1 //===-- TargetLoweringBase.cpp - Implement the TargetLoweringBase 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 TargetLoweringBase class. 11 // 12 //===----------------------------------------------------------------------===// 13 14 #include "llvm/Target/TargetLowering.h" 15 #include "llvm/ADT/BitVector.h" 16 #include "llvm/ADT/STLExtras.h" 17 #include "llvm/ADT/Triple.h" 18 #include "llvm/CodeGen/Analysis.h" 19 #include "llvm/CodeGen/MachineFrameInfo.h" 20 #include "llvm/CodeGen/MachineFunction.h" 21 #include "llvm/CodeGen/MachineInstrBuilder.h" 22 #include "llvm/CodeGen/MachineJumpTableInfo.h" 23 #include "llvm/CodeGen/StackMaps.h" 24 #include "llvm/IR/DataLayout.h" 25 #include "llvm/IR/DerivedTypes.h" 26 #include "llvm/IR/GlobalVariable.h" 27 #include "llvm/IR/Mangler.h" 28 #include "llvm/MC/MCAsmInfo.h" 29 #include "llvm/MC/MCContext.h" 30 #include "llvm/MC/MCExpr.h" 31 #include "llvm/Support/BranchProbability.h" 32 #include "llvm/Support/CommandLine.h" 33 #include "llvm/Support/ErrorHandling.h" 34 #include "llvm/Support/MathExtras.h" 35 #include "llvm/Target/TargetLoweringObjectFile.h" 36 #include "llvm/Target/TargetMachine.h" 37 #include "llvm/Target/TargetRegisterInfo.h" 38 #include "llvm/Target/TargetSubtargetInfo.h" 39 #include <cctype> 40 using namespace llvm; 41 42 static cl::opt<bool> JumpIsExpensiveOverride( 43 "jump-is-expensive", cl::init(false), 44 cl::desc("Do not create extra branches to split comparison logic."), 45 cl::Hidden); 46 47 // Although this default value is arbitrary, it is not random. It is assumed 48 // that a condition that evaluates the same way by a higher percentage than this 49 // is best represented as control flow. Therefore, the default value N should be 50 // set such that the win from N% correct executions is greater than the loss 51 // from (100 - N)% mispredicted executions for the majority of intended targets. 52 static cl::opt<int> MinPercentageForPredictableBranch( 53 "min-predictable-branch", cl::init(99), 54 cl::desc("Minimum percentage (0-100) that a condition must be either true " 55 "or false to assume that the condition is predictable"), 56 cl::Hidden); 57 58 /// InitLibcallNames - Set default libcall names. 59 /// 60 static void InitLibcallNames(const char **Names, const Triple &TT) { 61 Names[RTLIB::SHL_I16] = "__ashlhi3"; 62 Names[RTLIB::SHL_I32] = "__ashlsi3"; 63 Names[RTLIB::SHL_I64] = "__ashldi3"; 64 Names[RTLIB::SHL_I128] = "__ashlti3"; 65 Names[RTLIB::SRL_I16] = "__lshrhi3"; 66 Names[RTLIB::SRL_I32] = "__lshrsi3"; 67 Names[RTLIB::SRL_I64] = "__lshrdi3"; 68 Names[RTLIB::SRL_I128] = "__lshrti3"; 69 Names[RTLIB::SRA_I16] = "__ashrhi3"; 70 Names[RTLIB::SRA_I32] = "__ashrsi3"; 71 Names[RTLIB::SRA_I64] = "__ashrdi3"; 72 Names[RTLIB::SRA_I128] = "__ashrti3"; 73 Names[RTLIB::MUL_I8] = "__mulqi3"; 74 Names[RTLIB::MUL_I16] = "__mulhi3"; 75 Names[RTLIB::MUL_I32] = "__mulsi3"; 76 Names[RTLIB::MUL_I64] = "__muldi3"; 77 Names[RTLIB::MUL_I128] = "__multi3"; 78 Names[RTLIB::MULO_I32] = "__mulosi4"; 79 Names[RTLIB::MULO_I64] = "__mulodi4"; 80 Names[RTLIB::MULO_I128] = "__muloti4"; 81 Names[RTLIB::SDIV_I8] = "__divqi3"; 82 Names[RTLIB::SDIV_I16] = "__divhi3"; 83 Names[RTLIB::SDIV_I32] = "__divsi3"; 84 Names[RTLIB::SDIV_I64] = "__divdi3"; 85 Names[RTLIB::SDIV_I128] = "__divti3"; 86 Names[RTLIB::UDIV_I8] = "__udivqi3"; 87 Names[RTLIB::UDIV_I16] = "__udivhi3"; 88 Names[RTLIB::UDIV_I32] = "__udivsi3"; 89 Names[RTLIB::UDIV_I64] = "__udivdi3"; 90 Names[RTLIB::UDIV_I128] = "__udivti3"; 91 Names[RTLIB::SREM_I8] = "__modqi3"; 92 Names[RTLIB::SREM_I16] = "__modhi3"; 93 Names[RTLIB::SREM_I32] = "__modsi3"; 94 Names[RTLIB::SREM_I64] = "__moddi3"; 95 Names[RTLIB::SREM_I128] = "__modti3"; 96 Names[RTLIB::UREM_I8] = "__umodqi3"; 97 Names[RTLIB::UREM_I16] = "__umodhi3"; 98 Names[RTLIB::UREM_I32] = "__umodsi3"; 99 Names[RTLIB::UREM_I64] = "__umoddi3"; 100 Names[RTLIB::UREM_I128] = "__umodti3"; 101 102 Names[RTLIB::NEG_I32] = "__negsi2"; 103 Names[RTLIB::NEG_I64] = "__negdi2"; 104 Names[RTLIB::ADD_F32] = "__addsf3"; 105 Names[RTLIB::ADD_F64] = "__adddf3"; 106 Names[RTLIB::ADD_F80] = "__addxf3"; 107 Names[RTLIB::ADD_F128] = "__addtf3"; 108 Names[RTLIB::ADD_PPCF128] = "__gcc_qadd"; 109 Names[RTLIB::SUB_F32] = "__subsf3"; 110 Names[RTLIB::SUB_F64] = "__subdf3"; 111 Names[RTLIB::SUB_F80] = "__subxf3"; 112 Names[RTLIB::SUB_F128] = "__subtf3"; 113 Names[RTLIB::SUB_PPCF128] = "__gcc_qsub"; 114 Names[RTLIB::MUL_F32] = "__mulsf3"; 115 Names[RTLIB::MUL_F64] = "__muldf3"; 116 Names[RTLIB::MUL_F80] = "__mulxf3"; 117 Names[RTLIB::MUL_F128] = "__multf3"; 118 Names[RTLIB::MUL_PPCF128] = "__gcc_qmul"; 119 Names[RTLIB::DIV_F32] = "__divsf3"; 120 Names[RTLIB::DIV_F64] = "__divdf3"; 121 Names[RTLIB::DIV_F80] = "__divxf3"; 122 Names[RTLIB::DIV_F128] = "__divtf3"; 123 Names[RTLIB::DIV_PPCF128] = "__gcc_qdiv"; 124 Names[RTLIB::REM_F32] = "fmodf"; 125 Names[RTLIB::REM_F64] = "fmod"; 126 Names[RTLIB::REM_F80] = "fmodl"; 127 Names[RTLIB::REM_F128] = "fmodl"; 128 Names[RTLIB::REM_PPCF128] = "fmodl"; 129 Names[RTLIB::FMA_F32] = "fmaf"; 130 Names[RTLIB::FMA_F64] = "fma"; 131 Names[RTLIB::FMA_F80] = "fmal"; 132 Names[RTLIB::FMA_F128] = "fmal"; 133 Names[RTLIB::FMA_PPCF128] = "fmal"; 134 Names[RTLIB::POWI_F32] = "__powisf2"; 135 Names[RTLIB::POWI_F64] = "__powidf2"; 136 Names[RTLIB::POWI_F80] = "__powixf2"; 137 Names[RTLIB::POWI_F128] = "__powitf2"; 138 Names[RTLIB::POWI_PPCF128] = "__powitf2"; 139 Names[RTLIB::SQRT_F32] = "sqrtf"; 140 Names[RTLIB::SQRT_F64] = "sqrt"; 141 Names[RTLIB::SQRT_F80] = "sqrtl"; 142 Names[RTLIB::SQRT_F128] = "sqrtl"; 143 Names[RTLIB::SQRT_PPCF128] = "sqrtl"; 144 Names[RTLIB::LOG_F32] = "logf"; 145 Names[RTLIB::LOG_F64] = "log"; 146 Names[RTLIB::LOG_F80] = "logl"; 147 Names[RTLIB::LOG_F128] = "logl"; 148 Names[RTLIB::LOG_PPCF128] = "logl"; 149 Names[RTLIB::LOG2_F32] = "log2f"; 150 Names[RTLIB::LOG2_F64] = "log2"; 151 Names[RTLIB::LOG2_F80] = "log2l"; 152 Names[RTLIB::LOG2_F128] = "log2l"; 153 Names[RTLIB::LOG2_PPCF128] = "log2l"; 154 Names[RTLIB::LOG10_F32] = "log10f"; 155 Names[RTLIB::LOG10_F64] = "log10"; 156 Names[RTLIB::LOG10_F80] = "log10l"; 157 Names[RTLIB::LOG10_F128] = "log10l"; 158 Names[RTLIB::LOG10_PPCF128] = "log10l"; 159 Names[RTLIB::EXP_F32] = "expf"; 160 Names[RTLIB::EXP_F64] = "exp"; 161 Names[RTLIB::EXP_F80] = "expl"; 162 Names[RTLIB::EXP_F128] = "expl"; 163 Names[RTLIB::EXP_PPCF128] = "expl"; 164 Names[RTLIB::EXP2_F32] = "exp2f"; 165 Names[RTLIB::EXP2_F64] = "exp2"; 166 Names[RTLIB::EXP2_F80] = "exp2l"; 167 Names[RTLIB::EXP2_F128] = "exp2l"; 168 Names[RTLIB::EXP2_PPCF128] = "exp2l"; 169 Names[RTLIB::SIN_F32] = "sinf"; 170 Names[RTLIB::SIN_F64] = "sin"; 171 Names[RTLIB::SIN_F80] = "sinl"; 172 Names[RTLIB::SIN_F128] = "sinl"; 173 Names[RTLIB::SIN_PPCF128] = "sinl"; 174 Names[RTLIB::COS_F32] = "cosf"; 175 Names[RTLIB::COS_F64] = "cos"; 176 Names[RTLIB::COS_F80] = "cosl"; 177 Names[RTLIB::COS_F128] = "cosl"; 178 Names[RTLIB::COS_PPCF128] = "cosl"; 179 Names[RTLIB::POW_F32] = "powf"; 180 Names[RTLIB::POW_F64] = "pow"; 181 Names[RTLIB::POW_F80] = "powl"; 182 Names[RTLIB::POW_F128] = "powl"; 183 Names[RTLIB::POW_PPCF128] = "powl"; 184 Names[RTLIB::CEIL_F32] = "ceilf"; 185 Names[RTLIB::CEIL_F64] = "ceil"; 186 Names[RTLIB::CEIL_F80] = "ceill"; 187 Names[RTLIB::CEIL_F128] = "ceill"; 188 Names[RTLIB::CEIL_PPCF128] = "ceill"; 189 Names[RTLIB::TRUNC_F32] = "truncf"; 190 Names[RTLIB::TRUNC_F64] = "trunc"; 191 Names[RTLIB::TRUNC_F80] = "truncl"; 192 Names[RTLIB::TRUNC_F128] = "truncl"; 193 Names[RTLIB::TRUNC_PPCF128] = "truncl"; 194 Names[RTLIB::RINT_F32] = "rintf"; 195 Names[RTLIB::RINT_F64] = "rint"; 196 Names[RTLIB::RINT_F80] = "rintl"; 197 Names[RTLIB::RINT_F128] = "rintl"; 198 Names[RTLIB::RINT_PPCF128] = "rintl"; 199 Names[RTLIB::NEARBYINT_F32] = "nearbyintf"; 200 Names[RTLIB::NEARBYINT_F64] = "nearbyint"; 201 Names[RTLIB::NEARBYINT_F80] = "nearbyintl"; 202 Names[RTLIB::NEARBYINT_F128] = "nearbyintl"; 203 Names[RTLIB::NEARBYINT_PPCF128] = "nearbyintl"; 204 Names[RTLIB::ROUND_F32] = "roundf"; 205 Names[RTLIB::ROUND_F64] = "round"; 206 Names[RTLIB::ROUND_F80] = "roundl"; 207 Names[RTLIB::ROUND_F128] = "roundl"; 208 Names[RTLIB::ROUND_PPCF128] = "roundl"; 209 Names[RTLIB::FLOOR_F32] = "floorf"; 210 Names[RTLIB::FLOOR_F64] = "floor"; 211 Names[RTLIB::FLOOR_F80] = "floorl"; 212 Names[RTLIB::FLOOR_F128] = "floorl"; 213 Names[RTLIB::FLOOR_PPCF128] = "floorl"; 214 Names[RTLIB::FMIN_F32] = "fminf"; 215 Names[RTLIB::FMIN_F64] = "fmin"; 216 Names[RTLIB::FMIN_F80] = "fminl"; 217 Names[RTLIB::FMIN_F128] = "fminl"; 218 Names[RTLIB::FMIN_PPCF128] = "fminl"; 219 Names[RTLIB::FMAX_F32] = "fmaxf"; 220 Names[RTLIB::FMAX_F64] = "fmax"; 221 Names[RTLIB::FMAX_F80] = "fmaxl"; 222 Names[RTLIB::FMAX_F128] = "fmaxl"; 223 Names[RTLIB::FMAX_PPCF128] = "fmaxl"; 224 Names[RTLIB::ROUND_F32] = "roundf"; 225 Names[RTLIB::ROUND_F64] = "round"; 226 Names[RTLIB::ROUND_F80] = "roundl"; 227 Names[RTLIB::ROUND_F128] = "roundl"; 228 Names[RTLIB::ROUND_PPCF128] = "roundl"; 229 Names[RTLIB::COPYSIGN_F32] = "copysignf"; 230 Names[RTLIB::COPYSIGN_F64] = "copysign"; 231 Names[RTLIB::COPYSIGN_F80] = "copysignl"; 232 Names[RTLIB::COPYSIGN_F128] = "copysignl"; 233 Names[RTLIB::COPYSIGN_PPCF128] = "copysignl"; 234 Names[RTLIB::FPEXT_F32_PPCF128] = "__gcc_stoq"; 235 Names[RTLIB::FPEXT_F64_PPCF128] = "__gcc_dtoq"; 236 Names[RTLIB::FPEXT_F64_F128] = "__extenddftf2"; 237 Names[RTLIB::FPEXT_F32_F128] = "__extendsftf2"; 238 Names[RTLIB::FPEXT_F32_F64] = "__extendsfdf2"; 239 if (TT.isOSDarwin()) { 240 // For f16/f32 conversions, Darwin uses the standard naming scheme, instead 241 // of the gnueabi-style __gnu_*_ieee. 242 // FIXME: What about other targets? 