TargetLowering.cpp revision c698d3a2a40f0909d16cbe857685f0f22cb9ae43
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/ADT/BitVector.h" 16#include "llvm/ADT/STLExtras.h" 17#include "llvm/CodeGen/Analysis.h" 18#include "llvm/CodeGen/MachineFrameInfo.h" 19#include "llvm/CodeGen/MachineFunction.h" 20#include "llvm/CodeGen/MachineJumpTableInfo.h" 21#include "llvm/CodeGen/SelectionDAG.h" 22#include "llvm/DataLayout.h" 23#include "llvm/DerivedTypes.h" 24#include "llvm/GlobalVariable.h" 25#include "llvm/MC/MCAsmInfo.h" 26#include "llvm/MC/MCExpr.h" 27#include "llvm/Support/CommandLine.h" 28#include "llvm/Support/ErrorHandling.h" 29#include "llvm/Support/MathExtras.h" 30#include "llvm/Target/TargetLoweringObjectFile.h" 31#include "llvm/Target/TargetMachine.h" 32#include "llvm/Target/TargetRegisterInfo.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 if (RetVT == MVT::f64) 432 return SINTTOFP_I32_F64; 433 if (RetVT == MVT::f80) 434 return SINTTOFP_I32_F80; 435 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 if (RetVT == MVT::f64) 441 return SINTTOFP_I64_F64; 442 if (RetVT == MVT::f80) 443 return SINTTOFP_I64_F80; 444 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 if (RetVT == MVT::f64) 450 return SINTTOFP_I128_F64; 451 if (RetVT == MVT::f80) 452 return SINTTOFP_I128_F80; 453 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 if (RetVT == MVT::f64) 466 return UINTTOFP_I32_F64; 467 if (RetVT == MVT::f80) 468 return UINTTOFP_I32_F80; 469 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 if (RetVT == MVT::f64) 475 return UINTTOFP_I64_F64; 476 if (RetVT == MVT::f80) 477 return UINTTOFP_I64_F80; 478 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 if (RetVT == MVT::f64) 484 return UINTTOFP_I128_F64; 485 if (RetVT == MVT::f80) 486 return UINTTOFP_I128_F80; 487 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.getDataLayout()), 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 // On most systems, DEBUGTRAP and TRAP have no difference. The "Expand" 587 // here is to inform DAG Legalizer to replace DEBUGTRAP with TRAP. 588 // 589 setOperationAction(ISD::DEBUGTRAP, MVT::Other, Expand); 590 591 IsLittleEndian = TD->isLittleEndian(); 592 PointerTy = MVT::getIntegerVT(8*TD->getPointerSize(0)); 593 memset(RegClassForVT, 0,MVT::LAST_VALUETYPE*sizeof(TargetRegisterClass*)); 594 memset(TargetDAGCombineArray, 0, array_lengthof(TargetDAGCombineArray)); 595 maxStoresPerMemset = maxStoresPerMemcpy = maxStoresPerMemmove = 8; 596 maxStoresPerMemsetOptSize = maxStoresPerMemcpyOptSize 597 = maxStoresPerMemmoveOptSize = 4; 598 benefitFromCodePlacementOpt = false; 599 UseUnderscoreSetJmp = false; 600 UseUnderscoreLongJmp = false; 601 SelectIsExpensive = false; 602 IntDivIsCheap = false; 603 Pow2DivIsCheap = false; 604 JumpIsExpensive = false; 605 predictableSelectIsExpensive = false; 606 StackPointerRegisterToSaveRestore = 0; 607 ExceptionPointerRegister = 0; 608 ExceptionSelectorRegister = 0; 609 BooleanContents = UndefinedBooleanContent; 610 BooleanVectorContents = UndefinedBooleanContent; 611 SchedPreferenceInfo = Sched::ILP; 612 JumpBufSize = 0; 613 JumpBufAlignment = 0; 614 MinFunctionAlignment = 0; 615 PrefFunctionAlignment = 0; 616 PrefLoopAlignment = 0; 617 MinStackArgumentAlignment = 1; 618 ShouldFoldAtomicFences = false; 619 InsertFencesForAtomic = false; 620 SupportJumpTables = true; 621 MinimumJumpTableEntries = 4; 622 623 InitLibcallNames(LibcallRoutineNames); 624 InitCmpLibcallCCs(CmpLibcallCCs); 625 InitLibcallCallingConvs(LibcallCallingConvs); 626} 627 628TargetLowering::~TargetLowering() { 629 delete &TLOF; 630} 631 632MVT TargetLowering::getShiftAmountTy(EVT LHSTy) const { 633 return MVT::getIntegerVT(8*TD->getPointerSize(0)); 634} 635 636/// canOpTrap - Returns true if the operation can trap for the value type. 637/// VT must be a legal type. 638bool TargetLowering::canOpTrap(unsigned Op, EVT VT) const { 639 assert(isTypeLegal(VT)); 640 switch (Op) { 641 default: 642 return false; 643 case ISD::FDIV: 644 case ISD::FREM: 645 case ISD::SDIV: 646 case ISD::UDIV: 647 case ISD::SREM: 648 case ISD::UREM: 649 return true; 650 } 651} 652 653 654static unsigned getVectorTypeBreakdownMVT(MVT VT, MVT &IntermediateVT, 655 unsigned &NumIntermediates, 656 MVT &RegisterVT, 657 TargetLowering *TLI) { 658 // Figure out the right, legal destination reg to copy into. 659 unsigned NumElts = VT.getVectorNumElements(); 660 MVT EltTy = VT.getVectorElementType(); 661 662 unsigned NumVectorRegs = 1; 663 664 // FIXME: We don't support non-power-of-2-sized vectors for now. Ideally we 665 // could break down into LHS/RHS like LegalizeDAG does. 666 if (!isPowerOf2_32(NumElts)) { 667 NumVectorRegs = NumElts; 668 NumElts = 1; 669 } 670 671 // Divide the input until we get to a supported size. This will always 672 // end with a scalar if the target doesn't support vectors. 673 while (NumElts > 1 && !TLI->isTypeLegal(MVT::getVectorVT(EltTy, NumElts))) { 674 NumElts >>= 1; 675 NumVectorRegs <<= 1; 676 } 677 678 NumIntermediates = NumVectorRegs; 679 680 MVT NewVT = MVT::getVectorVT(EltTy, NumElts); 681 if (!TLI->isTypeLegal(NewVT)) 682 NewVT = EltTy; 683 IntermediateVT = NewVT; 684 685 unsigned NewVTSize = NewVT.getSizeInBits(); 686 687 // Convert sizes such as i33 to i64. 688 if (!isPowerOf2_32(NewVTSize)) 689 NewVTSize = NextPowerOf2(NewVTSize); 690 691 MVT DestVT = TLI->getRegisterType(NewVT); 692 RegisterVT = DestVT; 693 if (EVT(DestVT).bitsLT(NewVT)) // Value is expanded, e.g. i64 -> i16. 694 return NumVectorRegs*(NewVTSize/DestVT.getSizeInBits()); 695 696 // Otherwise, promotion or legal types use the same number of registers as 697 // the vector decimated to the appropriate level. 698 return NumVectorRegs; 699} 700 701/// isLegalRC - Return true if the value types that can be represented by the 702/// specified register class are all legal. 703bool TargetLowering::isLegalRC(const TargetRegisterClass *RC) const { 704 for (TargetRegisterClass::vt_iterator I = RC->vt_begin(), E = RC->vt_end(); 705 I != E; ++I) { 706 if (isTypeLegal(*I)) 707 return true; 708 } 709 return false; 710} 711 712/// findRepresentativeClass - Return the largest legal super-reg register class 713/// of the register class for the specified type and its associated "cost". 714std::pair<const TargetRegisterClass*, uint8_t> 715TargetLowering::findRepresentativeClass(MVT VT) const { 716 const TargetRegisterInfo *TRI = getTargetMachine().getRegisterInfo(); 717 const TargetRegisterClass *RC = RegClassForVT[VT.SimpleTy]; 718 if (!RC) 719 return std::make_pair(RC, 0); 720 721 // Compute the set of all super-register classes. 722 BitVector SuperRegRC(TRI->getNumRegClasses()); 723 for (SuperRegClassIterator RCI(RC, TRI); RCI.isValid(); ++RCI) 724 SuperRegRC.setBitsInMask(RCI.getMask()); 725 726 // Find the first legal register class with the largest spill size. 727 const TargetRegisterClass *BestRC = RC; 728 for (int i = SuperRegRC.find_first(); i >= 0; i = SuperRegRC.find_next(i)) { 729 const TargetRegisterClass *SuperRC = TRI->getRegClass(i); 730 // We want the largest possible spill size. 731 if (SuperRC->getSize() <= BestRC->getSize()) 732 continue; 733 if (!isLegalRC(SuperRC)) 734 continue; 735 BestRC = SuperRC; 736 } 737 return std::make_pair(BestRC, 1); 738} 739 740/// computeRegisterProperties - Once all of the register classes are added, 741/// this allows us to compute derived properties we expose. 742void TargetLowering::computeRegisterProperties() { 743 assert(MVT::LAST_VALUETYPE <= MVT::MAX_ALLOWED_VALUETYPE && 744 "Too many value types for ValueTypeActions to hold!"); 745 746 // Everything defaults to needing one register. 747 for (unsigned i = 0; i != MVT::LAST_VALUETYPE; ++i) { 748 NumRegistersForVT[i] = 1; 749 RegisterTypeForVT[i] = TransformToType[i] = (MVT::SimpleValueType)i; 750 } 751 // ...except isVoid, which doesn't need any registers. 752 NumRegistersForVT[MVT::isVoid] = 0; 753 754 // Find the largest integer register class. 755 unsigned LargestIntReg = MVT::LAST_INTEGER_VALUETYPE; 756 for (; RegClassForVT[LargestIntReg] == 0; --LargestIntReg) 757 assert(LargestIntReg != MVT::i1 && "No integer registers defined!"); 758 759 // Every integer value type larger than this largest register takes twice as 760 // many registers to represent as the previous ValueType. 761 for (unsigned ExpandedReg = LargestIntReg + 1; 762 ExpandedReg <= MVT::LAST_INTEGER_VALUETYPE; ++ExpandedReg) { 763 NumRegistersForVT[ExpandedReg] = 2*NumRegistersForVT[ExpandedReg-1]; 764 RegisterTypeForVT[ExpandedReg] = (MVT::SimpleValueType)LargestIntReg; 765 TransformToType[ExpandedReg] = (MVT::SimpleValueType)(ExpandedReg - 1); 766 ValueTypeActions.setTypeAction((MVT::SimpleValueType)ExpandedReg, 767 TypeExpandInteger); 768 } 769 770 // Inspect all of the ValueType's smaller than the largest integer 771 // register to see which ones need promotion. 772 unsigned LegalIntReg = LargestIntReg; 773 for (unsigned IntReg = LargestIntReg - 1; 774 IntReg >= (unsigned)MVT::i1; --IntReg) { 775 MVT IVT = (MVT::SimpleValueType)IntReg; 776 if (isTypeLegal(IVT)) { 777 LegalIntReg = IntReg; 778 } else { 779 RegisterTypeForVT[IntReg] = TransformToType[IntReg] = 780 (const MVT::SimpleValueType)LegalIntReg; 781 ValueTypeActions.setTypeAction(IVT, TypePromoteInteger); 782 } 783 } 784 785 // ppcf128 type is really two f64's. 786 if (!isTypeLegal(MVT::ppcf128)) { 787 NumRegistersForVT[MVT::ppcf128] = 2*NumRegistersForVT[MVT::f64]; 788 RegisterTypeForVT[MVT::ppcf128] = MVT::f64; 789 TransformToType[MVT::ppcf128] = MVT::f64; 790 ValueTypeActions.setTypeAction(MVT::ppcf128, TypeExpandFloat); 791 } 792 793 // Decide how to handle f64. If the target does not have native f64 support, 794 // expand it to i64 and we will be generating soft float library calls. 795 if (!isTypeLegal(MVT::f64)) { 796 NumRegistersForVT[MVT::f64] = NumRegistersForVT[MVT::i64]; 797 RegisterTypeForVT[MVT::f64] = RegisterTypeForVT[MVT::i64]; 798 TransformToType[MVT::f64] = MVT::i64; 799 ValueTypeActions.setTypeAction(MVT::f64, TypeSoftenFloat); 800 } 801 802 // Decide how to handle f32. If the target does not have native support for 803 // f32, promote it to f64 if it is legal. Otherwise, expand it to i32. 804 if (!isTypeLegal(MVT::f32)) { 805 if (isTypeLegal(MVT::f64)) { 806 NumRegistersForVT[MVT::f32] = NumRegistersForVT[MVT::f64]; 807 RegisterTypeForVT[MVT::f32] = RegisterTypeForVT[MVT::f64]; 808 TransformToType[MVT::f32] = MVT::f64; 809 ValueTypeActions.setTypeAction(MVT::f32, TypePromoteInteger); 810 } else { 811 NumRegistersForVT[MVT::f32] = NumRegistersForVT[MVT::i32]; 812 RegisterTypeForVT[MVT::f32] = RegisterTypeForVT[MVT::i32]; 813 TransformToType[MVT::f32] = MVT::i32; 814 ValueTypeActions.setTypeAction(MVT::f32, TypeSoftenFloat); 815 } 816 } 817 818 // Loop over all of the vector value types to see which need transformations. 819 for (unsigned i = MVT::FIRST_VECTOR_VALUETYPE; 820 i <= (unsigned)MVT::LAST_VECTOR_VALUETYPE; ++i) { 821 MVT VT = (MVT::SimpleValueType)i; 822 if (isTypeLegal(VT)) continue; 823 824 // Determine if there is a legal wider type. If so, we should promote to 825 // that wider vector type. 826 MVT EltVT = VT.getVectorElementType(); 827 unsigned NElts = VT.getVectorNumElements(); 828 if (NElts != 1 && !shouldSplitVectorElementType(EltVT)) { 829 bool IsLegalWiderType = false; 830 // First try to promote the elements of integer vectors. If no legal 831 // promotion was found, fallback to the widen-vector method. 832 for (unsigned nVT = i+1; nVT <= MVT::LAST_VECTOR_VALUETYPE; ++nVT) { 833 MVT SVT = (MVT::SimpleValueType)nVT; 834 // Promote vectors of integers to vectors with the same number 835 // of elements, with a wider element type. 836 if (SVT.getVectorElementType().getSizeInBits() > EltVT.getSizeInBits() 837 && SVT.getVectorNumElements() == NElts && 838 isTypeLegal(SVT) && SVT.getScalarType().isInteger()) { 839 TransformToType[i] = SVT; 840 RegisterTypeForVT[i] = SVT; 841 NumRegistersForVT[i] = 1; 842 ValueTypeActions.setTypeAction(VT, TypePromoteInteger); 843 IsLegalWiderType = true; 844 break; 845 } 846 } 847 848 if (IsLegalWiderType) continue; 849 850 // Try to widen the vector. 851 for (unsigned nVT = i+1; nVT <= MVT::LAST_VECTOR_VALUETYPE; ++nVT) { 852 MVT SVT = (MVT::SimpleValueType)nVT; 853 if (SVT.getVectorElementType() == EltVT && 854 SVT.getVectorNumElements() > NElts && 855 isTypeLegal(SVT)) { 856 TransformToType[i] = SVT; 857 RegisterTypeForVT[i] = SVT; 858 NumRegistersForVT[i] = 1; 859 ValueTypeActions.setTypeAction(VT, TypeWidenVector); 860 IsLegalWiderType = true; 861 break; 862 } 863 } 864 if (IsLegalWiderType) continue; 865 } 866 867 MVT IntermediateVT; 868 MVT RegisterVT; 869 unsigned NumIntermediates; 870 NumRegistersForVT[i] = 871 getVectorTypeBreakdownMVT(VT, IntermediateVT, NumIntermediates, 872 RegisterVT, this); 873 RegisterTypeForVT[i] = RegisterVT; 874 875 MVT NVT = VT.getPow2VectorType(); 876 if (NVT == VT) { 877 // Type is already a power of 2. The default action is to split. 878 TransformToType[i] = MVT::Other; 879 unsigned NumElts = VT.getVectorNumElements(); 880 ValueTypeActions.setTypeAction(VT, 881 NumElts > 1 ? TypeSplitVector : TypeScalarizeVector); 882 } else { 883 TransformToType[i] = NVT; 884 ValueTypeActions.setTypeAction(VT, TypeWidenVector); 885 } 886 } 887 888 // Determine the 'representative' register class for each value type. 889 // An representative register class is the largest (meaning one which is 890 // not a sub-register class / subreg register class) legal register class for 891 // a group of value types. For example, on i386, i8, i16, and i32 892 // representative would be GR32; while on x86_64 it's GR64. 893 for (unsigned i = 0; i != MVT::LAST_VALUETYPE; ++i) { 894 const TargetRegisterClass* RRC; 895 uint8_t Cost; 896 tie(RRC, Cost) = findRepresentativeClass((MVT::SimpleValueType)i); 897 RepRegClassForVT[i] = RRC; 898 RepRegClassCostForVT[i] = Cost; 899 } 900} 901 902const char *TargetLowering::getTargetNodeName(unsigned Opcode) const { 903 return NULL; 904} 905 906EVT TargetLowering::getSetCCResultType(EVT VT) const { 907 assert(!VT.isVector() && "No default SetCC type for vectors!"); 908 return getPointerTy(0).SimpleTy; 909} 910 911MVT::SimpleValueType TargetLowering::getCmpLibcallReturnType() const { 912 return MVT::i32; // return the default value 913} 914 915/// getVectorTypeBreakdown - Vector types are broken down into some number of 916/// legal first class types. For example, MVT::v8f32 maps to 2 MVT::v4f32 917/// with Altivec or SSE1, or 8 promoted MVT::f64 values with the X86 FP stack. 