELECT_CC: 1399 if (SimplifyDemandedBits(Op.getOperand(3), NewMask, KnownZero, 1400 KnownOne, TLO, Depth+1)) 1401 return true; 1402 if (SimplifyDemandedBits(Op.getOperand(2), NewMask, KnownZero2, 1403 KnownOne2, TLO, Depth+1)) 1404 return true; 1405 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?"); 1406 assert((KnownZero2 & KnownOne2) == 0 && "Bits known to be one AND zero?"); 1407 1408 // If the operands are constants, see if we can simplify them. 1409 if (TLO.ShrinkDemandedConstant(Op, NewMask)) 1410 return true; 1411 1412 // Only known if known in both the LHS and RHS. 1413 KnownOne &= KnownOne2; 1414 KnownZero &= KnownZero2; 1415 break; 1416 case ISD::SHL: 1417 if (ConstantSDNode *SA = dyn_cast<ConstantSDNode>(Op.getOperand(1))) { 1418 unsigned ShAmt = SA->getZExtValue(); 1419 SDValue InOp = Op.getOperand(0); 1420 1421 // If the shift count is an invalid immediate, don't do anything. 1422 if (ShAmt >= BitWidth) 1423 break; 1424 1425 // If this is ((X >>u C1) << ShAmt), see if we can simplify this into a 1426 // single shift. We can do this if the bottom bits (which are shifted 1427 // out) are never demanded. 1428 if (InOp.getOpcode() == ISD::SRL && 1429 isa<ConstantSDNode>(InOp.getOperand(1))) { 1430 if (ShAmt && (NewMask & APInt::getLowBitsSet(BitWidth, ShAmt)) == 0) { 1431 unsigned C1= cast<ConstantSDNode>(InOp.getOperand(1))->getZExtValue(); 1432 unsigned Opc = ISD::SHL; 1433 int Diff = ShAmt-C1; 1434 if (Diff < 0) { 1435 Diff = -Diff; 1436 Opc = ISD::SRL; 1437 } 1438 1439 SDValue NewSA = 1440 TLO.DAG.getConstant(Diff, Op.getOperand(1).getValueType()); 1441 EVT VT = Op.getValueType(); 1442 return TLO.CombineTo(Op, TLO.DAG.getNode(Opc, dl, VT, 1443 InOp.getOperand(0), NewSA)); 1444 } 1445 } 1446 1447 if (SimplifyDemandedBits(InOp, NewMask.lshr(ShAmt), 1448 KnownZero, KnownOne, TLO, Depth+1)) 1449 return true; 1450 1451 // Convert (shl (anyext x, c)) to (anyext (shl x, c)) if the high bits 1452 // are not demanded. This will likely allow the anyext to be folded away. 1453 if (InOp.getNode()->getOpcode() == ISD::ANY_EXTEND) { 1454 SDValue InnerOp = InOp.getNode()->getOperand(0); 1455 EVT InnerVT = InnerOp.getValueType(); 1456 unsigned InnerBits = InnerVT.getSizeInBits(); 1457 if (ShAmt < InnerBits && NewMask.lshr(InnerBits) == 0 && 1458 isTypeDesirableForOp(ISD::SHL, InnerVT)) { 1459 EVT ShTy = getShiftAmountTy(InnerVT); 1460 if (!APInt(BitWidth, ShAmt).isIntN(ShTy.getSizeInBits())) 1461 ShTy = InnerVT; 1462 SDValue NarrowShl = 1463 TLO.DAG.getNode(ISD::SHL, dl, InnerVT, InnerOp, 1464 TLO.DAG.getConstant(ShAmt, ShTy)); 1465 return 1466 TLO.CombineTo(Op, 1467 TLO.DAG.getNode(ISD::ANY_EXTEND, dl, Op.getValueType(), 1468 NarrowShl)); 1469 } 1470 } 1471 1472 KnownZero <<= SA->getZExtValue(); 1473 KnownOne <<= SA->getZExtValue(); 1474 // low bits known zero. 1475 KnownZero |= APInt::getLowBitsSet(BitWidth, SA->getZExtValue()); 1476 } 1477 break; 1478 case ISD::SRL: 1479 if (ConstantSDNode *SA = dyn_cast<ConstantSDNode>(Op.getOperand(1))) { 1480 EVT VT = Op.getValueType(); 1481 unsigned ShAmt = SA->getZExtValue(); 1482 unsigned VTSize = VT.getSizeInBits(); 1483 SDValue InOp = Op.getOperand(0); 1484 1485 // If the shift count is an invalid immediate, don't do anything. 1486 if (ShAmt >= BitWidth) 1487 break; 1488 1489 // If this is ((X << C1) >>u ShAmt), see if we can simplify this into a 1490 // single shift. We can do this if the top bits (which are shifted out) 1491 // are never demanded. 1492 if (InOp.getOpcode() == ISD::SHL && 1493 isa<ConstantSDNode>(InOp.getOperand(1))) { 1494 if (ShAmt && (NewMask & APInt::getHighBitsSet(VTSize, ShAmt)) == 0) { 1495 unsigned C1= cast<ConstantSDNode>(InOp.getOperand(1))->getZExtValue(); 1496 unsigned Opc = ISD::SRL; 1497 int Diff = ShAmt-C1; 1498 if (Diff < 0) { 1499 Diff = -Diff; 1500 Opc = ISD::SHL; 1501 } 1502 1503 SDValue NewSA = 1504 TLO.DAG.getConstant(Diff, Op.getOperand(1).getValueType()); 1505 return TLO.CombineTo(Op, TLO.DAG.getNode(Opc, dl, VT, 1506 InOp.getOperand(0), NewSA)); 1507 } 1508 } 1509 1510 // Compute the new bits that are at the top now. 1511 if (SimplifyDemandedBits(InOp, (NewMask << ShAmt), 1512 KnownZero, KnownOne, TLO, Depth+1)) 1513 return true; 1514 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?"); 1515 KnownZero = KnownZero.lshr(ShAmt); 1516 KnownOne = KnownOne.lshr(ShAmt); 1517 1518 APInt HighBits = APInt::getHighBitsSet(BitWidth, ShAmt); 1519 KnownZero |= HighBits; // High bits known zero. 1520 } 1521 break; 1522 case ISD::SRA: 1523 // If this is an arithmetic shift right and only the low-bit is set, we can 1524 // always convert this into a logical shr, even if the shift amount is 1525 // variable. The low bit of the shift cannot be an input sign bit unless 1526 // the shift amount is >= the size of the datatype, which is undefined. 1527 if (NewMask == 1) 1528 return TLO.CombineTo(Op, 1529 TLO.DAG.getNode(ISD::SRL, dl, Op.getValueType(), 1530 Op.getOperand(0), Op.getOperand(1))); 1531 1532 if (ConstantSDNode *SA = dyn_cast<ConstantSDNode>(Op.getOperand(1))) { 1533 EVT VT = Op.getValueType(); 1534 unsigned ShAmt = SA->getZExtValue(); 1535 1536 // If the shift count is an invalid immediate, don't do anything. 1537 if (ShAmt >= BitWidth) 1538 break; 1539 1540 APInt InDemandedMask = (NewMask << ShAmt); 1541 1542 // If any of the demanded bits are produced by the sign extension, we also 1543 // demand the input sign bit. 1544 APInt HighBits = APInt::getHighBitsSet(BitWidth, ShAmt); 1545 if (HighBits.intersects(NewMask)) 1546 InDemandedMask |= APInt::getSignBit(VT.getScalarType().getSizeInBits()); 1547 1548 if (SimplifyDemandedBits(Op.getOperand(0), InDemandedMask, 1549 KnownZero, KnownOne, TLO, Depth+1)) 1550 return true; 1551 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?"); 1552 KnownZero = KnownZero.lshr(ShAmt); 1553 KnownOne = KnownOne.lshr(ShAmt); 1554 1555 // Handle the sign bit, adjusted to where it is now in the mask. 1556 APInt SignBit = APInt::getSignBit(BitWidth).lshr(ShAmt); 1557 1558 // If the input sign bit is known to be zero, or if none of the top bits 1559 // are demanded, turn this into an unsigned shift right. 1560 if (KnownZero.intersects(SignBit) || (HighBits & ~NewMask) == HighBits) { 1561 return TLO.CombineTo(Op, TLO.DAG.getNode(ISD::SRL, dl, VT, 1562 Op.getOperand(0), 1563 Op.getOperand(1))); 1564 } else if (KnownOne.intersects(SignBit)) { // New bits are known one. 1565 KnownOne |= HighBits; 1566 } 1567 } 1568 break; 1569 case ISD::SIGN_EXTEND_INREG: { 1570 EVT ExVT = cast<VTSDNode>(Op.getOperand(1))->getVT(); 1571 1572 APInt MsbMask = APInt::getHighBitsSet(BitWidth, 1); 1573 // If we only care about the highest bit, don't bother shifting right. 1574 if (MsbMask == DemandedMask) { 1575 unsigned ShAmt = ExVT.getScalarType().getSizeInBits(); 1576 SDValue InOp = Op.getOperand(0); 1577 1578 // Compute the correct shift amount type, which must be getShiftAmountTy 1579 // for scalar types after legalization. 1580 EVT ShiftAmtTy = Op.getValueType(); 1581 if (TLO.LegalTypes() && !ShiftAmtTy.isVector()) 1582 ShiftAmtTy = getShiftAmountTy(ShiftAmtTy); 1583 1584 SDValue ShiftAmt = TLO.DAG.getConstant(BitWidth - ShAmt, ShiftAmtTy); 1585 return TLO.CombineTo(Op, TLO.DAG.getNode(ISD::SHL, dl, 1586 Op.getValueType(), InOp, ShiftAmt)); 1587 } 1588 1589 // Sign extension. Compute the demanded bits in the result that are not 1590 // present in the input. 1591 APInt NewBits = 1592 APInt::getHighBitsSet(BitWidth, 1593 BitWidth - ExVT.getScalarType().getSizeInBits()); 1594 1595 // If none of the extended bits are demanded, eliminate the sextinreg. 1596 if ((NewBits & NewMask) == 0) 1597 return TLO.CombineTo(Op, Op.getOperand(0)); 1598 1599 APInt InSignBit = 1600 APInt::getSignBit(ExVT.getScalarType().getSizeInBits()).zext(BitWidth); 1601 APInt InputDemandedBits = 1602 APInt::getLowBitsSet(BitWidth, 1603 ExVT.getScalarType().getSizeInBits()) & 1604 NewMask; 1605 1606 // Since the sign extended bits are demanded, we know that the sign 1607 // bit is demanded. 1608 InputDemandedBits |= InSignBit; 1609 1610 if (SimplifyDemandedBits(Op.getOperand(0), InputDemandedBits, 1611 KnownZero, KnownOne, TLO, Depth+1)) 1612 return true; 1613 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?"); 1614 1615 // If the sign bit of the input is known set or clear, then we know the 1616 // top bits of the result. 1617 1618 // If the input sign bit is known zero, convert this into a zero extension. 1619 if (KnownZero.intersects(InSignBit)) 1620 return TLO.CombineTo(Op, 1621 TLO.DAG.getZeroExtendInReg(Op.getOperand(0),dl,ExVT)); 1622 1623 if (KnownOne.intersects(InSignBit)) { // Input sign bit known set 1624 KnownOne |= NewBits; 1625 KnownZero &= ~NewBits; 1626 } else { // Input sign bit unknown 1627 KnownZero &= ~NewBits; 1628 KnownOne &= ~NewBits; 1629 } 1630 break; 1631 } 1632 case ISD::ZERO_EXTEND: { 1633 unsigned OperandBitWidth = 1634 Op.getOperand(0).getValueType().getScalarType().getSizeInBits(); 1635 APInt InMask = NewMask.trunc(OperandBitWidth); 1636 1637 // If none of the top bits are demanded, convert this into an any_extend. 1638 APInt NewBits = 1639 APInt::getHighBitsSet(BitWidth, BitWidth - OperandBitWidth) & NewMask; 1640 if (!NewBits.intersects(NewMask)) 1641 return TLO.CombineTo(Op, TLO.DAG.getNode(ISD::ANY_EXTEND, dl, 1642 Op.getValueType(), 1643 Op.getOperand(0))); 1644 1645 if (SimplifyDemandedBits(Op.getOperand(0), InMask, 1646 KnownZero, KnownOne, TLO, Depth+1)) 1647 return true; 1648 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?"); 1649 KnownZero = KnownZero.zext(BitWidth); 1650 KnownOne = KnownOne.zext(BitWidth); 1651 KnownZero |= NewBits; 1652 break; 1653 } 1654 case ISD::SIGN_EXTEND: { 1655 EVT InVT = Op.getOperand(0).getValueType(); 1656 unsigned InBits = InVT.getScalarType().getSizeInBits(); 1657 APInt InMask = APInt::getLowBitsSet(BitWidth, InBits); 1658 APInt InSignBit = APInt::getBitsSet(BitWidth, InBits - 1, InBits); 1659 APInt NewBits = ~InMask & NewMask; 1660 1661 // If none of the top bits are demanded, convert this into an any_extend. 