243 Names[RTLIB::FPEXT_F16_F32] = "__extendhfsf2"; 244 Names[RTLIB::FPROUND_F32_F16] = "__truncsfhf2"; 245 } else { 246 Names[RTLIB::FPEXT_F16_F32] = "__gnu_h2f_ieee"; 247 Names[RTLIB::FPROUND_F32_F16] = "__gnu_f2h_ieee"; 248 } 249 Names[RTLIB::FPROUND_F64_F16] = "__truncdfhf2"; 250 Names[RTLIB::FPROUND_F80_F16] = "__truncxfhf2"; 251 Names[RTLIB::FPROUND_F128_F16] = "__trunctfhf2"; 252 Names[RTLIB::FPROUND_PPCF128_F16] = "__trunctfhf2"; 253 Names[RTLIB::FPROUND_F64_F32] = "__truncdfsf2"; 254 Names[RTLIB::FPROUND_F80_F32] = "__truncxfsf2"; 255 Names[RTLIB::FPROUND_F128_F32] = "__trunctfsf2"; 256 Names[RTLIB::FPROUND_PPCF128_F32] = "__gcc_qtos"; 257 Names[RTLIB::FPROUND_F80_F64] = "__truncxfdf2"; 258 Names[RTLIB::FPROUND_F128_F64] = "__trunctfdf2"; 259 Names[RTLIB::FPROUND_PPCF128_F64] = "__gcc_qtod"; 260 Names[RTLIB::FPTOSINT_F32_I32] = "__fixsfsi"; 261 Names[RTLIB::FPTOSINT_F32_I64] = "__fixsfdi"; 262 Names[RTLIB::FPTOSINT_F32_I128] = "__fixsfti"; 263 Names[RTLIB::FPTOSINT_F64_I32] = "__fixdfsi"; 264 Names[RTLIB::FPTOSINT_F64_I64] = "__fixdfdi"; 265 Names[RTLIB::FPTOSINT_F64_I128] = "__fixdfti"; 266 Names[RTLIB::FPTOSINT_F80_I32] = "__fixxfsi"; 267 Names[RTLIB::FPTOSINT_F80_I64] = "__fixxfdi"; 268 Names[RTLIB::FPTOSINT_F80_I128] = "__fixxfti"; 269 Names[RTLIB::FPTOSINT_F128_I32] = "__fixtfsi"; 270 Names[RTLIB::FPTOSINT_F128_I64] = "__fixtfdi"; 271 Names[RTLIB::FPTOSINT_F128_I128] = "__fixtfti"; 272 Names[RTLIB::FPTOSINT_PPCF128_I32] = "__gcc_qtou"; 273 Names[RTLIB::FPTOSINT_PPCF128_I64] = "__fixtfdi"; 274 Names[RTLIB::FPTOSINT_PPCF128_I128] = "__fixtfti"; 275 Names[RTLIB::FPTOUINT_F32_I32] = "__fixunssfsi"; 276 Names[RTLIB::FPTOUINT_F32_I64] = "__fixunssfdi"; 277 Names[RTLIB::FPTOUINT_F32_I128] = "__fixunssfti"; 278 Names[RTLIB::FPTOUINT_F64_I32] = "__fixunsdfsi"; 279 Names[RTLIB::FPTOUINT_F64_I64] = "__fixunsdfdi"; 280 Names[RTLIB::FPTOUINT_F64_I128] = "__fixunsdfti"; 281 Names[RTLIB::FPTOUINT_F80_I32] = "__fixunsxfsi"; 282 Names[RTLIB::FPTOUINT_F80_I64] = "__fixunsxfdi"; 283 Names[RTLIB::FPTOUINT_F80_I128] = "__fixunsxfti"; 284 Names[RTLIB::FPTOUINT_F128_I32] = "__fixunstfsi"; 285 Names[RTLIB::FPTOUINT_F128_I64] = "__fixunstfdi"; 286 Names[RTLIB::FPTOUINT_F128_I128] = "__fixunstfti"; 287 Names[RTLIB::FPTOUINT_PPCF128_I32] = "__fixunstfsi"; 288 Names[RTLIB::FPTOUINT_PPCF128_I64] = "__fixunstfdi"; 289 Names[RTLIB::FPTOUINT_PPCF128_I128] = "__fixunstfti"; 290 Names[RTLIB::SINTTOFP_I32_F32] = "__floatsisf"; 291 Names[RTLIB::SINTTOFP_I32_F64] = "__floatsidf"; 292 Names[RTLIB::SINTTOFP_I32_F80] = "__floatsixf"; 293 Names[RTLIB::SINTTOFP_I32_F128] = "__floatsitf"; 294 Names[RTLIB::SINTTOFP_I32_PPCF128] = "__gcc_itoq"; 295 Names[RTLIB::SINTTOFP_I64_F32] = "__floatdisf"; 296 Names[RTLIB::SINTTOFP_I64_F64] = "__floatdidf"; 297 Names[RTLIB::SINTTOFP_I64_F80] = "__floatdixf"; 298 Names[RTLIB::SINTTOFP_I64_F128] = "__floatditf"; 299 Names[RTLIB::SINTTOFP_I64_PPCF128] = "__floatditf"; 300 Names[RTLIB::SINTTOFP_I128_F32] = "__floattisf"; 301 Names[RTLIB::SINTTOFP_I128_F64] = "__floattidf"; 302 Names[RTLIB::SINTTOFP_I128_F80] = "__floattixf"; 303 Names[RTLIB::SINTTOFP_I128_F128] = "__floattitf"; 304 Names[RTLIB::SINTTOFP_I128_PPCF128] = "__floattitf"; 305 Names[RTLIB::UINTTOFP_I32_F32] = "__floatunsisf"; 306 Names[RTLIB::UINTTOFP_I32_F64] = "__floatunsidf"; 307 Names[RTLIB::UINTTOFP_I32_F80] = "__floatunsixf"; 308 Names[RTLIB::UINTTOFP_I32_F128] = "__floatunsitf"; 309 Names[RTLIB::UINTTOFP_I32_PPCF128] = "__gcc_utoq"; 310 Names[RTLIB::UINTTOFP_I64_F32] = "__floatundisf"; 311 Names[RTLIB::UINTTOFP_I64_F64] = "__floatundidf"; 312 Names[RTLIB::UINTTOFP_I64_F80] = "__floatundixf"; 313 Names[RTLIB::UINTTOFP_I64_F128] = "__floatunditf"; 314 Names[RTLIB::UINTTOFP_I64_PPCF128] = "__floatunditf"; 315 Names[RTLIB::UINTTOFP_I128_F32] = "__floatuntisf"; 316 Names[RTLIB::UINTTOFP_I128_F64] = "__floatuntidf"; 317 Names[RTLIB::UINTTOFP_I128_F80] = "__floatuntixf"; 318 Names[RTLIB::UINTTOFP_I128_F128] = "__floatuntitf"; 319 Names[RTLIB::UINTTOFP_I128_PPCF128] = "__floatuntitf"; 320 Names[RTLIB::OEQ_F32] = "__eqsf2"; 321 Names[RTLIB::OEQ_F64] = "__eqdf2"; 322 Names[RTLIB::OEQ_F128] = "__eqtf2"; 323 Names[RTLIB::OEQ_PPCF128] = "__gcc_qeq"; 324 Names[RTLIB::UNE_F32] = "__nesf2"; 325 Names[RTLIB::UNE_F64] = "__nedf2"; 326 Names[RTLIB::UNE_F128] = "__netf2"; 327 Names[RTLIB::UNE_PPCF128] = "__gcc_qne"; 328 Names[RTLIB::OGE_F32] = "__gesf2"; 329 Names[RTLIB::OGE_F64] = "__gedf2"; 330 Names[RTLIB::OGE_F128] = "__getf2"; 331 Names[RTLIB::OGE_PPCF128] = "__gcc_qge"; 332 Names[RTLIB::OLT_F32] = "__ltsf2"; 333 Names[RTLIB::OLT_F64] = "__ltdf2"; 334 Names[RTLIB::OLT_F128] = "__lttf2"; 335 Names[RTLIB::OLT_PPCF128] = "__gcc_qlt"; 336 Names[RTLIB::OLE_F32] = "__lesf2"; 337 Names[RTLIB::OLE_F64] = "__ledf2"; 338 Names[RTLIB::OLE_F128] = "__letf2"; 339 Names[RTLIB::OLE_PPCF128] = "__gcc_qle"; 340 Names[RTLIB::OGT_F32] = "__gtsf2"; 341 Names[RTLIB::OGT_F64] = "__gtdf2"; 342 Names[RTLIB::OGT_F128] = "__gttf2"; 343 Names[RTLIB::OGT_PPCF128] = "__gcc_qgt"; 344 Names[RTLIB::UO_F32] = "__unordsf2"; 345 Names[RTLIB::UO_F64] = "__unorddf2"; 346 Names[RTLIB::UO_F128] = "__unordtf2"; 347 Names[RTLIB::UO_PPCF128] = "__gcc_qunord"; 348 Names[RTLIB::O_F32] = "__unordsf2"; 349 Names[RTLIB::O_F64] = "__unorddf2"; 350 Names[RTLIB::O_F128] = "__unordtf2"; 351 Names[RTLIB::O_PPCF128] = "__gcc_qunord"; 352 Names[RTLIB::MEMCPY] = "memcpy"; 353 Names[RTLIB::MEMMOVE] = "memmove"; 354 Names[RTLIB::MEMSET] = "memset"; 355 Names[RTLIB::UNWIND_RESUME] = "_Unwind_Resume"; 356 Names[RTLIB::SYNC_VAL_COMPARE_AND_SWAP_1] = "__sync_val_compare_and_swap_1"; 357 Names[RTLIB::SYNC_VAL_COMPARE_AND_SWAP_2] = "__sync_val_compare_and_swap_2"; 358 Names[RTLIB::SYNC_VAL_COMPARE_AND_SWAP_4] = "__sync_val_compare_and_swap_4"; 359 Names[RTLIB::SYNC_VAL_COMPARE_AND_SWAP_8] = "__sync_val_compare_and_swap_8"; 360 Names[RTLIB::SYNC_VAL_COMPARE_AND_SWAP_16] = "__sync_val_compare_and_swap_16"; 361 Names[RTLIB::SYNC_LOCK_TEST_AND_SET_1] = "__sync_lock_test_and_set_1"; 362 Names[RTLIB::SYNC_LOCK_TEST_AND_SET_2] = "__sync_lock_test_and_set_2"; 363 Names[RTLIB::SYNC_LOCK_TEST_AND_SET_4] = "__sync_lock_test_and_set_4"; 364 Names[RTLIB::SYNC_LOCK_TEST_AND_SET_8] = "__sync_lock_test_and_set_8"; 365 Names[RTLIB::SYNC_LOCK_TEST_AND_SET_16] = "__sync_lock_test_and_set_16"; 366 Names[RTLIB::SYNC_FETCH_AND_ADD_1] = "__sync_fetch_and_add_1"; 367 Names[RTLIB::SYNC_FETCH_AND_ADD_2] = "__sync_fetch_and_add_2"; 368 Names[RTLIB::SYNC_FETCH_AND_ADD_4] = "__sync_fetch_and_add_4"; 369 Names[RTLIB::SYNC_FETCH_AND_ADD_8] = "__sync_fetch_and_add_8"; 370 Names[RTLIB::SYNC_FETCH_AND_ADD_16] = "__sync_fetch_and_add_16"; 371 Names[RTLIB::SYNC_FETCH_AND_SUB_1] = "__sync_fetch_and_sub_1"; 372 Names[RTLIB::SYNC_FETCH_AND_SUB_2] = "__sync_fetch_and_sub_2"; 373 Names[RTLIB::SYNC_FETCH_AND_SUB_4] = "__sync_fetch_and_sub_4"; 374 Names[RTLIB::SYNC_FETCH_AND_SUB_8] = "__sync_fetch_and_sub_8"; 375 Names[RTLIB::SYNC_FETCH_AND_SUB_16] = "__sync_fetch_and_sub_16"; 376 Names[RTLIB::SYNC_FETCH_AND_AND_1] = "__sync_fetch_and_and_1"; 377 Names[RTLIB::SYNC_FETCH_AND_AND_2] = "__sync_fetch_and_and_2"; 378 Names[RTLIB::SYNC_FETCH_AND_AND_4] = "__sync_fetch_and_and_4"; 379 Names[RTLIB::SYNC_FETCH_AND_AND_8] = "__sync_fetch_and_and_8"; 380 Names[RTLIB::SYNC_FETCH_AND_AND_16] = "__sync_fetch_and_and_16"; 381 Names[RTLIB::SYNC_FETCH_AND_OR_1] = "__sync_fetch_and_or_1"; 382 Names[RTLIB::SYNC_FETCH_AND_OR_2] = "__sync_fetch_and_or_2"; 383 Names[RTLIB::SYNC_FETCH_AND_OR_4] = "__sync_fetch_and_or_4"; 384 Names[RTLIB::SYNC_FETCH_AND_OR_8] = "__sync_fetch_and_or_8"; 385 Names[RTLIB::SYNC_FETCH_AND_OR_16] = "__sync_fetch_and_or_16"; 386 Names[RTLIB::SYNC_FETCH_AND_XOR_1] = "__sync_fetch_and_xor_1"; 387 Names[RTLIB::SYNC_FETCH_AND_XOR_2] = "__sync_fetch_and_xor_2"; 388 Names[RTLIB::SYNC_FETCH_AND_XOR_4] = "__sync_fetch_and_xor_4"; 389 Names[RTLIB::SYNC_FETCH_AND_XOR_8] = "__sync_fetch_and_xor_8"; 390 Names[RTLIB::SYNC_FETCH_AND_XOR_16] = "__sync_fetch_and_xor_16"; 391 Names[RTLIB::SYNC_FETCH_AND_NAND_1] = "__sync_fetch_and_nand_1"; 392 Names[RTLIB::SYNC_FETCH_AND_NAND_2] = "__sync_fetch_and_nand_2"; 393 Names[RTLIB::SYNC_FETCH_AND_NAND_4] = "__sync_fetch_and_nand_4"; 394 Names[RTLIB::SYNC_FETCH_AND_NAND_8] = "__sync_fetch_and_nand_8"; 395 Names[RTLIB::SYNC_FETCH_AND_NAND_16] = "__sync_fetch_and_nand_16"; 396 Names[RTLIB::SYNC_FETCH_AND_MAX_1] = "__sync_fetch_and_max_1"; 397 Names[RTLIB::SYNC_FETCH_AND_MAX_2] = "__sync_fetch_and_max_2"; 398 Names[RTLIB::SYNC_FETCH_AND_MAX_4] = "__sync_fetch_and_max_4"; 399 Names[RTLIB::SYNC_FETCH_AND_MAX_8] = "__sync_fetch_and_max_8"; 400 Names[RTLIB::SYNC_FETCH_AND_MAX_16] = "__sync_fetch_and_max_16"; 401 Names[RTLIB::SYNC_FETCH_AND_UMAX_1] = "__sync_fetch_and_umax_1"; 402 Names[RTLIB::SYNC_FETCH_AND_UMAX_2] = "__sync_fetch_and_umax_2"; 403 Names[RTLIB::SYNC_FETCH_AND_UMAX_4] = "__sync_fetch_and_umax_4"; 404 Names[RTLIB::SYNC_FETCH_AND_UMAX_8] = "__sync_fetch_and_umax_8"; 405 Names[RTLIB::SYNC_FETCH_AND_UMAX_16] = "__sync_fetch_and_umax_16"; 406 