918/// Similarly, MVT::v2i64 turns into 4 MVT::i32 values with both PPC and X86. 919/// 920/// This method returns the number of registers needed, and the VT for each 921/// register. It also returns the VT and quantity of the intermediate values 922/// before they are promoted/expanded. 923/// 924unsigned TargetLowering::getVectorTypeBreakdown(LLVMContext &Context, EVT VT, 925 EVT &IntermediateVT, 926 unsigned &NumIntermediates, 927 MVT &RegisterVT) const { 928 unsigned NumElts = VT.getVectorNumElements(); 929 930 // If there is a wider vector type with the same element type as this one, 931 // or a promoted vector type that has the same number of elements which 932 // are wider, then we should convert to that legal vector type. 933 // This handles things like <2 x float> -> <4 x float> and 934 // <4 x i1> -> <4 x i32>. 935 LegalizeTypeAction TA = getTypeAction(Context, VT); 936 if (NumElts != 1 && (TA == TypeWidenVector || TA == TypePromoteInteger)) { 937 EVT RegisterEVT = getTypeToTransformTo(Context, VT); 938 if (isTypeLegal(RegisterEVT)) { 939 IntermediateVT = RegisterEVT; 940 RegisterVT = RegisterEVT.getSimpleVT(); 941 NumIntermediates = 1; 942 return 1; 943 } 944 } 945 946 // Figure out the right, legal destination reg to copy into. 947 EVT EltTy = VT.getVectorElementType(); 948 949 unsigned NumVectorRegs = 1; 950 951 // FIXME: We don't support non-power-of-2-sized vectors for now. Ideally we 952 // could break down into LHS/RHS like LegalizeDAG does. 953 if (!isPowerOf2_32(NumElts)) { 954 NumVectorRegs = NumElts; 955 NumElts = 1; 956 } 957 958 // Divide the input until we get to a supported size. This will always 959 // end with a scalar if the target doesn't support vectors. 960 while (NumElts > 1 && !isTypeLegal( 961 EVT::getVectorVT(Context, EltTy, NumElts))) { 962 NumElts >>= 1; 963 NumVectorRegs <<= 1; 964 } 965 966 NumIntermediates = NumVectorRegs; 967 968 EVT NewVT = EVT::getVectorVT(Context, EltTy, NumElts); 969 if (!isTypeLegal(NewVT)) 970 NewVT = EltTy; 971 IntermediateVT = NewVT; 972 973 MVT DestVT = getRegisterType(Context, NewVT); 974 RegisterVT = DestVT; 975 unsigned NewVTSize = NewVT.getSizeInBits(); 976 977 // Convert sizes such as i33 to i64. 978 if (!isPowerOf2_32(NewVTSize)) 979 NewVTSize = NextPowerOf2(NewVTSize); 980 981 if (EVT(DestVT).bitsLT(NewVT)) // Value is expanded, e.g. i64 -> i16. 982 return NumVectorRegs*(NewVTSize/DestVT.getSizeInBits()); 983 984 // Otherwise, promotion or legal types use the same number of registers as 985 // the vector decimated to the appropriate level. 986 return NumVectorRegs; 987} 988 989/// Get the EVTs and ArgFlags collections that represent the legalized return 990/// type of the given function. This does not require a DAG or a return value, 991/// and is suitable for use before any DAGs for the function are constructed. 992/// TODO: Move this out of TargetLowering.cpp. 993void llvm::GetReturnInfo(Type* ReturnType, Attribute attr, 994 SmallVectorImpl<ISD::OutputArg> &Outs, 995 const TargetLowering &TLI) { 996 SmallVector<EVT, 4> ValueVTs; 997 ComputeValueVTs(TLI, ReturnType, ValueVTs); 998 unsigned NumValues = ValueVTs.size(); 999 if (NumValues == 0) return; 1000 1001 for (unsigned j = 0, f = NumValues; j != f; ++j) { 1002 EVT VT = ValueVTs[j]; 1003 ISD::NodeType ExtendKind = ISD::ANY_EXTEND; 1004 1005 if (attr.hasAttribute(Attribute::SExt)) 1006 ExtendKind = ISD::SIGN_EXTEND; 1007 else if (attr.hasAttribute(Attribute::ZExt)) 1008 ExtendKind = ISD::ZERO_EXTEND; 1009 1010 // FIXME: C calling convention requires the return type to be promoted to 1011 // at least 32-bit. But this is not necessary for non-C calling 1012 // conventions. The frontend should mark functions whose return values 1013 // require promoting with signext or zeroext attributes. 1014 if (ExtendKind != ISD::ANY_EXTEND && VT.isInteger()) { 1015 MVT MinVT = TLI.getRegisterType(ReturnType->getContext(), MVT::i32); 1016 if (VT.bitsLT(MinVT)) 1017 VT = MinVT; 1018 } 1019 1020 unsigned NumParts = TLI.getNumRegisters(ReturnType->getContext(), VT); 1021 MVT PartVT = TLI.getRegisterType(ReturnType->getContext(), VT); 1022 1023 // 'inreg' on function refers to return value 1024 ISD::ArgFlagsTy Flags = ISD::ArgFlagsTy(); 1025 if (attr.hasAttribute(Attribute::InReg)) 1026 Flags.setInReg(); 1027 1028 // Propagate extension type if any 1029 if (attr.hasAttribute(Attribute::SExt)) 1030 Flags.setSExt(); 1031 else if (attr.hasAttribute(Attribute::ZExt)) 1032 Flags.setZExt(); 1033 1034 for (unsigned i = 0; i < NumParts; ++i) 1035 Outs.push_back(ISD::OutputArg(Flags, PartVT, /*isFixed=*/true, 0, 0)); 1036 } 1037} 1038 1039/// getByValTypeAlignment - Return the desired alignment for ByVal aggregate 1040/// function arguments in the caller parameter area. This is the actual 1041/// alignment, not its logarithm. 1042unsigned TargetLowering::getByValTypeAlignment(Type *Ty) const { 1043 return TD->getCallFrameTypeAlignment(Ty); 1044} 1045 1046/// getJumpTableEncoding - Return the entry encoding for a jump table in the 1047/// current function. The returned value is a member of the 1048/// MachineJumpTableInfo::JTEntryKind enum. 1049unsigned TargetLowering::getJumpTableEncoding() const { 1050 // In non-pic modes, just use the address of a block. 1051 if (getTargetMachine().getRelocationModel() != Reloc::PIC_) 1052 return MachineJumpTableInfo::EK_BlockAddress; 1053 1054 // In PIC mode, if the target supports a GPRel32 directive, use it. 1055 if (getTargetMachine().getMCAsmInfo()->getGPRel32Directive() != 0) 1056 return MachineJumpTableInfo::EK_GPRel32BlockAddress; 1057 1058 // Otherwise, use a label difference. 1059 return MachineJumpTableInfo::EK_LabelDifference32; 1060} 1061 1062SDValue TargetLowering::getPICJumpTableRelocBase(SDValue Table, 1063 SelectionDAG &DAG) const { 1064 // If our PIC model is GP relative, use the global offset table as the base. 1065 unsigned JTEncoding = getJumpTableEncoding(); 1066 1067 if ((JTEncoding == MachineJumpTableInfo::EK_GPRel64BlockAddress) || 1068 (JTEncoding == MachineJumpTableInfo::EK_GPRel32BlockAddress)) 1069 return DAG.getGLOBAL_OFFSET_TABLE(getPointerTy(0)); 1070 1071 return Table; 1072} 1073 1074/// getPICJumpTableRelocBaseExpr - This returns the relocation base for the 1075/// given PIC jumptable, the same as getPICJumpTableRelocBase, but as an 1076/// MCExpr. 1077const MCExpr * 1078TargetLowering::getPICJumpTableRelocBaseExpr(const MachineFunction *MF, 1079 unsigned JTI,MCContext &Ctx) const{ 1080 // The normal PIC reloc base is the label at the start of the jump table. 1081 return MCSymbolRefExpr::Create(MF->getJTISymbol(JTI, Ctx), Ctx); 1082} 1083 1084bool 1085TargetLowering::isOffsetFoldingLegal(const GlobalAddressSDNode *GA) const { 1086 // Assume that everything is safe in static mode. 1087 if (getTargetMachine().getRelocationModel() == Reloc::Static) 1088 return true; 1089 1090 // In dynamic-no-pic mode, assume that known defined values are safe. 1091 if (getTargetMachine().getRelocationModel() == Reloc::DynamicNoPIC && 1092 GA && 1093 !GA->getGlobal()->isDeclaration() && 1094 !GA->getGlobal()->isWeakForLinker()) 1095 return true; 1096 1097 // Otherwise assume nothing is safe. 1098 return false; 1099} 1100 1101//===----------------------------------------------------------------------===// 1102// Optimization Methods 1103//===----------------------------------------------------------------------===// 1104 1105/// ShrinkDemandedConstant - Check to see if the specified operand of the 1106/// specified instruction is a constant integer. If so, check to see if there 1107/// are any bits set in the constant that are not demanded. If so, shrink the 1108/// constant and return true. 1109bool TargetLowering::TargetLoweringOpt::ShrinkDemandedConstant(SDValue Op, 1110 const APInt &Demanded) { 1111 DebugLoc dl = Op.getDebugLoc(); 1112 1113 // FIXME: ISD::SELECT, ISD::SELECT_CC 1114 switch (Op.getOpcode()) { 1115 default: break; 1116 case ISD::XOR: 1117 case ISD::AND: 1118 case ISD::OR: { 1119 ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op.getOperand(1)); 1120 if (!C) return false; 1121 1122 if (Op.getOpcode() == ISD::XOR && 1123 (C->getAPIntValue() | (~Demanded)).isAllOnesValue()) 1124 return false; 1125 1126 // if we can expand it to have all bits set, do it 1127 if (C->getAPIntValue().intersects(~Demanded)) { 1128 EVT VT = Op.getValueType(); 1129 SDValue New = DAG.getNode(Op.getOpcode(), dl, VT, Op.getOperand(0), 1130 DAG.getConstant(Demanded & 1131 C->getAPIntValue(), 1132 VT)); 1133 return CombineTo(Op, New); 1134 } 1135 1136 break; 1137 } 1138 } 1139 1140 return false; 1141} 1142 1143/// ShrinkDemandedOp - Convert x+y to (VT)((SmallVT)x+(SmallVT)y) if the 1144/// casts are free. This uses isZExtFree and ZERO_EXTEND for the widening 1145/// cast, but it could be generalized for targets with other types of 1146/// implicit widening casts. 1147bool 1148TargetLowering::TargetLoweringOpt::ShrinkDemandedOp(SDValue Op, 1149 unsigned BitWidth, 1150 const APInt &Demanded, 1151 DebugLoc dl) { 1152 assert(Op.getNumOperands() == 2 && 1153 "ShrinkDemandedOp only supports binary operators!"); 1154 assert(Op.getNode()->getNumValues() == 1 && 1155 "ShrinkDemandedOp only supports nodes with one result!"); 1156 1157 // Don't do this if the node has another user, which may require the 1158 // full value. 1159 if (!Op.getNode()->hasOneUse()) 1160 return false; 1161 1162 // Search for the smallest integer type with free casts to and from 1163 // Op's type. For expedience, just check power-of-2 integer types. 1164 const TargetLowering &TLI = DAG.getTargetLoweringInfo(); 1165 unsigned DemandedSize = BitWidth - Demanded.countLeadingZeros(); 1166 unsigned SmallVTBits = DemandedSize; 1167 if (!isPowerOf2_32(SmallVTBits)) 1168 SmallVTBits = NextPowerOf2(SmallVTBits); 1169 for (; SmallVTBits < BitWidth; SmallVTBits = NextPowerOf2(SmallVTBits)) { 1170 EVT SmallVT = EVT::getIntegerVT(*DAG.getContext(), SmallVTBits); 1171 if (TLI.isTruncateFree(Op.getValueType(), SmallVT) && 1172 TLI.isZExtFree(SmallVT, Op.getValueType())) { 1173 // We found a type with free casts. 1174 SDValue X = DAG.getNode(Op.getOpcode(), dl, SmallVT, 1175 DAG.getNode(ISD::TRUNCATE, dl, SmallVT, 1176 Op.getNode()->getOperand(0)), 1177 DAG.getNode(ISD::TRUNCATE, dl, SmallVT, 1178 Op.getNode()->getOperand(1))); 1179 bool NeedZext = DemandedSize > SmallVTBits; 1180 SDValue Z = DAG.getNode(NeedZext ? ISD::ZERO_EXTEND : ISD::ANY_EXTEND, 1181 dl, Op.getValueType(), X); 1182 return CombineTo(Op, Z); 1183 } 1184 } 1185 return false; 1186} 1187 1188/// SimplifyDemandedBits - Look at Op. At this point, we know that only the 1189/// DemandedMask bits of the result of Op are ever used downstream. If we can 1190/// use this information to simplify Op, create a new simplified DAG node and 1191/// return true, returning the original and new nodes in Old and New. Otherwise, 1192/// analyze the expression and return a mask of KnownOne and KnownZero bits for 1193/// the expression (used to simplify the caller). The KnownZero/One bits may 1194/// only be accurate for those bits in the DemandedMask. 1195bool TargetLowering::SimplifyDemandedBits(SDValue Op, 1196 const APInt &DemandedMask, 1197 APInt &KnownZero, 1198 APInt &KnownOne, 1199 TargetLoweringOpt &TLO, 1200 unsigned Depth) const { 1201 unsigned BitWidth = DemandedMask.getBitWidth(); 1202 assert(Op.getValueType().getScalarType().getSizeInBits() == BitWidth && 1203 "Mask size mismatches value type size!"); 1204 APInt NewMask = DemandedMask; 1205 DebugLoc dl = Op.getDebugLoc(); 1206 1207 // Don't know anything. 1208 KnownZero = KnownOne = APInt(BitWidth, 0); 1209 1210 // Other users may use these bits. 1211 if (!Op.getNode()->hasOneUse()) { 1212 if (Depth != 0) { 1213 // If not at the root, Just compute the KnownZero/KnownOne bits to 1214 // simplify things downstream. 1215 TLO.DAG.ComputeMaskedBits(Op, KnownZero, KnownOne, Depth); 1216 return false; 1217 } 1218 // If this is the root being simplified, allow it to have multiple uses, 1219 // just set the NewMask to all bits. 1220 NewMask = APInt::getAllOnesValue(BitWidth); 1221 } else if (DemandedMask == 0) { 1222 // Not demanding any bits from Op. 1223 if (Op.getOpcode() != ISD::UNDEF) 1224 return TLO.CombineTo(Op, TLO.DAG.getUNDEF(Op.getValueType())); 1225 return false; 1226 } else if (Depth == 6) { // Limit search depth. 1227 return false; 1228 } 1229 1230 APInt KnownZero2, KnownOne2, KnownZeroOut, KnownOneOut; 1231 switch (Op.getOpcode()) { 1232 case ISD::Constant: 1233 // We know all of the bits for a constant! 1234 KnownOne = cast<ConstantSDNode>(Op)->getAPIntValue(); 1235 KnownZero = ~KnownOne; 1236 return false; // Don't fall through, will infinitely loop. 1237 case ISD::AND: 1238 // If the RHS is a constant, check to see if the LHS would be zero without 1239 // using the bits from the RHS. Below, we use knowledge about the RHS to 1240 // simplify the LHS, here we're using information from the LHS to simplify 1241 // the RHS. 1242 if (ConstantSDNode *RHSC = dyn_cast<ConstantSDNode>(Op.getOperand(1))) { 1243 APInt LHSZero, LHSOne; 1244 // Do not increment Depth here; that can cause an infinite loop. 1245 TLO.DAG.ComputeMaskedBits(Op.getOperand(0), LHSZero, LHSOne, Depth); 1246 // If the LHS already has zeros where RHSC does, this and is dead. 1247 if ((LHSZero & NewMask) == (~RHSC->getAPIntValue() & NewMask)) 1248 return TLO.CombineTo(Op, Op.getOperand(0)); 1249 // If any of the set bits in the RHS are known zero on the LHS, shrink 1250 // the constant. 1251 if (TLO.ShrinkDemandedConstant(Op, ~LHSZero & NewMask)) 1252 return true; 1253 } 1254 1255 if (SimplifyDemandedBits(Op.getOperand(1), NewMask, KnownZero, 1256 KnownOne, TLO, Depth+1)) 1257 return true; 1258 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?"); 1259 if (SimplifyDemandedBits(Op.getOperand(0), ~KnownZero & NewMask, 1260 KnownZero2, KnownOne2, TLO, Depth+1)) 1261 return true; 1262 assert((KnownZero2 & KnownOne2) == 0 && "Bits known to be one AND zero?"); 1263 1264 // If all of the demanded bits are known one on one side, return the other. 1265 // These bits cannot contribute to the result of the 'and'. 