1662 if (NewBits == 0) 1663 return TLO.CombineTo(Op,TLO.DAG.getNode(ISD::ANY_EXTEND, dl, 1664 Op.getValueType(), 1665 Op.getOperand(0))); 1666 1667 // Since some of the sign extended bits are demanded, we know that the sign 1668 // bit is demanded. 1669 APInt InDemandedBits = InMask & NewMask; 1670 InDemandedBits |= InSignBit; 1671 InDemandedBits = InDemandedBits.trunc(InBits); 1672 1673 if (SimplifyDemandedBits(Op.getOperand(0), InDemandedBits, KnownZero, 1674 KnownOne, TLO, Depth+1)) 1675 return true; 1676 KnownZero = KnownZero.zext(BitWidth); 1677 KnownOne = KnownOne.zext(BitWidth); 1678 1679 // If the sign bit is known zero, convert this to a zero extend. 1680 if (KnownZero.intersects(InSignBit)) 1681 return TLO.CombineTo(Op, TLO.DAG.getNode(ISD::ZERO_EXTEND, dl, 1682 Op.getValueType(), 1683 Op.getOperand(0))); 1684 1685 // If the sign bit is known one, the top bits match. 1686 if (KnownOne.intersects(InSignBit)) { 1687 KnownOne |= NewBits; 1688 assert((KnownZero & NewBits) == 0); 1689 } else { // Otherwise, top bits aren't known. 1690 assert((KnownOne & NewBits) == 0); 1691 assert((KnownZero & NewBits) == 0); 1692 } 1693 break; 1694 } 1695 case ISD::ANY_EXTEND: { 1696 unsigned OperandBitWidth = 1697 Op.getOperand(0).getValueType().getScalarType().getSizeInBits(); 1698 APInt InMask = NewMask.trunc(OperandBitWidth); 1699 if (SimplifyDemandedBits(Op.getOperand(0), InMask, 1700 KnownZero, KnownOne, TLO, Depth+1)) 1701 return true; 1702 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?"); 1703 KnownZero = KnownZero.zext(BitWidth); 1704 KnownOne = KnownOne.zext(BitWidth); 1705 break; 1706 } 1707 case ISD::TRUNCATE: { 1708 // Simplify the input, using demanded bit information, and compute the known 1709 // zero/one bits live out. 1710 unsigned OperandBitWidth = 1711 Op.getOperand(0).getValueType().getScalarType().getSizeInBits(); 1712 APInt TruncMask = NewMask.zext(OperandBitWidth); 1713 if (SimplifyDemandedBits(Op.getOperand(0), TruncMask, 1714 KnownZero, KnownOne, TLO, Depth+1)) 1715 return true; 1716 KnownZero = KnownZero.trunc(BitWidth); 1717 KnownOne = KnownOne.trunc(BitWidth); 1718 1719 // If the input is only used by this truncate, see if we can shrink it based 1720 // on the known demanded bits. 1721 if (Op.getOperand(0).getNode()->hasOneUse()) { 1722 SDValue In = Op.getOperand(0); 1723 switch (In.getOpcode()) { 1724 default: break; 1725 case ISD::SRL: 1726 // Shrink SRL by a constant if none of the high bits shifted in are 1727 // demanded. 1728 if (TLO.LegalTypes() && 1729 !isTypeDesirableForOp(ISD::SRL, Op.getValueType())) 1730 // Do not turn (vt1 truncate (vt2 srl)) into (vt1 srl) if vt1 is 1731 // undesirable. 1732 break; 1733 ConstantSDNode *ShAmt = dyn_cast<ConstantSDNode>(In.getOperand(1)); 1734 if (!ShAmt) 1735 break; 1736 SDValue Shift = In.getOperand(1); 1737 if (TLO.LegalTypes()) { 1738 uint64_t ShVal = ShAmt->getZExtValue(); 1739 Shift = 1740 TLO.DAG.getConstant(ShVal, getShiftAmountTy(Op.getValueType())); 1741 } 1742 1743 APInt HighBits = APInt::getHighBitsSet(OperandBitWidth, 1744 OperandBitWidth - BitWidth); 1745 HighBits = HighBits.lshr(ShAmt->getZExtValue()).trunc(BitWidth); 1746 1747 if (ShAmt->getZExtValue() < BitWidth && !(HighBits & NewMask)) { 1748 // None of the shifted in bits are needed. Add a truncate of the 1749 // shift input, then shift it. 1750 SDValue NewTrunc = TLO.DAG.getNode(ISD::TRUNCATE, dl, 1751 Op.getValueType(), 1752 In.getOperand(0)); 1753 return TLO.CombineTo(Op, TLO.DAG.getNode(ISD::SRL, dl, 1754 Op.getValueType(), 1755 NewTrunc, 1756 Shift)); 1757 } 1758 break; 1759 } 1760 } 1761 1762 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?"); 1763 break; 1764 } 1765 case ISD::AssertZext: { 1766 // AssertZext demands all of the high bits, plus any of the low bits 1767 // demanded by its users. 1768 EVT VT = cast<VTSDNode>(Op.getOperand(1))->getVT(); 1769 APInt InMask = APInt::getLowBitsSet(BitWidth, 1770 VT.getSizeInBits()); 1771 if (SimplifyDemandedBits(Op.getOperand(0), ~InMask | NewMask, 1772 KnownZero, KnownOne, TLO, Depth+1)) 1773 return true; 1774 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?"); 1775 1776 KnownZero |= ~InMask & NewMask; 1777 break; 1778 } 1779 case ISD::BITCAST: 1780 // If this is an FP->Int bitcast and if the sign bit is the only 1781 // thing demanded, turn this into a FGETSIGN. 1782 if (!TLO.LegalOperations() && 1783 !Op.getValueType().isVector() && 1784 !Op.getOperand(0).getValueType().isVector() && 1785 NewMask == APInt::getSignBit(Op.getValueType().getSizeInBits()) && 1786 Op.getOperand(0).getValueType().isFloatingPoint()) { 1787 bool OpVTLegal = isOperationLegalOrCustom(ISD::FGETSIGN, Op.getValueType()); 1788 bool i32Legal = isOperationLegalOrCustom(ISD::FGETSIGN, MVT::i32); 1789 if ((OpVTLegal || i32Legal) && Op.getValueType().isSimple()) { 1790 EVT Ty = OpVTLegal ? Op.getValueType() : MVT::i32; 1791 // Make a FGETSIGN + SHL to move the sign bit into the appropriate 1792 // place. We expect the SHL to be eliminated by other optimizations. 1793 SDValue Sign = TLO.DAG.getNode(ISD::FGETSIGN, dl, Ty, Op.getOperand(0)); 1794 unsigned OpVTSizeInBits = Op.getValueType().getSizeInBits(); 1795 if (!OpVTLegal && OpVTSizeInBits > 32) 1796 Sign = TLO.DAG.getNode(ISD::ZERO_EXTEND, dl, Op.getValueType(), Sign); 1797 unsigned ShVal = Op.getValueType().getSizeInBits()-1; 1798 SDValue ShAmt = TLO.DAG.getConstant(ShVal, Op.getValueType()); 1799 return TLO.CombineTo(Op, TLO.DAG.getNode(ISD::SHL, dl, 1800 Op.getValueType(), 1801 Sign, ShAmt)); 1802 } 1803 } 1804 break; 1805 case ISD::ADD: 1806 case ISD::MUL: 1807 case ISD::SUB: { 1808 // Add, Sub, and Mul don't demand any bits in positions beyond that 1809 // of the highest bit demanded of them. 1810 APInt LoMask = APInt::getLowBitsSet(BitWidth, 1811 BitWidth - NewMask.countLeadingZeros()); 1812 if (SimplifyDemandedBits(Op.getOperand(0), LoMask, KnownZero2, 1813 KnownOne2, TLO, Depth+1)) 1814 return true; 1815 if (SimplifyDemandedBits(Op.getOperand(1), LoMask, KnownZero2, 1816 KnownOne2, TLO, Depth+1)) 1817 return true; 1818 // See if the operation should be performed at a smaller bit width. 1819 if (TLO.ShrinkDemandedOp(Op, BitWidth, NewMask, dl)) 1820 return true; 1821 } 1822 // FALL THROUGH 1823 default: 1824 // Just use ComputeMaskedBits to compute output bits. 1825 TLO.DAG.ComputeMaskedBits(Op, KnownZero, KnownOne, Depth); 1826 break; 1827 } 1828 1829 // If we know the value of all of the demanded bits, return this as a 1830 // constant. 1831 if ((NewMask & (KnownZero|KnownOne)) == NewMask) 1832 return TLO.CombineTo(Op, TLO.DAG.getConstant(KnownOne, Op.getValueType())); 1833 1834 return false; 1835} 1836 1837/// computeMaskedBitsForTargetNode - Determine which of the bits specified 1838/// in Mask are known to be either zero or one and return them in the 1839/// KnownZero/KnownOne bitsets. 1840void TargetLowering::computeMaskedBitsForTargetNode(const SDValue Op, 1841 APInt &KnownZero, 1842 APInt &KnownOne, 1843 const SelectionDAG &DAG, 1844 unsigned Depth) const { 1845 assert((Op.getOpcode() >= ISD::BUILTIN_OP_END || 1846 Op.getOpcode() == ISD::INTRINSIC_WO_CHAIN || 1847 Op.getOpcode() == ISD::INTRINSIC_W_CHAIN || 1848 Op.getOpcode() == ISD::INTRINSIC_VOID) && 1849 "Should use MaskedValueIsZero if you don't know whether Op" 1850 " is a target node!"); 1851 KnownZero = KnownOne = APInt(KnownOne.getBitWidth(), 0); 1852} 1853 1854/// ComputeNumSignBitsForTargetNode - This method can be implemented by 1855/// targets that want to expose additional information about sign bits to the 1856/// DAG Combiner. 1857unsigned TargetLowering::ComputeNumSignBitsForTargetNode(SDValue Op, 1858 unsigned Depth) const { 1859 assert((Op.getOpcode() >= ISD::BUILTIN_OP_END || 1860 Op.getOpcode() == ISD::INTRINSIC_WO_CHAIN || 1861 Op.getOpcode() == ISD::INTRINSIC_W_CHAIN || 1862 Op.getOpcode() == ISD::INTRINSIC_VOID) && 1863 "Should use ComputeNumSignBits if you don't know whether Op" 1864 " is a target node!"); 1865 return 1; 1866} 1867 1868/// ValueHasExactlyOneBitSet - Test if the given value is known to have exactly 1869/// one bit set. This differs from ComputeMaskedBits in that it doesn't need to 1870/// determine which bit is set. 1871/// 1872static bool ValueHasExactlyOneBitSet(SDValue Val, const SelectionDAG &DAG) { 1873 // A left-shift of a constant one will have exactly one bit set, because 1874 // shifting the bit off the end is undefined. 1875 if (Val.getOpcode() == ISD::SHL) 1876 if (ConstantSDNode *C = 1877 dyn_cast<ConstantSDNode>(Val.getNode()->getOperand(0))) 1878 if (C->getAPIntValue() == 1) 1879 return true; 1880 1881 // Similarly, a right-shift of a constant sign-bit will have exactly 1882 // one bit set. 1883 if (Val.getOpcode() == ISD::SRL) 1884 if (ConstantSDNode *C = 1885 dyn_cast<ConstantSDNode>(Val.getNode()->getOperand(0))) 1886 if (C->getAPIntValue().isSignBit()) 1887 return true; 1888 1889 // More could be done here, though the above checks are enough 1890 // to handle some common cases. 1891 1892 // Fall back to ComputeMaskedBits to catch other known cases. 1893 EVT OpVT = Val.getValueType(); 1894 unsigned BitWidth = OpVT.getScalarType().getSizeInBits(); 1895 APInt KnownZero, KnownOne; 1896 DAG.ComputeMaskedBits(Val, KnownZero, KnownOne); 1897 return (KnownZero.countPopulation() == BitWidth - 1) && 1898 (KnownOne.countPopulation() == 1); 1899} 1900 1901/// SimplifySetCC - Try to simplify a setcc built with the specified operands 1902/// and cc. If it is unable to simplify it, return a null SDValue. 