Names[RTLIB::SYNC_FETCH_AND_MIN_1] = "__sync_fetch_and_min_1"; 407 Names[RTLIB::SYNC_FETCH_AND_MIN_2] = "__sync_fetch_and_min_2"; 408 Names[RTLIB::SYNC_FETCH_AND_MIN_4] = "__sync_fetch_and_min_4"; 409 Names[RTLIB::SYNC_FETCH_AND_MIN_8] = "__sync_fetch_and_min_8"; 410 Names[RTLIB::SYNC_FETCH_AND_MIN_16] = "__sync_fetch_and_min_16"; 411 Names[RTLIB::SYNC_FETCH_AND_UMIN_1] = "__sync_fetch_and_umin_1"; 412 Names[RTLIB::SYNC_FETCH_AND_UMIN_2] = "__sync_fetch_and_umin_2"; 413 Names[RTLIB::SYNC_FETCH_AND_UMIN_4] = "__sync_fetch_and_umin_4"; 414 Names[RTLIB::SYNC_FETCH_AND_UMIN_8] = "__sync_fetch_and_umin_8"; 415 Names[RTLIB::SYNC_FETCH_AND_UMIN_16] = "__sync_fetch_and_umin_16"; 416 417 Names[RTLIB::ATOMIC_LOAD] = "__atomic_load"; 418 Names[RTLIB::ATOMIC_LOAD_1] = "__atomic_load_1"; 419 Names[RTLIB::ATOMIC_LOAD_2] = "__atomic_load_2"; 420 Names[RTLIB::ATOMIC_LOAD_4] = "__atomic_load_4"; 421 Names[RTLIB::ATOMIC_LOAD_8] = "__atomic_load_8"; 422 Names[RTLIB::ATOMIC_LOAD_16] = "__atomic_load_16"; 423 424 Names[RTLIB::ATOMIC_STORE] = "__atomic_store"; 425 Names[RTLIB::ATOMIC_STORE_1] = "__atomic_store_1"; 426 Names[RTLIB::ATOMIC_STORE_2] = "__atomic_store_2"; 427 Names[RTLIB::ATOMIC_STORE_4] = "__atomic_store_4"; 428 Names[RTLIB::ATOMIC_STORE_8] = "__atomic_store_8"; 429 Names[RTLIB::ATOMIC_STORE_16] = "__atomic_store_16"; 430 431 Names[RTLIB::ATOMIC_EXCHANGE] = "__atomic_exchange"; 432 Names[RTLIB::ATOMIC_EXCHANGE_1] = "__atomic_exchange_1"; 433 Names[RTLIB::ATOMIC_EXCHANGE_2] = "__atomic_exchange_2"; 434 Names[RTLIB::ATOMIC_EXCHANGE_4] = "__atomic_exchange_4"; 435 Names[RTLIB::ATOMIC_EXCHANGE_8] = "__atomic_exchange_8"; 436 Names[RTLIB::ATOMIC_EXCHANGE_16] = "__atomic_exchange_16"; 437 438 Names[RTLIB::ATOMIC_COMPARE_EXCHANGE] = "__atomic_compare_exchange"; 439 Names[RTLIB::ATOMIC_COMPARE_EXCHANGE_1] = "__atomic_compare_exchange_1"; 440 Names[RTLIB::ATOMIC_COMPARE_EXCHANGE_2] = "__atomic_compare_exchange_2"; 441 Names[RTLIB::ATOMIC_COMPARE_EXCHANGE_4] = "__atomic_compare_exchange_4"; 442 Names[RTLIB::ATOMIC_COMPARE_EXCHANGE_8] = "__atomic_compare_exchange_8"; 443 Names[RTLIB::ATOMIC_COMPARE_EXCHANGE_16] = "__atomic_compare_exchange_16"; 444 445 Names[RTLIB::ATOMIC_FETCH_ADD_1] = "__atomic_fetch_add_1"; 446 Names[RTLIB::ATOMIC_FETCH_ADD_2] = "__atomic_fetch_add_2"; 447 Names[RTLIB::ATOMIC_FETCH_ADD_4] = "__atomic_fetch_add_4"; 448 Names[RTLIB::ATOMIC_FETCH_ADD_8] = "__atomic_fetch_add_8"; 449 Names[RTLIB::ATOMIC_FETCH_ADD_16] = "__atomic_fetch_add_16"; 450 Names[RTLIB::ATOMIC_FETCH_SUB_1] = "__atomic_fetch_sub_1"; 451 Names[RTLIB::ATOMIC_FETCH_SUB_2] = "__atomic_fetch_sub_2"; 452 Names[RTLIB::ATOMIC_FETCH_SUB_4] = "__atomic_fetch_sub_4"; 453 Names[RTLIB::ATOMIC_FETCH_SUB_8] = "__atomic_fetch_sub_8"; 454 Names[RTLIB::ATOMIC_FETCH_SUB_16] = "__atomic_fetch_sub_16"; 455 Names[RTLIB::ATOMIC_FETCH_AND_1] = "__atomic_fetch_and_1"; 456 Names[RTLIB::ATOMIC_FETCH_AND_2] = "__atomic_fetch_and_2"; 457 Names[RTLIB::ATOMIC_FETCH_AND_4] = "__atomic_fetch_and_4"; 458 Names[RTLIB::ATOMIC_FETCH_AND_8] = "__atomic_fetch_and_8"; 459 Names[RTLIB::ATOMIC_FETCH_AND_16] = "__atomic_fetch_and_16"; 460 Names[RTLIB::ATOMIC_FETCH_OR_1] = "__atomic_fetch_or_1"; 461 Names[RTLIB::ATOMIC_FETCH_OR_2] = "__atomic_fetch_or_2"; 462 Names[RTLIB::ATOMIC_FETCH_OR_4] = "__atomic_fetch_or_4"; 463 Names[RTLIB::ATOMIC_FETCH_OR_8] = "__atomic_fetch_or_8"; 464 Names[RTLIB::ATOMIC_FETCH_OR_16] = "__atomic_fetch_or_16"; 465 Names[RTLIB::ATOMIC_FETCH_XOR_1] = "__atomic_fetch_xor_1"; 466 Names[RTLIB::ATOMIC_FETCH_XOR_2] = "__atomic_fetch_xor_2"; 467 Names[RTLIB::ATOMIC_FETCH_XOR_4] = "__atomic_fetch_xor_4"; 468 Names[RTLIB::ATOMIC_FETCH_XOR_8] = "__atomic_fetch_xor_8"; 469 Names[RTLIB::ATOMIC_FETCH_XOR_16] = "__atomic_fetch_xor_16"; 470 Names[RTLIB::ATOMIC_FETCH_NAND_1] = "__atomic_fetch_nand_1"; 471 Names[RTLIB::ATOMIC_FETCH_NAND_2] = "__atomic_fetch_nand_2"; 472 Names[RTLIB::ATOMIC_FETCH_NAND_4] = "__atomic_fetch_nand_4"; 473 Names[RTLIB::ATOMIC_FETCH_NAND_8] = "__atomic_fetch_nand_8"; 474 Names[RTLIB::ATOMIC_FETCH_NAND_16] = "__atomic_fetch_nand_16"; 475 476 if (TT.getEnvironment() == Triple::GNU) { 477 Names[RTLIB::SINCOS_F32] = "sincosf"; 478 Names[RTLIB::SINCOS_F64] = "sincos"; 479 Names[RTLIB::SINCOS_F80] = "sincosl"; 480 Names[RTLIB::SINCOS_F128] = "sincosl"; 481 Names[RTLIB::SINCOS_PPCF128] = "sincosl"; 482 } 483 484 if (!TT.isOSOpenBSD()) { 485 Names[RTLIB::STACKPROTECTOR_CHECK_FAIL] = "__stack_chk_fail"; 486 } 487 488 Names[RTLIB::DEOPTIMIZE] = "__llvm_deoptimize"; 489 } 490 491 /// InitLibcallCallingConvs - Set default libcall CallingConvs. 492 /// 493 static void InitLibcallCallingConvs(CallingConv::ID *CCs) { 494 for (int i = 0; i < RTLIB::UNKNOWN_LIBCALL; ++i) { 495 CCs[i] = CallingConv::C; 496 } 497 } 498 499 /// getFPEXT - Return the FPEXT_*_* value for the given types, or 500 /// UNKNOWN_LIBCALL if there is none. 501 RTLIB::Libcall RTLIB::getFPEXT(EVT OpVT, EVT RetVT) { 502 if (OpVT == MVT::f16) { 503 if (RetVT == MVT::f32) 504 return FPEXT_F16_F32; 505 } else if (OpVT == MVT::f32) { 506 if (RetVT == MVT::f64) 507 return FPEXT_F32_F64; 508 if (RetVT == MVT::f128) 509 return FPEXT_F32_F128; 510 if (RetVT == MVT::ppcf128) 511 return FPEXT_F32_PPCF128; 512 } else if (OpVT == MVT::f64) { 513 if (RetVT == MVT::f128) 514 return FPEXT_F64_F128; 515 else if (RetVT == MVT::ppcf128) 516 return FPEXT_F64_PPCF128; 517 } 518 519 return UNKNOWN_LIBCALL; 520 } 521 522 /// getFPROUND - Return the FPROUND_*_* value for the given types, or 523 /// UNKNOWN_LIBCALL if there is none. 524 RTLIB::Libcall RTLIB::getFPROUND(EVT OpVT, EVT RetVT) { 525 if (RetVT == MVT::f16) { 526 if (OpVT == MVT::f32) 527 return FPROUND_F32_F16; 528 if (OpVT == MVT::f64) 529 return FPROUND_F64_F16; 530 if (OpVT == MVT::f80) 531 return FPROUND_F80_F16; 532 if (OpVT == MVT::f128) 533 return FPROUND_F128_F16; 534 if (OpVT == MVT::ppcf128) 535 return FPROUND_PPCF128_F16; 536 } else if (RetVT == MVT::f32) { 537 if (OpVT == MVT::f64) 538 return FPROUND_F64_F32; 539 if (OpVT == MVT::f80) 540 return FPROUND_F80_F32; 541 if (OpVT == MVT::f128) 542 return FPROUND_F128_F32; 543 if (OpVT == MVT::ppcf128) 544 return FPROUND_PPCF128_F32; 545 } else if (RetVT == MVT::f64) { 546 if (OpVT == MVT::f80) 547 return FPROUND_F80_F64; 548 if (OpVT == MVT::f128) 549 return FPROUND_F128_F64; 550 if (OpVT == MVT::ppcf128) 551 return FPROUND_PPCF128_F64; 552 } 553 554 return UNKNOWN_LIBCALL; 555 } 556 557 /// getFPTOSINT - Return the FPTOSINT_*_* value for the given types, or 558 /// UNKNOWN_LIBCALL if there is none. 559 RTLIB::Libcall RTLIB::getFPTOSINT(EVT OpVT, EVT RetVT) { 560 if (OpVT == MVT::f32) { 561 if (RetVT == MVT::i32) 562 return FPTOSINT_F32_I32; 563 if (RetVT == MVT::i64) 564 return FPTOSINT_F32_I64; 565 if (RetVT == MVT::i128) 566 return FPTOSINT_F32_I128; 567 } else if (OpVT == MVT::f64) { 568 if (RetVT == MVT::i32) 569 return FPTOSINT_F64_I32; 570 if (RetVT == MVT::i64) 571 return FPTOSINT_F64_I64; 572 if (RetVT == MVT::i128) 573 return FPTOSINT_F64_I128; 574 } else if (OpVT == MVT::f80) { 575 if (RetVT == MVT::i32) 576 return FPTOSINT_F80_I32; 577 if (RetVT == MVT::i64) 578 return FPTOSINT_F80_I64; 579 if (RetVT == MVT::i128) 580 return FPTOSINT_F80_I128; 581 } else if (OpVT == MVT::f128) { 582 if (RetVT == MVT::i32) 583 return FPTOSINT_F128_I32; 584 if (RetVT == MVT::i64) 585 return FPTOSINT_F128_I64; 586 if (RetVT == MVT::i128) 587 return FPTOSINT_F128_I128; 588 } else if (OpVT == MVT::ppcf128) { 589 if (RetVT == MVT::i32) 590 return FPTOSINT_PPCF128_I32; 591 if (RetVT == MVT::i64) 592 return FPTOSINT_PPCF128_I64; 593 if (RetVT == MVT::i128) 594 return FPTOSINT_PPCF128_I128; 595 } 596 return UNKNOWN_LIBCALL; 597 } 598 599 /// getFPTOUINT - Return the FPTOUINT_*_* value for the given types, or 600 /// UNKNOWN_LIBCALL if there is none. 601 RTLIB::Libcall RTLIB::getFPTOUINT(EVT OpVT, EVT RetVT) { 602 if (OpVT == MVT::f32) { 603 if (RetVT == MVT::i32) 604 return FPTOUINT_F32_I32; 605 if (RetVT == MVT::i64) 606 return FPTOUINT_F32_I64; 607 if (RetVT == MVT::i128) 608 return FPTOUINT_F32_I128; 609 } else if (OpVT == MVT::f64) { 610 if (RetVT == MVT::i32) 611 return FPTOUINT_F64_I32; 612 if (RetVT == MVT::i64) 613 return FPTOUINT_F64_I64; 614 if (RetVT == MVT::i128) 615 return FPTOUINT_F64_I128; 616 } else if (OpVT == MVT::f80) { 617 if (RetVT == MVT::i32) 618 return FPTOUINT_F80_I32; 619 if (RetVT == MVT::i64) 620 return FPTOUINT_F80_I64; 621 if (RetVT == MVT::i128) 622 return FPTOUINT_F80_I128; 623 } else if (OpVT == MVT::f128) { 624 if (RetVT == MVT::i32) 625 return FPTOUINT_F128_I32; 626 if (RetVT == MVT::i64) 627 return FPTOUINT_F128_I64; 628 if (RetVT == MVT::i128) 629 return FPTOUINT_F128_I128; 630 } else if (OpVT == MVT::ppcf128) { 631 if (RetVT == MVT::i32) 632 return FPTOUINT_PPCF128_I32; 633 if (RetVT == MVT::i64) 634 return FPTOUINT_PPCF128_I64; 635 if (RetVT == MVT::i128) 636 return FPTOUINT_PPCF128_I128; 637 } 638 return UNKNOWN_LIBCALL; 639 } 640 641 /// getSINTTOFP - Return the SINTTOFP_*_* value for the given types, or 642 /// UNKNOWN_LIBCALL if there is none. 643 RTLIB::Libcall RTLIB::getSINTTOFP(EVT OpVT, EVT RetVT) { 644 if (OpVT == MVT::i32) { 645 if (RetVT == MVT::f32) 646 return SINTTOFP_I32_F32; 647 if (RetVT == MVT::f64) 648 return SINTTOFP_I32_F64; 649 if (RetVT == MVT::f80) 650 return SINTTOFP_I32_F80; 651 if (RetVT == MVT::f128) 