1266 if ((NewMask & ~KnownZero2 & KnownOne) == (~KnownZero2 & NewMask)) 1267 return TLO.CombineTo(Op, Op.getOperand(0)); 1268 if ((NewMask & ~KnownZero & KnownOne2) == (~KnownZero & NewMask)) 1269 return TLO.CombineTo(Op, Op.getOperand(1)); 1270 // If all of the demanded bits in the inputs are known zeros, return zero. 1271 if ((NewMask & (KnownZero|KnownZero2)) == NewMask) 1272 return TLO.CombineTo(Op, TLO.DAG.getConstant(0, Op.getValueType())); 1273 // If the RHS is a constant, see if we can simplify it. 1274 if (TLO.ShrinkDemandedConstant(Op, ~KnownZero2 & NewMask)) 1275 return true; 1276 // If the operation can be done in a smaller type, do so. 1277 if (TLO.ShrinkDemandedOp(Op, BitWidth, NewMask, dl)) 1278 return true; 1279 1280 // Output known-1 bits are only known if set in both the LHS & RHS. 1281 KnownOne &= KnownOne2; 1282 // Output known-0 are known to be clear if zero in either the LHS | RHS. 1283 KnownZero |= KnownZero2; 1284 break; 1285 case ISD::OR: 1286 if (SimplifyDemandedBits(Op.getOperand(1), NewMask, KnownZero, 1287 KnownOne, TLO, Depth+1)) 1288 return true; 1289 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?"); 1290 if (SimplifyDemandedBits(Op.getOperand(0), ~KnownOne & NewMask, 1291 KnownZero2, KnownOne2, TLO, Depth+1)) 1292 return true; 1293 assert((KnownZero2 & KnownOne2) == 0 && "Bits known to be one AND zero?"); 1294 1295 // If all of the demanded bits are known zero on one side, return the other. 1296 // These bits cannot contribute to the result of the 'or'. 1297 if ((NewMask & ~KnownOne2 & KnownZero) == (~KnownOne2 & NewMask)) 1298 return TLO.CombineTo(Op, Op.getOperand(0)); 1299 if ((NewMask & ~KnownOne & KnownZero2) == (~KnownOne & NewMask)) 1300 return TLO.CombineTo(Op, Op.getOperand(1)); 1301 // If all of the potentially set bits on one side are known to be set on 1302 // the other side, just use the 'other' side. 1303 if ((NewMask & ~KnownZero & KnownOne2) == (~KnownZero & NewMask)) 1304 return TLO.CombineTo(Op, Op.getOperand(0)); 1305 if ((NewMask & ~KnownZero2 & KnownOne) == (~KnownZero2 & NewMask)) 1306 return TLO.CombineTo(Op, Op.getOperand(1)); 1307 // If the RHS is a constant, see if we can simplify it. 1308 if (TLO.ShrinkDemandedConstant(Op, NewMask)) 1309 return true; 1310 // If the operation can be done in a smaller type, do so. 1311 if (TLO.ShrinkDemandedOp(Op, BitWidth, NewMask, dl)) 1312 return true; 1313 1314 // Output known-0 bits are only known if clear in both the LHS & RHS. 1315 KnownZero &= KnownZero2; 1316 // Output known-1 are known to be set if set in either the LHS | RHS. 1317 KnownOne |= KnownOne2; 1318 break; 1319 case ISD::XOR: 1320 if (SimplifyDemandedBits(Op.getOperand(1), NewMask, KnownZero, 1321 KnownOne, TLO, Depth+1)) 1322 return true; 1323 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?"); 1324 if (SimplifyDemandedBits(Op.getOperand(0), NewMask, KnownZero2, 1325 KnownOne2, TLO, Depth+1)) 1326 return true; 1327 assert((KnownZero2 & KnownOne2) == 0 && "Bits known to be one AND zero?"); 1328 1329 // If all of the demanded bits are known zero on one side, return the other. 1330 // These bits cannot contribute to the result of the 'xor'. 1331 if ((KnownZero & NewMask) == NewMask) 1332 return TLO.CombineTo(Op, Op.getOperand(0)); 1333 if ((KnownZero2 & NewMask) == NewMask) 1334 return TLO.CombineTo(Op, Op.getOperand(1)); 1335 // If the operation can be done in a smaller type, do so. 1336 if (TLO.ShrinkDemandedOp(Op, BitWidth, NewMask, dl)) 1337 return true; 1338 1339 // If all of the unknown bits are known to be zero on one side or the other 1340 // (but not both) turn this into an *inclusive* or. 1341 // e.g. (A & C1)^(B & C2) -> (A & C1)|(B & C2) iff C1&C2 == 0 1342 if ((NewMask & ~KnownZero & ~KnownZero2) == 0) 1343 return TLO.CombineTo(Op, TLO.DAG.getNode(ISD::OR, dl, Op.getValueType(), 1344 Op.getOperand(0), 1345 Op.getOperand(1))); 1346 1347 // Output known-0 bits are known if clear or set in both the LHS & RHS. 1348 KnownZeroOut = (KnownZero & KnownZero2) | (KnownOne & KnownOne2); 1349 // Output known-1 are known to be set if set in only one of the LHS, RHS. 1350 KnownOneOut = (KnownZero & KnownOne2) | (KnownOne & KnownZero2); 1351 1352 // If all of the demanded bits on one side are known, and all of the set 1353 // bits on that side are also known to be set on the other side, turn this 1354 // into an AND, as we know the bits will be cleared. 1355 // e.g. (X | C1) ^ C2 --> (X | C1) & ~C2 iff (C1&C2) == C2 1356 // NB: it is okay if more bits are known than are requested 1357 if ((NewMask & (KnownZero|KnownOne)) == NewMask) { // all known on one side 1358 if (KnownOne == KnownOne2) { // set bits are the same on both sides 1359 EVT VT = Op.getValueType(); 1360 SDValue ANDC = TLO.DAG.getConstant(~KnownOne & NewMask, VT); 1361 return TLO.CombineTo(Op, TLO.DAG.getNode(ISD::AND, dl, VT, 1362 Op.getOperand(0), ANDC)); 1363 } 1364 } 1365 1366 // If the RHS is a constant, see if we can simplify it. 1367 // for XOR, we prefer to force bits to 1 if they will make a -1. 1368 // if we can't force bits, try to shrink constant 1369 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op.getOperand(1))) { 1370 APInt Expanded = C->getAPIntValue() | (~NewMask); 1371 // if we can expand it to have all bits set, do it 1372 if (Expanded.isAllOnesValue()) { 1373 if (Expanded != C->getAPIntValue()) { 1374 EVT VT = Op.getValueType(); 1375 SDValue New = TLO.DAG.getNode(Op.getOpcode(), dl,VT, Op.getOperand(0), 1376 TLO.DAG.getConstant(Expanded, VT)); 1377 return TLO.CombineTo(Op, New); 1378 } 1379 // if it already has all the bits set, nothing to change 1380 // but don't shrink either! 1381 } else if (TLO.ShrinkDemandedConstant(Op, NewMask)) { 1382 return true; 1383 } 1384 } 1385 1386 KnownZero = KnownZeroOut; 1387 KnownOne = KnownOneOut; 1388 break; 1389 case ISD::SELECT: 1390 if (SimplifyDemandedBits(Op.getOperand(2), NewMask, KnownZero, 1391 KnownOne, TLO, Depth+1)) 1392 return true; 1393 if (SimplifyDemandedBits(Op.getOperand(1), NewMask, KnownZero2, 1394 KnownOne2, TLO, Depth+1)) 1395 return true; 1396 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?"); 1397 assert((KnownZero2 & KnownOne2) == 0 && "Bits known to be one AND zero?"); 1398 1399 // If the operands are constants, see if we can simplify them. 1400 if (TLO.ShrinkDemandedConstant(Op, NewMask)) 1401 return true; 1402 1403 // Only known if known in both the LHS and RHS. 1404 KnownOne &= KnownOne2; 1405 KnownZero &= KnownZero2; 1406 break; 1407 case ISD::SELECT_CC: 1408 if (SimplifyDemandedBits(Op.getOperand(3), NewMask, KnownZero, 1409 KnownOne, TLO, Depth+1)) 1410 return true; 1411 if (SimplifyDemandedBits(Op.getOperand(2), NewMask, KnownZero2, 1412 KnownOne2, TLO, Depth+1)) 1413 return true; 1414 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?"); 1415 assert((KnownZero2 & KnownOne2) == 0 && "Bits known to be one AND zero?"); 1416 1417 // If the operands are constants, see if we can simplify them. 1418 if (TLO.ShrinkDemandedConstant(Op, NewMask)) 1419 return true; 1420 1421 // Only known if known in both the LHS and RHS. 1422 KnownOne &= KnownOne2; 1423 KnownZero &= KnownZero2; 1424 break; 1425 case ISD::SHL: 1426 if (ConstantSDNode *SA = dyn_cast<ConstantSDNode>(Op.getOperand(1))) { 1427 unsigned ShAmt = SA->getZExtValue(); 1428 SDValue InOp = Op.getOperand(0); 1429 1430 // If the shift count is an invalid immediate, don't do anything. 1431 if (ShAmt >= BitWidth) 1432 break; 1433 1434 // If this is ((X >>u C1) << ShAmt), see if we can simplify this into a 1435 // single shift. We can do this if the bottom bits (which are shifted 1436 // out) are never demanded. 1437 if (InOp.getOpcode() == ISD::SRL && 1438 isa<ConstantSDNode>(InOp.getOperand(1))) { 1439 if (ShAmt && (NewMask & APInt::getLowBitsSet(BitWidth, ShAmt)) == 0) { 1440 unsigned C1= cast<ConstantSDNode>(InOp.getOperand(1))->getZExtValue(); 1441 unsigned Opc = ISD::SHL; 1442 int Diff = ShAmt-C1; 1443 if (Diff < 0) { 1444 Diff = -Diff; 1445 Opc = ISD::SRL; 1446 } 1447 1448 SDValue NewSA = 1449 TLO.DAG.getConstant(Diff, Op.getOperand(1).getValueType()); 1450 EVT VT = Op.getValueType(); 1451 return TLO.CombineTo(Op, TLO.DAG.getNode(Opc, dl, VT, 1452 InOp.getOperand(0), NewSA)); 1453 } 1454 } 1455 1456 if (SimplifyDemandedBits(InOp, NewMask.lshr(ShAmt), 1457 KnownZero, KnownOne, TLO, Depth+1)) 1458 return true; 1459 1460 // Convert (shl (anyext x, c)) to (anyext (shl x, c)) if the high bits 1461 // are not demanded. This will likely allow the anyext to be folded away. 1462 if (InOp.getNode()->getOpcode() == ISD::ANY_EXTEND) { 1463 SDValue InnerOp = InOp.getNode()->getOperand(0); 1464 EVT InnerVT = InnerOp.getValueType(); 1465 unsigned InnerBits = InnerVT.getSizeInBits(); 1466 if (ShAmt < InnerBits && NewMask.lshr(InnerBits) == 0 && 1467 isTypeDesirableForOp(ISD::SHL, InnerVT)) { 1468 EVT ShTy = getShiftAmountTy(InnerVT); 1469 if (!APInt(BitWidth, ShAmt).isIntN(ShTy.getSizeInBits())) 1470 ShTy = InnerVT; 1471 SDValue NarrowShl = 1472 TLO.DAG.getNode(ISD::SHL, dl, InnerVT, InnerOp, 1473 TLO.DAG.getConstant(ShAmt, ShTy)); 1474 return 1475 TLO.CombineTo(Op, 1476 TLO.DAG.getNode(ISD::ANY_EXTEND, dl, Op.getValueType(), 1477 NarrowShl)); 1478 } 1479 } 1480 1481 KnownZero <<= SA->getZExtValue(); 1482 KnownOne <<= SA->getZExtValue(); 1483 // low bits known zero. 1484 KnownZero |= APInt::getLowBitsSet(BitWidth, SA->getZExtValue()); 1485 } 1486 break; 1487 case ISD::SRL: 1488 if (ConstantSDNode *SA = dyn_cast<ConstantSDNode>(Op.getOperand(1))) { 1489 EVT VT = Op.getValueType(); 1490 unsigned ShAmt = SA->getZExtValue(); 1491 unsigned VTSize = VT.getSizeInBits(); 1492 SDValue InOp = Op.getOperand(0); 1493 1494 // If the shift count is an invalid immediate, don't do anything. 1495 if (ShAmt >= BitWidth) 1496 break; 1497 1498 // If this is ((X << C1) >>u ShAmt), see if we can simplify this into a 1499 // single shift. We can do this if the top bits (which are shifted out) 1500 // are never demanded. 1501 if (InOp.getOpcode() == ISD::SHL && 1502 isa<ConstantSDNode>(InOp.getOperand(1))) { 1503 if (ShAmt && (NewMask & APInt::getHighBitsSet(VTSize, ShAmt)) == 0) { 1504 unsigned C1= cast<ConstantSDNode>(InOp.getOperand(1))->getZExtValue(); 1505 unsigned Opc = ISD::SRL; 1506 int Diff = ShAmt-C1; 1507 if (Diff < 0) { 1508 Diff = -Diff; 1509 Opc = ISD::SHL; 1510 } 1511 1512 SDValue NewSA = 1513 TLO.DAG.getConstant(Diff, Op.getOperand(1).getValueType()); 1514 return TLO.CombineTo(Op, TLO.DAG.getNode(Opc, dl, VT, 1515 InOp.getOperand(0), NewSA)); 1516 } 1517 } 1518 1519 // Compute the new bits that are at the top now. 1520 if (SimplifyDemandedBits(InOp, (NewMask << ShAmt), 1521 KnownZero, KnownOne, TLO, Depth+1)) 1522 return true; 1523 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?"); 1524 KnownZero = KnownZero.lshr(ShAmt); 1525 KnownOne = KnownOne.lshr(ShAmt); 1526 1527 APInt HighBits = APInt::getHighBitsSet(BitWidth, ShAmt); 1528 KnownZero |= HighBits; // High bits known zero. 1529 } 1530 break; 1531 case ISD::SRA: 1532 // If this is an arithmetic shift right and only the low-bit is set, we can 1533 // always convert this into a logical shr, even if the shift amount is 1534 // variable. The low bit of the shift cannot be an input sign bit unless 1535 // the shift amount is >= the size of the datatype, which is undefined. 1536 if (NewMask == 1) 1537 return TLO.CombineTo(Op, 1538 TLO.DAG.getNode(ISD::SRL, dl, Op.getValueType(), 1539 Op.getOperand(0), Op.getOperand(1))); 1540 1541 if (ConstantSDNode *SA = dyn_cast<ConstantSDNode>(Op.getOperand(1))) { 1542 EVT VT = Op.getValueType(); 1543 unsigned ShAmt = SA->getZExtValue(); 1544 1545 // If the shift count is an invalid immediate, don't do anything. 1546 if (ShAmt >= BitWidth) 1547 break; 1548 1549 APInt InDemandedMask = (NewMask << ShAmt); 1550 1551 // If any of the demanded bits are produced by the sign extension, we also 1552 // demand the input sign bit. 1553 APInt HighBits = APInt::getHighBitsSet(BitWidth, ShAmt); 1554 if (HighBits.intersects(NewMask)) 1555 InDemandedMask |= APInt::getSignBit(VT.getScalarType().getSizeInBits()); 1556 1557 if (SimplifyDemandedBits(Op.getOperand(0), InDemandedMask, 1558 KnownZero, KnownOne, TLO, Depth+1)) 1559 return true; 1560 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?"); 1561 KnownZero = KnownZero.lshr(ShAmt); 1562 KnownOne = KnownOne.lshr(ShAmt); 1563 1564 // Handle the sign bit, adjusted to where it is now in the mask. 1565 APInt SignBit = APInt::getSignBit(BitWidth).lshr(ShAmt); 1566 1567 // If the input sign bit is known to be zero, or if none of the top bits 1568 // are demanded, turn this into an unsigned shift right. 1569 if (KnownZero.intersects(SignBit) || (HighBits & ~NewMask) == HighBits) { 1570 return TLO.CombineTo(Op, TLO.DAG.getNode(ISD::SRL, dl, VT, 1571 Op.getOperand(0), 1572 Op.getOperand(1))); 1573 } else if (KnownOne.intersects(SignBit)) { // New bits are known one. 1574 KnownOne |= HighBits; 1575 } 1576 } 1577 break; 1578 case ISD::SIGN_EXTEND_INREG: { 1579 EVT ExVT = cast<VTSDNode>(Op.getOperand(1))->getVT(); 1580 1581 APInt MsbMask = APInt::getHighBitsSet(BitWidth, 1); 1582 // If we only care about the highest bit, don't bother shifting right. 1583 if (MsbMask == DemandedMask) { 1584 unsigned ShAmt = ExVT.getScalarType().getSizeInBits(); 1585 SDValue InOp = Op.getOperand(0); 1586 1587 // Compute the correct shift amount type, which must be getShiftAmountTy 1588 // for scalar types after legalization. 1589 EVT ShiftAmtTy = Op.getValueType(); 1590 if (TLO.LegalTypes() && !ShiftAmtTy.isVector()) 1591 ShiftAmtTy = getShiftAmountTy(ShiftAmtTy); 1592 1593 SDValue ShiftAmt = TLO.DAG.getConstant(BitWidth - ShAmt, ShiftAmtTy); 1594 return TLO.CombineTo(Op, TLO.DAG.getNode(ISD::SHL, dl, 1595 Op.getValueType(), InOp, ShiftAmt)); 1596 } 1597 1598 // Sign extension. Compute the demanded bits in the result that are not 1599 // present in the input. 1600 APInt NewBits = 1601 APInt::getHighBitsSet(BitWidth, 1602 BitWidth - ExVT.getScalarType().getSizeInBits()); 1603 1604 // If none of the extended bits are demanded, eliminate the sextinreg. 1605 if ((NewBits & NewMask) == 0) 1606 return TLO.CombineTo(Op, Op.getOperand(0)); 1607 1608 APInt InSignBit = 1609 APInt::getSignBit(ExVT.getScalarType().getSizeInBits()).zext(BitWidth); 1610 APInt InputDemandedBits = 1611 APInt::getLowBitsSet(BitWidth, 1612 ExVT.getScalarType().getSizeInBits()) & 1613 NewMask; 1614 1615 // Since the sign extended bits are demanded, we know that the sign 1616 // bit is demanded. 