1903SDValue 1904TargetLowering::SimplifySetCC(EVT VT, SDValue N0, SDValue N1, 1905 ISD::CondCode Cond, bool foldBooleans, 1906 DAGCombinerInfo &DCI, DebugLoc dl) const { 1907 SelectionDAG &DAG = DCI.DAG; 1908 1909 // These setcc operations always fold. 1910 switch (Cond) { 1911 default: break; 1912 case ISD::SETFALSE: 1913 case ISD::SETFALSE2: return DAG.getConstant(0, VT); 1914 case ISD::SETTRUE: 1915 case ISD::SETTRUE2: return DAG.getConstant(1, VT); 1916 } 1917 1918 // Ensure that the constant occurs on the RHS, and fold constant 1919 // comparisons. 1920 if (isa<ConstantSDNode>(N0.getNode())) 1921 return DAG.getSetCC(dl, VT, N1, N0, ISD::getSetCCSwappedOperands(Cond)); 1922 1923 if (ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1.getNode())) { 1924 const APInt &C1 = N1C->getAPIntValue(); 1925 1926 // If the LHS is '(srl (ctlz x), 5)', the RHS is 0/1, and this is an 1927 // equality comparison, then we're just comparing whether X itself is 1928 // zero. 1929 if (N0.getOpcode() == ISD::SRL && (C1 == 0 || C1 == 1) && 1930 N0.getOperand(0).getOpcode() == ISD::CTLZ && 1931 N0.getOperand(1).getOpcode() == ISD::Constant) { 1932 const APInt &ShAmt 1933 = cast<ConstantSDNode>(N0.getOperand(1))->getAPIntValue(); 1934 if ((Cond == ISD::SETEQ || Cond == ISD::SETNE) && 1935 ShAmt == Log2_32(N0.getValueType().getSizeInBits())) { 1936 if ((C1 == 0) == (Cond == ISD::SETEQ)) { 1937 // (srl (ctlz x), 5) == 0 -> X != 0 1938 // (srl (ctlz x), 5) != 1 -> X != 0 1939 Cond = ISD::SETNE; 1940 } else { 1941 // (srl (ctlz x), 5) != 0 -> X == 0 1942 // (srl (ctlz x), 5) == 1 -> X == 0 1943 Cond = ISD::SETEQ; 1944 } 1945 SDValue Zero = DAG.getConstant(0, N0.getValueType()); 1946 return DAG.getSetCC(dl, VT, N0.getOperand(0).getOperand(0), 1947 Zero, Cond); 1948 } 1949 } 1950 1951 SDValue CTPOP = N0; 1952 // Look through truncs that don't change the value of a ctpop. 1953 if (N0.hasOneUse() && N0.getOpcode() == ISD::TRUNCATE) 1954 CTPOP = N0.getOperand(0); 1955 1956 if (CTPOP.hasOneUse() && CTPOP.getOpcode() == ISD::CTPOP && 1957 (N0 == CTPOP || N0.getValueType().getSizeInBits() > 1958 Log2_32_Ceil(CTPOP.getValueType().getSizeInBits()))) { 1959 EVT CTVT = CTPOP.getValueType(); 1960 SDValue CTOp = CTPOP.getOperand(0); 1961 1962 // (ctpop x) u< 2 -> (x & x-1) == 0 1963 // (ctpop x) u> 1 -> (x & x-1) != 0 1964 if ((Cond == ISD::SETULT && C1 == 2) || (Cond == ISD::SETUGT && C1 == 1)){ 1965 SDValue Sub = DAG.getNode(ISD::SUB, dl, CTVT, CTOp, 1966 DAG.getConstant(1, CTVT)); 1967 SDValue And = DAG.getNode(ISD::AND, dl, CTVT, CTOp, Sub); 1968 ISD::CondCode CC = Cond == ISD::SETULT ? ISD::SETEQ : ISD::SETNE; 1969 return DAG.getSetCC(dl, VT, And, DAG.getConstant(0, CTVT), CC); 1970 } 1971 1972 // TODO: (ctpop x) == 1 -> x && (x & x-1) == 0 iff ctpop is illegal. 1973 } 1974 1975 // (zext x) == C --> x == (trunc C) 1976 if (DCI.isBeforeLegalize() && N0->hasOneUse() && 1977 (Cond == ISD::SETEQ || Cond == ISD::SETNE)) { 1978 unsigned MinBits = N0.getValueSizeInBits(); 1979 SDValue PreZExt; 1980 if (N0->getOpcode() == ISD::ZERO_EXTEND) { 1981 // ZExt 1982 MinBits = N0->getOperand(0).getValueSizeInBits(); 1983 PreZExt = N0->getOperand(0); 1984 } else if (N0->getOpcode() == ISD::AND) { 1985 // DAGCombine turns costly ZExts into ANDs 1986 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(N0->getOperand(1))) 1987 if ((C->getAPIntValue()+1).isPowerOf2()) { 1988 MinBits = C->getAPIntValue().countTrailingOnes(); 1989 PreZExt = N0->getOperand(0); 1990 } 1991 } else if (LoadSDNode *LN0 = dyn_cast<LoadSDNode>(N0)) { 1992 // ZEXTLOAD 1993 if (LN0->getExtensionType() == ISD::ZEXTLOAD) { 1994 MinBits = LN0->getMemoryVT().getSizeInBits(); 1995 PreZExt = N0; 1996 } 1997 } 1998 1999 // Make sure we're not losing bits from the constant. 2000 if (MinBits < C1.getBitWidth() && MinBits > C1.getActiveBits()) { 2001 EVT MinVT = EVT::getIntegerVT(*DAG.getContext(), MinBits); 2002 if (isTypeDesirableForOp(ISD::SETCC, MinVT)) { 2003 // Will get folded away. 2004 SDValue Trunc = DAG.getNode(ISD::TRUNCATE, dl, MinVT, PreZExt); 2005 SDValue C = DAG.getConstant(C1.trunc(MinBits), MinVT); 2006 return DAG.getSetCC(dl, VT, Trunc, C, Cond); 2007 } 2008 } 2009 } 2010 2011 // If the LHS is '(and load, const)', the RHS is 0, 2012 // the test is for equality or unsigned, and all 1 bits of the const are 2013 // in the same partial word, see if we can shorten the load. 2014 if (DCI.isBeforeLegalize() && 2015 N0.getOpcode() == ISD::AND && C1 == 0 && 2016 N0.getNode()->hasOneUse() && 2017 isa<LoadSDNode>(N0.getOperand(0)) && 2018 N0.getOperand(0).getNode()->hasOneUse() && 2019 isa<ConstantSDNode>(N0.getOperand(1))) { 2020 LoadSDNode *Lod = cast<LoadSDNode>(N0.getOperand(0)); 2021 APInt bestMask; 2022 unsigned bestWidth = 0, bestOffset = 0; 2023 if (!Lod->isVolatile() && Lod->isUnindexed()) { 2024 unsigned origWidth = N0.getValueType().getSizeInBits(); 2025 unsigned maskWidth = origWidth; 2026 // We can narrow (e.g.) 16-bit extending loads on 32-bit target to 2027 // 8 bits, but have to be careful... 2028 if (Lod->getExtensionType() != ISD::NON_EXTLOAD) 2029 origWidth = Lod->getMemoryVT().getSizeInBits(); 2030 const APInt &Mask = 2031 cast<ConstantSDNode>(N0.getOperand(1))->getAPIntValue(); 2032 for (unsigned width = origWidth / 2; width>=8; width /= 2) { 2033 APInt newMask = APInt::getLowBitsSet(maskWidth, width); 2034 for (unsigned offset=0; offset<origWidth/width; offset++) { 2035 if ((newMask & Mask) == Mask) { 2036 if (!TD->isLittleEndian()) 2037 bestOffset = (origWidth/width - offset - 1) * (width/8); 2038 else 2039 bestOffset = (uint64_t)offset * (width/8); 2040 bestMask = Mask.lshr(offset * (width/8) * 8); 2041 bestWidth = width; 2042 break; 2043 } 2044 newMask = newMask << width; 2045 } 2046 } 2047 } 2048 if (bestWidth) { 2049 EVT newVT = EVT::getIntegerVT(*DAG.getContext(), bestWidth); 2050 if (newVT.isRound()) { 2051 EVT PtrType = Lod->getOperand(1).getValueType(); 2052 SDValue Ptr = Lod->getBasePtr(); 2053 if (bestOffset != 0) 2054 Ptr = DAG.getNode(ISD::ADD, dl, PtrType, Lod->getBasePtr(), 2055 DAG.getConstant(bestOffset, PtrType)); 2056 unsigned NewAlign = MinAlign(Lod->getAlignment(), bestOffset); 2057 SDValue NewLoad = DAG.getLoad(newVT, dl, Lod->getChain(), Ptr, 2058 Lod->getPointerInfo().getWithOffset(bestOffset), 2059 false, false, false, NewAlign); 2060 return DAG.getSetCC(dl, VT, 2061 DAG.getNode(ISD::AND, dl, newVT, NewLoad, 2062 DAG.getConstant(bestMask.trunc(bestWidth), 2063 newVT)), 2064 DAG.getConstant(0LL, newVT), Cond); 2065 } 2066 } 2067 } 2068 2069 // If the LHS is a ZERO_EXTEND, perform the comparison on the input. 2070 if (N0.getOpcode() == ISD::ZERO_EXTEND) { 2071 unsigned InSize = N0.getOperand(0).getValueType().getSizeInBits(); 2072 2073 // If the comparison constant has bits in the upper part, the 2074 // zero-extended value could never match. 2075 if (C1.intersects(APInt::getHighBitsSet(C1.getBitWidth(), 2076 C1.getBitWidth() - InSize))) { 2077 switch (Cond) { 2078 case ISD::SETUGT: 2079 case ISD::SETUGE: 2080 case ISD::SETEQ: return DAG.getConstant(0, VT); 2081 case ISD::SETULT: 2082 case ISD::SETULE: 2083 case ISD::SETNE: return DAG.getConstant(1, VT); 2084 case ISD::SETGT: 2085 case ISD::SETGE: 2086 // True if the sign bit of C1 is set. 2087 return DAG.getConstant(C1.isNegative(), VT); 2088 case ISD::SETLT: 2089 case ISD::SETLE: 2090 // True if the sign bit of C1 isn't set. 2091 return DAG.getConstant(C1.isNonNegative(), VT); 2092 default: 2093 break; 2094 } 2095 } 2096 2097 // Otherwise, we can perform the comparison with the low bits. 2098 switch (Cond) { 2099 case ISD::SETEQ: 2100 case ISD::SETNE: 2101 case ISD::SETUGT: 2102 case ISD::SETUGE: 2103 case ISD::SETULT: 2104 case ISD::SETULE: { 2105 EVT newVT = N0.getOperand(0).getValueType(); 2106 if (DCI.isBeforeLegalizeOps() || 2107 (isOperationLegal(ISD::SETCC, newVT) && 2108 getCondCodeAction(Cond, newVT)==Legal)) 2109 return DAG.getSetCC(dl, VT, N0.getOperand(0), 2110 DAG.getConstant(C1.trunc(InSize), newVT), 2111 Cond); 2112 break; 2113 } 2114 default: 2115 break; // todo, be more careful with signed comparisons 2116 } 2117 } else if (N0.getOpcode() == ISD::SIGN_EXTEND_INREG && 2118 (Cond == ISD::SETEQ || Cond == ISD::SETNE)) { 2119 EVT ExtSrcTy = cast<VTSDNode>(N0.getOperand(1))->getVT(); 2120 unsigned ExtSrcTyBits = ExtSrcTy.getSizeInBits(); 2121 EVT ExtDstTy = N0.getValueType(); 2122 unsigned ExtDstTyBits = ExtDstTy.getSizeInBits(); 2123 2124 // If the constant doesn't fit into the number of bits for the source of 2125 // the sign extension, it is impossible for both sides to be equal. 2126 if (C1.getMinSignedBits() > ExtSrcTyBits) 2127 return DAG.getConstant(Cond == ISD::SETNE, VT); 2128 2129 SDValue ZextOp; 2130 EVT Op0Ty = N0.getOperand(0).getValueType(); 2131 if (Op0Ty == ExtSrcTy) { 2132 ZextOp = N0.getOperand(0); 2133 } else { 2134 APInt Imm = APInt::getLowBitsSet(ExtDstTyBits, ExtSrcTyBits); 2135 ZextOp = DAG.getNode(ISD::AND, dl, Op0Ty, N0.getOperand(0), 2136 DAG.getConstant(Imm, Op0Ty)); 2137 } 2138 if (!DCI.isCalledByLegalizer()) 2139 DCI.AddToWorklist(ZextOp.getNode()); 2140 // Otherwise, make this a use of a zext. 2141 return DAG.getSetCC(dl, VT, ZextOp, 2142 DAG.getConstant(C1 & APInt::getLowBitsSet( 2143 ExtDstTyBits, 2144 ExtSrcTyBits), 2145 ExtDstTy), 2146 Cond); 2147 } else if ((N1C->isNullValue() || N1C->getAPIntValue() == 1) && 2148 (Cond == ISD::SETEQ || Cond == ISD::SETNE)) { 2149 // SETCC (SETCC), [0|1], [EQ|NE] -> SETCC 2150 if (N0.getOpcode() == ISD::SETCC && 2151 isTypeLegal(VT) && VT.bitsLE(N0.getValueType())) { 2152 bool TrueWhenTrue = (Cond == ISD::SETEQ) ^ (N1C->getAPIntValue() != 1); 2153 if (TrueWhenTrue) 2154 return DAG.getNode(ISD::TRUNCATE, dl, VT, N0); 2155 // Invert the condition. 