652 return SINTTOFP_I32_F128; 653 if (RetVT == MVT::ppcf128) 654 return SINTTOFP_I32_PPCF128; 655 } else if (OpVT == MVT::i64) { 656 if (RetVT == MVT::f32) 657 return SINTTOFP_I64_F32; 658 if (RetVT == MVT::f64) 659 return SINTTOFP_I64_F64; 660 if (RetVT == MVT::f80) 661 return SINTTOFP_I64_F80; 662 if (RetVT == MVT::f128) 663 return SINTTOFP_I64_F128; 664 if (RetVT == MVT::ppcf128) 665 return SINTTOFP_I64_PPCF128; 666 } else if (OpVT == MVT::i128) { 667 if (RetVT == MVT::f32) 668 return SINTTOFP_I128_F32; 669 if (RetVT == MVT::f64) 670 return SINTTOFP_I128_F64; 671 if (RetVT == MVT::f80) 672 return SINTTOFP_I128_F80; 673 if (RetVT == MVT::f128) 674 return SINTTOFP_I128_F128; 675 if (RetVT == MVT::ppcf128) 676 return SINTTOFP_I128_PPCF128; 677 } 678 return UNKNOWN_LIBCALL; 679 } 680 681 /// getUINTTOFP - Return the UINTTOFP_*_* value for the given types, or 682 /// UNKNOWN_LIBCALL if there is none. 683 RTLIB::Libcall RTLIB::getUINTTOFP(EVT OpVT, EVT RetVT) { 684 if (OpVT == MVT::i32) { 685 if (RetVT == MVT::f32) 686 return UINTTOFP_I32_F32; 687 if (RetVT == MVT::f64) 688 return UINTTOFP_I32_F64; 689 if (RetVT == MVT::f80) 690 return UINTTOFP_I32_F80; 691 if (RetVT == MVT::f128) 692 return UINTTOFP_I32_F128; 693 if (RetVT == MVT::ppcf128) 694 return UINTTOFP_I32_PPCF128; 695 } else if (OpVT == MVT::i64) { 696 if (RetVT == MVT::f32) 697 return UINTTOFP_I64_F32; 698 if (RetVT == MVT::f64) 699 return UINTTOFP_I64_F64; 700 if (RetVT == MVT::f80) 701 return UINTTOFP_I64_F80; 702 if (RetVT == MVT::f128) 703 return UINTTOFP_I64_F128; 704 if (RetVT == MVT::ppcf128) 705 return UINTTOFP_I64_PPCF128; 706 } else if (OpVT == MVT::i128) { 707 if (RetVT == MVT::f32) 708 return UINTTOFP_I128_F32; 709 if (RetVT == MVT::f64) 710 return UINTTOFP_I128_F64; 711 if (RetVT == MVT::f80) 712 return UINTTOFP_I128_F80; 713 if (RetVT == MVT::f128) 714 return UINTTOFP_I128_F128; 715 if (RetVT == MVT::ppcf128) 716 return UINTTOFP_I128_PPCF128; 717 } 718 return UNKNOWN_LIBCALL; 719 } 720 721 RTLIB::Libcall RTLIB::getSYNC(unsigned Opc, MVT VT) { 722 #define OP_TO_LIBCALL(Name, Enum) \ 723 case Name: \ 724 switch (VT.SimpleTy) { \ 725 default: \ 726 return UNKNOWN_LIBCALL; \ 727 case MVT::i8: \ 728 return Enum##_1; \ 729 case MVT::i16: \ 730 return Enum##_2; \ 731 case MVT::i32: \ 732 return Enum##_4; \ 733 case MVT::i64: \ 734 return Enum##_8; \ 735 case MVT::i128: \ 736 return Enum##_16; \ 737 } 738 739 switch (Opc) { 740 OP_TO_LIBCALL(ISD::ATOMIC_SWAP, SYNC_LOCK_TEST_AND_SET) 741 OP_TO_LIBCALL(ISD::ATOMIC_CMP_SWAP, SYNC_VAL_COMPARE_AND_SWAP) 742 OP_TO_LIBCALL(ISD::ATOMIC_LOAD_ADD, SYNC_FETCH_AND_ADD) 743 OP_TO_LIBCALL(ISD::ATOMIC_LOAD_SUB, SYNC_FETCH_AND_SUB) 744 OP_TO_LIBCALL(ISD::ATOMIC_LOAD_AND, SYNC_FETCH_AND_AND) 745 OP_TO_LIBCALL(ISD::ATOMIC_LOAD_OR, SYNC_FETCH_AND_OR) 746 OP_TO_LIBCALL(ISD::ATOMIC_LOAD_XOR, SYNC_FETCH_AND_XOR) 747 OP_TO_LIBCALL(ISD::ATOMIC_LOAD_NAND, SYNC_FETCH_AND_NAND) 748 OP_TO_LIBCALL(ISD::ATOMIC_LOAD_MAX, SYNC_FETCH_AND_MAX) 749 OP_TO_LIBCALL(ISD::ATOMIC_LOAD_UMAX, SYNC_FETCH_AND_UMAX) 750 OP_TO_LIBCALL(ISD::ATOMIC_LOAD_MIN, SYNC_FETCH_AND_MIN) 751 OP_TO_LIBCALL(ISD::ATOMIC_LOAD_UMIN, SYNC_FETCH_AND_UMIN) 752 } 753 754 #undef OP_TO_LIBCALL 755 756 return UNKNOWN_LIBCALL; 757 } 758 759 /// InitCmpLibcallCCs - Set default comparison libcall CC. 760 /// 761 static void InitCmpLibcallCCs(ISD::CondCode *CCs) { 762 memset(CCs, ISD::SETCC_INVALID, sizeof(ISD::CondCode)*RTLIB::UNKNOWN_LIBCALL); 763 CCs[RTLIB::OEQ_F32] = ISD::SETEQ; 764 CCs[RTLIB::OEQ_F64] = ISD::SETEQ; 765 CCs[RTLIB::OEQ_F128] = ISD::SETEQ; 766 CCs[RTLIB::OEQ_PPCF128] = ISD::SETEQ; 767 CCs[RTLIB::UNE_F32] = ISD::SETNE; 768 CCs[RTLIB::UNE_F64] = ISD::SETNE; 769 CCs[RTLIB::UNE_F128] = ISD::SETNE; 770 CCs[RTLIB::UNE_PPCF128] = ISD::SETNE; 771 CCs[RTLIB::OGE_F32] = ISD::SETGE; 772 CCs[RTLIB::OGE_F64] = ISD::SETGE; 773 CCs[RTLIB::OGE_F128] = ISD::SETGE; 774 CCs[RTLIB::OGE_PPCF128] = ISD::SETGE; 775 CCs[RTLIB::OLT_F32] = ISD::SETLT; 776 CCs[RTLIB::OLT_F64] = ISD::SETLT; 777 CCs[RTLIB::OLT_F128] = ISD::SETLT; 778 CCs[RTLIB::OLT_PPCF128] = ISD::SETLT; 779 CCs[RTLIB::OLE_F32] = ISD::SETLE; 780 CCs[RTLIB::OLE_F64] = ISD::SETLE; 781 CCs[RTLIB::OLE_F128] = ISD::SETLE; 782 CCs[RTLIB::OLE_PPCF128] = ISD::SETLE; 783 CCs[RTLIB::OGT_F32] = ISD::SETGT; 784 CCs[RTLIB::OGT_F64] = ISD::SETGT; 785 CCs[RTLIB::OGT_F128] = ISD::SETGT; 786 CCs[RTLIB::OGT_PPCF128] = ISD::SETGT; 787 CCs[RTLIB::UO_F32] = ISD::SETNE; 788 CCs[RTLIB::UO_F64] = ISD::SETNE; 789 CCs[RTLIB::UO_F128] = ISD::SETNE; 790 CCs[RTLIB::UO_PPCF128] = ISD::SETNE; 791 CCs[RTLIB::O_F32] = ISD::SETEQ; 792 CCs[RTLIB::O_F64] = ISD::SETEQ; 793 CCs[RTLIB::O_F128] = ISD::SETEQ; 794 CCs[RTLIB::O_PPCF128] = ISD::SETEQ; 795 } 796 797 /// NOTE: The TargetMachine owns TLOF. 798 TargetLoweringBase::TargetLoweringBase(const TargetMachine &tm) : TM(tm) { 799 initActions(); 800 801 // Perform these initializations only once. 802 MaxStoresPerMemset = MaxStoresPerMemcpy = MaxStoresPerMemmove = 8; 803 MaxStoresPerMemsetOptSize = MaxStoresPerMemcpyOptSize 804 = MaxStoresPerMemmoveOptSize = 4; 805 UseUnderscoreSetJmp = false; 806 UseUnderscoreLongJmp = false; 807 SelectIsExpensive = false; 808 HasMultipleConditionRegisters = false; 809 HasExtractBitsInsn = false; 810 FsqrtIsCheap = false; 811 JumpIsExpensive = JumpIsExpensiveOverride; 812 PredictableSelectIsExpensive = false; 813 MaskAndBranchFoldingIsLegal = false; 814 EnableExtLdPromotion = false; 815 HasFloatingPointExceptions = true; 816 StackPointerRegisterToSaveRestore = 0; 817 BooleanContents = UndefinedBooleanContent; 818 BooleanFloatContents = UndefinedBooleanContent; 819 BooleanVectorContents = UndefinedBooleanContent; 820 SchedPreferenceInfo = Sched::ILP; 821 JumpBufSize = 0; 822 JumpBufAlignment = 0; 823 MinFunctionAlignment = 0; 824 PrefFunctionAlignment = 0; 825 PrefLoopAlignment = 0; 826 GatherAllAliasesMaxDepth = 6; 827 MinStackArgumentAlignment = 1; 828 MinimumJumpTableEntries = 4; 829 // TODO: the default will be switched to 0 in the next commit, along 830 // with the Target-specific changes necessary. 831 MaxAtomicSizeInBitsSupported = 1024; 832 833 std::fill(std::begin(LibcallRoutineNames), std::end(LibcallRoutineNames), nullptr); 834 835 InitLibcallNames(LibcallRoutineNames, TM.getTargetTriple()); 836 InitCmpLibcallCCs(CmpLibcallCCs); 837 InitLibcallCallingConvs(LibcallCallingConvs); 838 } 839 840 void TargetLoweringBase::initActions() { 841 // All operations default to being supported. 842 memset(OpActions, 0, sizeof(OpActions)); 843 memset(LoadExtActions, 0, sizeof(LoadExtActions)); 844 memset(TruncStoreActions, 0, sizeof(TruncStoreActions)); 845 memset(IndexedModeActions, 0, sizeof(IndexedModeActions)); 846 memset(CondCodeActions, 0, sizeof(CondCodeActions)); 847 std::fill(std::begin(RegClassForVT), std::end(RegClassForVT), nullptr); 848 std::fill(std::begin(TargetDAGCombineArray), 849 std::end(TargetDAGCombineArray), 0); 850 851 // Set default actions for various operations. 852 for (MVT VT : MVT::all_valuetypes()) { 853 // Default all indexed load / store to expand. 854 for (unsigned IM = (unsigned)ISD::PRE_INC; 855 IM != (unsigned)ISD::LAST_INDEXED_MODE; ++IM) { 856 setIndexedLoadAction(IM, VT, Expand); 857 setIndexedStoreAction(IM, VT, Expand); 858 } 859 860 // Most backends expect to see the node which just returns the value loaded. 861 setOperationAction(ISD::ATOMIC_CMP_SWAP_WITH_SUCCESS, VT, Expand); 862 863 // These operations default to expand. 864 setOperationAction(ISD::FGETSIGN, VT, Expand); 865 setOperationAction(ISD::CONCAT_VECTORS, VT, Expand); 866 setOperationAction(ISD::FMINNUM, VT, Expand); 867 setOperationAction(ISD::FMAXNUM, VT, Expand); 868 setOperationAction(ISD::FMINNAN, VT, Expand); 869 setOperationAction(ISD::FMAXNAN, VT, Expand); 870 setOperationAction(ISD::FMAD, VT, Expand); 871 setOperationAction(ISD::SMIN, VT, Expand); 872 setOperationAction(ISD::SMAX, VT, Expand); 873 setOperationAction(ISD::UMIN, VT, Expand); 874 setOperationAction(ISD::UMAX, VT, Expand); 875 876 // Overflow operations default to expand 877 setOperationAction(ISD::SADDO, VT, Expand); 878 setOperationAction(ISD::SSUBO, VT, Expand); 879 setOperationAction(ISD::UADDO, VT, Expand); 880 setOperationAction(ISD::USUBO, VT, Expand); 881 setOperationAction(ISD::SMULO, VT, Expand); 882 setOperationAction(ISD::UMULO, VT, Expand); 883 884 // These default to Expand so they will be expanded to CTLZ/CTTZ by default. 885 setOperationAction(ISD::CTLZ_ZERO_UNDEF, VT, Expand); 886 setOperationAction(ISD::CTTZ_ZERO_UNDEF, VT, Expand); 887 888 setOperationAction(ISD::BITREVERSE, VT, Expand); 889 890 // These library functions default to expand. 891 setOperationAction(ISD::FROUND, VT, Expand); 892 893 // These operations default to expand for vector types. 894 if (VT.isVector()) { 895 setOperationAction(ISD::FCOPYSIGN, VT, Expand); 896 setOperationAction(ISD::ANY_EXTEND_VECTOR_INREG, VT, Expand); 897 setOperationAction(ISD::SIGN_EXTEND_VECTOR_INREG, VT, Expand); 898 setOperationAction(ISD::ZERO_EXTEND_VECTOR_INREG, VT, Expand); 899 } 900 901 // For most targets @llvm.get.dynamic.area.offest just returns 0. 902 setOperationAction(ISD::GET_DYNAMIC_AREA_OFFSET, VT, Expand); 903 } 904 905 // Most targets ignore the @llvm.prefetch intrinsic. 906 setOperationAction(ISD::PREFETCH, MVT::Other, Expand); 907 908 // Most targets also ignore the @llvm.readcyclecounter intrinsic. 