1617 InputDemandedBits |= InSignBit; 1618 1619 if (SimplifyDemandedBits(Op.getOperand(0), InputDemandedBits, 1620 KnownZero, KnownOne, TLO, Depth+1)) 1621 return true; 1622 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?"); 1623 1624 // If the sign bit of the input is known set or clear, then we know the 1625 // top bits of the result. 1626 1627 // If the input sign bit is known zero, convert this into a zero extension. 1628 if (KnownZero.intersects(InSignBit)) 1629 return TLO.CombineTo(Op, 1630 TLO.DAG.getZeroExtendInReg(Op.getOperand(0),dl,ExVT)); 1631 1632 if (KnownOne.intersects(InSignBit)) { // Input sign bit known set 1633 KnownOne |= NewBits; 1634 KnownZero &= ~NewBits; 1635 } else { // Input sign bit unknown 1636 KnownZero &= ~NewBits; 1637 KnownOne &= ~NewBits; 1638 } 1639 break; 1640 } 1641 case ISD::ZERO_EXTEND: { 1642 unsigned OperandBitWidth = 1643 Op.getOperand(0).getValueType().getScalarType().getSizeInBits(); 1644 APInt InMask = NewMask.trunc(OperandBitWidth); 1645 1646 // If none of the top bits are demanded, convert this into an any_extend. 1647 APInt NewBits = 1648 APInt::getHighBitsSet(BitWidth, BitWidth - OperandBitWidth) & NewMask; 1649 if (!NewBits.intersects(NewMask)) 1650 return TLO.CombineTo(Op, TLO.DAG.getNode(ISD::ANY_EXTEND, dl, 1651 Op.getValueType(), 1652 Op.getOperand(0))); 1653 1654 if (SimplifyDemandedBits(Op.getOperand(0), InMask, 1655 KnownZero, KnownOne, TLO, Depth+1)) 1656 return true; 1657 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?"); 1658 KnownZero = KnownZero.zext(BitWidth); 1659 KnownOne = KnownOne.zext(BitWidth); 1660 KnownZero |= NewBits; 1661 break; 1662 } 1663 case ISD::SIGN_EXTEND: { 1664 EVT InVT = Op.getOperand(0).getValueType(); 1665 unsigned InBits = InVT.getScalarType().getSizeInBits(); 1666 APInt InMask = APInt::getLowBitsSet(BitWidth, InBits); 1667 APInt InSignBit = APInt::getBitsSet(BitWidth, InBits - 1, InBits); 1668 APInt NewBits = ~InMask & NewMask; 1669 1670 // If none of the top bits are demanded, convert this into an any_extend. 1671 if (NewBits == 0) 1672 return TLO.CombineTo(Op,TLO.DAG.getNode(ISD::ANY_EXTEND, dl, 1673 Op.getValueType(), 1674 Op.getOperand(0))); 1675 1676 // Since some of the sign extended bits are demanded, we know that the sign 1677 // bit is demanded. 1678 APInt InDemandedBits = InMask & NewMask; 1679 InDemandedBits |= InSignBit; 1680 InDemandedBits = InDemandedBits.trunc(InBits); 1681 1682 if (SimplifyDemandedBits(Op.getOperand(0), InDemandedBits, KnownZero, 1683 KnownOne, TLO, Depth+1)) 1684 return true; 1685 KnownZero = KnownZero.zext(BitWidth); 1686 KnownOne = KnownOne.zext(BitWidth); 1687 1688 // If the sign bit is known zero, convert this to a zero extend. 1689 if (KnownZero.intersects(InSignBit)) 1690 return TLO.CombineTo(Op, TLO.DAG.getNode(ISD::ZERO_EXTEND, dl, 1691 Op.getValueType(), 1692 Op.getOperand(0))); 1693 1694 // If the sign bit is known one, the top bits match. 1695 if (KnownOne.intersects(InSignBit)) { 1696 KnownOne |= NewBits; 1697 assert((KnownZero & NewBits) == 0); 1698 } else { // Otherwise, top bits aren't known. 1699 assert((KnownOne & NewBits) == 0); 1700 assert((KnownZero & NewBits) == 0); 1701 } 1702 break; 1703 } 1704 case ISD::ANY_EXTEND: { 1705 unsigned OperandBitWidth = 1706 Op.getOperand(0).getValueType().getScalarType().getSizeInBits(); 1707 APInt InMask = NewMask.trunc(OperandBitWidth); 1708 if (SimplifyDemandedBits(Op.getOperand(0), InMask, 1709 KnownZero, KnownOne, TLO, Depth+1)) 1710 return true; 1711 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?"); 1712 KnownZero = KnownZero.zext(BitWidth); 1713 KnownOne = KnownOne.zext(BitWidth); 1714 break; 1715 } 1716 case ISD::TRUNCATE: { 1717 // Simplify the input, using demanded bit information, and compute the known 1718 // zero/one bits live out. 1719 unsigned OperandBitWidth = 1720 Op.getOperand(0).getValueType().getScalarType().getSizeInBits(); 1721 APInt TruncMask = NewMask.zext(OperandBitWidth); 1722 if (SimplifyDemandedBits(Op.getOperand(0), TruncMask, 1723 KnownZero, KnownOne, TLO, Depth+1)) 1724 return true; 1725 KnownZero = KnownZero.trunc(BitWidth); 1726 KnownOne = KnownOne.trunc(BitWidth); 1727 1728 // If the input is only used by this truncate, see if we can shrink it based 1729 // on the known demanded bits. 1730 if (Op.getOperand(0).getNode()->hasOneUse()) { 1731 SDValue In = Op.getOperand(0); 1732 switch (In.getOpcode()) { 1733 default: break; 1734 case ISD::SRL: 1735 // Shrink SRL by a constant if none of the high bits shifted in are 1736 // demanded. 1737 if (TLO.LegalTypes() && 1738 !isTypeDesirableForOp(ISD::SRL, Op.getValueType())) 1739 // Do not turn (vt1 truncate (vt2 srl)) into (vt1 srl) if vt1 is 1740 // undesirable. 1741 break; 1742 ConstantSDNode *ShAmt = dyn_cast<ConstantSDNode>(In.getOperand(1)); 1743 if (!ShAmt) 1744 break; 1745 SDValue Shift = In.getOperand(1); 1746 if (TLO.LegalTypes()) { 1747 uint64_t ShVal = ShAmt->getZExtValue(); 1748 Shift = 1749 TLO.DAG.getConstant(ShVal, getShiftAmountTy(Op.getValueType())); 1750 } 1751 1752 APInt HighBits = APInt::getHighBitsSet(OperandBitWidth, 1753 OperandBitWidth - BitWidth); 1754 HighBits = HighBits.lshr(ShAmt->getZExtValue()).trunc(BitWidth); 1755 1756 if (ShAmt->getZExtValue() < BitWidth && !(HighBits & NewMask)) { 1757 // None of the shifted in bits are needed. Add a truncate of the 1758 // shift input, then shift it. 1759 SDValue NewTrunc = TLO.DAG.getNode(ISD::TRUNCATE, dl, 1760 Op.getValueType(), 1761 In.getOperand(0)); 1762 return TLO.CombineTo(Op, TLO.DAG.getNode(ISD::SRL, dl, 1763 Op.getValueType(), 1764 NewTrunc, 1765 Shift)); 1766 } 1767 break; 1768 } 1769 } 1770 1771 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?"); 1772 break; 1773 } 1774 case ISD::AssertZext: { 1775 // AssertZext demands all of the high bits, plus any of the low bits 1776 // demanded by its users. 1777 EVT VT = cast<VTSDNode>(Op.getOperand(1))->getVT(); 1778 APInt InMask = APInt::getLowBitsSet(BitWidth, 1779 VT.getSizeInBits()); 1780 if (SimplifyDemandedBits(Op.getOperand(0), ~InMask | NewMask, 1781 KnownZero, KnownOne, TLO, Depth+1)) 1782 return true; 1783 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?"); 1784 1785 KnownZero |= ~InMask & NewMask; 1786 break; 1787 } 1788 case ISD::BITCAST: 1789 // If this is an FP->Int bitcast and if the sign bit is the only 1790 // thing demanded, turn this into a FGETSIGN. 1791 if (!TLO.LegalOperations() && 1792 !Op.getValueType().isVector() && 1793 !Op.getOperand(0).getValueType().isVector() && 1794 NewMask == APInt::getSignBit(Op.getValueType().getSizeInBits()) && 1795 Op.getOperand(0).getValueType().isFloatingPoint()) { 1796 bool OpVTLegal = isOperationLegalOrCustom(ISD::FGETSIGN, Op.getValueType()); 1797 bool i32Legal = isOperationLegalOrCustom(ISD::FGETSIGN, MVT::i32); 1798 if ((OpVTLegal || i32Legal) && Op.getValueType().isSimple()) { 1799 EVT Ty = OpVTLegal ? Op.getValueType() : MVT::i32; 1800 // Make a FGETSIGN + SHL to move the sign bit into the appropriate 1801 // place. We expect the SHL to be eliminated by other optimizations. 1802 SDValue Sign = TLO.DAG.getNode(ISD::FGETSIGN, dl, Ty, Op.getOperand(0)); 1803 unsigned OpVTSizeInBits = Op.getValueType().getSizeInBits(); 1804 if (!OpVTLegal && OpVTSizeInBits > 32) 1805 Sign = TLO.DAG.getNode(ISD::ZERO_EXTEND, dl, Op.getValueType(), Sign); 1806 unsigned ShVal = Op.getValueType().getSizeInBits()-1; 1807 SDValue ShAmt = TLO.DAG.getConstant(ShVal, Op.getValueType()); 1808 return TLO.CombineTo(Op, TLO.DAG.getNode(ISD::SHL, dl, 1809 Op.getValueType(), 1810 Sign, ShAmt)); 1811 } 1812 } 1813 break; 1814 case ISD::ADD: 1815 case ISD::MUL: 1816 case ISD::SUB: { 1817 // Add, Sub, and Mul don't demand any bits in positions beyond that 1818 // of the highest bit demanded of them. 1819 APInt LoMask = APInt::getLowBitsSet(BitWidth, 1820 BitWidth - NewMask.countLeadingZeros()); 1821 if (SimplifyDemandedBits(Op.getOperand(0), LoMask, KnownZero2, 1822 KnownOne2, TLO, Depth+1)) 1823 return true; 1824 if (SimplifyDemandedBits(Op.getOperand(1), LoMask, KnownZero2, 1825 KnownOne2, TLO, Depth+1)) 1826 return true; 1827 // See if the operation should be performed at a smaller bit width. 1828 if (TLO.ShrinkDemandedOp(Op, BitWidth, NewMask, dl)) 1829 return true; 1830 } 1831 // FALL THROUGH 1832 default: 1833 // Just use ComputeMaskedBits to compute output bits. 1834 TLO.DAG.ComputeMaskedBits(Op, KnownZero, KnownOne, Depth); 1835 break; 1836 } 1837 1838 // If we know the value of all of the demanded bits, return this as a 1839 // constant. 1840 if ((NewMask & (KnownZero|KnownOne)) == NewMask) 1841 return TLO.CombineTo(Op, TLO.DAG.getConstant(KnownOne, Op.getValueType())); 1842 1843 return false; 1844} 1845 1846/// computeMaskedBitsForTargetNode - Determine which of the bits specified 1847/// in Mask are known to be either zero or one and return them in the 1848/// KnownZero/KnownOne bitsets. 1849void TargetLowering::computeMaskedBitsForTargetNode(const SDValue Op, 1850 APInt &KnownZero, 1851 APInt &KnownOne, 1852 const SelectionDAG &DAG, 1853 unsigned Depth) const { 1854 assert((Op.getOpcode() >= ISD::BUILTIN_OP_END || 1855 Op.getOpcode() == ISD::INTRINSIC_WO_CHAIN || 1856 Op.getOpcode() == ISD::INTRINSIC_W_CHAIN || 1857 Op.getOpcode() == ISD::INTRINSIC_VOID) && 1858 "Should use MaskedValueIsZero if you don't know whether Op" 1859 " is a target node!"); 1860 KnownZero = KnownOne = APInt(KnownOne.getBitWidth(), 0); 1861} 1862 1863/// ComputeNumSignBitsForTargetNode - This method can be implemented by 1864/// targets that want to expose additional information about sign bits to the 1865/// DAG Combiner. 1866unsigned TargetLowering::ComputeNumSignBitsForTargetNode(SDValue Op, 1867 unsigned Depth) const { 1868 assert((Op.getOpcode() >= ISD::BUILTIN_OP_END || 1869 Op.getOpcode() == ISD::INTRINSIC_WO_CHAIN || 1870 Op.getOpcode() == ISD::INTRINSIC_W_CHAIN || 1871 Op.getOpcode() == ISD::INTRINSIC_VOID) && 1872 "Should use ComputeNumSignBits if you don't know whether Op" 1873 " is a target node!"); 1874 return 1; 1875} 1876 1877/// ValueHasExactlyOneBitSet - Test if the given value is known to have exactly 1878/// one bit set. This differs from ComputeMaskedBits in that it doesn't need to 1879/// determine which bit is set. 1880/// 1881static bool ValueHasExactlyOneBitSet(SDValue Val, const SelectionDAG &DAG) { 1882 // A left-shift of a constant one will have exactly one bit set, because 1883 // shifting the bit off the end is undefined. 1884 if (Val.getOpcode() == ISD::SHL) 1885 if (ConstantSDNode *C = 1886 dyn_cast<ConstantSDNode>(Val.getNode()->getOperand(0))) 1887 if (C->getAPIntValue() == 1) 1888 return true; 1889 1890 // Similarly, a right-shift of a constant sign-bit will have exactly 1891 // one bit set. 1892 if (Val.getOpcode() == ISD::SRL) 1893 if (ConstantSDNode *C = 1894 dyn_cast<ConstantSDNode>(Val.getNode()->getOperand(0))) 1895 if (C->getAPIntValue().isSignBit()) 1896 return true; 1897 1898 // More could be done here, though the above checks are enough 1899 // to handle some common cases. 1900 1901 // Fall back to ComputeMaskedBits to catch other known cases. 1902 EVT OpVT = Val.getValueType(); 1903 unsigned BitWidth = OpVT.getScalarType().getSizeInBits(); 1904 APInt KnownZero, KnownOne; 1905 DAG.ComputeMaskedBits(Val, KnownZero, KnownOne); 1906 return (KnownZero.countPopulation() == BitWidth - 1) && 1907 (KnownOne.countPopulation() == 1); 1908} 1909 1910/// SimplifySetCC - Try to simplify a setcc built with the specified operands 1911/// and cc. If it is unable to simplify it, return a null SDValue. 1912SDValue 1913TargetLowering::SimplifySetCC(EVT VT, SDValue N0, SDValue N1, 1914 ISD::CondCode Cond, bool foldBooleans, 1915 DAGCombinerInfo &DCI, DebugLoc dl) const { 1916 SelectionDAG &DAG = DCI.DAG; 1917 1918 // These setcc operations always fold. 1919 switch (Cond) { 1920 default: break; 1921 case ISD::SETFALSE: 1922 case ISD::SETFALSE2: return DAG.getConstant(0, VT); 1923 case ISD::SETTRUE: 1924 case ISD::SETTRUE2: return DAG.getConstant(1, VT); 1925 } 1926 1927 // Ensure that the constant occurs on the RHS, and fold constant 1928 // comparisons. 1929 if (isa<ConstantSDNode>(N0.getNode())) 1930 return DAG.getSetCC(dl, VT, N1, N0, ISD::getSetCCSwappedOperands(Cond)); 1931 1932 if (ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1.getNode())) { 1933 const APInt &C1 = N1C->getAPIntValue(); 1934 1935 // If the LHS is '(srl (ctlz x), 5)', the RHS is 0/1, and this is an 1936 // equality comparison, then we're just comparing whether X itself is 1937 // zero. 1938 if (N0.getOpcode() == ISD::SRL && (C1 == 0 || C1 == 1) && 1939 N0.getOperand(0).getOpcode() == ISD::CTLZ && 1940 N0.getOperand(1).getOpcode() == ISD::Constant) { 1941 const APInt &ShAmt 1942 = cast<ConstantSDNode>(N0.getOperand(1))->getAPIntValue(); 1943 if ((Cond == ISD::SETEQ || Cond == ISD::SETNE) && 1944 ShAmt == Log2_32(N0.getValueType().getSizeInBits())) { 1945 if ((C1 == 0) == (Cond == ISD::SETEQ)) { 1946 // (srl (ctlz x), 5) == 0 -> X != 0 1947 // (srl (ctlz x), 5) != 1 -> X != 0 1948 Cond = ISD::SETNE; 1949 } else { 1950 // (srl (ctlz x), 5) != 0 -> X == 0 1951 // (srl (ctlz x), 5) == 1 -> X == 0 1952 Cond = ISD::SETEQ; 1953 } 1954 SDValue Zero = DAG.getConstant(0, N0.getValueType()); 1955 return DAG.getSetCC(dl, VT, N0.getOperand(0).getOperand(0), 1956 Zero, Cond); 1957 } 1958 } 1959 1960 SDValue CTPOP = N0; 1961 // Look through truncs that don't change the value of a ctpop. 1962 if (N0.hasOneUse() && N0.getOpcode() == ISD::TRUNCATE) 1963 CTPOP = N0.getOperand(0); 1964 1965 if (CTPOP.hasOneUse() && CTPOP.getOpcode() == ISD::CTPOP && 1966 (N0 == CTPOP || N0.getValueType().getSizeInBits() > 1967 Log2_32_Ceil(CTPOP.getValueType().getSizeInBits()))) { 1968 EVT CTVT = CTPOP.getValueType(); 1969 SDValue CTOp = CTPOP.getOperand(0); 1970 1971 // (ctpop x) u< 2 -> (x & x-1) == 0 1972 // (ctpop x) u> 1 -> (x & x-1) != 0 1973 if ((Cond == ISD::SETULT && C1 == 2) || (Cond == ISD::SETUGT && C1 == 1)){ 1974 SDValue Sub = DAG.getNode(ISD::SUB, dl, CTVT, CTOp, 1975 DAG.getConstant(1, CTVT)); 1976 SDValue And = DAG.getNode(ISD::AND, dl, CTVT, CTOp, Sub); 1977 ISD::CondCode CC = Cond == ISD::SETULT ? ISD::SETEQ : ISD::SETNE; 1978 return DAG.getSetCC(dl, VT, And, DAG.getConstant(0, CTVT), CC); 1979 } 1980 1981 // TODO: (ctpop x) == 1 -> x && (x & x-1) == 0 iff ctpop is illegal. 