2156 ISD::CondCode CC = cast<CondCodeSDNode>(N0.getOperand(2))->get(); 2157 CC = ISD::getSetCCInverse(CC, 2158 N0.getOperand(0).getValueType().isInteger()); 2159 return DAG.getSetCC(dl, VT, N0.getOperand(0), N0.getOperand(1), CC); 2160 } 2161 2162 if ((N0.getOpcode() == ISD::XOR || 2163 (N0.getOpcode() == ISD::AND && 2164 N0.getOperand(0).getOpcode() == ISD::XOR && 2165 N0.getOperand(1) == N0.getOperand(0).getOperand(1))) && 2166 isa<ConstantSDNode>(N0.getOperand(1)) && 2167 cast<ConstantSDNode>(N0.getOperand(1))->getAPIntValue() == 1) { 2168 // If this is (X^1) == 0/1, swap the RHS and eliminate the xor. We 2169 // can only do this if the top bits are known zero. 2170 unsigned BitWidth = N0.getValueSizeInBits(); 2171 if (DAG.MaskedValueIsZero(N0, 2172 APInt::getHighBitsSet(BitWidth, 2173 BitWidth-1))) { 2174 // Okay, get the un-inverted input value. 2175 SDValue Val; 2176 if (N0.getOpcode() == ISD::XOR) 2177 Val = N0.getOperand(0); 2178 else { 2179 assert(N0.getOpcode() == ISD::AND && 2180 N0.getOperand(0).getOpcode() == ISD::XOR); 2181 // ((X^1)&1)^1 -> X & 1 2182 Val = DAG.getNode(ISD::AND, dl, N0.getValueType(), 2183 N0.getOperand(0).getOperand(0), 2184 N0.getOperand(1)); 2185 } 2186 2187 return DAG.getSetCC(dl, VT, Val, N1, 2188 Cond == ISD::SETEQ ? ISD::SETNE : ISD::SETEQ); 2189 } 2190 } else if (N1C->getAPIntValue() == 1 && 2191 (VT == MVT::i1 || 2192 getBooleanContents(false) == ZeroOrOneBooleanContent)) { 2193 SDValue Op0 = N0; 2194 if (Op0.getOpcode() == ISD::TRUNCATE) 2195 Op0 = Op0.getOperand(0); 2196 2197 if ((Op0.getOpcode() == ISD::XOR) && 2198 Op0.getOperand(0).getOpcode() == ISD::SETCC && 2199 Op0.getOperand(1).getOpcode() == ISD::SETCC) { 2200 // (xor (setcc), (setcc)) == / != 1 -> (setcc) != / == (setcc) 2201 Cond = (Cond == ISD::SETEQ) ? ISD::SETNE : ISD::SETEQ; 2202 return DAG.getSetCC(dl, VT, Op0.getOperand(0), Op0.getOperand(1), 2203 Cond); 2204 } else if (Op0.getOpcode() == ISD::AND && 2205 isa<ConstantSDNode>(Op0.getOperand(1)) && 2206 cast<ConstantSDNode>(Op0.getOperand(1))->getAPIntValue() == 1) { 2207 // If this is (X&1) == / != 1, normalize it to (X&1) != / == 0. 2208 if (Op0.getValueType().bitsGT(VT)) 2209 Op0 = DAG.getNode(ISD::AND, dl, VT, 2210 DAG.getNode(ISD::TRUNCATE, dl, VT, Op0.getOperand(0)), 2211 DAG.getConstant(1, VT)); 2212 else if (Op0.getValueType().bitsLT(VT)) 2213 Op0 = DAG.getNode(ISD::AND, dl, VT, 2214 DAG.getNode(ISD::ANY_EXTEND, dl, VT, Op0.getOperand(0)), 2215 DAG.getConstant(1, VT)); 2216 2217 return DAG.getSetCC(dl, VT, Op0, 2218 DAG.getConstant(0, Op0.getValueType()), 2219 Cond == ISD::SETEQ ? ISD::SETNE : ISD::SETEQ); 2220 } 2221 } 2222 } 2223 2224 APInt MinVal, MaxVal; 2225 unsigned OperandBitSize = N1C->getValueType(0).getSizeInBits(); 2226 if (ISD::isSignedIntSetCC(Cond)) { 2227 MinVal = APInt::getSignedMinValue(OperandBitSize); 2228 MaxVal = APInt::getSignedMaxValue(OperandBitSize); 2229 } else { 2230 MinVal = APInt::getMinValue(OperandBitSize); 2231 MaxVal = APInt::getMaxValue(OperandBitSize); 2232 } 2233 2234 // Canonicalize GE/LE comparisons to use GT/LT comparisons. 2235 if (Cond == ISD::SETGE || Cond == ISD::SETUGE) { 2236 if (C1 == MinVal) return DAG.getConstant(1, VT); // X >= MIN --> true 2237 // X >= C0 --> X > (C0-1) 2238 return DAG.getSetCC(dl, VT, N0, 2239 DAG.getConstant(C1-1, N1.getValueType()), 2240 (Cond == ISD::SETGE) ? ISD::SETGT : ISD::SETUGT); 2241 } 2242 2243 if (Cond == ISD::SETLE || Cond == ISD::SETULE) { 2244 if (C1 == MaxVal) return DAG.getConstant(1, VT); // X <= MAX --> true 2245 // X <= C0 --> X < (C0+1) 2246 return DAG.getSetCC(dl, VT, N0, 2247 DAG.getConstant(C1+1, N1.getValueType()), 2248 (Cond == ISD::SETLE) ? ISD::SETLT : ISD::SETULT); 2249 } 2250 2251 if ((Cond == ISD::SETLT || Cond == ISD::SETULT) && C1 == MinVal) 2252 return DAG.getConstant(0, VT); // X < MIN --> false 2253 if ((Cond == ISD::SETGE || Cond == ISD::SETUGE) && C1 == MinVal) 2254 return DAG.getConstant(1, VT); // X >= MIN --> true 2255 if ((Cond == ISD::SETGT || Cond == ISD::SETUGT) && C1 == MaxVal) 2256 return DAG.getConstant(0, VT); // X > MAX --> false 2257 if ((Cond == ISD::SETLE || Cond == ISD::SETULE) && C1 == MaxVal) 2258 return DAG.getConstant(1, VT); // X <= MAX --> true 2259 2260 // Canonicalize setgt X, Min --> setne X, Min 2261 if ((Cond == ISD::SETGT || Cond == ISD::SETUGT) && C1 == MinVal) 2262 return DAG.getSetCC(dl, VT, N0, N1, ISD::SETNE); 2263 // Canonicalize setlt X, Max --> setne X, Max 2264 if ((Cond == ISD::SETLT || Cond == ISD::SETULT) && C1 == MaxVal) 2265 return DAG.getSetCC(dl, VT, N0, N1, ISD::SETNE); 2266 2267 // If we have setult X, 1, turn it into seteq X, 0 2268 if ((Cond == ISD::SETLT || Cond == ISD::SETULT) && C1 == MinVal+1) 2269 return DAG.getSetCC(dl, VT, N0, 2270 DAG.getConstant(MinVal, N0.getValueType()), 2271 ISD::SETEQ); 2272 // If we have setugt X, Max-1, turn it into seteq X, Max 2273 else if ((Cond == ISD::SETGT || Cond == ISD::SETUGT) && C1 == MaxVal-1) 2274 return DAG.getSetCC(dl, VT, N0, 2275 DAG.getConstant(MaxVal, N0.getValueType()), 2276 ISD::SETEQ); 2277 2278 // If we have "setcc X, C0", check to see if we can shrink the immediate 2279 // by changing cc. 2280 2281 // SETUGT X, SINTMAX -> SETLT X, 0 2282 if (Cond == ISD::SETUGT && 2283 C1 == APInt::getSignedMaxValue(OperandBitSize)) 2284 return DAG.getSetCC(dl, VT, N0, 2285 DAG.getConstant(0, N1.getValueType()), 2286 ISD::SETLT); 2287 2288 // SETULT X, SINTMIN -> SETGT X, -1 2289 if (Cond == ISD::SETULT && 2290 C1 == APInt::getSignedMinValue(OperandBitSize)) { 2291 SDValue ConstMinusOne = 2292 DAG.getConstant(APInt::getAllOnesValue(OperandBitSize), 2293 N1.getValueType()); 2294 return DAG.getSetCC(dl, VT, N0, ConstMinusOne, ISD::SETGT); 2295 } 2296 2297 // Fold bit comparisons when we can. 2298 if ((Cond == ISD::SETEQ || Cond == ISD::SETNE) && 2299 (VT == N0.getValueType() || 2300 (isTypeLegal(VT) && VT.bitsLE(N0.getValueType()))) && 2301 N0.getOpcode() == ISD::AND) 2302 if (ConstantSDNode *AndRHS = 2303 dyn_cast<ConstantSDNode>(N0.getOperand(1))) { 2304 EVT ShiftTy = DCI.isBeforeLegalizeOps() ? 2305 getPointerTy() : getShiftAmountTy(N0.getValueType()); 2306 if (Cond == ISD::SETNE && C1 == 0) {// (X & 8) != 0 --> (X & 8) >> 3 2307 // Perform the xform if the AND RHS is a single bit. 2308 if (AndRHS->getAPIntValue().isPowerOf2()) { 2309 return DAG.getNode(ISD::TRUNCATE, dl, VT, 2310 DAG.getNode(ISD::SRL, dl, N0.getValueType(), N0, 2311 DAG.getConstant(AndRHS->getAPIntValue().logBase2(), ShiftTy))); 2312 } 2313 } else if (Cond == ISD::SETEQ && C1 == AndRHS->getAPIntValue()) { 2314 // (X & 8) == 8 --> (X & 8) >> 3 2315 // Perform the xform if C1 is a single bit. 2316 if (C1.isPowerOf2()) { 2317 return DAG.getNode(ISD::TRUNCATE, dl, VT, 2318 DAG.getNode(ISD::SRL, dl, N0.getValueType(), N0, 2319 DAG.getConstant(C1.logBase2(), ShiftTy))); 2320 } 2321 } 2322 } 2323 2324 if (C1.getMinSignedBits() <= 64 && 2325 !isLegalICmpImmediate(C1.getSExtValue())) { 2326 // (X & -256) == 256 -> (X >> 8) == 1 2327 if ((Cond == ISD::SETEQ || Cond == ISD::SETNE) && 2328 N0.getOpcode() == ISD::AND && N0.hasOneUse()) { 2329 if (ConstantSDNode *AndRHS = 2330 dyn_cast<ConstantSDNode>(N0.getOperand(1))) { 2331 const APInt &AndRHSC = AndRHS->getAPIntValue(); 2332 if ((-AndRHSC).isPowerOf2() && (AndRHSC & C1) == C1) { 2333 unsigned ShiftBits = AndRHSC.countTrailingZeros(); 2334 EVT ShiftTy = DCI.isBeforeLegalizeOps() ? 2335 getPointerTy() : getShiftAmountTy(N0.getValueType()); 2336 EVT CmpTy = N0.getValueType(); 2337 SDValue Shift = DAG.getNode(ISD::SRL, dl, CmpTy, N0.getOperand(0), 2338 DAG.getConstant(ShiftBits, ShiftTy)); 2339 SDValue CmpRHS = DAG.getConstant(C1.lshr(ShiftBits), CmpTy); 2340 return DAG.getSetCC(dl, VT, Shift, CmpRHS, Cond); 2341 } 2342 } 2343 } else if (Cond == ISD::SETULT || Cond == ISD::SETUGE || 2344 Cond == ISD::SETULE || Cond == ISD::SETUGT) { 2345 bool AdjOne = (Cond == ISD::SETULE || Cond == ISD::SETUGT); 2346 // X < 0x100000000 -> (X >> 32) < 1 2347 // X >= 0x100000000 -> (X >> 32) >= 1 2348 // X <= 0x0ffffffff -> (X >> 32) < 1 2349 // X > 0x0ffffffff -> (X >> 32) >= 1 2350 unsigned ShiftBits; 2351 APInt NewC = C1; 2352 ISD::CondCode NewCond = Cond; 2353 if (AdjOne) { 2354 ShiftBits = C1.countTrailingOnes(); 2355 NewC = NewC + 1; 2356 NewCond = (Cond == ISD::SETULE) ? ISD::SETULT : ISD::SETUGE; 2357 } else { 2358 ShiftBits = C1.countTrailingZeros(); 2359 } 2360 NewC = NewC.lshr(ShiftBits); 2361 if (ShiftBits && isLegalICmpImmediate(NewC.getSExtValue())) { 2362 EVT ShiftTy = DCI.isBeforeLegalizeOps() ? 2363 getPointerTy() : getShiftAmountTy(N0.getValueType()); 2364 EVT CmpTy = N0.getValueType(); 2365 SDValue Shift = DAG.getNode(ISD::SRL, dl, CmpTy, N0, 2366 DAG.getConstant(ShiftBits, ShiftTy)); 2367 SDValue CmpRHS = DAG.getConstant(NewC, CmpTy); 2368 return DAG.getSetCC(dl, VT, Shift, CmpRHS, NewCond); 2369 } 2370 } 2371 } 2372 } 2373 2374 if (isa<ConstantFPSDNode>(N0.getNode())) { 2375 // Constant fold or commute setcc. 2376 SDValue O = DAG.FoldSetCC(VT, N0, N1, Cond, dl); 2377 if (O.getNode()) return O; 2378 } else if (ConstantFPSDNode *CFP = dyn_cast<ConstantFPSDNode>(N1.getNode())) { 2379 // If the RHS of an FP comparison is a constant, simplify it away in 2380 // some cases. 2381 if (CFP->getValueAPF().isNaN()) { 2382 // If an operand is known to be a nan, we can fold it. 2383 switch (ISD::getUnorderedFlavor(Cond)) { 2384 default: llvm_unreachable("Unknown flavor!"); 2385 case 0: // Known false. 2386 return DAG.getConstant(0, VT); 2387 case 1: // Known true. 2388 return DAG.getConstant(1, VT); 2389 case 2: // Undefined. 2390 return DAG.getUNDEF(VT); 2391 } 2392 } 2393 2394 // Otherwise, we know the RHS is not a NaN. Simplify the node to drop the 2395 // constant if knowing that the operand is non-nan is enough. We prefer to 2396 // have SETO(x,x) instead of SETO(x, 0.0) because this avoids having to 2397 // materialize 0.0. 2398 if (Cond == ISD::SETO || Cond == ISD::SETUO) 2399 return DAG.getSetCC(dl, VT, N0, N0, Cond); 2400 2401 // If the condition is not legal, see if we can find an equivalent one 2402 // which is legal. 2403 if (!isCondCodeLegal(Cond, N0.getValueType())) { 2404 // If the comparison was an awkward floating-point == or != and one of 2405 // the comparison operands is infinity or negative infinity, convert the 2406 // condition to a less-awkward <= or >=. 