909 setOperationAction(ISD::READCYCLECOUNTER, MVT::i64, Expand); 910 911 // ConstantFP nodes default to expand. Targets can either change this to 912 // Legal, in which case all fp constants are legal, or use isFPImmLegal() 913 // to optimize expansions for certain constants. 914 setOperationAction(ISD::ConstantFP, MVT::f16, Expand); 915 setOperationAction(ISD::ConstantFP, MVT::f32, Expand); 916 setOperationAction(ISD::ConstantFP, MVT::f64, Expand); 917 setOperationAction(ISD::ConstantFP, MVT::f80, Expand); 918 setOperationAction(ISD::ConstantFP, MVT::f128, Expand); 919 920 // These library functions default to expand. 921 for (MVT VT : {MVT::f32, MVT::f64, MVT::f128}) { 922 setOperationAction(ISD::FLOG , VT, Expand); 923 setOperationAction(ISD::FLOG2, VT, Expand); 924 setOperationAction(ISD::FLOG10, VT, Expand); 925 setOperationAction(ISD::FEXP , VT, Expand); 926 setOperationAction(ISD::FEXP2, VT, Expand); 927 setOperationAction(ISD::FFLOOR, VT, Expand); 928 setOperationAction(ISD::FNEARBYINT, VT, Expand); 929 setOperationAction(ISD::FCEIL, VT, Expand); 930 setOperationAction(ISD::FRINT, VT, Expand); 931 setOperationAction(ISD::FTRUNC, VT, Expand); 932 setOperationAction(ISD::FROUND, VT, Expand); 933 } 934 935 // Default ISD::TRAP to expand (which turns it into abort). 936 setOperationAction(ISD::TRAP, MVT::Other, Expand); 937 938 // On most systems, DEBUGTRAP and TRAP have no difference. The "Expand" 939 // here is to inform DAG Legalizer to replace DEBUGTRAP with TRAP. 940 // 941 setOperationAction(ISD::DEBUGTRAP, MVT::Other, Expand); 942 } 943 944 MVT TargetLoweringBase::getScalarShiftAmountTy(const DataLayout &DL, 945 EVT) const { 946 return MVT::getIntegerVT(8 * DL.getPointerSize(0)); 947 } 948 949 EVT TargetLoweringBase::getShiftAmountTy(EVT LHSTy, 950 const DataLayout &DL) const { 951 assert(LHSTy.isInteger() && "Shift amount is not an integer type!"); 952 if (LHSTy.isVector()) 953 return LHSTy; 954 return getScalarShiftAmountTy(DL, LHSTy); 955 } 956 957 /// canOpTrap - Returns true if the operation can trap for the value type. 958 /// VT must be a legal type. 959 bool TargetLoweringBase::canOpTrap(unsigned Op, EVT VT) const { 960 assert(isTypeLegal(VT)); 961 switch (Op) { 962 default: 963 return false; 964 case ISD::FDIV: 965 case ISD::FREM: 966 case ISD::SDIV: 967 case ISD::UDIV: 968 case ISD::SREM: 969 case ISD::UREM: 970 return true; 971 } 972 } 973 974 void TargetLoweringBase::setJumpIsExpensive(bool isExpensive) { 975 // If the command-line option was specified, ignore this request. 976 if (!JumpIsExpensiveOverride.getNumOccurrences()) 977 JumpIsExpensive = isExpensive; 978 } 979 980 TargetLoweringBase::LegalizeKind 981 TargetLoweringBase::getTypeConversion(LLVMContext &Context, EVT VT) const { 982 // If this is a simple type, use the ComputeRegisterProp mechanism. 983 if (VT.isSimple()) { 984 MVT SVT = VT.getSimpleVT(); 985 assert((unsigned)SVT.SimpleTy < array_lengthof(TransformToType)); 986 MVT NVT = TransformToType[SVT.SimpleTy]; 987 LegalizeTypeAction LA = ValueTypeActions.getTypeAction(SVT); 988 989 assert((LA == TypeLegal || LA == TypeSoftenFloat || 990 ValueTypeActions.getTypeAction(NVT) != TypePromoteInteger) && 991 "Promote may not follow Expand or Promote"); 992 993 if (LA == TypeSplitVector) 994 return LegalizeKind(LA, 995 EVT::getVectorVT(Context, SVT.getVectorElementType(), 996 SVT.getVectorNumElements() / 2)); 997 if (LA == TypeScalarizeVector) 998 return LegalizeKind(LA, SVT.getVectorElementType()); 999 return LegalizeKind(LA, NVT); 1000 } 1001 1002 // Handle Extended Scalar Types. 1003 if (!VT.isVector()) { 1004 assert(VT.isInteger() && "Float types must be simple"); 1005 unsigned BitSize = VT.getSizeInBits(); 1006 // First promote to a power-of-two size, then expand if necessary. 1007 if (BitSize < 8 || !isPowerOf2_32(BitSize)) { 1008 EVT NVT = VT.getRoundIntegerType(Context); 1009 assert(NVT != VT && "Unable to round integer VT"); 1010 LegalizeKind NextStep = getTypeConversion(Context, NVT); 1011 // Avoid multi-step promotion. 1012 if (NextStep.first == TypePromoteInteger) 1013 return NextStep; 1014 // Return rounded integer type. 1015 return LegalizeKind(TypePromoteInteger, NVT); 1016 } 1017 1018 return LegalizeKind(TypeExpandInteger, 1019 EVT::getIntegerVT(Context, VT.getSizeInBits() / 2)); 1020 } 1021 1022 // Handle vector types. 1023 unsigned NumElts = VT.getVectorNumElements(); 1024 EVT EltVT = VT.getVectorElementType(); 1025 1026 // Vectors with only one element are always scalarized. 1027 if (NumElts == 1) 1028 return LegalizeKind(TypeScalarizeVector, EltVT); 1029 1030 // Try to widen vector elements until the element type is a power of two and 1031 // promote it to a legal type later on, for example: 1032 // <3 x i8> -> <4 x i8> -> <4 x i32> 1033 if (EltVT.isInteger()) { 1034 // Vectors with a number of elements that is not a power of two are always 1035 // widened, for example <3 x i8> -> <4 x i8>. 1036 if (!VT.isPow2VectorType()) { 1037 NumElts = (unsigned)NextPowerOf2(NumElts); 1038 EVT NVT = EVT::getVectorVT(Context, EltVT, NumElts); 1039 return LegalizeKind(TypeWidenVector, NVT); 1040 } 1041 1042 // Examine the element type. 1043 LegalizeKind LK = getTypeConversion(Context, EltVT); 1044 1045 // If type is to be expanded, split the vector. 1046 // <4 x i140> -> <2 x i140> 1047 if (LK.first == TypeExpandInteger) 1048 return LegalizeKind(TypeSplitVector, 1049 EVT::getVectorVT(Context, EltVT, NumElts / 2)); 1050 1051 // Promote the integer element types until a legal vector type is found 1052 // or until the element integer type is too big. If a legal type was not 1053 // found, fallback to the usual mechanism of widening/splitting the 1054 // vector. 1055 EVT OldEltVT = EltVT; 1056 while (1) { 1057 // Increase the bitwidth of the element to the next pow-of-two 1058 // (which is greater than 8 bits). 1059 EltVT = EVT::getIntegerVT(Context, 1 + EltVT.getSizeInBits()) 1060 .getRoundIntegerType(Context); 1061 1062 // Stop trying when getting a non-simple element type. 1063 // Note that vector elements may be greater than legal vector element 1064 // types. Example: X86 XMM registers hold 64bit element on 32bit 1065 // systems. 1066 if (!EltVT.isSimple()) 1067 break; 1068 1069 // Build a new vector type and check if it is legal. 1070 MVT NVT = MVT::getVectorVT(EltVT.getSimpleVT(), NumElts); 1071 // Found a legal promoted vector type. 1072 if (NVT != MVT() && ValueTypeActions.getTypeAction(NVT) == TypeLegal) 1073 return LegalizeKind(TypePromoteInteger, 1074 EVT::getVectorVT(Context, EltVT, NumElts)); 1075 } 1076 1077 // Reset the type to the unexpanded type if we did not find a legal vector 1078 // type with a promoted vector element type. 1079 EltVT = OldEltVT; 1080 } 1081 1082 // Try to widen the vector until a legal type is found. 1083 // If there is no wider legal type, split the vector. 1084 while (1) { 1085 // Round up to the next power of 2. 1086 NumElts = (unsigned)NextPowerOf2(NumElts); 1087 1088 // If there is no simple vector type with this many elements then there 1089 // cannot be a larger legal vector type. Note that this assumes that 1090 // there are no skipped intermediate vector types in the simple types. 1091 if (!EltVT.isSimple()) 1092 break; 1093 MVT LargerVector = MVT::getVectorVT(EltVT.getSimpleVT(), NumElts); 1094 if (LargerVector == MVT()) 1095 break; 1096 1097 // If this type is legal then widen the vector. 1098 if (ValueTypeActions.getTypeAction(LargerVector) == TypeLegal) 1099 return LegalizeKind(TypeWidenVector, LargerVector); 1100 } 1101 1102 // Widen odd vectors to next power of two. 1103 if (!VT.isPow2VectorType()) { 1104 EVT NVT = VT.getPow2VectorType(Context); 1105 return LegalizeKind(TypeWidenVector, NVT); 1106 } 1107 1108 // Vectors with illegal element types are expanded. 1109 EVT NVT = EVT::getVectorVT(Context, EltVT, VT.getVectorNumElements() / 2); 1110 return LegalizeKind(TypeSplitVector, NVT); 1111 } 1112 1113 static unsigned getVectorTypeBreakdownMVT(MVT VT, MVT &IntermediateVT, 1114 unsigned &NumIntermediates, 1115 MVT &RegisterVT, 1116 TargetLoweringBase *TLI) { 1117 // Figure out the right, legal destination reg to copy into. 1118 unsigned NumElts = VT.getVectorNumElements(); 1119 MVT EltTy = VT.getVectorElementType(); 1120 1121 unsigned NumVectorRegs = 1; 1122 1123 // FIXME: We don't support non-power-of-2-sized vectors for now. Ideally we 1124 // could break down into LHS/RHS like LegalizeDAG does. 1125 if (!isPowerOf2_32(NumElts)) { 1126 NumVectorRegs = NumElts; 1127 NumElts = 1; 1128 } 1129 1130 // Divide the input until we get to a supported size. This will always 1131 // end with a scalar if the target doesn't support vectors. 1132 while (NumElts > 1 && !TLI->isTypeLegal(MVT::getVectorVT(EltTy, NumElts))) { 1133 NumElts >>= 1; 1134 NumVectorRegs <<= 1; 1135 } 1136 1137 NumIntermediates = NumVectorRegs; 1138 1139 MVT NewVT = MVT::getVectorVT(EltTy, NumElts); 1140 if (!TLI->isTypeLegal(NewVT)) 1141 NewVT = EltTy; 1142 IntermediateVT = NewVT; 1143 1144 unsigned NewVTSize = NewVT.getSizeInBits(); 1145 1146 // Convert sizes such as i33 to i64. 1147 if (!isPowerOf2_32(NewVTSize)) 1148 NewVTSize = NextPowerOf2(NewVTSize); 1149 1150 MVT DestVT = TLI->getRegisterType(NewVT); 1151 RegisterVT = DestVT; 1152 if (EVT(DestVT).bitsLT(NewVT)) // Value is expanded, e.g. i64 -> i16. 