1982 } 1983 1984 // (zext x) == C --> x == (trunc C) 1985 if (DCI.isBeforeLegalize() && N0->hasOneUse() && 1986 (Cond == ISD::SETEQ || Cond == ISD::SETNE)) { 1987 unsigned MinBits = N0.getValueSizeInBits(); 1988 SDValue PreZExt; 1989 if (N0->getOpcode() == ISD::ZERO_EXTEND) { 1990 // ZExt 1991 MinBits = N0->getOperand(0).getValueSizeInBits(); 1992 PreZExt = N0->getOperand(0); 1993 } else if (N0->getOpcode() == ISD::AND) { 1994 // DAGCombine turns costly ZExts into ANDs 1995 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(N0->getOperand(1))) 1996 if ((C->getAPIntValue()+1).isPowerOf2()) { 1997 MinBits = C->getAPIntValue().countTrailingOnes(); 1998 PreZExt = N0->getOperand(0); 1999 } 2000 } else if (LoadSDNode *LN0 = dyn_cast<LoadSDNode>(N0)) { 2001 // ZEXTLOAD 2002 if (LN0->getExtensionType() == ISD::ZEXTLOAD) { 2003 MinBits = LN0->getMemoryVT().getSizeInBits(); 2004 PreZExt = N0; 2005 } 2006 } 2007 2008 // Make sure we're not losing bits from the constant. 2009 if (MinBits < C1.getBitWidth() && MinBits > C1.getActiveBits()) { 2010 EVT MinVT = EVT::getIntegerVT(*DAG.getContext(), MinBits); 2011 if (isTypeDesirableForOp(ISD::SETCC, MinVT)) { 2012 // Will get folded away. 2013 SDValue Trunc = DAG.getNode(ISD::TRUNCATE, dl, MinVT, PreZExt); 2014 SDValue C = DAG.getConstant(C1.trunc(MinBits), MinVT); 2015 return DAG.getSetCC(dl, VT, Trunc, C, Cond); 2016 } 2017 } 2018 } 2019 2020 // If the LHS is '(and load, const)', the RHS is 0, 2021 // the test is for equality or unsigned, and all 1 bits of the const are 2022 // in the same partial word, see if we can shorten the load. 2023 if (DCI.isBeforeLegalize() && 2024 N0.getOpcode() == ISD::AND && C1 == 0 && 2025 N0.getNode()->hasOneUse() && 2026 isa<LoadSDNode>(N0.getOperand(0)) && 2027 N0.getOperand(0).getNode()->hasOneUse() && 2028 isa<ConstantSDNode>(N0.getOperand(1))) { 2029 LoadSDNode *Lod = cast<LoadSDNode>(N0.getOperand(0)); 2030 APInt bestMask; 2031 unsigned bestWidth = 0, bestOffset = 0; 2032 if (!Lod->isVolatile() && Lod->isUnindexed()) { 2033 unsigned origWidth = N0.getValueType().getSizeInBits(); 2034 unsigned maskWidth = origWidth; 2035 // We can narrow (e.g.) 16-bit extending loads on 32-bit target to 2036 // 8 bits, but have to be careful... 2037 if (Lod->getExtensionType() != ISD::NON_EXTLOAD) 2038 origWidth = Lod->getMemoryVT().getSizeInBits(); 2039 const APInt &Mask = 2040 cast<ConstantSDNode>(N0.getOperand(1))->getAPIntValue(); 2041 for (unsigned width = origWidth / 2; width>=8; width /= 2) { 2042 APInt newMask = APInt::getLowBitsSet(maskWidth, width); 2043 for (unsigned offset=0; offset<origWidth/width; offset++) { 2044 if ((newMask & Mask) == Mask) { 2045 if (!TD->isLittleEndian()) 2046 bestOffset = (origWidth/width - offset - 1) * (width/8); 2047 else 2048 bestOffset = (uint64_t)offset * (width/8); 2049 bestMask = Mask.lshr(offset * (width/8) * 8); 2050 bestWidth = width; 2051 break; 2052 } 2053 newMask = newMask << width; 2054 } 2055 } 2056 } 2057 if (bestWidth) { 2058 EVT newVT = EVT::getIntegerVT(*DAG.getContext(), bestWidth); 2059 if (newVT.isRound()) { 2060 EVT PtrType = Lod->getOperand(1).getValueType(); 2061 SDValue Ptr = Lod->getBasePtr(); 2062 if (bestOffset != 0) 2063 Ptr = DAG.getNode(ISD::ADD, dl, PtrType, Lod->getBasePtr(), 2064 DAG.getConstant(bestOffset, PtrType)); 2065 unsigned NewAlign = MinAlign(Lod->getAlignment(), bestOffset); 2066 SDValue NewLoad = DAG.getLoad(newVT, dl, Lod->getChain(), Ptr, 2067 Lod->getPointerInfo().getWithOffset(bestOffset), 2068 false, false, false, NewAlign); 2069 return DAG.getSetCC(dl, VT, 2070 DAG.getNode(ISD::AND, dl, newVT, NewLoad, 2071 DAG.getConstant(bestMask.trunc(bestWidth), 2072 newVT)), 2073 DAG.getConstant(0LL, newVT), Cond); 2074 } 2075 } 2076 } 2077 2078 // If the LHS is a ZERO_EXTEND, perform the comparison on the input. 2079 if (N0.getOpcode() == ISD::ZERO_EXTEND) { 2080 unsigned InSize = N0.getOperand(0).getValueType().getSizeInBits(); 2081 2082 // If the comparison constant has bits in the upper part, the 2083 // zero-extended value could never match. 2084 if (C1.intersects(APInt::getHighBitsSet(C1.getBitWidth(), 2085 C1.getBitWidth() - InSize))) { 2086 switch (Cond) { 2087 case ISD::SETUGT: 2088 case ISD::SETUGE: 2089 case ISD::SETEQ: return DAG.getConstant(0, VT); 2090 case ISD::SETULT: 2091 case ISD::SETULE: 2092 case ISD::SETNE: return DAG.getConstant(1, VT); 2093 case ISD::SETGT: 2094 case ISD::SETGE: 2095 // True if the sign bit of C1 is set. 2096 return DAG.getConstant(C1.isNegative(), VT); 2097 case ISD::SETLT: 2098 case ISD::SETLE: 2099 // True if the sign bit of C1 isn't set. 2100 return DAG.getConstant(C1.isNonNegative(), VT); 2101 default: 2102 break; 2103 } 2104 } 2105 2106 // Otherwise, we can perform the comparison with the low bits. 2107 switch (Cond) { 2108 case ISD::SETEQ: 2109 case ISD::SETNE: 2110 case ISD::SETUGT: 2111 case ISD::SETUGE: 2112 case ISD::SETULT: 2113 case ISD::SETULE: { 2114 EVT newVT = N0.getOperand(0).getValueType(); 2115 if (DCI.isBeforeLegalizeOps() || 2116 (isOperationLegal(ISD::SETCC, newVT) && 2117 getCondCodeAction(Cond, newVT.getSimpleVT())==Legal)) 2118 return DAG.getSetCC(dl, VT, N0.getOperand(0), 2119 DAG.getConstant(C1.trunc(InSize), newVT), 2120 Cond); 2121 break; 2122 } 2123 default: 2124 break; // todo, be more careful with signed comparisons 2125 } 2126 } else if (N0.getOpcode() == ISD::SIGN_EXTEND_INREG && 2127 (Cond == ISD::SETEQ || Cond == ISD::SETNE)) { 2128 EVT ExtSrcTy = cast<VTSDNode>(N0.getOperand(1))->getVT(); 2129 unsigned ExtSrcTyBits = ExtSrcTy.getSizeInBits(); 2130 EVT ExtDstTy = N0.getValueType(); 2131 unsigned ExtDstTyBits = ExtDstTy.getSizeInBits(); 2132 2133 // If the constant doesn't fit into the number of bits for the source of 2134 // the sign extension, it is impossible for both sides to be equal. 2135 if (C1.getMinSignedBits() > ExtSrcTyBits) 2136 return DAG.getConstant(Cond == ISD::SETNE, VT); 2137 2138 SDValue ZextOp; 2139 EVT Op0Ty = N0.getOperand(0).getValueType(); 2140 if (Op0Ty == ExtSrcTy) { 2141 ZextOp = N0.getOperand(0); 2142 } else { 2143 APInt Imm = APInt::getLowBitsSet(ExtDstTyBits, ExtSrcTyBits); 2144 ZextOp = DAG.getNode(ISD::AND, dl, Op0Ty, N0.getOperand(0), 2145 DAG.getConstant(Imm, Op0Ty)); 2146 } 2147 if (!DCI.isCalledByLegalizer()) 2148 DCI.AddToWorklist(ZextOp.getNode()); 2149 // Otherwise, make this a use of a zext. 2150 return DAG.getSetCC(dl, VT, ZextOp, 2151 DAG.getConstant(C1 & APInt::getLowBitsSet( 2152 ExtDstTyBits, 2153 ExtSrcTyBits), 2154 ExtDstTy), 2155 Cond); 2156 } else if ((N1C->isNullValue() || N1C->getAPIntValue() == 1) && 2157 (Cond == ISD::SETEQ || Cond == ISD::SETNE)) { 2158 // SETCC (SETCC), [0|1], [EQ|NE] -> SETCC 2159 if (N0.getOpcode() == ISD::SETCC && 2160 isTypeLegal(VT) && VT.bitsLE(N0.getValueType())) { 2161 bool TrueWhenTrue = (Cond == ISD::SETEQ) ^ (N1C->getAPIntValue() != 1); 2162 if (TrueWhenTrue) 2163 return DAG.getNode(ISD::TRUNCATE, dl, VT, N0); 2164 // Invert the condition. 2165 ISD::CondCode CC = cast<CondCodeSDNode>(N0.getOperand(2))->get(); 2166 CC = ISD::getSetCCInverse(CC, 2167 N0.getOperand(0).getValueType().isInteger()); 2168 return DAG.getSetCC(dl, VT, N0.getOperand(0), N0.getOperand(1), CC); 2169 } 2170 2171 if ((N0.getOpcode() == ISD::XOR || 2172 (N0.getOpcode() == ISD::AND && 2173 N0.getOperand(0).getOpcode() == ISD::XOR && 2174 N0.getOperand(1) == N0.getOperand(0).getOperand(1))) && 2175 isa<ConstantSDNode>(N0.getOperand(1)) && 2176 cast<ConstantSDNode>(N0.getOperand(1))->getAPIntValue() == 1) { 2177 // If this is (X^1) == 0/1, swap the RHS and eliminate the xor. We 2178 // can only do this if the top bits are known zero. 2179 unsigned BitWidth = N0.getValueSizeInBits(); 2180 if (DAG.MaskedValueIsZero(N0, 2181 APInt::getHighBitsSet(BitWidth, 2182 BitWidth-1))) { 2183 // Okay, get the un-inverted input value. 2184 SDValue Val; 2185 if (N0.getOpcode() == ISD::XOR) 2186 Val = N0.getOperand(0); 2187 else { 2188 assert(N0.getOpcode() == ISD::AND && 2189 N0.getOperand(0).getOpcode() == ISD::XOR); 2190 // ((X^1)&1)^1 -> X & 1 2191 Val = DAG.getNode(ISD::AND, dl, N0.getValueType(), 2192 N0.getOperand(0).getOperand(0), 2193 N0.getOperand(1)); 2194 } 2195 2196 return DAG.getSetCC(dl, VT, Val, N1, 2197 Cond == ISD::SETEQ ? ISD::SETNE : ISD::SETEQ); 2198 } 2199 } else if (N1C->getAPIntValue() == 1 && 2200 (VT == MVT::i1 || 2201 getBooleanContents(false) == ZeroOrOneBooleanContent)) { 2202 SDValue Op0 = N0; 2203 if (Op0.getOpcode() == ISD::TRUNCATE) 2204 Op0 = Op0.getOperand(0); 2205 2206 if ((Op0.getOpcode() == ISD::XOR) && 2207 Op0.getOperand(0).getOpcode() == ISD::SETCC && 2208 Op0.getOperand(1).getOpcode() == ISD::SETCC) { 2209 // (xor (setcc), (setcc)) == / != 1 -> (setcc) != / == (setcc) 2210 Cond = (Cond == ISD::SETEQ) ? ISD::SETNE : ISD::SETEQ; 2211 return DAG.getSetCC(dl, VT, Op0.getOperand(0), Op0.getOperand(1), 2212 Cond); 2213 } 2214 if (Op0.getOpcode() == ISD::AND && 2215 isa<ConstantSDNode>(Op0.getOperand(1)) && 2216 cast<ConstantSDNode>(Op0.getOperand(1))->getAPIntValue() == 1) { 2217 // If this is (X&1) == / != 1, normalize it to (X&1) != / == 0. 2218 if (Op0.getValueType().bitsGT(VT)) 2219 Op0 = DAG.getNode(ISD::AND, dl, VT, 2220 DAG.getNode(ISD::TRUNCATE, dl, VT, Op0.getOperand(0)), 2221 DAG.getConstant(1, VT)); 2222 else if (Op0.getValueType().bitsLT(VT)) 2223 Op0 = DAG.getNode(ISD::AND, dl, VT, 2224 DAG.getNode(ISD::ANY_EXTEND, dl, VT, Op0.getOperand(0)), 2225 DAG.getConstant(1, VT)); 2226 2227 return DAG.getSetCC(dl, VT, Op0, 2228 DAG.getConstant(0, Op0.getValueType()), 2229 Cond == ISD::SETEQ ? ISD::SETNE : ISD::SETEQ); 2230 } 2231 if (Op0.getOpcode() == ISD::AssertZext && 2232 cast<VTSDNode>(Op0.getOperand(1))->getVT() == MVT::i1) 2233 return DAG.getSetCC(dl, VT, Op0, 2234 DAG.getConstant(0, Op0.getValueType()), 2235 Cond == ISD::SETEQ ? ISD::SETNE : ISD::SETEQ); 2236 } 2237 } 2238 2239 APInt MinVal, MaxVal; 2240 unsigned OperandBitSize = N1C->getValueType(0).getSizeInBits(); 2241 if (ISD::isSignedIntSetCC(Cond)) { 2242 MinVal = APInt::getSignedMinValue(OperandBitSize); 2243 MaxVal = APInt::getSignedMaxValue(OperandBitSize); 2244 } else { 2245 MinVal = APInt::getMinValue(OperandBitSize); 2246 MaxVal = APInt::getMaxValue(OperandBitSize); 2247 } 2248 2249 // Canonicalize GE/LE comparisons to use GT/LT comparisons. 2250 if (Cond == ISD::SETGE || Cond == ISD::SETUGE) { 2251 if (C1 == MinVal) return DAG.getConstant(1, VT); // X >= MIN --> true 2252 // X >= C0 --> X > (C0-1) 2253 return DAG.getSetCC(dl, VT, N0, 2254 DAG.getConstant(C1-1, N1.getValueType()), 2255 (Cond == ISD::SETGE) ? ISD::SETGT : ISD::SETUGT); 2256 } 2257 2258 if (Cond == ISD::SETLE || Cond == ISD::SETULE) { 2259 if (C1 == MaxVal) return DAG.getConstant(1, VT); // X <= MAX --> true 2260 // X <= C0 --> X < (C0+1) 2261 return DAG.getSetCC(dl, VT, N0, 2262 DAG.getConstant(C1+1, N1.getValueType()), 2263 (Cond == ISD::SETLE) ? ISD::SETLT : ISD::SETULT); 2264 } 2265 2266 if ((Cond == ISD::SETLT || Cond == ISD::SETULT) && C1 == MinVal) 2267 return DAG.getConstant(0, VT); // X < MIN --> false 2268 if ((Cond == ISD::SETGE || Cond == ISD::SETUGE) && C1 == MinVal) 2269 return DAG.getConstant(1, VT); // X >= MIN --> true 2270 if ((Cond == ISD::SETGT || Cond == ISD::SETUGT) && C1 == MaxVal) 2271 return DAG.getConstant(0, VT); // X > MAX --> false 2272 if ((Cond == ISD::SETLE || Cond == ISD::SETULE) && C1 == MaxVal) 2273 return DAG.getConstant(1, VT); // X <= MAX --> true 2274 2275 // Canonicalize setgt X, Min --> setne X, Min 2276 if ((Cond == ISD::SETGT || Cond == ISD::SETUGT) && C1 == MinVal) 2277 return DAG.getSetCC(dl, VT, N0, N1, ISD::SETNE); 2278 // Canonicalize setlt X, Max --> setne X, Max 2279 if ((Cond == ISD::SETLT || Cond == ISD::SETULT) && C1 == MaxVal) 2280 return DAG.getSetCC(dl, VT, N0, N1, ISD::SETNE); 2281 2282 // If we have setult X, 1, turn it into seteq X, 0 2283 if ((Cond == ISD::SETLT || Cond == ISD::SETULT) && C1 == MinVal+1) 2284 return DAG.getSetCC(dl, VT, N0, 2285 DAG.getConstant(MinVal, N0.getValueType()), 2286 ISD::SETEQ); 2287 // If we have setugt X, Max-1, turn it into seteq X, Max 2288 if ((Cond == ISD::SETGT || Cond == ISD::SETUGT) && C1 == MaxVal-1) 2289 return DAG.getSetCC(dl, VT, N0, 2290 DAG.getConstant(MaxVal, N0.getValueType()), 2291 ISD::SETEQ); 2292 2293 // If we have "setcc X, C0", check to see if we can shrink the immediate 2294 // by changing cc. 2295 2296 // SETUGT X, SINTMAX -> SETLT X, 0 2297 if (Cond == ISD::SETUGT && 2298 C1 == APInt::getSignedMaxValue(OperandBitSize)) 2299 return DAG.getSetCC(dl, VT, N0, 2300 DAG.getConstant(0, N1.getValueType()), 2301 ISD::SETLT); 2302 2303 // SETULT X, SINTMIN -> SETGT X, -1 2304 if (Cond == ISD::SETULT && 2305 C1 == APInt::getSignedMinValue(OperandBitSize)) { 2306 SDValue ConstMinusOne = 2307 DAG.getConstant(APInt::getAllOnesValue(OperandBitSize), 2308 N1.getValueType()); 2309 return DAG.getSetCC(dl, VT, N0, ConstMinusOne, ISD::SETGT); 2310 } 2311 2312 // Fold bit comparisons when we can. 2313 if ((Cond == ISD::SETEQ || Cond == ISD::SETNE) && 2314 (VT == N0.getValueType() || 2315 (isTypeLegal(VT) && VT.bitsLE(N0.getValueType()))) && 2316 N0.getOpcode() == ISD::AND) 2317 if (ConstantSDNode *AndRHS = 2318 dyn_cast<ConstantSDNode>(N0.getOperand(1))) { 2319 EVT ShiftTy = DCI.isBeforeLegalizeOps() ? 2320 getPointerTy() : getShiftAmountTy(N0.getValueType()); 2321 if (Cond == ISD::SETNE && C1 == 0) {// (X & 8) != 0 --> (X & 8) >> 3 2322 // Perform the xform if the AND RHS is a single bit. 2323 if (AndRHS->getAPIntValue().isPowerOf2()) { 2324 return DAG.getNode(ISD::TRUNCATE, dl, VT, 2325 DAG.getNode(ISD::SRL, dl, N0.getValueType(), N0, 2326 DAG.getConstant(AndRHS->getAPIntValue().logBase2(), ShiftTy))); 2327 } 2328 } else if (Cond == ISD::SETEQ && C1 == AndRHS->getAPIntValue()) { 2329 // (X & 8) == 8 --> (X & 8) >> 3 2330 // Perform the xform if C1 is a single bit. 2331 if (C1.isPowerOf2()) { 2332 return DAG.getNode(ISD::TRUNCATE, dl, VT, 2333 DAG.getNode(ISD::SRL, dl, N0.getValueType(), N0, 2334 DAG.getConstant(C1.logBase2(), ShiftTy))); 2335 } 2336 } 2337 } 2338 2339 if (C1.getMinSignedBits() <= 64 && 2340 !isLegalICmpImmediate(C1.getSExtValue())) { 2341 // (X & -256) == 256 -> (X >> 8) == 1 2342 if ((Cond == ISD::SETEQ || Cond == ISD::SETNE) && 2343 N0.getOpcode() == ISD::AND && N0.hasOneUse()) { 2344 if (ConstantSDNode *AndRHS = 2345 dyn_cast<ConstantSDNode>(N0.getOperand(1))) { 2346 const APInt &AndRHSC = AndRHS->getAPIntValue(); 2347 if ((-AndRHSC).isPowerOf2() && (AndRHSC & C1) == C1) { 2348 unsigned ShiftBits = AndRHSC.countTrailingZeros(); 2349 EVT ShiftTy = DCI.isBeforeLegalizeOps() ? 