2407 if (CFP->getValueAPF().isInfinity()) { 2408 if (CFP->getValueAPF().isNegative()) { 2409 if (Cond == ISD::SETOEQ && 2410 isCondCodeLegal(ISD::SETOLE, N0.getValueType())) 2411 return DAG.getSetCC(dl, VT, N0, N1, ISD::SETOLE); 2412 if (Cond == ISD::SETUEQ && 2413 isCondCodeLegal(ISD::SETOLE, N0.getValueType())) 2414 return DAG.getSetCC(dl, VT, N0, N1, ISD::SETULE); 2415 if (Cond == ISD::SETUNE && 2416 isCondCodeLegal(ISD::SETUGT, N0.getValueType())) 2417 return DAG.getSetCC(dl, VT, N0, N1, ISD::SETUGT); 2418 if (Cond == ISD::SETONE && 2419 isCondCodeLegal(ISD::SETUGT, N0.getValueType())) 2420 return DAG.getSetCC(dl, VT, N0, N1, ISD::SETOGT); 2421 } else { 2422 if (Cond == ISD::SETOEQ && 2423 isCondCodeLegal(ISD::SETOGE, N0.getValueType())) 2424 return DAG.getSetCC(dl, VT, N0, N1, ISD::SETOGE); 2425 if (Cond == ISD::SETUEQ && 2426 isCondCodeLegal(ISD::SETOGE, N0.getValueType())) 2427 return DAG.getSetCC(dl, VT, N0, N1, ISD::SETUGE); 2428 if (Cond == ISD::SETUNE && 2429 isCondCodeLegal(ISD::SETULT, N0.getValueType())) 2430 return DAG.getSetCC(dl, VT, N0, N1, ISD::SETULT); 2431 if (Cond == ISD::SETONE && 2432 isCondCodeLegal(ISD::SETULT, N0.getValueType())) 2433 return DAG.getSetCC(dl, VT, N0, N1, ISD::SETOLT); 2434 } 2435 } 2436 } 2437 } 2438 2439 if (N0 == N1) { 2440 // The sext(setcc()) => setcc() optimization relies on the appropriate 2441 // constant being emitted. 2442 uint64_t EqVal = 0; 2443 switch (getBooleanContents(N0.getValueType().isVector())) { 2444 case UndefinedBooleanContent: 2445 case ZeroOrOneBooleanContent: 2446 EqVal = ISD::isTrueWhenEqual(Cond); 2447 break; 2448 case ZeroOrNegativeOneBooleanContent: 2449 EqVal = ISD::isTrueWhenEqual(Cond) ? -1 : 0; 2450 break; 2451 } 2452 2453 // We can always fold X == X for integer setcc's. 2454 if (N0.getValueType().isInteger()) { 2455 return DAG.getConstant(EqVal, VT); 2456 } 2457 unsigned UOF = ISD::getUnorderedFlavor(Cond); 2458 if (UOF == 2) // FP operators that are undefined on NaNs. 2459 return DAG.getConstant(EqVal, VT); 2460 if (UOF == unsigned(ISD::isTrueWhenEqual(Cond))) 2461 return DAG.getConstant(EqVal, VT); 2462 // Otherwise, we can't fold it. However, we can simplify it to SETUO/SETO 2463 // if it is not already. 2464 ISD::CondCode NewCond = UOF == 0 ? ISD::SETO : ISD::SETUO; 2465 if (NewCond != Cond && (DCI.isBeforeLegalizeOps() || 2466 getCondCodeAction(NewCond, N0.getValueType()) == Legal)) 2467 return DAG.getSetCC(dl, VT, N0, N1, NewCond); 2468 } 2469 2470 if ((Cond == ISD::SETEQ || Cond == ISD::SETNE) && 2471 N0.getValueType().isInteger()) { 2472 if (N0.getOpcode() == ISD::ADD || N0.getOpcode() == ISD::SUB || 2473 N0.getOpcode() == ISD::XOR) { 2474 // Simplify (X+Y) == (X+Z) --> Y == Z 2475 if (N0.getOpcode() == N1.getOpcode()) { 2476 if (N0.getOperand(0) == N1.getOperand(0)) 2477 return DAG.getSetCC(dl, VT, N0.getOperand(1), N1.getOperand(1), Cond); 2478 if (N0.getOperand(1) == N1.getOperand(1)) 2479 return DAG.getSetCC(dl, VT, N0.getOperand(0), N1.getOperand(0), Cond); 2480 if (DAG.isCommutativeBinOp(N0.getOpcode())) { 2481 // If X op Y == Y op X, try other combinations. 2482 if (N0.getOperand(0) == N1.getOperand(1)) 2483 return DAG.getSetCC(dl, VT, N0.getOperand(1), N1.getOperand(0), 2484 Cond); 2485 if (N0.getOperand(1) == N1.getOperand(0)) 2486 return DAG.getSetCC(dl, VT, N0.getOperand(0), N1.getOperand(1), 2487 Cond); 2488 } 2489 } 2490 2491 // If RHS is a legal immediate value for a compare instruction, we need 2492 // to be careful about increasing register pressure needlessly. 2493 bool LegalRHSImm = false; 2494 2495 if (ConstantSDNode *RHSC = dyn_cast<ConstantSDNode>(N1)) { 2496 if (ConstantSDNode *LHSR = dyn_cast<ConstantSDNode>(N0.getOperand(1))) { 2497 // Turn (X+C1) == C2 --> X == C2-C1 2498 if (N0.getOpcode() == ISD::ADD && N0.getNode()->hasOneUse()) { 2499 return DAG.getSetCC(dl, VT, N0.getOperand(0), 2500 DAG.getConstant(RHSC->getAPIntValue()- 2501 LHSR->getAPIntValue(), 2502 N0.getValueType()), Cond); 2503 } 2504 2505 // Turn (X^C1) == C2 into X == C1^C2 iff X&~C1 = 0. 2506 if (N0.getOpcode() == ISD::XOR) 2507 // If we know that all of the inverted bits are zero, don't bother 2508 // performing the inversion. 2509 if (DAG.MaskedValueIsZero(N0.getOperand(0), ~LHSR->getAPIntValue())) 2510 return 2511 DAG.getSetCC(dl, VT, N0.getOperand(0), 2512 DAG.getConstant(LHSR->getAPIntValue() ^ 2513 RHSC->getAPIntValue(), 2514 N0.getValueType()), 2515 Cond); 2516 } 2517 2518 // Turn (C1-X) == C2 --> X == C1-C2 2519 if (ConstantSDNode *SUBC = dyn_cast<ConstantSDNode>(N0.getOperand(0))) { 2520 if (N0.getOpcode() == ISD::SUB && N0.getNode()->hasOneUse()) { 2521 return 2522 DAG.getSetCC(dl, VT, N0.getOperand(1), 2523 DAG.getConstant(SUBC->getAPIntValue() - 2524 RHSC->getAPIntValue(), 2525 N0.getValueType()), 2526 Cond); 2527 } 2528 } 2529 2530 // Could RHSC fold directly into a compare? 2531 if (RHSC->getValueType(0).getSizeInBits() <= 64) 2532 LegalRHSImm = isLegalICmpImmediate(RHSC->getSExtValue()); 2533 } 2534 2535 // Simplify (X+Z) == X --> Z == 0 2536 // Don't do this if X is an immediate that can fold into a cmp 2537 // instruction and X+Z has other uses. It could be an induction variable 2538 // chain, and the transform would increase register pressure. 2539 if (!LegalRHSImm || N0.getNode()->hasOneUse()) { 2540 if (N0.getOperand(0) == N1) 2541 return DAG.getSetCC(dl, VT, N0.getOperand(1), 2542 DAG.getConstant(0, N0.getValueType()), Cond); 2543 if (N0.getOperand(1) == N1) { 2544 if (DAG.isCommutativeBinOp(N0.getOpcode())) 2545 return DAG.getSetCC(dl, VT, N0.getOperand(0), 2546 DAG.getConstant(0, N0.getValueType()), Cond); 2547 else if (N0.getNode()->hasOneUse()) { 2548 assert(N0.getOpcode() == ISD::SUB && "Unexpected operation!"); 2549 // (Z-X) == X --> Z == X<<1 2550 SDValue SH = DAG.getNode(ISD::SHL, dl, N1.getValueType(), N1, 2551 DAG.getConstant(1, getShiftAmountTy(N1.getValueType()))); 2552 if (!DCI.isCalledByLegalizer()) 2553 DCI.AddToWorklist(SH.getNode()); 2554 return DAG.getSetCC(dl, VT, N0.getOperand(0), SH, Cond); 2555 } 2556 } 2557 } 2558 } 2559 2560 if (N1.getOpcode() == ISD::ADD || N1.getOpcode() == ISD::SUB || 2561 N1.getOpcode() == ISD::XOR) { 2562 // Simplify X == (X+Z) --> Z == 0 2563 if (N1.getOperand(0) == N0) { 2564 return DAG.getSetCC(dl, VT, N1.getOperand(1), 2565 DAG.getConstant(0, N1.getValueType()), Cond); 2566 } else if (N1.getOperand(1) == N0) { 2567 if (DAG.isCommutativeBinOp(N1.getOpcode())) { 2568 return DAG.getSetCC(dl, VT, N1.getOperand(0), 2569 DAG.getConstant(0, N1.getValueType()), Cond); 2570 } else if (N1.getNode()->hasOneUse()) { 2571 assert(N1.getOpcode() == ISD::SUB && "Unexpected operation!"); 2572 // X == (Z-X) --> X<<1 == Z 2573 SDValue SH = DAG.getNode(ISD::SHL, dl, N1.getValueType(), N0, 2574 DAG.getConstant(1, getShiftAmountTy(N0.getValueType()))); 2575 if (!DCI.isCalledByLegalizer()) 2576 DCI.AddToWorklist(SH.getNode()); 2577 return DAG.getSetCC(dl, VT, SH, N1.getOperand(0), Cond); 2578 } 2579 } 2580 } 2581 2582 // Simplify x&y == y to x&y != 0 if y has exactly one bit set. 2583 // Note that where y is variable and is known to have at most 2584 // one bit set (for example, if it is z&1) we cannot do this; 2585 // the expressions are not equivalent when y==0. 2586 if (N0.getOpcode() == ISD::AND) 2587 if (N0.getOperand(0) == N1 || N0.getOperand(1) == N1) { 2588 if (ValueHasExactlyOneBitSet(N1, DAG)) { 2589 Cond = ISD::getSetCCInverse(Cond, /*isInteger=*/true); 2590 SDValue Zero = DAG.getConstant(0, N1.getValueType()); 2591 return DAG.getSetCC(dl, VT, N0, Zero, Cond); 2592 } 2593 } 2594 if (N1.getOpcode() == ISD::AND) 2595 if (N1.getOperand(0) == N0 || N1.getOperand(1) == N0) { 2596 if (ValueHasExactlyOneBitSet(N0, DAG)) { 2597 Cond = ISD::getSetCCInverse(Cond, /*isInteger=*/true); 2598 SDValue Zero = DAG.getConstant(0, N0.getValueType()); 2599 return DAG.getSetCC(dl, VT, N1, Zero, Cond); 2600 } 2601 } 2602 } 2603 2604 // Fold away ALL boolean setcc's. 2605 SDValue Temp; 2606 if (N0.getValueType() == MVT::i1 && foldBooleans) { 2607 switch (Cond) { 2608 default: llvm_unreachable("Unknown integer setcc!"); 2609 case ISD::SETEQ: // X == Y -> ~(X^Y) 2610 Temp = DAG.getNode(ISD::XOR, dl, MVT::i1, N0, N1); 2611 N0 = DAG.getNOT(dl, Temp, MVT::i1); 2612 if (!DCI.isCalledByLegalizer()) 2613 DCI.AddToWorklist(Temp.getNode()); 2614 break; 2615 case ISD::SETNE: // X != Y --> (X^Y) 2616 N0 = DAG.getNode(ISD::XOR, dl, MVT::i1, N0, N1); 2617 break; 2618 case ISD::SETGT: // X >s Y --> X == 0 & Y == 1 --> ~X & Y 2619 case ISD::SETULT: // X <u Y --> X == 0 & Y == 1 --> ~X & Y 2620 Temp = DAG.getNOT(dl, N0, MVT::i1); 2621 N0 = DAG.getNode(ISD::AND, dl, MVT::i1, N1, Temp); 2622 if (!DCI.isCalledByLegalizer()) 2623 DCI.AddToWorklist(Temp.getNode()); 2624 break; 2625 case ISD::SETLT: // X <s Y --> X == 1 & Y == 0 --> ~Y & X 2626 case ISD::SETUGT: // X >u Y --> X == 1 & Y == 0 --> ~Y & X 2627 Temp = DAG.getNOT(dl, N1, MVT::i1); 2628 N0 = DAG.getNode(ISD::AND, dl, MVT::i1, N0, Temp); 2629 if (!DCI.isCalledByLegalizer()) 2630 DCI.AddToWorklist(Temp.getNode()); 2631 break; 2632 case ISD::SETULE: // X <=u Y --> X == 0 | Y == 1 --> ~X | Y 2633 case ISD::SETGE: // X >=s Y --> X == 0 | Y == 1 --> ~X | Y 2634 Temp = DAG.getNOT(dl, N0, MVT::i1); 2635 N0 = DAG.getNode(ISD::OR, dl, MVT::i1, N1, Temp); 2636 if (!DCI.isCalledByLegalizer()) 2637 DCI.AddToWorklist(Temp.getNode()); 2638 break; 2639 case ISD::SETUGE: // X >=u Y --> X == 1 | Y == 0 --> ~Y | X 2640 case ISD::SETLE: // X <=s Y --> X == 1 | Y == 0 --> ~Y | X 2641 Temp = DAG.getNOT(dl, N1, MVT::i1); 2642 N0 = DAG.getNode(ISD::OR, dl, MVT::i1, N0, Temp); 2643 break; 2644 } 2645 if (VT != MVT::i1) { 2646 if (!DCI.isCalledByLegalizer()) 2647 DCI.AddToWorklist(N0.getNode()); 2648 // FIXME: If running after legalize, we probably can't do this. 2649 N0 = DAG.getNode(ISD::ZERO_EXTEND, dl, VT, N0); 2650 } 2651 return N0; 2652 } 2653 2654 // Could not fold it. 2655 return SDValue(); 2656} 2657 2658/// isGAPlusOffset - Returns true (and the GlobalValue and the offset) if the 2659/// node is a GlobalAddress + offset. 