1153 return NumVectorRegs*(NewVTSize/DestVT.getSizeInBits()); 1154 1155 // Otherwise, promotion or legal types use the same number of registers as 1156 // the vector decimated to the appropriate level. 1157 return NumVectorRegs; 1158 } 1159 1160 /// isLegalRC - Return true if the value types that can be represented by the 1161 /// specified register class are all legal. 1162 bool TargetLoweringBase::isLegalRC(const TargetRegisterClass *RC) const { 1163 for (TargetRegisterClass::vt_iterator I = RC->vt_begin(), E = RC->vt_end(); 1164 I != E; ++I) { 1165 if (isTypeLegal(*I)) 1166 return true; 1167 } 1168 return false; 1169 } 1170 1171 /// Replace/modify any TargetFrameIndex operands with a targte-dependent 1172 /// sequence of memory operands that is recognized by PrologEpilogInserter. 1173 MachineBasicBlock* 1174 TargetLoweringBase::emitPatchPoint(MachineInstr *MI, 1175 MachineBasicBlock *MBB) const { 1176 MachineFunction &MF = *MI->getParent()->getParent(); 1177 MachineFrameInfo &MFI = *MF.getFrameInfo(); 1178 1179 // We're handling multiple types of operands here: 1180 // PATCHPOINT MetaArgs - live-in, read only, direct 1181 // STATEPOINT Deopt Spill - live-through, read only, indirect 1182 // STATEPOINT Deopt Alloca - live-through, read only, direct 1183 // (We're currently conservative and mark the deopt slots read/write in 1184 // practice.) 1185 // STATEPOINT GC Spill - live-through, read/write, indirect 1186 // STATEPOINT GC Alloca - live-through, read/write, direct 1187 // The live-in vs live-through is handled already (the live through ones are 1188 // all stack slots), but we need to handle the different type of stackmap 1189 // operands and memory effects here. 1190 1191 // MI changes inside this loop as we grow operands. 1192 for(unsigned OperIdx = 0; OperIdx != MI->getNumOperands(); ++OperIdx) { 1193 MachineOperand &MO = MI->getOperand(OperIdx); 1194 if (!MO.isFI()) 1195 continue; 1196 1197 // foldMemoryOperand builds a new MI after replacing a single FI operand 1198 // with the canonical set of five x86 addressing-mode operands. 1199 int FI = MO.getIndex(); 1200 MachineInstrBuilder MIB = BuildMI(MF, MI->getDebugLoc(), MI->getDesc()); 1201 1202 // Copy operands before the frame-index. 1203 for (unsigned i = 0; i < OperIdx; ++i) 1204 MIB.addOperand(MI->getOperand(i)); 1205 // Add frame index operands recognized by stackmaps.cpp 1206 if (MFI.isStatepointSpillSlotObjectIndex(FI)) { 1207 // indirect-mem-ref tag, size, #FI, offset. 1208 // Used for spills inserted by StatepointLowering. This codepath is not 1209 // used for patchpoints/stackmaps at all, for these spilling is done via 1210 // foldMemoryOperand callback only. 1211 assert(MI->getOpcode() == TargetOpcode::STATEPOINT && "sanity"); 1212 MIB.addImm(StackMaps::IndirectMemRefOp); 1213 MIB.addImm(MFI.getObjectSize(FI)); 1214 MIB.addOperand(MI->getOperand(OperIdx)); 1215 MIB.addImm(0); 1216 } else { 1217 // direct-mem-ref tag, #FI, offset. 1218 // Used by patchpoint, and direct alloca arguments to statepoints 1219 MIB.addImm(StackMaps::DirectMemRefOp); 1220 MIB.addOperand(MI->getOperand(OperIdx)); 1221 MIB.addImm(0); 1222 } 1223 // Copy the operands after the frame index. 1224 for (unsigned i = OperIdx + 1; i != MI->getNumOperands(); ++i) 1225 MIB.addOperand(MI->getOperand(i)); 1226 1227 // Inherit previous memory operands. 1228 MIB->setMemRefs(MI->memoperands_begin(), MI->memoperands_end()); 1229 assert(MIB->mayLoad() && "Folded a stackmap use to a non-load!"); 1230 1231 // Add a new memory operand for this FI. 1232 assert(MFI.getObjectOffset(FI) != -1); 1233 1234 unsigned Flags = MachineMemOperand::MOLoad; 1235 if (MI->getOpcode() == TargetOpcode::STATEPOINT) { 1236 Flags |= MachineMemOperand::MOStore; 1237 Flags |= MachineMemOperand::MOVolatile; 1238 } 1239 MachineMemOperand *MMO = MF.getMachineMemOperand( 1240 MachinePointerInfo::getFixedStack(MF, FI), Flags, 1241 MF.getDataLayout().getPointerSize(), MFI.getObjectAlignment(FI)); 1242 MIB->addMemOperand(MF, MMO); 1243 1244 // Replace the instruction and update the operand index. 1245 MBB->insert(MachineBasicBlock::iterator(MI), MIB); 1246 OperIdx += (MIB->getNumOperands() - MI->getNumOperands()) - 1; 1247 MI->eraseFromParent(); 1248 MI = MIB; 1249 } 1250 return MBB; 1251 } 1252 1253 /// findRepresentativeClass - Return the largest legal super-reg register class 1254 /// of the register class for the specified type and its associated "cost". 1255 // This function is in TargetLowering because it uses RegClassForVT which would 1256 // need to be moved to TargetRegisterInfo and would necessitate moving 1257 // isTypeLegal over as well - a massive change that would just require 1258 // TargetLowering having a TargetRegisterInfo class member that it would use. 1259 std::pair<const TargetRegisterClass *, uint8_t> 1260 TargetLoweringBase::findRepresentativeClass(const TargetRegisterInfo *TRI, 1261 MVT VT) const { 1262 const TargetRegisterClass *RC = RegClassForVT[VT.SimpleTy]; 1263 if (!RC) 1264 return std::make_pair(RC, 0); 1265 1266 // Compute the set of all super-register classes. 1267 BitVector SuperRegRC(TRI->getNumRegClasses()); 1268 for (SuperRegClassIterator RCI(RC, TRI); RCI.isValid(); ++RCI) 1269 SuperRegRC.setBitsInMask(RCI.getMask()); 1270 1271 // Find the first legal register class with the largest spill size. 1272 const TargetRegisterClass *BestRC = RC; 1273 for (int i = SuperRegRC.find_first(); i >= 0; i = SuperRegRC.find_next(i)) { 1274 const TargetRegisterClass *SuperRC = TRI->getRegClass(i); 1275 // We want the largest possible spill size. 1276 if (SuperRC->getSize() <= BestRC->getSize()) 1277 continue; 1278 if (!isLegalRC(SuperRC)) 1279 continue; 1280 BestRC = SuperRC; 1281 } 1282 return std::make_pair(BestRC, 1); 1283 } 1284 1285 /// computeRegisterProperties - Once all of the register classes are added, 1286 /// this allows us to compute derived properties we expose. 1287 void TargetLoweringBase::computeRegisterProperties( 1288 const TargetRegisterInfo *TRI) { 1289 static_assert(MVT::LAST_VALUETYPE <= MVT::MAX_ALLOWED_VALUETYPE, 1290 "Too many value types for ValueTypeActions to hold!"); 1291 1292 // Everything defaults to needing one register. 1293 for (unsigned i = 0; i != MVT::LAST_VALUETYPE; ++i) { 1294 NumRegistersForVT[i] = 1; 1295 RegisterTypeForVT[i] = TransformToType[i] = (MVT::SimpleValueType)i; 1296 } 1297 // ...except isVoid, which doesn't need any registers. 1298 NumRegistersForVT[MVT::isVoid] = 0; 1299 1300 // Find the largest integer register class. 1301 unsigned LargestIntReg = MVT::LAST_INTEGER_VALUETYPE; 1302 for (; RegClassForVT[LargestIntReg] == nullptr; --LargestIntReg) 1303 assert(LargestIntReg != MVT::i1 && "No integer registers defined!"); 1304 1305 // Every integer value type larger than this largest register takes twice as 1306 // many registers to represent as the previous ValueType. 1307 for (unsigned ExpandedReg = LargestIntReg + 1; 1308 ExpandedReg <= MVT::LAST_INTEGER_VALUETYPE; ++ExpandedReg) { 1309 NumRegistersForVT[ExpandedReg] = 2*NumRegistersForVT[ExpandedReg-1]; 1310 RegisterTypeForVT[ExpandedReg] = (MVT::SimpleValueType)LargestIntReg; 1311 TransformToType[ExpandedReg] = (MVT::SimpleValueType)(ExpandedReg - 1); 1312 ValueTypeActions.setTypeAction((MVT::SimpleValueType)ExpandedReg, 1313 TypeExpandInteger); 1314 } 1315 1316 // Inspect all of the ValueType's smaller than the largest integer 1317 // register to see which ones need promotion. 1318 unsigned LegalIntReg = LargestIntReg; 1319 for (unsigned IntReg = LargestIntReg - 1; 1320 IntReg >= (unsigned)MVT::i1; --IntReg) { 1321 MVT IVT = (MVT::SimpleValueType)IntReg; 1322 if (isTypeLegal(IVT)) { 1323 LegalIntReg = IntReg; 1324 } else { 1325 RegisterTypeForVT[IntReg] = TransformToType[IntReg] = 1326 (const MVT::SimpleValueType)LegalIntReg; 1327 ValueTypeActions.setTypeAction(IVT, TypePromoteInteger); 1328 } 1329 } 1330 1331 // ppcf128 type is really two f64's. 1332 if (!isTypeLegal(MVT::ppcf128)) { 1333 if (isTypeLegal(MVT::f64)) { 1334 NumRegistersForVT[MVT::ppcf128] = 2*NumRegistersForVT[MVT::f64]; 1335 RegisterTypeForVT[MVT::ppcf128] = MVT::f64; 1336 TransformToType[MVT::ppcf128] = MVT::f64; 1337 ValueTypeActions.setTypeAction(MVT::ppcf128, TypeExpandFloat); 1338 } else { 1339 NumRegistersForVT[MVT::ppcf128] = NumRegistersForVT[MVT::i128]; 1340 RegisterTypeForVT[MVT::ppcf128] = RegisterTypeForVT[MVT::i128]; 1341 TransformToType[MVT::ppcf128] = MVT::i128; 1342 ValueTypeActions.setTypeAction(MVT::ppcf128, TypeSoftenFloat); 1343 } 1344 } 1345 1346 // Decide how to handle f128. If the target does not have native f128 support, 1347 // expand it to i128 and we will be generating soft float library calls. 1348 if (!isTypeLegal(MVT::f128)) { 1349 NumRegistersForVT[MVT::f128] = NumRegistersForVT[MVT::i128]; 1350 RegisterTypeForVT[MVT::f128] = RegisterTypeForVT[MVT::i128]; 1351 TransformToType[MVT::f128] = MVT::i128; 1352 ValueTypeActions.setTypeAction(MVT::f128, TypeSoftenFloat); 1353 } 1354 1355 // Decide how to handle f64. If the target does not have native f64 support, 1356 // expand it to i64 and we will be generating soft float library calls. 