2350 getPointerTy() : getShiftAmountTy(N0.getValueType()); 2351 EVT CmpTy = N0.getValueType(); 2352 SDValue Shift = DAG.getNode(ISD::SRL, dl, CmpTy, N0.getOperand(0), 2353 DAG.getConstant(ShiftBits, ShiftTy)); 2354 SDValue CmpRHS = DAG.getConstant(C1.lshr(ShiftBits), CmpTy); 2355 return DAG.getSetCC(dl, VT, Shift, CmpRHS, Cond); 2356 } 2357 } 2358 } else if (Cond == ISD::SETULT || Cond == ISD::SETUGE || 2359 Cond == ISD::SETULE || Cond == ISD::SETUGT) { 2360 bool AdjOne = (Cond == ISD::SETULE || Cond == ISD::SETUGT); 2361 // X < 0x100000000 -> (X >> 32) < 1 2362 // X >= 0x100000000 -> (X >> 32) >= 1 2363 // X <= 0x0ffffffff -> (X >> 32) < 1 2364 // X > 0x0ffffffff -> (X >> 32) >= 1 2365 unsigned ShiftBits; 2366 APInt NewC = C1; 2367 ISD::CondCode NewCond = Cond; 2368 if (AdjOne) { 2369 ShiftBits = C1.countTrailingOnes(); 2370 NewC = NewC + 1; 2371 NewCond = (Cond == ISD::SETULE) ? ISD::SETULT : ISD::SETUGE; 2372 } else { 2373 ShiftBits = C1.countTrailingZeros(); 2374 } 2375 NewC = NewC.lshr(ShiftBits); 2376 if (ShiftBits && isLegalICmpImmediate(NewC.getSExtValue())) { 2377 EVT ShiftTy = DCI.isBeforeLegalizeOps() ? 2378 getPointerTy() : getShiftAmountTy(N0.getValueType()); 2379 EVT CmpTy = N0.getValueType(); 2380 SDValue Shift = DAG.getNode(ISD::SRL, dl, CmpTy, N0, 2381 DAG.getConstant(ShiftBits, ShiftTy)); 2382 SDValue CmpRHS = DAG.getConstant(NewC, CmpTy); 2383 return DAG.getSetCC(dl, VT, Shift, CmpRHS, NewCond); 2384 } 2385 } 2386 } 2387 } 2388 2389 if (isa<ConstantFPSDNode>(N0.getNode())) { 2390 // Constant fold or commute setcc. 2391 SDValue O = DAG.FoldSetCC(VT, N0, N1, Cond, dl); 2392 if (O.getNode()) return O; 2393 } else if (ConstantFPSDNode *CFP = dyn_cast<ConstantFPSDNode>(N1.getNode())) { 2394 // If the RHS of an FP comparison is a constant, simplify it away in 2395 // some cases. 2396 if (CFP->getValueAPF().isNaN()) { 2397 // If an operand is known to be a nan, we can fold it. 2398 switch (ISD::getUnorderedFlavor(Cond)) { 2399 default: llvm_unreachable("Unknown flavor!"); 2400 case 0: // Known false. 2401 return DAG.getConstant(0, VT); 2402 case 1: // Known true. 2403 return DAG.getConstant(1, VT); 2404 case 2: // Undefined. 2405 return DAG.getUNDEF(VT); 2406 } 2407 } 2408 2409 // Otherwise, we know the RHS is not a NaN. Simplify the node to drop the 2410 // constant if knowing that the operand is non-nan is enough. We prefer to 2411 // have SETO(x,x) instead of SETO(x, 0.0) because this avoids having to 2412 // materialize 0.0. 2413 if (Cond == ISD::SETO || Cond == ISD::SETUO) 2414 return DAG.getSetCC(dl, VT, N0, N0, Cond); 2415 2416 // If the condition is not legal, see if we can find an equivalent one 2417 // which is legal. 2418 if (!isCondCodeLegal(Cond, N0.getSimpleValueType())) { 2419 // If the comparison was an awkward floating-point == or != and one of 2420 // the comparison operands is infinity or negative infinity, convert the 2421 // condition to a less-awkward <= or >=. 2422 if (CFP->getValueAPF().isInfinity()) { 2423 if (CFP->getValueAPF().isNegative()) { 2424 if (Cond == ISD::SETOEQ && 2425 isCondCodeLegal(ISD::SETOLE, N0.getSimpleValueType())) 2426 return DAG.getSetCC(dl, VT, N0, N1, ISD::SETOLE); 2427 if (Cond == ISD::SETUEQ && 2428 isCondCodeLegal(ISD::SETOLE, N0.getSimpleValueType())) 2429 return DAG.getSetCC(dl, VT, N0, N1, ISD::SETULE); 2430 if (Cond == ISD::SETUNE && 2431 isCondCodeLegal(ISD::SETUGT, N0.getSimpleValueType())) 2432 return DAG.getSetCC(dl, VT, N0, N1, ISD::SETUGT); 2433 if (Cond == ISD::SETONE && 2434 isCondCodeLegal(ISD::SETUGT, N0.getSimpleValueType())) 2435 return DAG.getSetCC(dl, VT, N0, N1, ISD::SETOGT); 2436 } else { 2437 if (Cond == ISD::SETOEQ && 2438 isCondCodeLegal(ISD::SETOGE, N0.getSimpleValueType())) 2439 return DAG.getSetCC(dl, VT, N0, N1, ISD::SETOGE); 2440 if (Cond == ISD::SETUEQ && 2441 isCondCodeLegal(ISD::SETOGE, N0.getSimpleValueType())) 2442 return DAG.getSetCC(dl, VT, N0, N1, ISD::SETUGE); 2443 if (Cond == ISD::SETUNE && 2444 isCondCodeLegal(ISD::SETULT, N0.getSimpleValueType())) 2445 return DAG.getSetCC(dl, VT, N0, N1, ISD::SETULT); 2446 if (Cond == ISD::SETONE && 2447 isCondCodeLegal(ISD::SETULT, N0.getSimpleValueType())) 2448 return DAG.getSetCC(dl, VT, N0, N1, ISD::SETOLT); 2449 } 2450 } 2451 } 2452 } 2453 2454 if (N0 == N1) { 2455 // The sext(setcc()) => setcc() optimization relies on the appropriate 2456 // constant being emitted. 2457 uint64_t EqVal = 0; 2458 switch (getBooleanContents(N0.getValueType().isVector())) { 2459 case UndefinedBooleanContent: 2460 case ZeroOrOneBooleanContent: 2461 EqVal = ISD::isTrueWhenEqual(Cond); 2462 break; 2463 case ZeroOrNegativeOneBooleanContent: 2464 EqVal = ISD::isTrueWhenEqual(Cond) ? -1 : 0; 2465 break; 2466 } 2467 2468 // We can always fold X == X for integer setcc's. 2469 if (N0.getValueType().isInteger()) { 2470 return DAG.getConstant(EqVal, VT); 2471 } 2472 unsigned UOF = ISD::getUnorderedFlavor(Cond); 2473 if (UOF == 2) // FP operators that are undefined on NaNs. 2474 return DAG.getConstant(EqVal, VT); 2475 if (UOF == unsigned(ISD::isTrueWhenEqual(Cond))) 2476 return DAG.getConstant(EqVal, VT); 2477 // Otherwise, we can't fold it. However, we can simplify it to SETUO/SETO 2478 // if it is not already. 2479 ISD::CondCode NewCond = UOF == 0 ? ISD::SETO : ISD::SETUO; 2480 if (NewCond != Cond && (DCI.isBeforeLegalizeOps() || 2481 getCondCodeAction(NewCond, N0.getSimpleValueType()) == Legal)) 2482 return DAG.getSetCC(dl, VT, N0, N1, NewCond); 2483 } 2484 2485 if ((Cond == ISD::SETEQ || Cond == ISD::SETNE) && 2486 N0.getValueType().isInteger()) { 2487 if (N0.getOpcode() == ISD::ADD || N0.getOpcode() == ISD::SUB || 2488 N0.getOpcode() == ISD::XOR) { 2489 // Simplify (X+Y) == (X+Z) --> Y == Z 2490 if (N0.getOpcode() == N1.getOpcode()) { 2491 if (N0.getOperand(0) == N1.getOperand(0)) 2492 return DAG.getSetCC(dl, VT, N0.getOperand(1), N1.getOperand(1), Cond); 2493 if (N0.getOperand(1) == N1.getOperand(1)) 2494 return DAG.getSetCC(dl, VT, N0.getOperand(0), N1.getOperand(0), Cond); 2495 if (DAG.isCommutativeBinOp(N0.getOpcode())) { 2496 // If X op Y == Y op X, try other combinations. 2497 if (N0.getOperand(0) == N1.getOperand(1)) 2498 return DAG.getSetCC(dl, VT, N0.getOperand(1), N1.getOperand(0), 2499 Cond); 2500 if (N0.getOperand(1) == N1.getOperand(0)) 2501 return DAG.getSetCC(dl, VT, N0.getOperand(0), N1.getOperand(1), 2502 Cond); 2503 } 2504 } 2505 2506 // If RHS is a legal immediate value for a compare instruction, we need 2507 // to be careful about increasing register pressure needlessly. 2508 bool LegalRHSImm = false; 2509 2510 if (ConstantSDNode *RHSC = dyn_cast<ConstantSDNode>(N1)) { 2511 if (ConstantSDNode *LHSR = dyn_cast<ConstantSDNode>(N0.getOperand(1))) { 2512 // Turn (X+C1) == C2 --> X == C2-C1 2513 if (N0.getOpcode() == ISD::ADD && N0.getNode()->hasOneUse()) { 2514 return DAG.getSetCC(dl, VT, N0.getOperand(0), 2515 DAG.getConstant(RHSC->getAPIntValue()- 2516 LHSR->getAPIntValue(), 2517 N0.getValueType()), Cond); 2518 } 2519 2520 // Turn (X^C1) == C2 into X == C1^C2 iff X&~C1 = 0. 2521 if (N0.getOpcode() == ISD::XOR) 2522 // If we know that all of the inverted bits are zero, don't bother 2523 // performing the inversion. 2524 if (DAG.MaskedValueIsZero(N0.getOperand(0), ~LHSR->getAPIntValue())) 2525 return 2526 DAG.getSetCC(dl, VT, N0.getOperand(0), 2527 DAG.getConstant(LHSR->getAPIntValue() ^ 2528 RHSC->getAPIntValue(), 2529 N0.getValueType()), 2530 Cond); 2531 } 2532 2533 // Turn (C1-X) == C2 --> X == C1-C2 2534 if (ConstantSDNode *SUBC = dyn_cast<ConstantSDNode>(N0.getOperand(0))) { 2535 if (N0.getOpcode() == ISD::SUB && N0.getNode()->hasOneUse()) { 2536 return 2537 DAG.getSetCC(dl, VT, N0.getOperand(1), 2538 DAG.getConstant(SUBC->getAPIntValue() - 2539 RHSC->getAPIntValue(), 2540 N0.getValueType()), 2541 Cond); 2542 } 2543 } 2544 2545 // Could RHSC fold directly into a compare? 2546 if (RHSC->getValueType(0).getSizeInBits() <= 64) 2547 LegalRHSImm = isLegalICmpImmediate(RHSC->getSExtValue()); 2548 } 2549 2550 // Simplify (X+Z) == X --> Z == 0 2551 // Don't do this if X is an immediate that can fold into a cmp 2552 // instruction and X+Z has other uses. It could be an induction variable 2553 // chain, and the transform would increase register pressure. 2554 if (!LegalRHSImm || N0.getNode()->hasOneUse()) { 2555 if (N0.getOperand(0) == N1) 2556 return DAG.getSetCC(dl, VT, N0.getOperand(1), 2557 DAG.getConstant(0, N0.getValueType()), Cond); 2558 if (N0.getOperand(1) == N1) { 2559 if (DAG.isCommutativeBinOp(N0.getOpcode())) 2560 return DAG.getSetCC(dl, VT, N0.getOperand(0), 2561 DAG.getConstant(0, N0.getValueType()), Cond); 2562 if (N0.getNode()->hasOneUse()) { 2563 assert(N0.getOpcode() == ISD::SUB && "Unexpected operation!"); 2564 // (Z-X) == X --> Z == X<<1 2565 SDValue SH = DAG.getNode(ISD::SHL, dl, N1.getValueType(), N1, 2566 DAG.getConstant(1, getShiftAmountTy(N1.getValueType()))); 2567 if (!DCI.isCalledByLegalizer()) 2568 DCI.AddToWorklist(SH.getNode()); 2569 return DAG.getSetCC(dl, VT, N0.getOperand(0), SH, Cond); 2570 } 2571 } 2572 } 2573 } 2574 2575 if (N1.getOpcode() == ISD::ADD || N1.getOpcode() == ISD::SUB || 2576 N1.getOpcode() == ISD::XOR) { 2577 // Simplify X == (X+Z) --> Z == 0 2578 if (N1.getOperand(0) == N0) 2579 return DAG.getSetCC(dl, VT, N1.getOperand(1), 2580 DAG.getConstant(0, N1.getValueType()), Cond); 2581 if (N1.getOperand(1) == N0) { 2582 if (DAG.isCommutativeBinOp(N1.getOpcode())) 2583 return DAG.getSetCC(dl, VT, N1.getOperand(0), 2584 DAG.getConstant(0, N1.getValueType()), Cond); 2585 if (N1.getNode()->hasOneUse()) { 2586 assert(N1.getOpcode() == ISD::SUB && "Unexpected operation!"); 2587 // X == (Z-X) --> X<<1 == Z 2588 SDValue SH = DAG.getNode(ISD::SHL, dl, N1.getValueType(), N0, 2589 DAG.getConstant(1, getShiftAmountTy(N0.getValueType()))); 2590 if (!DCI.isCalledByLegalizer()) 2591 DCI.AddToWorklist(SH.getNode()); 2592 return DAG.getSetCC(dl, VT, SH, N1.getOperand(0), Cond); 2593 } 2594 } 2595 } 2596 2597 // Simplify x&y == y to x&y != 0 if y has exactly one bit set. 2598 // Note that where y is variable and is known to have at most 2599 // one bit set (for example, if it is z&1) we cannot do this; 2600 // the expressions are not equivalent when y==0. 2601 if (N0.getOpcode() == ISD::AND) 2602 if (N0.getOperand(0) == N1 || N0.getOperand(1) == N1) { 2603 if (ValueHasExactlyOneBitSet(N1, DAG)) { 2604 Cond = ISD::getSetCCInverse(Cond, /*isInteger=*/true); 2605 SDValue Zero = DAG.getConstant(0, N1.getValueType()); 2606 return DAG.getSetCC(dl, VT, N0, Zero, Cond); 2607 } 2608 } 2609 if (N1.getOpcode() == ISD::AND) 2610 if (N1.getOperand(0) == N0 || N1.getOperand(1) == N0) { 2611 if (ValueHasExactlyOneBitSet(N0, DAG)) { 2612 Cond = ISD::getSetCCInverse(Cond, /*isInteger=*/true); 2613 SDValue Zero = DAG.getConstant(0, N0.getValueType()); 2614 return DAG.getSetCC(dl, VT, N1, Zero, Cond); 2615 } 2616 } 2617 } 2618 2619 // Fold away ALL boolean setcc's. 2620 SDValue Temp; 2621 if (N0.getValueType() == MVT::i1 && foldBooleans) { 2622 switch (Cond) { 2623 default: llvm_unreachable("Unknown integer setcc!"); 2624 case ISD::SETEQ: // X == Y -> ~(X^Y) 2625 Temp = DAG.getNode(ISD::XOR, dl, MVT::i1, N0, N1); 2626 N0 = DAG.getNOT(dl, Temp, MVT::i1); 2627 if (!DCI.isCalledByLegalizer()) 2628 DCI.AddToWorklist(Temp.getNode()); 2629 break; 2630 case ISD::SETNE: // X != Y --> (X^Y) 2631 N0 = DAG.getNode(ISD::XOR, dl, MVT::i1, N0, N1); 2632 break; 2633 case ISD::SETGT: // X >s Y --> X == 0 & Y == 1 --> ~X & Y 2634 case ISD::SETULT: // X <u Y --> X == 0 & Y == 1 --> ~X & Y 2635 Temp = DAG.getNOT(dl, N0, MVT::i1); 2636 N0 = DAG.getNode(ISD::AND, dl, MVT::i1, N1, Temp); 2637 if (!DCI.isCalledByLegalizer()) 2638 DCI.AddToWorklist(Temp.getNode()); 2639 break; 2640 case ISD::SETLT: // X <s Y --> X == 1 & Y == 0 --> ~Y & X 2641 case ISD::SETUGT: // X >u Y --> X == 1 & Y == 0 --> ~Y & X 2642 Temp = DAG.getNOT(dl, N1, MVT::i1); 2643 N0 = DAG.getNode(ISD::AND, dl, MVT::i1, N0, Temp); 2644 if (!DCI.isCalledByLegalizer()) 2645 DCI.AddToWorklist(Temp.getNode()); 2646 break; 2647 case ISD::SETULE: // X <=u Y --> X == 0 | Y == 1 --> ~X | Y 2648 case ISD::SETGE: // X >=s Y --> X == 0 | Y == 1 --> ~X | Y 2649 Temp = DAG.getNOT(dl, N0, MVT::i1); 2650 N0 = DAG.getNode(ISD::OR, dl, MVT::i1, N1, Temp); 2651 if (!DCI.isCalledByLegalizer()) 2652 DCI.AddToWorklist(Temp.getNode()); 2653 break; 2654 case ISD::SETUGE: // X >=u Y --> X == 1 | Y == 0 --> ~Y | X 2655 case ISD::SETLE: // X <=s Y --> X == 1 | Y == 0 --> ~Y | X 2656 Temp = DAG.getNOT(dl, N1, MVT::i1); 2657 N0 = DAG.getNode(ISD::OR, dl, MVT::i1, N0, Temp); 2658 break; 2659 } 2660 if (VT != MVT::i1) { 2661 if (!DCI.isCalledByLegalizer()) 2662 DCI.AddToWorklist(N0.getNode()); 2663 // FIXME: If running after legalize, we probably can't do this. 2664 N0 = DAG.getNode(ISD::ZERO_EXTEND, dl, VT, N0); 2665 } 2666 return N0; 2667 } 2668 2669 // Could not fold it. 2670 return SDValue(); 2671} 2672 2673/// isGAPlusOffset - Returns true (and the GlobalValue and the offset) if the 2674/// node is a GlobalAddress + offset. 2675bool TargetLowering::isGAPlusOffset(SDNode *N, const GlobalValue *&GA, 2676 int64_t &Offset) const { 2677 if (isa<GlobalAddressSDNode>(N)) { 2678 GlobalAddressSDNode *GASD = cast<GlobalAddressSDNode>(N); 2679 GA = GASD->getGlobal(); 2680 Offset += GASD->getOffset(); 2681 return true; 2682 } 2683 2684 if (N->getOpcode() == ISD::ADD) { 2685 SDValue N1 = N->getOperand(0); 2686 SDValue N2 = N->getOperand(1); 2687 if (isGAPlusOffset(N1.getNode(), GA, Offset)) { 2688 ConstantSDNode *V = dyn_cast<ConstantSDNode>(N2); 2689 if (V) { 2690 Offset += V->getSExtValue(); 2691 return true; 2692 } 2693 } else if (isGAPlusOffset(N2.getNode(), GA, Offset)) { 2694 ConstantSDNode *V = dyn_cast<ConstantSDNode>(N1); 2695 if (V) { 2696 Offset += V->getSExtValue(); 2697 return true; 2698 } 2699 } 2700 } 2701 2702 return false; 2703} 2704 2705 2706SDValue TargetLowering:: 2707PerformDAGCombine(SDNode *N, DAGCombinerInfo &DCI) const { 2708 // Default implementation: no optimization. 