2660bool TargetLowering::isGAPlusOffset(SDNode *N, const GlobalValue *&GA, 2661 int64_t &Offset) const { 2662 if (isa<GlobalAddressSDNode>(N)) { 2663 GlobalAddressSDNode *GASD = cast<GlobalAddressSDNode>(N); 2664 GA = GASD->getGlobal(); 2665 Offset += GASD->getOffset(); 2666 return true; 2667 } 2668 2669 if (N->getOpcode() == ISD::ADD) { 2670 SDValue N1 = N->getOperand(0); 2671 SDValue N2 = N->getOperand(1); 2672 if (isGAPlusOffset(N1.getNode(), GA, Offset)) { 2673 ConstantSDNode *V = dyn_cast<ConstantSDNode>(N2); 2674 if (V) { 2675 Offset += V->getSExtValue(); 2676 return true; 2677 } 2678 } else if (isGAPlusOffset(N2.getNode(), GA, Offset)) { 2679 ConstantSDNode *V = dyn_cast<ConstantSDNode>(N1); 2680 if (V) { 2681 Offset += V->getSExtValue(); 2682 return true; 2683 } 2684 } 2685 } 2686 2687 return false; 2688} 2689 2690 2691SDValue TargetLowering:: 2692PerformDAGCombine(SDNode *N, DAGCombinerInfo &DCI) const { 2693 // Default implementation: no optimization. 2694 return SDValue(); 2695} 2696 2697//===----------------------------------------------------------------------===// 2698// Inline Assembler Implementation Methods 2699//===----------------------------------------------------------------------===// 2700 2701 2702TargetLowering::ConstraintType 2703TargetLowering::getConstraintType(const std::string &Constraint) const { 2704 if (Constraint.size() == 1) { 2705 switch (Constraint[0]) { 2706 default: break; 2707 case 'r': return C_RegisterClass; 2708 case 'm': // memory 2709 case 'o': // offsetable 2710 case 'V': // not offsetable 2711 return C_Memory; 2712 case 'i': // Simple Integer or Relocatable Constant 2713 case 'n': // Simple Integer 2714 case 'E': // Floating Point Constant 2715 case 'F': // Floating Point Constant 2716 case 's': // Relocatable Constant 2717 case 'p': // Address. 2718 case 'X': // Allow ANY value. 2719 case 'I': // Target registers. 2720 case 'J': 2721 case 'K': 2722 case 'L': 2723 case 'M': 2724 case 'N': 2725 case 'O': 2726 case 'P': 2727 case '<': 2728 case '>': 2729 return C_Other; 2730 } 2731 } 2732 2733 if (Constraint.size() > 1 && Constraint[0] == '{' && 2734 Constraint[Constraint.size()-1] == '}') 2735 return C_Register; 2736 return C_Unknown; 2737} 2738 2739/// LowerXConstraint - try to replace an X constraint, which matches anything, 2740/// with another that has more specific requirements based on the type of the 2741/// corresponding operand. 2742const char *TargetLowering::LowerXConstraint(EVT ConstraintVT) const{ 2743 if (ConstraintVT.isInteger()) 2744 return "r"; 2745 if (ConstraintVT.isFloatingPoint()) 2746 return "f"; // works for many targets 2747 return 0; 2748} 2749 2750/// LowerAsmOperandForConstraint - Lower the specified operand into the Ops 2751/// vector. If it is invalid, don't add anything to Ops. 2752void TargetLowering::LowerAsmOperandForConstraint(SDValue Op, 2753 std::string &Constraint, 2754 std::vector<SDValue> &Ops, 2755 SelectionDAG &DAG) const { 2756 2757 if (Constraint.length() > 1) return; 2758 2759 char ConstraintLetter = Constraint[0]; 2760 switch (ConstraintLetter) { 2761 default: break; 2762 case 'X': // Allows any operand; labels (basic block) use this. 2763 if (Op.getOpcode() == ISD::BasicBlock) { 2764 Ops.push_back(Op); 2765 return; 2766 } 2767 // fall through 2768 case 'i': // Simple Integer or Relocatable Constant 2769 case 'n': // Simple Integer 2770 case 's': { // Relocatable Constant 2771 // These operands are interested in values of the form (GV+C), where C may 2772 // be folded in as an offset of GV, or it may be explicitly added. Also, it 2773 // is possible and fine if either GV or C are missing. 2774 ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op); 2775 GlobalAddressSDNode *GA = dyn_cast<GlobalAddressSDNode>(Op); 2776 2777 // If we have "(add GV, C)", pull out GV/C 2778 if (Op.getOpcode() == ISD::ADD) { 2779 C = dyn_cast<ConstantSDNode>(Op.getOperand(1)); 2780 GA = dyn_cast<GlobalAddressSDNode>(Op.getOperand(0)); 2781 if (C == 0 || GA == 0) { 2782 C = dyn_cast<ConstantSDNode>(Op.getOperand(0)); 2783 GA = dyn_cast<GlobalAddressSDNode>(Op.getOperand(1)); 2784 } 2785 if (C == 0 || GA == 0) 2786 C = 0, GA = 0; 2787 } 2788 2789 // If we find a valid operand, map to the TargetXXX version so that the 2790 // value itself doesn't get selected. 2791 if (GA) { // Either &GV or &GV+C 2792 if (ConstraintLetter != 'n') { 2793 int64_t Offs = GA->getOffset(); 2794 if (C) Offs += C->getZExtValue(); 2795 Ops.push_back(DAG.getTargetGlobalAddress(GA->getGlobal(), 2796 C ? C->getDebugLoc() : DebugLoc(), 2797 Op.getValueType(), Offs)); 2798 return; 2799 } 2800 } 2801 if (C) { // just C, no GV. 2802 // Simple constants are not allowed for 's'. 2803 if (ConstraintLetter != 's') { 2804 // gcc prints these as sign extended. Sign extend value to 64 bits 2805 // now; without this it would get ZExt'd later in 2806 // ScheduleDAGSDNodes::EmitNode, which is very generic. 2807 Ops.push_back(DAG.getTargetConstant(C->getAPIntValue().getSExtValue(), 2808 MVT::i64)); 2809 return; 2810 } 2811 } 2812 break; 2813 } 2814 } 2815} 2816 2817std::pair<unsigned, const TargetRegisterClass*> TargetLowering:: 2818getRegForInlineAsmConstraint(const std::string &Constraint, 2819 EVT VT) const { 2820 if (Constraint[0] != '{') 2821 return std::make_pair(0u, static_cast<TargetRegisterClass*>(0)); 2822 assert(*(Constraint.end()-1) == '}' && "Not a brace enclosed constraint?"); 2823 2824 // Remove the braces from around the name. 2825 StringRef RegName(Constraint.data()+1, Constraint.size()-2); 2826 2827 // Figure out which register class contains this reg. 2828 const TargetRegisterInfo *RI = TM.getRegisterInfo(); 2829 for (TargetRegisterInfo::regclass_iterator RCI = RI->regclass_begin(), 2830 E = RI->regclass_end(); RCI != E; ++RCI) { 2831 const TargetRegisterClass *RC = *RCI; 2832 2833 // If none of the value types for this register class are valid, we 2834 // can't use it. For example, 64-bit reg classes on 32-bit targets. 2835 if (!isLegalRC(RC)) 2836 continue; 2837 2838 for (TargetRegisterClass::iterator I = RC->begin(), E = RC->end(); 2839 I != E; ++I) { 2840 if (RegName.equals_lower(RI->getName(*I))) 2841 return std::make_pair(*I, RC); 2842 } 2843 } 2844 2845 return std::make_pair(0u, static_cast<const TargetRegisterClass*>(0)); 2846} 2847 2848//===----------------------------------------------------------------------===// 2849// Constraint Selection. 2850 2851/// isMatchingInputConstraint - Return true of this is an input operand that is 2852/// a matching constraint like "4". 2853bool TargetLowering::AsmOperandInfo::isMatchingInputConstraint() const { 2854 assert(!ConstraintCode.empty() && "No known constraint!"); 2855 return isdigit(ConstraintCode[0]); 2856} 2857 2858/// getMatchedOperand - If this is an input matching constraint, this method 2859/// returns the output operand it matches. 2860unsigned TargetLowering::AsmOperandInfo::getMatchedOperand() const { 2861 assert(!ConstraintCode.empty() && "No known constraint!"); 2862 return atoi(ConstraintCode.c_str()); 2863} 2864 2865 2866/// ParseConstraints - Split up the constraint string from the inline 2867/// assembly value into the specific constraints and their prefixes, 2868/// and also tie in the associated operand values. 2869/// If this returns an empty vector, and if the constraint string itself 2870/// isn't empty, there was an error parsing. 2871TargetLowering::AsmOperandInfoVector TargetLowering::ParseConstraints( 2872 ImmutableCallSite CS) const { 2873 /// ConstraintOperands - Information about all of the constraints. 2874 AsmOperandInfoVector ConstraintOperands; 2875 const InlineAsm *IA = cast<InlineAsm>(CS.getCalledValue()); 2876 unsigned maCount = 0; // Largest number of multiple alternative constraints. 2877 2878 // Do a prepass over the constraints, canonicalizing them, and building up the 2879 // ConstraintOperands list. 2880 InlineAsm::ConstraintInfoVector 2881 ConstraintInfos = IA->ParseConstraints(); 2882 2883 unsigned ArgNo = 0; // ArgNo - The argument of the CallInst. 2884 unsigned ResNo = 0; // ResNo - The result number of the next output. 2885 2886 for (unsigned i = 0, e = ConstraintInfos.size(); i != e; ++i) { 2887 ConstraintOperands.push_back(AsmOperandInfo(ConstraintInfos[i])); 2888 AsmOperandInfo &OpInfo = ConstraintOperands.back(); 2889 2890 // Update multiple alternative constraint count. 2891 if (OpInfo.multipleAlternatives.size() > maCount) 2892 maCount = OpInfo.multipleAlternatives.size(); 2893 2894 OpInfo.ConstraintVT = MVT::Other; 2895 2896 // Compute the value type for each operand. 2897 switch (OpInfo.Type) { 2898 case InlineAsm::isOutput: 2899 // Indirect outputs just consume an argument. 2900 if (OpInfo.isIndirect) { 2901 OpInfo.CallOperandVal = const_cast<Value *>(CS.getArgument(ArgNo++)); 2902 break; 2903 } 2904 2905 // The return value of the call is this value. As such, there is no 2906 // corresponding argument. 2907 assert(!CS.getType()->isVoidTy() && 2908 "Bad inline asm!"); 2909 if (StructType *STy = dyn_cast<StructType>(CS.getType())) { 2910 OpInfo.ConstraintVT = getValueType(STy->getElementType(ResNo)); 2911 } else { 2912 assert(ResNo == 0 && "Asm only has one result!"); 2913 OpInfo.ConstraintVT = getValueType(CS.getType()); 2914 } 2915 ++ResNo; 2916 break; 2917 case InlineAsm::isInput: 2918 OpInfo.CallOperandVal = const_cast<Value *>(CS.getArgument(ArgNo++)); 2919 break; 2920 case InlineAsm::isClobber: 2921 // Nothing to do. 2922 break; 2923 } 2924 2925 if (OpInfo.CallOperandVal) { 2926 llvm::Type *OpTy = OpInfo.CallOperandVal->getType(); 2927 if (OpInfo.isIndirect) { 2928 llvm::PointerType *PtrTy = dyn_cast<PointerType>(OpTy); 2929 if (!PtrTy) 2930 report_fatal_error("Indirect operand for inline asm not a pointer!"); 2931 OpTy = PtrTy->getElementType(); 2932 } 2933 2934 // Look for vector wrapped in a struct. e.g. { <16 x i8> }. 