1357 if (!isTypeLegal(MVT::f64)) { 1358 NumRegistersForVT[MVT::f64] = NumRegistersForVT[MVT::i64]; 1359 RegisterTypeForVT[MVT::f64] = RegisterTypeForVT[MVT::i64]; 1360 TransformToType[MVT::f64] = MVT::i64; 1361 ValueTypeActions.setTypeAction(MVT::f64, TypeSoftenFloat); 1362 } 1363 1364 // Decide how to handle f32. If the target does not have native f32 support, 1365 // expand it to i32 and we will be generating soft float library calls. 1366 if (!isTypeLegal(MVT::f32)) { 1367 NumRegistersForVT[MVT::f32] = NumRegistersForVT[MVT::i32]; 1368 RegisterTypeForVT[MVT::f32] = RegisterTypeForVT[MVT::i32]; 1369 TransformToType[MVT::f32] = MVT::i32; 1370 ValueTypeActions.setTypeAction(MVT::f32, TypeSoftenFloat); 1371 } 1372 1373 // Decide how to handle f16. If the target does not have native f16 support, 1374 // promote it to f32, because there are no f16 library calls (except for 1375 // conversions). 1376 if (!isTypeLegal(MVT::f16)) { 1377 NumRegistersForVT[MVT::f16] = NumRegistersForVT[MVT::f32]; 1378 RegisterTypeForVT[MVT::f16] = RegisterTypeForVT[MVT::f32]; 1379 TransformToType[MVT::f16] = MVT::f32; 1380 ValueTypeActions.setTypeAction(MVT::f16, TypePromoteFloat); 1381 } 1382 1383 // Loop over all of the vector value types to see which need transformations. 1384 for (unsigned i = MVT::FIRST_VECTOR_VALUETYPE; 1385 i <= (unsigned)MVT::LAST_VECTOR_VALUETYPE; ++i) { 1386 MVT VT = (MVT::SimpleValueType) i; 1387 if (isTypeLegal(VT)) 1388 continue; 1389 1390 MVT EltVT = VT.getVectorElementType(); 1391 unsigned NElts = VT.getVectorNumElements(); 1392 bool IsLegalWiderType = false; 1393 LegalizeTypeAction PreferredAction = getPreferredVectorAction(VT); 1394 switch (PreferredAction) { 1395 case TypePromoteInteger: { 1396 // Try to promote the elements of integer vectors. If no legal 1397 // promotion was found, fall through to the widen-vector method. 1398 for (unsigned nVT = i + 1; nVT <= MVT::LAST_INTEGER_VECTOR_VALUETYPE; ++nVT) { 1399 MVT SVT = (MVT::SimpleValueType) nVT; 1400 // Promote vectors of integers to vectors with the same number 1401 // of elements, with a wider element type. 1402 if (SVT.getVectorElementType().getSizeInBits() > EltVT.getSizeInBits() && 1403 SVT.getVectorNumElements() == NElts && isTypeLegal(SVT)) { 1404 TransformToType[i] = SVT; 1405 RegisterTypeForVT[i] = SVT; 1406 NumRegistersForVT[i] = 1; 1407 ValueTypeActions.setTypeAction(VT, TypePromoteInteger); 1408 IsLegalWiderType = true; 1409 break; 1410 } 1411 } 1412 if (IsLegalWiderType) 1413 break; 1414 } 1415 case TypeWidenVector: { 1416 // Try to widen the vector. 1417 for (unsigned nVT = i + 1; nVT <= MVT::LAST_VECTOR_VALUETYPE; ++nVT) { 1418 MVT SVT = (MVT::SimpleValueType) nVT; 1419 if (SVT.getVectorElementType() == EltVT 1420 && SVT.getVectorNumElements() > NElts && isTypeLegal(SVT)) { 1421 TransformToType[i] = SVT; 1422 RegisterTypeForVT[i] = SVT; 1423 NumRegistersForVT[i] = 1; 1424 ValueTypeActions.setTypeAction(VT, TypeWidenVector); 1425 IsLegalWiderType = true; 1426 break; 1427 } 1428 } 1429 if (IsLegalWiderType) 1430 break; 1431 } 1432 case TypeSplitVector: 1433 case TypeScalarizeVector: { 1434 MVT IntermediateVT; 1435 MVT RegisterVT; 1436 unsigned NumIntermediates; 1437 NumRegistersForVT[i] = getVectorTypeBreakdownMVT(VT, IntermediateVT, 1438 NumIntermediates, RegisterVT, this); 1439 RegisterTypeForVT[i] = RegisterVT; 1440 1441 MVT NVT = VT.getPow2VectorType(); 1442 if (NVT == VT) { 1443 // Type is already a power of 2. The default action is to split. 1444 TransformToType[i] = MVT::Other; 1445 if (PreferredAction == TypeScalarizeVector) 1446 ValueTypeActions.setTypeAction(VT, TypeScalarizeVector); 1447 else if (PreferredAction == TypeSplitVector) 1448 ValueTypeActions.setTypeAction(VT, TypeSplitVector); 1449 else 1450 // Set type action according to the number of elements. 1451 ValueTypeActions.setTypeAction(VT, NElts == 1 ? TypeScalarizeVector 1452 : TypeSplitVector); 1453 } else { 1454 TransformToType[i] = NVT; 1455 ValueTypeActions.setTypeAction(VT, TypeWidenVector); 1456 } 1457 break; 1458 } 1459 default: 1460 llvm_unreachable("Unknown vector legalization action!"); 1461 } 1462 } 1463 1464 // Determine the 'representative' register class for each value type. 1465 // An representative register class is the largest (meaning one which is 1466 // not a sub-register class / subreg register class) legal register class for 1467 // a group of value types. For example, on i386, i8, i16, and i32 1468 // representative would be GR32; while on x86_64 it's GR64. 1469 for (unsigned i = 0; i != MVT::LAST_VALUETYPE; ++i) { 1470 const TargetRegisterClass* RRC; 1471 uint8_t Cost; 1472 std::tie(RRC, Cost) = findRepresentativeClass(TRI, (MVT::SimpleValueType)i); 1473 RepRegClassForVT[i] = RRC; 1474 RepRegClassCostForVT[i] = Cost; 1475 } 1476 } 1477 1478 EVT TargetLoweringBase::getSetCCResultType(const DataLayout &DL, LLVMContext &, 1479 EVT VT) const { 1480 assert(!VT.isVector() && "No default SetCC type for vectors!"); 1481 return getPointerTy(DL).SimpleTy; 1482 } 1483 1484 MVT::SimpleValueType TargetLoweringBase::getCmpLibcallReturnType() const { 1485 return MVT::i32; // return the default value 1486 } 1487 1488 /// getVectorTypeBreakdown - Vector types are broken down into some number of 1489 /// legal first class types. For example, MVT::v8f32 maps to 2 MVT::v4f32 1490 /// with Altivec or SSE1, or 8 promoted MVT::f64 values with the X86 FP stack. 1491 /// Similarly, MVT::v2i64 turns into 4 MVT::i32 values with both PPC and X86. 1492 /// 1493 /// This method returns the number of registers needed, and the VT for each 1494 /// register. It also returns the VT and quantity of the intermediate values 1495 /// before they are promoted/expanded. 1496 /// 1497 unsigned TargetLoweringBase::getVectorTypeBreakdown(LLVMContext &Context, EVT VT, 1498 EVT &IntermediateVT, 1499 unsigned &NumIntermediates, 1500 MVT &RegisterVT) const { 1501 unsigned NumElts = VT.getVectorNumElements(); 1502 1503 // If there is a wider vector type with the same element type as this one, 1504 // or a promoted vector type that has the same number of elements which 1505 // are wider, then we should convert to that legal vector type. 1506 // This handles things like <2 x float> -> <4 x float> and 1507 // <4 x i1> -> <4 x i32>. 1508 LegalizeTypeAction TA = getTypeAction(Context, VT); 1509 if (NumElts != 1 && (TA == TypeWidenVector || TA == TypePromoteInteger)) { 1510 EVT RegisterEVT = getTypeToTransformTo(Context, VT); 1511 if (isTypeLegal(RegisterEVT)) { 1512 IntermediateVT = RegisterEVT; 1513 RegisterVT = RegisterEVT.getSimpleVT(); 1514 NumIntermediates = 1; 1515 return 1; 1516 } 1517 } 1518 1519 // Figure out the right, legal destination reg to copy into. 1520 EVT EltTy = VT.getVectorElementType(); 1521 1522 unsigned NumVectorRegs = 1; 1523 1524 // FIXME: We don't support non-power-of-2-sized vectors for now. Ideally we 1525 // could break down into LHS/RHS like LegalizeDAG does. 1526 if (!isPowerOf2_32(NumElts)) { 1527 NumVectorRegs = NumElts; 1528 NumElts = 1; 1529 } 1530 1531 // Divide the input until we get to a supported size. This will always 1532 // end with a scalar if the target doesn't support vectors. 1533 while (NumElts > 1 && !isTypeLegal( 1534 EVT::getVectorVT(Context, EltTy, NumElts))) { 1535 NumElts >>= 1; 1536 NumVectorRegs <<= 1; 1537 } 1538 1539 NumIntermediates = NumVectorRegs; 1540 1541 EVT NewVT = EVT::getVectorVT(Context, EltTy, NumElts); 1542 if (!isTypeLegal(NewVT)) 1543 NewVT = EltTy; 1544 IntermediateVT = NewVT; 1545 1546 MVT DestVT = getRegisterType(Context, NewVT); 1547 RegisterVT = DestVT; 1548 unsigned NewVTSize = NewVT.getSizeInBits(); 1549 1550 // Convert sizes such as i33 to i64. 1551 if (!isPowerOf2_32(NewVTSize)) 1552 NewVTSize = NextPowerOf2(NewVTSize); 1553 1554 if (EVT(DestVT).bitsLT(NewVT)) // Value is expanded, e.g. i64 -> i16. 1555 return NumVectorRegs*(NewVTSize/DestVT.getSizeInBits()); 1556 1557 // Otherwise, promotion or legal types use the same number of registers as 1558 // the vector decimated to the appropriate level. 1559 return NumVectorRegs; 1560 } 1561 1562 /// Get the EVTs and ArgFlags collections that represent the legalized return 1563 /// type of the given function. This does not require a DAG or a return value, 1564 /// and is suitable for use before any DAGs for the function are constructed. 1565 /// TODO: Move this out of TargetLowering.cpp. 1566 void llvm::GetReturnInfo(Type *ReturnType, AttributeSet attr, 1567 SmallVectorImpl<ISD::OutputArg> &Outs, 1568 const TargetLowering &TLI, const DataLayout &DL) { 1569 SmallVector<EVT, 4> ValueVTs; 1570 ComputeValueVTs(TLI, DL, ReturnType, ValueVTs); 1571 unsigned NumValues = ValueVTs.size(); 1572 if (NumValues == 0) return; 1573 1574 for (unsigned j = 0, f = NumValues; j != f; ++j) { 1575 EVT VT = ValueVTs[j]; 1576 ISD::NodeType ExtendKind = ISD::ANY_EXTEND; 1577 1578 if (attr.hasAttribute(AttributeSet::ReturnIndex, Attribute::SExt)) 1579 ExtendKind = ISD::SIGN_EXTEND; 1580 else if (attr.hasAttribute(AttributeSet::ReturnIndex, Attribute::ZExt)) 1581 ExtendKind = ISD::ZERO_EXTEND; 1582 1583 // FIXME: C calling convention requires the return type to be promoted to 1584 // at least 32-bit. But this is not necessary for non-C calling 1585 // conventions. The frontend should mark functions whose return values 1586 // require promoting with signext or zeroext attributes. 