2709 return SDValue(); 2710} 2711 2712//===----------------------------------------------------------------------===// 2713// Inline Assembler Implementation Methods 2714//===----------------------------------------------------------------------===// 2715 2716 2717TargetLowering::ConstraintType 2718TargetLowering::getConstraintType(const std::string &Constraint) const { 2719 if (Constraint.size() == 1) { 2720 switch (Constraint[0]) { 2721 default: break; 2722 case 'r': return C_RegisterClass; 2723 case 'm': // memory 2724 case 'o': // offsetable 2725 case 'V': // not offsetable 2726 return C_Memory; 2727 case 'i': // Simple Integer or Relocatable Constant 2728 case 'n': // Simple Integer 2729 case 'E': // Floating Point Constant 2730 case 'F': // Floating Point Constant 2731 case 's': // Relocatable Constant 2732 case 'p': // Address. 2733 case 'X': // Allow ANY value. 2734 case 'I': // Target registers. 2735 case 'J': 2736 case 'K': 2737 case 'L': 2738 case 'M': 2739 case 'N': 2740 case 'O': 2741 case 'P': 2742 case '<': 2743 case '>': 2744 return C_Other; 2745 } 2746 } 2747 2748 if (Constraint.size() > 1 && Constraint[0] == '{' && 2749 Constraint[Constraint.size()-1] == '}') 2750 return C_Register; 2751 return C_Unknown; 2752} 2753 2754/// LowerXConstraint - try to replace an X constraint, which matches anything, 2755/// with another that has more specific requirements based on the type of the 2756/// corresponding operand. 2757const char *TargetLowering::LowerXConstraint(EVT ConstraintVT) const{ 2758 if (ConstraintVT.isInteger()) 2759 return "r"; 2760 if (ConstraintVT.isFloatingPoint()) 2761 return "f"; // works for many targets 2762 return 0; 2763} 2764 2765/// LowerAsmOperandForConstraint - Lower the specified operand into the Ops 2766/// vector. If it is invalid, don't add anything to Ops. 2767void TargetLowering::LowerAsmOperandForConstraint(SDValue Op, 2768 std::string &Constraint, 2769 std::vector<SDValue> &Ops, 2770 SelectionDAG &DAG) const { 2771 2772 if (Constraint.length() > 1) return; 2773 2774 char ConstraintLetter = Constraint[0]; 2775 switch (ConstraintLetter) { 2776 default: break; 2777 case 'X': // Allows any operand; labels (basic block) use this. 2778 if (Op.getOpcode() == ISD::BasicBlock) { 2779 Ops.push_back(Op); 2780 return; 2781 } 2782 // fall through 2783 case 'i': // Simple Integer or Relocatable Constant 2784 case 'n': // Simple Integer 2785 case 's': { // Relocatable Constant 2786 // These operands are interested in values of the form (GV+C), where C may 2787 // be folded in as an offset of GV, or it may be explicitly added. Also, it 2788 // is possible and fine if either GV or C are missing. 2789 ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op); 2790 GlobalAddressSDNode *GA = dyn_cast<GlobalAddressSDNode>(Op); 2791 2792 // If we have "(add GV, C)", pull out GV/C 2793 if (Op.getOpcode() == ISD::ADD) { 2794 C = dyn_cast<ConstantSDNode>(Op.getOperand(1)); 2795 GA = dyn_cast<GlobalAddressSDNode>(Op.getOperand(0)); 2796 if (C == 0 || GA == 0) { 2797 C = dyn_cast<ConstantSDNode>(Op.getOperand(0)); 2798 GA = dyn_cast<GlobalAddressSDNode>(Op.getOperand(1)); 2799 } 2800 if (C == 0 || GA == 0) 2801 C = 0, GA = 0; 2802 } 2803 2804 // If we find a valid operand, map to the TargetXXX version so that the 2805 // value itself doesn't get selected. 2806 if (GA) { // Either &GV or &GV+C 2807 if (ConstraintLetter != 'n') { 2808 int64_t Offs = GA->getOffset(); 2809 if (C) Offs += C->getZExtValue(); 2810 Ops.push_back(DAG.getTargetGlobalAddress(GA->getGlobal(), 2811 C ? C->getDebugLoc() : DebugLoc(), 2812 Op.getValueType(), Offs)); 2813 return; 2814 } 2815 } 2816 if (C) { // just C, no GV. 2817 // Simple constants are not allowed for 's'. 2818 if (ConstraintLetter != 's') { 2819 // gcc prints these as sign extended. Sign extend value to 64 bits 2820 // now; without this it would get ZExt'd later in 2821 // ScheduleDAGSDNodes::EmitNode, which is very generic. 2822 Ops.push_back(DAG.getTargetConstant(C->getAPIntValue().getSExtValue(), 2823 MVT::i64)); 2824 return; 2825 } 2826 } 2827 break; 2828 } 2829 } 2830} 2831 2832std::pair<unsigned, const TargetRegisterClass*> TargetLowering:: 2833getRegForInlineAsmConstraint(const std::string &Constraint, 2834 EVT VT) const { 2835 if (Constraint[0] != '{') 2836 return std::make_pair(0u, static_cast<TargetRegisterClass*>(0)); 2837 assert(*(Constraint.end()-1) == '}' && "Not a brace enclosed constraint?"); 2838 2839 // Remove the braces from around the name. 2840 StringRef RegName(Constraint.data()+1, Constraint.size()-2); 2841 2842 std::pair<unsigned, const TargetRegisterClass*> R = 2843 std::make_pair(0u, static_cast<const TargetRegisterClass*>(0)); 2844 2845 // Figure out which register class contains this reg. 2846 const TargetRegisterInfo *RI = TM.getRegisterInfo(); 2847 for (TargetRegisterInfo::regclass_iterator RCI = RI->regclass_begin(), 2848 E = RI->regclass_end(); RCI != E; ++RCI) { 2849 const TargetRegisterClass *RC = *RCI; 2850 2851 // If none of the value types for this register class are valid, we 2852 // can't use it. For example, 64-bit reg classes on 32-bit targets. 2853 if (!isLegalRC(RC)) 2854 continue; 2855 2856 for (TargetRegisterClass::iterator I = RC->begin(), E = RC->end(); 2857 I != E; ++I) { 2858 if (RegName.equals_lower(RI->getName(*I))) { 2859 std::pair<unsigned, const TargetRegisterClass*> S = 2860 std::make_pair(*I, RC); 2861 2862 // If this register class has the requested value type, return it, 2863 // otherwise keep searching and return the first class found 2864 // if no other is found which explicitly has the requested type. 2865 if (RC->hasType(VT)) 2866 return S; 2867 else if (!R.second) 2868 R = S; 2869 } 2870 } 2871 } 2872 2873 return R; 2874} 2875 2876//===----------------------------------------------------------------------===// 2877// Constraint Selection. 2878 2879/// isMatchingInputConstraint - Return true of this is an input operand that is 2880/// a matching constraint like "4". 2881bool TargetLowering::AsmOperandInfo::isMatchingInputConstraint() const { 2882 assert(!ConstraintCode.empty() && "No known constraint!"); 2883 return isdigit(ConstraintCode[0]); 2884} 2885 2886/// getMatchedOperand - If this is an input matching constraint, this method 2887/// returns the output operand it matches. 2888unsigned TargetLowering::AsmOperandInfo::getMatchedOperand() const { 2889 assert(!ConstraintCode.empty() && "No known constraint!"); 2890 return atoi(ConstraintCode.c_str()); 2891} 2892 2893 2894/// ParseConstraints - Split up the constraint string from the inline 2895/// assembly value into the specific constraints and their prefixes, 2896/// and also tie in the associated operand values. 2897/// If this returns an empty vector, and if the constraint string itself 2898/// isn't empty, there was an error parsing. 2899TargetLowering::AsmOperandInfoVector TargetLowering::ParseConstraints( 2900 ImmutableCallSite CS) const { 2901 /// ConstraintOperands - Information about all of the constraints. 2902 AsmOperandInfoVector ConstraintOperands; 2903 const InlineAsm *IA = cast<InlineAsm>(CS.getCalledValue()); 2904 unsigned maCount = 0; // Largest number of multiple alternative constraints. 2905 2906 // Do a prepass over the constraints, canonicalizing them, and building up the 2907 // ConstraintOperands list. 2908 InlineAsm::ConstraintInfoVector 2909 ConstraintInfos = IA->ParseConstraints(); 2910 2911 unsigned ArgNo = 0; // ArgNo - The argument of the CallInst. 2912 unsigned ResNo = 0; // ResNo - The result number of the next output. 2913 2914 for (unsigned i = 0, e = ConstraintInfos.size(); i != e; ++i) { 2915 ConstraintOperands.push_back(AsmOperandInfo(ConstraintInfos[i])); 2916 AsmOperandInfo &OpInfo = ConstraintOperands.back(); 2917 2918 // Update multiple alternative constraint count. 2919 if (OpInfo.multipleAlternatives.size() > maCount) 2920 maCount = OpInfo.multipleAlternatives.size(); 2921 2922 OpInfo.ConstraintVT = MVT::Other; 2923 2924 // Compute the value type for each operand. 2925 switch (OpInfo.Type) { 2926 case InlineAsm::isOutput: 2927 // Indirect outputs just consume an argument. 2928 if (OpInfo.isIndirect) { 2929 OpInfo.CallOperandVal = const_cast<Value *>(CS.getArgument(ArgNo++)); 2930 break; 2931 } 2932 2933 // The return value of the call is this value. As such, there is no 2934 // corresponding argument. 2935 assert(!CS.getType()->isVoidTy() && 2936 "Bad inline asm!"); 2937 if (StructType *STy = dyn_cast<StructType>(CS.getType())) { 2938 OpInfo.ConstraintVT = getSimpleValueType(STy->getElementType(ResNo)); 2939 } else { 2940 assert(ResNo == 0 && "Asm only has one result!"); 2941 OpInfo.ConstraintVT = getSimpleValueType(CS.getType()); 2942 } 2943 ++ResNo; 2944 break; 2945 case InlineAsm::isInput: 2946 OpInfo.CallOperandVal = const_cast<Value *>(CS.getArgument(ArgNo++)); 2947 break; 2948 case InlineAsm::isClobber: 2949 // Nothing to do. 2950 break; 2951 } 2952 2953 if (OpInfo.CallOperandVal) { 2954 llvm::Type *OpTy = OpInfo.CallOperandVal->getType(); 2955 if (OpInfo.isIndirect) { 2956 llvm::PointerType *PtrTy = dyn_cast<PointerType>(OpTy); 2957 if (!PtrTy) 2958 report_fatal_error("Indirect operand for inline asm not a pointer!"); 2959 OpTy = PtrTy->getElementType(); 2960 } 2961 2962 // Look for vector wrapped in a struct. e.g. { <16 x i8> }. 2963 if (StructType *STy = dyn_cast<StructType>(OpTy)) 2964 if (STy->getNumElements() == 1) 2965 OpTy = STy->getElementType(0); 2966 2967 // If OpTy is not a single value, it may be a struct/union that we 2968 // can tile with integers. 2969 if (!OpTy->isSingleValueType() && OpTy->isSized()) { 2970 unsigned BitSize = TD->getTypeSizeInBits(OpTy); 2971 switch (BitSize) { 2972 default: break; 2973 case 1: 2974 case 8: 2975 case 16: 2976 case 32: 2977 case 64: 2978 case 128: 2979 OpInfo.ConstraintVT = 2980 MVT::getVT(IntegerType::get(OpTy->getContext(), BitSize), true); 2981 break; 2982 } 2983 } else if (PointerType *PT = dyn_cast<PointerType>(OpTy)) { 2984 OpInfo.ConstraintVT = MVT::getIntegerVT( 2985 8*TD->getPointerSize(PT->getAddressSpace())); 2986 } else { 2987 OpInfo.ConstraintVT = MVT::getVT(OpTy, true); 2988 } 2989 } 2990 } 2991 2992 // If we have multiple alternative constraints, select the best alternative. 2993 if (ConstraintInfos.size()) { 2994 if (maCount) { 2995 unsigned bestMAIndex = 0; 2996 int bestWeight = -1; 2997 // weight: -1 = invalid match, and 0 = so-so match to 5 = good match. 2998 int weight = -1; 2999 unsigned maIndex; 3000 // Compute the sums of the weights for each alternative, keeping track 3001 // of the best (highest weight) one so far. 3002 for (maIndex = 0; maIndex < maCount; ++maIndex) { 3003 int weightSum = 0; 3004 for (unsigned cIndex = 0, eIndex = ConstraintOperands.size(); 3005 cIndex != eIndex; ++cIndex) { 3006 AsmOperandInfo& OpInfo = ConstraintOperands[cIndex]; 3007 if (OpInfo.Type == InlineAsm::isClobber) 3008 continue; 3009 3010 // If this is an output operand with a matching input operand, 3011 // look up the matching input. If their types mismatch, e.g. one 3012 // is an integer, the other is floating point, or their sizes are 3013 // different, flag it as an maCantMatch. 3014 if (OpInfo.hasMatchingInput()) { 3015 AsmOperandInfo &Input = ConstraintOperands[OpInfo.MatchingInput]; 3016 if (OpInfo.ConstraintVT != Input.ConstraintVT) { 3017 if ((OpInfo.ConstraintVT.isInteger() != 3018 Input.ConstraintVT.isInteger()) || 3019 (OpInfo.ConstraintVT.getSizeInBits() != 3020 Input.ConstraintVT.getSizeInBits())) { 3021 weightSum = -1; // Can't match. 3022 break; 3023 } 3024 } 3025 } 3026 weight = getMultipleConstraintMatchWeight(OpInfo, maIndex); 3027 if (weight == -1) { 3028 weightSum = -1; 3029 break; 3030 } 3031 weightSum += weight; 3032 } 3033 // Update best. 3034 if (weightSum > bestWeight) { 3035 bestWeight = weightSum; 3036 bestMAIndex = maIndex; 3037 } 3038 } 3039 3040 // Now select chosen alternative in each constraint. 3041 for (unsigned cIndex = 0, eIndex = ConstraintOperands.size(); 3042 cIndex != eIndex; ++cIndex) { 3043 AsmOperandInfo& cInfo = ConstraintOperands[cIndex]; 3044 if (cInfo.Type == InlineAsm::isClobber) 3045 continue; 3046 cInfo.selectAlternative(bestMAIndex); 3047 } 3048 } 3049 } 3050 3051 // Check and hook up tied operands, choose constraint code to use. 3052 for (unsigned cIndex = 0, eIndex = ConstraintOperands.size(); 3053 cIndex != eIndex; ++cIndex) { 3054 AsmOperandInfo& OpInfo = ConstraintOperands[cIndex]; 3055 3056 // If this is an output operand with a matching input operand, look up the 3057 // matching input. If their types mismatch, e.g. one is an integer, the 3058 // other is floating point, or their sizes are different, flag it as an 3059 // error. 3060 if (OpInfo.hasMatchingInput()) { 3061 AsmOperandInfo &Input = ConstraintOperands[OpInfo.MatchingInput]; 3062 3063 if (OpInfo.ConstraintVT != Input.ConstraintVT) { 3064 std::pair<unsigned, const TargetRegisterClass*> MatchRC = 3065 getRegForInlineAsmConstraint(OpInfo.ConstraintCode, 3066 OpInfo.ConstraintVT); 3067 std::pair<unsigned, const TargetRegisterClass*> InputRC = 3068 getRegForInlineAsmConstraint(Input.ConstraintCode, 3069 Input.ConstraintVT); 3070 if ((OpInfo.ConstraintVT.isInteger() != 3071 Input.ConstraintVT.isInteger()) || 3072 (MatchRC.second != InputRC.second)) { 3073 report_fatal_error("Unsupported asm: input constraint" 3074 " with a matching output constraint of" 3075 " incompatible type!"); 3076 } 3077 } 3078 3079 } 3080 } 3081 3082 return ConstraintOperands; 3083} 3084 3085 3086/// getConstraintGenerality - Return an integer indicating how general CT 3087/// is. 3088static unsigned getConstraintGenerality(TargetLowering::ConstraintType CT) { 3089 switch (CT) { 3090 case TargetLowering::C_Other: 3091 case TargetLowering::C_Unknown: 3092 return 0; 3093 case TargetLowering::C_Register: 3094 return 1; 3095 case TargetLowering::C_RegisterClass: 3096 return 2; 3097 case TargetLowering::C_Memory: 3098 return 3; 3099 } 3100 llvm_unreachable("Invalid constraint type"); 3101} 3102 3103/// Examine constraint type and operand type and determine a weight value. 3104/// This object must already have been set up with the operand type 3105/// and the current alternative constraint selected. 