2935 if (StructType *STy = dyn_cast<StructType>(OpTy)) 2936 if (STy->getNumElements() == 1) 2937 OpTy = STy->getElementType(0); 2938 2939 // If OpTy is not a single value, it may be a struct/union that we 2940 // can tile with integers. 2941 if (!OpTy->isSingleValueType() && OpTy->isSized()) { 2942 unsigned BitSize = TD->getTypeSizeInBits(OpTy); 2943 switch (BitSize) { 2944 default: break; 2945 case 1: 2946 case 8: 2947 case 16: 2948 case 32: 2949 case 64: 2950 case 128: 2951 OpInfo.ConstraintVT = 2952 EVT::getEVT(IntegerType::get(OpTy->getContext(), BitSize), true); 2953 break; 2954 } 2955 } else if (dyn_cast<PointerType>(OpTy)) { 2956 OpInfo.ConstraintVT = MVT::getIntegerVT(8*TD->getPointerSize()); 2957 } else { 2958 OpInfo.ConstraintVT = EVT::getEVT(OpTy, true); 2959 } 2960 } 2961 } 2962 2963 // If we have multiple alternative constraints, select the best alternative. 2964 if (ConstraintInfos.size()) { 2965 if (maCount) { 2966 unsigned bestMAIndex = 0; 2967 int bestWeight = -1; 2968 // weight: -1 = invalid match, and 0 = so-so match to 5 = good match. 2969 int weight = -1; 2970 unsigned maIndex; 2971 // Compute the sums of the weights for each alternative, keeping track 2972 // of the best (highest weight) one so far. 2973 for (maIndex = 0; maIndex < maCount; ++maIndex) { 2974 int weightSum = 0; 2975 for (unsigned cIndex = 0, eIndex = ConstraintOperands.size(); 2976 cIndex != eIndex; ++cIndex) { 2977 AsmOperandInfo& OpInfo = ConstraintOperands[cIndex]; 2978 if (OpInfo.Type == InlineAsm::isClobber) 2979 continue; 2980 2981 // If this is an output operand with a matching input operand, 2982 // look up the matching input. If their types mismatch, e.g. one 2983 // is an integer, the other is floating point, or their sizes are 2984 // different, flag it as an maCantMatch. 2985 if (OpInfo.hasMatchingInput()) { 2986 AsmOperandInfo &Input = ConstraintOperands[OpInfo.MatchingInput]; 2987 if (OpInfo.ConstraintVT != Input.ConstraintVT) { 2988 if ((OpInfo.ConstraintVT.isInteger() != 2989 Input.ConstraintVT.isInteger()) || 2990 (OpInfo.ConstraintVT.getSizeInBits() != 2991 Input.ConstraintVT.getSizeInBits())) { 2992 weightSum = -1; // Can't match. 2993 break; 2994 } 2995 } 2996 } 2997 weight = getMultipleConstraintMatchWeight(OpInfo, maIndex); 2998 if (weight == -1) { 2999 weightSum = -1; 3000 break; 3001 } 3002 weightSum += weight; 3003 } 3004 // Update best. 3005 if (weightSum > bestWeight) { 3006 bestWeight = weightSum; 3007 bestMAIndex = maIndex; 3008 } 3009 } 3010 3011 // Now select chosen alternative in each constraint. 3012 for (unsigned cIndex = 0, eIndex = ConstraintOperands.size(); 3013 cIndex != eIndex; ++cIndex) { 3014 AsmOperandInfo& cInfo = ConstraintOperands[cIndex]; 3015 if (cInfo.Type == InlineAsm::isClobber) 3016 continue; 3017 cInfo.selectAlternative(bestMAIndex); 3018 } 3019 } 3020 } 3021 3022 // Check and hook up tied operands, choose constraint code to use. 3023 for (unsigned cIndex = 0, eIndex = ConstraintOperands.size(); 3024 cIndex != eIndex; ++cIndex) { 3025 AsmOperandInfo& OpInfo = ConstraintOperands[cIndex]; 3026 3027 // If this is an output operand with a matching input operand, look up the 3028 // matching input. If their types mismatch, e.g. one is an integer, the 3029 // other is floating point, or their sizes are different, flag it as an 3030 // error. 3031 if (OpInfo.hasMatchingInput()) { 3032 AsmOperandInfo &Input = ConstraintOperands[OpInfo.MatchingInput]; 3033 3034 if (OpInfo.ConstraintVT != Input.ConstraintVT) { 3035 std::pair<unsigned, const TargetRegisterClass*> MatchRC = 3036 getRegForInlineAsmConstraint(OpInfo.ConstraintCode, 3037 OpInfo.ConstraintVT); 3038 std::pair<unsigned, const TargetRegisterClass*> InputRC = 3039 getRegForInlineAsmConstraint(Input.ConstraintCode, 3040 Input.ConstraintVT); 3041 if ((OpInfo.ConstraintVT.isInteger() != 3042 Input.ConstraintVT.isInteger()) || 3043 (MatchRC.second != InputRC.second)) { 3044 report_fatal_error("Unsupported asm: input constraint" 3045 " with a matching output constraint of" 3046 " incompatible type!"); 3047 } 3048 } 3049 3050 } 3051 } 3052 3053 return ConstraintOperands; 3054} 3055 3056 3057/// getConstraintGenerality - Return an integer indicating how general CT 3058/// is. 3059static unsigned getConstraintGenerality(TargetLowering::ConstraintType CT) { 3060 switch (CT) { 3061 case TargetLowering::C_Other: 3062 case TargetLowering::C_Unknown: 3063 return 0; 3064 case TargetLowering::C_Register: 3065 return 1; 3066 case TargetLowering::C_RegisterClass: 3067 return 2; 3068 case TargetLowering::C_Memory: 3069 return 3; 3070 } 3071 llvm_unreachable("Invalid constraint type"); 3072} 3073 3074/// Examine constraint type and operand type and determine a weight value. 3075/// This object must already have been set up with the operand type 3076/// and the current alternative constraint selected. 3077TargetLowering::ConstraintWeight 3078 TargetLowering::getMultipleConstraintMatchWeight( 3079 AsmOperandInfo &info, int maIndex) const { 3080 InlineAsm::ConstraintCodeVector *rCodes; 3081 if (maIndex >= (int)info.multipleAlternatives.size()) 3082 rCodes = &info.Codes; 3083 else 3084 rCodes = &info.multipleAlternatives[maIndex].Codes; 3085 ConstraintWeight BestWeight = CW_Invalid; 3086 3087 // Loop over the options, keeping track of the most general one. 3088 for (unsigned i = 0, e = rCodes->size(); i != e; ++i) { 3089 ConstraintWeight weight = 3090 getSingleConstraintMatchWeight(info, (*rCodes)[i].c_str()); 3091 if (weight > BestWeight) 3092 BestWeight = weight; 3093 } 3094 3095 return BestWeight; 3096} 3097 3098/// Examine constraint type and operand type and determine a weight value. 3099/// This object must already have been set up with the operand type 3100/// and the current alternative constraint selected. 3101TargetLowering::ConstraintWeight 3102 TargetLowering::getSingleConstraintMatchWeight( 3103 AsmOperandInfo &info, const char *constraint) const { 3104 ConstraintWeight weight = CW_Invalid; 3105 Value *CallOperandVal = info.CallOperandVal; 3106 // If we don't have a value, we can't do a match, 3107 // but allow it at the lowest weight. 3108 if (CallOperandVal == NULL) 3109 return CW_Default; 3110 // Look at the constraint type. 3111 switch (*constraint) { 3112 case 'i': // immediate integer. 3113 case 'n': // immediate integer with a known value. 3114 if (isa<ConstantInt>(CallOperandVal)) 3115 weight = CW_Constant; 3116 break; 3117 case 's': // non-explicit intregal immediate. 3118 if (isa<GlobalValue>(CallOperandVal)) 3119 weight = CW_Constant; 3120 break; 3121 case 'E': // immediate float if host format. 3122 case 'F': // immediate float. 3123 if (isa<ConstantFP>(CallOperandVal)) 3124 weight = CW_Constant; 3125 break; 3126 case '<': // memory operand with autodecrement. 3127 case '>': // memory operand with autoincrement. 3128 case 'm': // memory operand. 3129 case 'o': // offsettable memory operand 3130 case 'V': // non-offsettable memory operand 3131 weight = CW_Memory; 3132 break; 3133 case 'r': // general register. 3134 case 'g': // general register, memory operand or immediate integer. 3135 // note: Clang converts "g" to "imr". 3136 if (CallOperandVal->getType()->isIntegerTy()) 3137 weight = CW_Register; 3138 break; 3139 case 'X': // any operand. 3140 default: 3141 weight = CW_Default; 3142 break; 3143 } 3144 return weight; 3145} 3146 3147/// ChooseConstraint - If there are multiple different constraints that we 3148/// could pick for this operand (e.g. "imr") try to pick the 'best' one. 3149/// This is somewhat tricky: constraints fall into four classes: 3150/// Other -> immediates and magic values 3151/// Register -> one specific register 3152/// RegisterClass -> a group of regs 3153/// Memory -> memory 3154/// Ideally, we would pick the most specific constraint possible: if we have 3155/// something that fits into a register, we would pick it. The problem here 3156/// is that if we have something that could either be in a register or in 3157/// memory that use of the register could cause selection of *other* 3158/// operands to fail: they might only succeed if we pick memory. Because of 3159/// this the heuristic we use is: 3160/// 3161/// 1) If there is an 'other' constraint, and if the operand is valid for 3162/// that constraint, use it. This makes us take advantage of 'i' 3163/// constraints when available. 3164/// 2) Otherwise, pick the most general constraint present. This prefers 3165/// 'm' over 'r', for example. 3166/// 3167static void ChooseConstraint(TargetLowering::AsmOperandInfo &OpInfo, 3168 const TargetLowering &TLI, 3169 SDValue Op, SelectionDAG *DAG) { 3170 assert(OpInfo.Codes.size() > 1 && "Doesn't have multiple constraint options"); 3171 unsigned BestIdx = 0; 3172 TargetLowering::ConstraintType BestType = TargetLowering::C_Unknown; 3173 int BestGenerality = -1; 3174 3175 // Loop over the options, keeping track of the most general one. 3176 for (unsigned i = 0, e = OpInfo.Codes.size(); i != e; ++i) { 3177 TargetLowering::ConstraintType CType = 3178 TLI.getConstraintType(OpInfo.Codes[i]); 3179 3180 // If this is an 'other' constraint, see if the operand is valid for it. 3181 // For example, on X86 we might have an 'rI' constraint. If the operand 3182 // is an integer in the range [0..31] we want to use I (saving a load 3183 // of a register), otherwise we must use 'r'. 3184 if (CType == TargetLowering::C_Other && Op.getNode()) { 3185 assert(OpInfo.Codes[i].size() == 1 && 3186 "Unhandled multi-letter 'other' constraint"); 3187 std::vector<SDValue> ResultOps; 3188 TLI.LowerAsmOperandForConstraint(Op, OpInfo.Codes[i], 3189 ResultOps, *DAG); 3190 if (!ResultOps.empty()) { 3191 BestType = CType; 3192 BestIdx = i; 3193 break; 3194 } 3195 } 3196 3197 // Things with matching constraints can only be registers, per gcc 3198 // documentation. This mainly affects "g" constraints. 