1587 if (ExtendKind != ISD::ANY_EXTEND && VT.isInteger()) { 1588 MVT MinVT = TLI.getRegisterType(ReturnType->getContext(), MVT::i32); 1589 if (VT.bitsLT(MinVT)) 1590 VT = MinVT; 1591 } 1592 1593 unsigned NumParts = TLI.getNumRegisters(ReturnType->getContext(), VT); 1594 MVT PartVT = TLI.getRegisterType(ReturnType->getContext(), VT); 1595 1596 // 'inreg' on function refers to return value 1597 ISD::ArgFlagsTy Flags = ISD::ArgFlagsTy(); 1598 if (attr.hasAttribute(AttributeSet::ReturnIndex, Attribute::InReg)) 1599 Flags.setInReg(); 1600 1601 // Propagate extension type if any 1602 if (attr.hasAttribute(AttributeSet::ReturnIndex, Attribute::SExt)) 1603 Flags.setSExt(); 1604 else if (attr.hasAttribute(AttributeSet::ReturnIndex, Attribute::ZExt)) 1605 Flags.setZExt(); 1606 1607 for (unsigned i = 0; i < NumParts; ++i) 1608 Outs.push_back(ISD::OutputArg(Flags, PartVT, VT, /*isFixed=*/true, 0, 0)); 1609 } 1610 } 1611 1612 /// getByValTypeAlignment - Return the desired alignment for ByVal aggregate 1613 /// function arguments in the caller parameter area. This is the actual 1614 /// alignment, not its logarithm. 1615 unsigned TargetLoweringBase::getByValTypeAlignment(Type *Ty, 1616 const DataLayout &DL) const { 1617 return DL.getABITypeAlignment(Ty); 1618 } 1619 1620 bool TargetLoweringBase::allowsMemoryAccess(LLVMContext &Context, 1621 const DataLayout &DL, EVT VT, 1622 unsigned AddrSpace, 1623 unsigned Alignment, 1624 bool *Fast) const { 1625 // Check if the specified alignment is sufficient based on the data layout. 1626 // TODO: While using the data layout works in practice, a better solution 1627 // would be to implement this check directly (make this a virtual function). 1628 // For example, the ABI alignment may change based on software platform while 1629 // this function should only be affected by hardware implementation. 1630 Type *Ty = VT.getTypeForEVT(Context); 1631 if (Alignment >= DL.getABITypeAlignment(Ty)) { 1632 // Assume that an access that meets the ABI-specified alignment is fast. 1633 if (Fast != nullptr) 1634 *Fast = true; 1635 return true; 1636 } 1637 1638 // This is a misaligned access. 1639 return allowsMisalignedMemoryAccesses(VT, AddrSpace, Alignment, Fast); 1640 } 1641 1642 BranchProbability TargetLoweringBase::getPredictableBranchThreshold() const { 1643 return BranchProbability(MinPercentageForPredictableBranch, 100); 1644 } 1645 1646 //===----------------------------------------------------------------------===// 1647 // TargetTransformInfo Helpers 1648 //===----------------------------------------------------------------------===// 1649 1650 int TargetLoweringBase::InstructionOpcodeToISD(unsigned Opcode) const { 1651 enum InstructionOpcodes { 1652 #define HANDLE_INST(NUM, OPCODE, CLASS) OPCODE = NUM, 1653 #define LAST_OTHER_INST(NUM) InstructionOpcodesCount = NUM 1654 #include "llvm/IR/Instruction.def" 1655 }; 1656 switch (static_cast<InstructionOpcodes>(Opcode)) { 1657 case Ret: return 0; 1658 case Br: return 0; 1659 case Switch: return 0; 1660 case IndirectBr: return 0; 1661 case Invoke: return 0; 1662 case Resume: return 0; 1663 case Unreachable: return 0; 1664 case CleanupRet: return 0; 1665 case CatchRet: return 0; 1666 case CatchPad: return 0; 1667 case CatchSwitch: return 0; 1668 case CleanupPad: return 0; 1669 case Add: return ISD::ADD; 1670 case FAdd: return ISD::FADD; 1671 case Sub: return ISD::SUB; 1672 case FSub: return ISD::FSUB; 1673 case Mul: return ISD::MUL; 1674 case FMul: return ISD::FMUL; 1675 case UDiv: return ISD::UDIV; 1676 case SDiv: return ISD::SDIV; 1677 case FDiv: return ISD::FDIV; 1678 case URem: return ISD::UREM; 1679 case SRem: return ISD::SREM; 1680 case FRem: return ISD::FREM; 1681 case Shl: return ISD::SHL; 1682 case LShr: return ISD::SRL; 1683 case AShr: return ISD::SRA; 1684 case And: return ISD::AND; 1685 case Or: return ISD::OR; 1686 case Xor: return ISD::XOR; 1687 case Alloca: return 0; 1688 case Load: return ISD::LOAD; 1689 case Store: return ISD::STORE; 1690 case GetElementPtr: return 0; 1691 case Fence: return 0; 1692 case AtomicCmpXchg: return 0; 1693 case AtomicRMW: return 0; 1694 case Trunc: return ISD::TRUNCATE; 1695 case ZExt: return ISD::ZERO_EXTEND; 1696 case SExt: return ISD::SIGN_EXTEND; 1697 case FPToUI: return ISD::FP_TO_UINT; 1698 case FPToSI: return ISD::FP_TO_SINT; 1699 case UIToFP: return ISD::UINT_TO_FP; 1700 case SIToFP: return ISD::SINT_TO_FP; 1701 case FPTrunc: return ISD::FP_ROUND; 1702 case FPExt: return ISD::FP_EXTEND; 1703 case PtrToInt: return ISD::BITCAST; 1704 case IntToPtr: return ISD::BITCAST; 1705 case BitCast: return ISD::BITCAST; 1706 case AddrSpaceCast: return ISD::ADDRSPACECAST; 1707 case ICmp: return ISD::SETCC; 1708 case FCmp: return ISD::SETCC; 1709 case PHI: return 0; 1710 case Call: return 0; 1711 case Select: return ISD::SELECT; 1712 case UserOp1: return 0; 1713 case UserOp2: return 0; 1714 case VAArg: return 0; 1715 case ExtractElement: return ISD::EXTRACT_VECTOR_ELT; 1716 case InsertElement: return ISD::INSERT_VECTOR_ELT; 1717 case ShuffleVector: return ISD::VECTOR_SHUFFLE; 1718 case ExtractValue: return ISD::MERGE_VALUES; 1719 case InsertValue: return ISD::MERGE_VALUES; 1720 case LandingPad: return 0; 1721 } 1722 1723 llvm_unreachable("Unknown instruction type encountered!"); 1724 } 1725 1726 std::pair<int, MVT> 1727 TargetLoweringBase::getTypeLegalizationCost(const DataLayout &DL, 1728 Type *Ty) const { 1729 LLVMContext &C = Ty->getContext(); 1730 EVT MTy = getValueType(DL, Ty); 1731 1732 int Cost = 1; 1733 // We keep legalizing the type until we find a legal kind. We assume that 1734 // the only operation that costs anything is the split. After splitting 1735 // we need to handle two types. 1736 while (true) { 1737 LegalizeKind LK = getTypeConversion(C, MTy); 1738 1739 if (LK.first == TypeLegal) 1740 return std::make_pair(Cost, MTy.getSimpleVT()); 1741 1742 if (LK.first == TypeSplitVector || LK.first == TypeExpandInteger) 1743 Cost *= 2; 1744 1745 // Do not loop with f128 type. 1746 if (MTy == LK.second) 1747 return std::make_pair(Cost, MTy.getSimpleVT()); 1748 1749 // Keep legalizing the type. 1750 MTy = LK.second; 1751 } 1752 } 1753 1754 Value *TargetLoweringBase::getSafeStackPointerLocation(IRBuilder<> &IRB) const { 1755 if (!TM.getTargetTriple().isAndroid()) 1756 return nullptr; 1757 1758 // Android provides a libc function to retrieve the address of the current 1759 // thread's unsafe stack pointer. 1760 Module *M = IRB.GetInsertBlock()->getParent()->getParent(); 1761 Type *StackPtrTy = Type::getInt8PtrTy(M->getContext()); 1762 Value *Fn = M->getOrInsertFunction("__safestack_pointer_address", 1763 StackPtrTy->getPointerTo(0), nullptr); 1764 return IRB.CreateCall(Fn); 1765 } 1766 1767 //===----------------------------------------------------------------------===// 1768 // Loop Strength Reduction hooks 1769 //===----------------------------------------------------------------------===// 1770 1771 /// isLegalAddressingMode - Return true if the addressing mode represented 1772 /// by AM is legal for this target, for a load/store of the specified type. 1773 bool TargetLoweringBase::isLegalAddressingMode(const DataLayout &DL, 1774 const AddrMode &AM, Type *Ty, 1775 unsigned AS) const { 1776 // The default implementation of this implements a conservative RISCy, r+r and 1777 // r+i addr mode. 1778 1779 // Allows a sign-extended 16-bit immediate field. 1780 if (AM.BaseOffs <= -(1LL << 16) || AM.BaseOffs >= (1LL << 16)-1) 1781 return false; 1782 1783 // No global is ever allowed as a base. 1784 if (AM.BaseGV) 1785 return false; 1786 1787 // Only support r+r, 1788 switch (AM.Scale) { 1789 case 0: // "r+i" or just "i", depending on HasBaseReg. 1790 break; 1791 case 1: 1792 if (AM.HasBaseReg && AM.BaseOffs) // "r+r+i" is not allowed. 1793 return false; 1794 // Otherwise we have r+r or r+i. 1795 break; 1796 case 2: 1797 if (AM.HasBaseReg || AM.BaseOffs) // 2*r+r or 2*r+i is not allowed. 1798 return false; 1799 // Allow 2*r as r+r. 1800 break; 1801 default: // Don't allow n * r 1802 return false; 1803 } 1804 1805 return true; 1806 } 1807 1808 //===----------------------------------------------------------------------===// 1809 // Stack Protector 1810 //===----------------------------------------------------------------------===// 1811 1812 // For OpenBSD return its special guard variable. Otherwise return nullptr, 1813 // so that SelectionDAG handle SSP. 1814 Value *TargetLoweringBase::getIRStackGuard(IRBuilder<> &IRB) const { 1815 if (getTargetMachine().getTargetTriple().isOSOpenBSD()) { 1816 Module &M = *IRB.GetInsertBlock()->getParent()->getParent(); 1817 PointerType *PtrTy = Type::getInt8PtrTy(M.getContext()); 1818 auto Guard = cast<GlobalValue>(M.getOrInsertGlobal("__guard_local", PtrTy)); 1819 Guard->setVisibility(GlobalValue::HiddenVisibility); 1820 return Guard; 1821 } 1822 return nullptr; 1823 } 1824 1825 // Currently only support "standard" __stack_chk_guard. 1826 // TODO: add LOAD_STACK_GUARD support. 1827 void TargetLoweringBase::insertSSPDeclarations(Module &M) const { 1828 M.getOrInsertGlobal("__stack_chk_guard", Type::getInt8PtrTy(M.getContext())); 1829 } 1830 1831 // Currently only support "standard" __stack_chk_guard. 1832 // TODO: add LOAD_STACK_GUARD support. 1833 Value *TargetLoweringBase::getSDagStackGuard(const Module &M) const { 1834 return M.getGlobalVariable("__stack_chk_guard"); 1835 } 1836