3106TargetLowering::ConstraintWeight 3107 TargetLowering::getMultipleConstraintMatchWeight( 3108 AsmOperandInfo &info, int maIndex) const { 3109 InlineAsm::ConstraintCodeVector *rCodes; 3110 if (maIndex >= (int)info.multipleAlternatives.size()) 3111 rCodes = &info.Codes; 3112 else 3113 rCodes = &info.multipleAlternatives[maIndex].Codes; 3114 ConstraintWeight BestWeight = CW_Invalid; 3115 3116 // Loop over the options, keeping track of the most general one. 3117 for (unsigned i = 0, e = rCodes->size(); i != e; ++i) { 3118 ConstraintWeight weight = 3119 getSingleConstraintMatchWeight(info, (*rCodes)[i].c_str()); 3120 if (weight > BestWeight) 3121 BestWeight = weight; 3122 } 3123 3124 return BestWeight; 3125} 3126 3127/// Examine constraint type and operand type and determine a weight value. 3128/// This object must already have been set up with the operand type 3129/// and the current alternative constraint selected. 3130TargetLowering::ConstraintWeight 3131 TargetLowering::getSingleConstraintMatchWeight( 3132 AsmOperandInfo &info, const char *constraint) const { 3133 ConstraintWeight weight = CW_Invalid; 3134 Value *CallOperandVal = info.CallOperandVal; 3135 // If we don't have a value, we can't do a match, 3136 // but allow it at the lowest weight. 3137 if (CallOperandVal == NULL) 3138 return CW_Default; 3139 // Look at the constraint type. 3140 switch (*constraint) { 3141 case 'i': // immediate integer. 3142 case 'n': // immediate integer with a known value. 3143 if (isa<ConstantInt>(CallOperandVal)) 3144 weight = CW_Constant; 3145 break; 3146 case 's': // non-explicit intregal immediate. 3147 if (isa<GlobalValue>(CallOperandVal)) 3148 weight = CW_Constant; 3149 break; 3150 case 'E': // immediate float if host format. 3151 case 'F': // immediate float. 3152 if (isa<ConstantFP>(CallOperandVal)) 3153 weight = CW_Constant; 3154 break; 3155 case '<': // memory operand with autodecrement. 3156 case '>': // memory operand with autoincrement. 3157 case 'm': // memory operand. 3158 case 'o': // offsettable memory operand 3159 case 'V': // non-offsettable memory operand 3160 weight = CW_Memory; 3161 break; 3162 case 'r': // general register. 3163 case 'g': // general register, memory operand or immediate integer. 3164 // note: Clang converts "g" to "imr". 3165 if (CallOperandVal->getType()->isIntegerTy()) 3166 weight = CW_Register; 3167 break; 3168 case 'X': // any operand. 3169 default: 3170 weight = CW_Default; 3171 break; 3172 } 3173 return weight; 3174} 3175 3176/// ChooseConstraint - If there are multiple different constraints that we 3177/// could pick for this operand (e.g. "imr") try to pick the 'best' one. 3178/// This is somewhat tricky: constraints fall into four classes: 3179/// Other -> immediates and magic values 3180/// Register -> one specific register 3181/// RegisterClass -> a group of regs 3182/// Memory -> memory 3183/// Ideally, we would pick the most specific constraint possible: if we have 3184/// something that fits into a register, we would pick it. The problem here 3185/// is that if we have something that could either be in a register or in 3186/// memory that use of the register could cause selection of *other* 3187/// operands to fail: they might only succeed if we pick memory. Because of 3188/// this the heuristic we use is: 3189/// 3190/// 1) If there is an 'other' constraint, and if the operand is valid for 3191/// that constraint, use it. This makes us take advantage of 'i' 3192/// constraints when available. 3193/// 2) Otherwise, pick the most general constraint present. This prefers 3194/// 'm' over 'r', for example. 3195/// 3196static void ChooseConstraint(TargetLowering::AsmOperandInfo &OpInfo, 3197 const TargetLowering &TLI, 3198 SDValue Op, SelectionDAG *DAG) { 3199 assert(OpInfo.Codes.size() > 1 && "Doesn't have multiple constraint options"); 3200 unsigned BestIdx = 0; 3201 TargetLowering::ConstraintType BestType = TargetLowering::C_Unknown; 3202 int BestGenerality = -1; 3203 3204 // Loop over the options, keeping track of the most general one. 3205 for (unsigned i = 0, e = OpInfo.Codes.size(); i != e; ++i) { 3206 TargetLowering::ConstraintType CType = 3207 TLI.getConstraintType(OpInfo.Codes[i]); 3208 3209 // If this is an 'other' constraint, see if the operand is valid for it. 3210 // For example, on X86 we might have an 'rI' constraint. If the operand 3211 // is an integer in the range [0..31] we want to use I (saving a load 3212 // of a register), otherwise we must use 'r'. 3213 if (CType == TargetLowering::C_Other && Op.getNode()) { 3214 assert(OpInfo.Codes[i].size() == 1 && 3215 "Unhandled multi-letter 'other' constraint"); 3216 std::vector<SDValue> ResultOps; 3217 TLI.LowerAsmOperandForConstraint(Op, OpInfo.Codes[i], 3218 ResultOps, *DAG); 3219 if (!ResultOps.empty()) { 3220 BestType = CType; 3221 BestIdx = i; 3222 break; 3223 } 3224 } 3225 3226 // Things with matching constraints can only be registers, per gcc 3227 // documentation. This mainly affects "g" constraints. 3228 if (CType == TargetLowering::C_Memory && OpInfo.hasMatchingInput()) 3229 continue; 3230 3231 // This constraint letter is more general than the previous one, use it. 3232 int Generality = getConstraintGenerality(CType); 3233 if (Generality > BestGenerality) { 3234 BestType = CType; 3235 BestIdx = i; 3236 BestGenerality = Generality; 3237 } 3238 } 3239 3240 OpInfo.ConstraintCode = OpInfo.Codes[BestIdx]; 3241 OpInfo.ConstraintType = BestType; 3242} 3243 3244/// ComputeConstraintToUse - Determines the constraint code and constraint 3245/// type to use for the specific AsmOperandInfo, setting 3246/// OpInfo.ConstraintCode and OpInfo.ConstraintType. 3247void TargetLowering::ComputeConstraintToUse(AsmOperandInfo &OpInfo, 3248 SDValue Op, 3249 SelectionDAG *DAG) const { 3250 assert(!OpInfo.Codes.empty() && "Must have at least one constraint"); 3251 3252 // Single-letter constraints ('r') are very common. 3253 if (OpInfo.Codes.size() == 1) { 3254 OpInfo.ConstraintCode = OpInfo.Codes[0]; 3255 OpInfo.ConstraintType = getConstraintType(OpInfo.ConstraintCode); 3256 } else { 3257 ChooseConstraint(OpInfo, *this, Op, DAG); 3258 } 3259 3260 // 'X' matches anything. 3261 if (OpInfo.ConstraintCode == "X" && OpInfo.CallOperandVal) { 3262 // Labels and constants are handled elsewhere ('X' is the only thing 3263 // that matches labels). For Functions, the type here is the type of 3264 // the result, which is not what we want to look at; leave them alone. 3265 Value *v = OpInfo.CallOperandVal; 3266 if (isa<BasicBlock>(v) || isa<ConstantInt>(v) || isa<Function>(v)) { 3267 OpInfo.CallOperandVal = v; 3268 return; 3269 } 3270 3271 // Otherwise, try to resolve it to something we know about by looking at 3272 // the actual operand type. 3273 if (const char *Repl = LowerXConstraint(OpInfo.ConstraintVT)) { 3274 OpInfo.ConstraintCode = Repl; 3275 OpInfo.ConstraintType = getConstraintType(OpInfo.ConstraintCode); 3276 } 3277 } 3278} 3279 3280//===----------------------------------------------------------------------===// 3281// Loop Strength Reduction hooks 3282//===----------------------------------------------------------------------===// 3283 3284/// isLegalAddressingMode - Return true if the addressing mode represented 3285/// by AM is legal for this target, for a load/store of the specified type. 3286bool TargetLowering::isLegalAddressingMode(const AddrMode &AM, 3287 Type *Ty) const { 3288 // The default implementation of this implements a conservative RISCy, r+r and 3289 // r+i addr mode. 3290 3291 // Allows a sign-extended 16-bit immediate field. 3292 if (AM.BaseOffs <= -(1LL << 16) || AM.BaseOffs >= (1LL << 16)-1) 3293 return false; 3294 3295 // No global is ever allowed as a base. 3296 if (AM.BaseGV) 3297 return false; 3298 3299 // Only support r+r, 3300 switch (AM.Scale) { 3301 case 0: // "r+i" or just "i", depending on HasBaseReg. 3302 break; 3303 case 1: 3304 if (AM.HasBaseReg && AM.BaseOffs) // "r+r+i" is not allowed. 3305 return false; 3306 // Otherwise we have r+r or r+i. 3307 break; 3308 case 2: 3309 if (AM.HasBaseReg || AM.BaseOffs) // 2*r+r or 2*r+i is not allowed. 3310 return false; 3311 // Allow 2*r as r+r. 3312 break; 3313 } 3314 3315 return true; 3316} 3317 3318/// BuildExactDiv - Given an exact SDIV by a constant, create a multiplication 3319/// with the multiplicative inverse of the constant. 3320SDValue TargetLowering::BuildExactSDIV(SDValue Op1, SDValue Op2, DebugLoc dl, 3321 SelectionDAG &DAG) const { 3322 ConstantSDNode *C = cast<ConstantSDNode>(Op2); 3323 APInt d = C->getAPIntValue(); 3324 assert(d != 0 && "Division by zero!"); 3325 3326 // Shift the value upfront if it is even, so the LSB is one. 3327 unsigned ShAmt = d.countTrailingZeros(); 3328 if (ShAmt) { 3329 // TODO: For UDIV use SRL instead of SRA. 3330 SDValue Amt = DAG.getConstant(ShAmt, getShiftAmountTy(Op1.getValueType())); 3331 Op1 = DAG.getNode(ISD::SRA, dl, Op1.getValueType(), Op1, Amt); 3332 d = d.ashr(ShAmt); 3333 } 3334 3335 // Calculate the multiplicative inverse, using Newton's method. 3336 APInt t, xn = d; 3337 while ((t = d*xn) != 1) 3338 xn *= APInt(d.getBitWidth(), 2) - t; 3339 3340 Op2 = DAG.getConstant(xn, Op1.getValueType()); 3341 return DAG.getNode(ISD::MUL, dl, Op1.getValueType(), Op1, Op2); 3342} 3343 3344/// BuildSDIVSequence - Given an ISD::SDIV node expressing a divide by constant, 3345/// return a DAG expression to select that will generate the same value by 3346/// multiplying by a magic number. See: 3347/// <http://the.wall.riscom.net/books/proc/ppc/cwg/code2.html> 3348SDValue TargetLowering:: 3349BuildSDIV(SDNode *N, SelectionDAG &DAG, bool IsAfterLegalization, 3350 std::vector<SDNode*> *Created) const { 3351 EVT VT = N->getValueType(0); 3352 DebugLoc dl= N->getDebugLoc(); 3353 3354 // Check to see if we can do this. 3355 // FIXME: We should be more aggressive here. 3356 if (!isTypeLegal(VT)) 3357 return SDValue(); 3358 3359 APInt d = cast<ConstantSDNode>(N->getOperand(1))->getAPIntValue(); 3360 APInt::ms magics = d.magic(); 3361 3362 // Multiply the numerator (operand 0) by the magic value 3363 // FIXME: We should support doing a MUL in a wider type 3364 SDValue Q; 3365 if (IsAfterLegalization ? isOperationLegal(ISD::MULHS, VT) : 3366 isOperationLegalOrCustom(ISD::MULHS, VT)) 3367 Q = DAG.getNode(ISD::MULHS, dl, VT, N->getOperand(0), 3368 DAG.getConstant(magics.m, VT)); 3369 else if (IsAfterLegalization ? isOperationLegal(ISD::SMUL_LOHI, VT) : 3370 isOperationLegalOrCustom(ISD::SMUL_LOHI, VT)) 3371 Q = SDValue(DAG.getNode(ISD::SMUL_LOHI, dl, DAG.getVTList(VT, VT), 3372 N->getOperand(0), 3373 DAG.getConstant(magics.m, VT)).getNode(), 1); 3374 else 3375 return SDValue(); // No mulhs or equvialent 3376 // If d > 0 and m < 0, add the numerator 3377 if (d.isStrictlyPositive() && magics.m.isNegative()) { 3378 Q = DAG.getNode(ISD::ADD, dl, VT, Q, N->getOperand(0)); 3379 if (Created) 3380 Created->push_back(Q.getNode()); 3381 } 3382 // If d < 0 and m > 0, subtract the numerator. 3383 if (d.isNegative() && magics.m.isStrictlyPositive()) { 3384 Q = DAG.getNode(ISD::SUB, dl, VT, Q, N->getOperand(0)); 3385 if (Created) 3386 Created->push_back(Q.getNode()); 3387 } 3388 // Shift right algebraic if shift value is nonzero 3389 if (magics.s > 0) { 3390 Q = DAG.getNode(ISD::SRA, dl, VT, Q, 3391 DAG.getConstant(magics.s, getShiftAmountTy(Q.getValueType()))); 3392 if (Created) 3393 Created->push_back(Q.getNode()); 3394 } 3395 // Extract the sign bit and add it to the quotient 3396 SDValue T = 3397 DAG.getNode(ISD::SRL, dl, VT, Q, DAG.getConstant(VT.getSizeInBits()-1, 3398 getShiftAmountTy(Q.getValueType()))); 3399 if (Created) 3400 Created->push_back(T.getNode()); 3401 return DAG.getNode(ISD::ADD, dl, VT, Q, T); 3402} 3403 3404/// BuildUDIVSequence - Given an ISD::UDIV node expressing a divide by constant, 3405/// return a DAG expression to select that will generate the same value by 3406/// multiplying by a magic number. See: 3407/// <http://the.wall.riscom.net/books/proc/ppc/cwg/code2.html> 3408SDValue TargetLowering:: 3409BuildUDIV(SDNode *N, SelectionDAG &DAG, bool IsAfterLegalization, 3410 std::vector<SDNode*> *Created) const { 3411 EVT VT = N->getValueType(0); 3412 DebugLoc dl = N->getDebugLoc(); 3413 3414 // Check to see if we can do this. 3415 // FIXME: We should be more aggressive here. 3416 if (!isTypeLegal(VT)) 3417 return SDValue(); 3418 3419 // FIXME: We should use a narrower constant when the upper 3420 // bits are known to be zero. 3421 const APInt &N1C = cast<ConstantSDNode>(N->getOperand(1))->getAPIntValue(); 3422 APInt::mu magics = N1C.magicu(); 3423 3424 SDValue Q = N->getOperand(0); 3425 3426 // If the divisor is even, we can avoid using the expensive fixup by shifting 3427 // the divided value upfront. 3428 if (magics.a != 0 && !N1C[0]) { 3429 unsigned Shift = N1C.countTrailingZeros(); 3430 Q = DAG.getNode(ISD::SRL, dl, VT, Q, 3431 DAG.getConstant(Shift, getShiftAmountTy(Q.getValueType()))); 3432 if (Created) 3433 Created->push_back(Q.getNode()); 3434 3435 // Get magic number for the shifted divisor. 3436 magics = N1C.lshr(Shift).magicu(Shift); 3437 assert(magics.a == 0 && "Should use cheap fixup now"); 3438 } 3439 3440 // Multiply the numerator (operand 0) by the magic value 3441 // FIXME: We should support doing a MUL in a wider type 3442 if (IsAfterLegalization ? isOperationLegal(ISD::MULHU, VT) : 3443 isOperationLegalOrCustom(ISD::MULHU, VT)) 3444 Q = DAG.getNode(ISD::MULHU, dl, VT, Q, DAG.getConstant(magics.m, VT)); 3445 else if (IsAfterLegalization ? isOperationLegal(ISD::UMUL_LOHI, VT) : 3446 isOperationLegalOrCustom(ISD::UMUL_LOHI, VT)) 3447 Q = SDValue(DAG.getNode(ISD::UMUL_LOHI, dl, DAG.getVTList(VT, VT), Q, 3448 DAG.getConstant(magics.m, VT)).getNode(), 1); 3449 else 3450 return SDValue(); // No mulhu or equvialent 3451 if (Created) 3452 Created->push_back(Q.getNode()); 3453 3454 if (magics.a == 0) { 3455 assert(magics.s < N1C.getBitWidth() && 3456 "We shouldn't generate an undefined shift!"); 3457 return DAG.getNode(ISD::SRL, dl, VT, Q, 3458 DAG.getConstant(magics.s, getShiftAmountTy(Q.getValueType()))); 3459 } else { 3460 SDValue NPQ = DAG.getNode(ISD::SUB, dl, VT, N->getOperand(0), Q); 3461 if (Created) 3462 Created->push_back(NPQ.getNode()); 3463 NPQ = DAG.getNode(ISD::SRL, dl, VT, NPQ, 3464 DAG.getConstant(1, getShiftAmountTy(NPQ.getValueType()))); 3465 if (Created) 3466 Created->push_back(NPQ.getNode()); 3467 NPQ = DAG.getNode(ISD::ADD, dl, VT, NPQ, Q); 3468 if (Created) 3469 Created->push_back(NPQ.getNode()); 3470 return DAG.getNode(ISD::SRL, dl, VT, NPQ, 3471 DAG.getConstant(magics.s-1, getShiftAmountTy(NPQ.getValueType()))); 3472 } 3473} 3474