3199 if (CType == TargetLowering::C_Memory && OpInfo.hasMatchingInput()) 3200 continue; 3201 3202 // This constraint letter is more general than the previous one, use it. 3203 int Generality = getConstraintGenerality(CType); 3204 if (Generality > BestGenerality) { 3205 BestType = CType; 3206 BestIdx = i; 3207 BestGenerality = Generality; 3208 } 3209 } 3210 3211 OpInfo.ConstraintCode = OpInfo.Codes[BestIdx]; 3212 OpInfo.ConstraintType = BestType; 3213} 3214 3215/// ComputeConstraintToUse - Determines the constraint code and constraint 3216/// type to use for the specific AsmOperandInfo, setting 3217/// OpInfo.ConstraintCode and OpInfo.ConstraintType. 3218void TargetLowering::ComputeConstraintToUse(AsmOperandInfo &OpInfo, 3219 SDValue Op, 3220 SelectionDAG *DAG) const { 3221 assert(!OpInfo.Codes.empty() && "Must have at least one constraint"); 3222 3223 // Single-letter constraints ('r') are very common. 3224 if (OpInfo.Codes.size() == 1) { 3225 OpInfo.ConstraintCode = OpInfo.Codes[0]; 3226 OpInfo.ConstraintType = getConstraintType(OpInfo.ConstraintCode); 3227 } else { 3228 ChooseConstraint(OpInfo, *this, Op, DAG); 3229 } 3230 3231 // 'X' matches anything. 3232 if (OpInfo.ConstraintCode == "X" && OpInfo.CallOperandVal) { 3233 // Labels and constants are handled elsewhere ('X' is the only thing 3234 // that matches labels). For Functions, the type here is the type of 3235 // the result, which is not what we want to look at; leave them alone. 3236 Value *v = OpInfo.CallOperandVal; 3237 if (isa<BasicBlock>(v) || isa<ConstantInt>(v) || isa<Function>(v)) { 3238 OpInfo.CallOperandVal = v; 3239 return; 3240 } 3241 3242 // Otherwise, try to resolve it to something we know about by looking at 3243 // the actual operand type. 3244 if (const char *Repl = LowerXConstraint(OpInfo.ConstraintVT)) { 3245 OpInfo.ConstraintCode = Repl; 3246 OpInfo.ConstraintType = getConstraintType(OpInfo.ConstraintCode); 3247 } 3248 } 3249} 3250 3251//===----------------------------------------------------------------------===// 3252// Loop Strength Reduction hooks 3253//===----------------------------------------------------------------------===// 3254 3255/// isLegalAddressingMode - Return true if the addressing mode represented 3256/// by AM is legal for this target, for a load/store of the specified type. 3257bool TargetLowering::isLegalAddressingMode(const AddrMode &AM, 3258 Type *Ty) const { 3259 // The default implementation of this implements a conservative RISCy, r+r and 3260 // r+i addr mode. 3261 3262 // Allows a sign-extended 16-bit immediate field. 3263 if (AM.BaseOffs <= -(1LL << 16) || AM.BaseOffs >= (1LL << 16)-1) 3264 return false; 3265 3266 // No global is ever allowed as a base. 3267 if (AM.BaseGV) 3268 return false; 3269 3270 // Only support r+r, 3271 switch (AM.Scale) { 3272 case 0: // "r+i" or just "i", depending on HasBaseReg. 3273 break; 3274 case 1: 3275 if (AM.HasBaseReg && AM.BaseOffs) // "r+r+i" is not allowed. 3276 return false; 3277 // Otherwise we have r+r or r+i. 3278 break; 3279 case 2: 3280 if (AM.HasBaseReg || AM.BaseOffs) // 2*r+r or 2*r+i is not allowed. 3281 return false; 3282 // Allow 2*r as r+r. 3283 break; 3284 } 3285 3286 return true; 3287} 3288 3289/// BuildExactDiv - Given an exact SDIV by a constant, create a multiplication 3290/// with the multiplicative inverse of the constant. 3291SDValue TargetLowering::BuildExactSDIV(SDValue Op1, SDValue Op2, DebugLoc dl, 3292 SelectionDAG &DAG) const { 3293 ConstantSDNode *C = cast<ConstantSDNode>(Op2); 3294 APInt d = C->getAPIntValue(); 3295 assert(d != 0 && "Division by zero!"); 3296 3297 // Shift the value upfront if it is even, so the LSB is one. 3298 unsigned ShAmt = d.countTrailingZeros(); 3299 if (ShAmt) { 3300 // TODO: For UDIV use SRL instead of SRA. 3301 SDValue Amt = DAG.getConstant(ShAmt, getShiftAmountTy(Op1.getValueType())); 3302 Op1 = DAG.getNode(ISD::SRA, dl, Op1.getValueType(), Op1, Amt); 3303 d = d.ashr(ShAmt); 3304 } 3305 3306 // Calculate the multiplicative inverse, using Newton's method. 3307 APInt t, xn = d; 3308 while ((t = d*xn) != 1) 3309 xn *= APInt(d.getBitWidth(), 2) - t; 3310 3311 Op2 = DAG.getConstant(xn, Op1.getValueType()); 3312 return DAG.getNode(ISD::MUL, dl, Op1.getValueType(), Op1, Op2); 3313} 3314 3315/// BuildSDIVSequence - Given an ISD::SDIV node expressing a divide by constant, 3316/// return a DAG expression to select that will generate the same value by 3317/// multiplying by a magic number. See: 3318/// <http://the.wall.riscom.net/books/proc/ppc/cwg/code2.html> 3319SDValue TargetLowering:: 3320BuildSDIV(SDNode *N, SelectionDAG &DAG, bool IsAfterLegalization, 3321 std::vector<SDNode*>* Created) const { 3322 EVT VT = N->getValueType(0); 3323 DebugLoc dl= N->getDebugLoc(); 3324 3325 // Check to see if we can do this. 3326 // FIXME: We should be more aggressive here. 3327 if (!isTypeLegal(VT)) 3328 return SDValue(); 3329 3330 APInt d = cast<ConstantSDNode>(N->getOperand(1))->getAPIntValue(); 3331 APInt::ms magics = d.magic(); 3332 3333 // Multiply the numerator (operand 0) by the magic value 3334 // FIXME: We should support doing a MUL in a wider type 3335 SDValue Q; 3336 if (IsAfterLegalization ? isOperationLegal(ISD::MULHS, VT) : 3337 isOperationLegalOrCustom(ISD::MULHS, VT)) 3338 Q = DAG.getNode(ISD::MULHS, dl, VT, N->getOperand(0), 3339 DAG.getConstant(magics.m, VT)); 3340 else if (IsAfterLegalization ? isOperationLegal(ISD::SMUL_LOHI, VT) : 3341 isOperationLegalOrCustom(ISD::SMUL_LOHI, VT)) 3342 Q = SDValue(DAG.getNode(ISD::SMUL_LOHI, dl, DAG.getVTList(VT, VT), 3343 N->getOperand(0), 3344 DAG.getConstant(magics.m, VT)).getNode(), 1); 3345 else 3346 return SDValue(); // No mulhs or equvialent 3347 // If d > 0 and m < 0, add the numerator 3348 if (d.isStrictlyPositive() && magics.m.isNegative()) { 3349 Q = DAG.getNode(ISD::ADD, dl, VT, Q, N->getOperand(0)); 3350 if (Created) 3351 Created->push_back(Q.getNode()); 3352 } 3353 // If d < 0 and m > 0, subtract the numerator. 3354 if (d.isNegative() && magics.m.isStrictlyPositive()) { 3355 Q = DAG.getNode(ISD::SUB, dl, VT, Q, N->getOperand(0)); 3356 if (Created) 3357 Created->push_back(Q.getNode()); 3358 } 3359 // Shift right algebraic if shift value is nonzero 3360 if (magics.s > 0) { 3361 Q = DAG.getNode(ISD::SRA, dl, VT, Q, 3362 DAG.getConstant(magics.s, getShiftAmountTy(Q.getValueType()))); 3363 if (Created) 3364 Created->push_back(Q.getNode()); 3365 } 3366 // Extract the sign bit and add it to the quotient 3367 SDValue T = 3368 DAG.getNode(ISD::SRL, dl, VT, Q, DAG.getConstant(VT.getSizeInBits()-1, 3369 getShiftAmountTy(Q.getValueType()))); 3370 if (Created) 3371 Created->push_back(T.getNode()); 3372 return DAG.getNode(ISD::ADD, dl, VT, Q, T); 3373} 3374 3375/// BuildUDIVSequence - Given an ISD::UDIV node expressing a divide by constant, 3376/// return a DAG expression to select that will generate the same value by 3377/// multiplying by a magic number. See: 3378/// <http://the.wall.riscom.net/books/proc/ppc/cwg/code2.html> 3379SDValue TargetLowering:: 3380BuildUDIV(SDNode *N, SelectionDAG &DAG, bool IsAfterLegalization, 3381 std::vector<SDNode*>* Created) const { 3382 EVT VT = N->getValueType(0); 3383 DebugLoc dl = N->getDebugLoc(); 3384 3385 // Check to see if we can do this. 3386 // FIXME: We should be more aggressive here. 3387 if (!isTypeLegal(VT)) 3388 return SDValue(); 3389 3390 // FIXME: We should use a narrower constant when the upper 3391 // bits are known to be zero. 3392 const APInt &N1C = cast<ConstantSDNode>(N->getOperand(1))->getAPIntValue(); 3393 APInt::mu magics = N1C.magicu(); 3394 3395 SDValue Q = N->getOperand(0); 3396 3397 // If the divisor is even, we can avoid using the expensive fixup by shifting 3398 // the divided value upfront. 3399 if (magics.a != 0 && !N1C[0]) { 3400 unsigned Shift = N1C.countTrailingZeros(); 3401 Q = DAG.getNode(ISD::SRL, dl, VT, Q, 3402 DAG.getConstant(Shift, getShiftAmountTy(Q.getValueType()))); 3403 if (Created) 3404 Created->push_back(Q.getNode()); 3405 3406 // Get magic number for the shifted divisor. 3407 magics = N1C.lshr(Shift).magicu(Shift); 3408 assert(magics.a == 0 && "Should use cheap fixup now"); 3409 } 3410 3411 // Multiply the numerator (operand 0) by the magic value 3412 // FIXME: We should support doing a MUL in a wider type 3413 if (IsAfterLegalization ? isOperationLegal(ISD::MULHU, VT) : 3414 isOperationLegalOrCustom(ISD::MULHU, VT)) 3415 Q = DAG.getNode(ISD::MULHU, dl, VT, Q, DAG.getConstant(magics.m, VT)); 3416 else if (IsAfterLegalization ? isOperationLegal(ISD::UMUL_LOHI, VT) : 3417 isOperationLegalOrCustom(ISD::UMUL_LOHI, VT)) 3418 Q = SDValue(DAG.getNode(ISD::UMUL_LOHI, dl, DAG.getVTList(VT, VT), Q, 3419 DAG.getConstant(magics.m, VT)).getNode(), 1); 3420 else 3421 return SDValue(); // No mulhu or equvialent 3422 if (Created) 3423 Created->push_back(Q.getNode()); 3424 3425 if (magics.a == 0) { 3426 assert(magics.s < N1C.getBitWidth() && 3427 "We shouldn't generate an undefined shift!"); 3428 return DAG.getNode(ISD::SRL, dl, VT, Q, 3429 DAG.getConstant(magics.s, getShiftAmountTy(Q.getValueType()))); 3430 } else { 3431 SDValue NPQ = DAG.getNode(ISD::SUB, dl, VT, N->getOperand(0), Q); 3432 if (Created) 3433 Created->push_back(NPQ.getNode()); 3434 NPQ = DAG.getNode(ISD::SRL, dl, VT, NPQ, 3435 DAG.getConstant(1, getShiftAmountTy(NPQ.getValueType()))); 3436 if (Created) 3437 Created->push_back(NPQ.getNode()); 3438 NPQ = DAG.getNode(ISD::ADD, dl, VT, NPQ, Q); 3439 if (Created) 3440 Created->push_back(NPQ.getNode()); 3441 return DAG.getNode(ISD::SRL, dl, VT, NPQ, 3442 DAG.getConstant(magics.s-1, getShiftAmountTy(NPQ.getValueType()))); 3443 } 3444} 3445