TargetLowering.cpp revision 34525f9ac098c1c6bc9002886d6da3039a284fd2
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 else if (RetVT == MVT::f64) 432 return SINTTOFP_I32_F64; 433 else if (RetVT == MVT::f80) 434 return SINTTOFP_I32_F80; 435 else if (RetVT == MVT::ppcf128) 436 return SINTTOFP_I32_PPCF128; 437 } else if (OpVT == MVT::i64) { 438 if (RetVT == MVT::f32) 439 return SINTTOFP_I64_F32; 440 else if (RetVT == MVT::f64) 441 return SINTTOFP_I64_F64; 442 else if (RetVT == MVT::f80) 443 return SINTTOFP_I64_F80; 444 else if (RetVT == MVT::ppcf128) 445 return SINTTOFP_I64_PPCF128; 446 } else if (OpVT == MVT::i128) { 447 if (RetVT == MVT::f32) 448 return SINTTOFP_I128_F32; 449 else if (RetVT == MVT::f64) 450 return SINTTOFP_I128_F64; 451 else if (RetVT == MVT::f80) 452 return SINTTOFP_I128_F80; 453 else if (RetVT == MVT::ppcf128) 454 return SINTTOFP_I128_PPCF128; 455 } 456 return UNKNOWN_LIBCALL; 457} 458 459/// getUINTTOFP - Return the UINTTOFP_*_* value for the given types, or 460/// UNKNOWN_LIBCALL if there is none. 461RTLIB::Libcall RTLIB::getUINTTOFP(EVT OpVT, EVT RetVT) { 462 if (OpVT == MVT::i32) { 463 if (RetVT == MVT::f32) 464 return UINTTOFP_I32_F32; 465 else if (RetVT == MVT::f64) 466 return UINTTOFP_I32_F64; 467 else if (RetVT == MVT::f80) 468 return UINTTOFP_I32_F80; 469 else if (RetVT == MVT::ppcf128) 470 return UINTTOFP_I32_PPCF128; 471 } else if (OpVT == MVT::i64) { 472 if (RetVT == MVT::f32) 473 return UINTTOFP_I64_F32; 474 else if (RetVT == MVT::f64) 475 return UINTTOFP_I64_F64; 476 else if (RetVT == MVT::f80) 477 return UINTTOFP_I64_F80; 478 else if (RetVT == MVT::ppcf128) 479 return UINTTOFP_I64_PPCF128; 480 } else if (OpVT == MVT::i128) { 481 if (RetVT == MVT::f32) 482 return UINTTOFP_I128_F32; 483 else if (RetVT == MVT::f64) 484 return UINTTOFP_I128_F64; 485 else if (RetVT == MVT::f80) 486 return UINTTOFP_I128_F80; 487 else if (RetVT == MVT::ppcf128) 488 return UINTTOFP_I128_PPCF128; 489 } 490 return UNKNOWN_LIBCALL; 491} 492 493/// InitCmpLibcallCCs - Set default comparison libcall CC. 494/// 495static void InitCmpLibcallCCs(ISD::CondCode *CCs) { 496 memset(CCs, ISD::SETCC_INVALID, sizeof(ISD::CondCode)*RTLIB::UNKNOWN_LIBCALL); 497 CCs[RTLIB::OEQ_F32] = ISD::SETEQ; 498 CCs[RTLIB::OEQ_F64] = ISD::SETEQ; 499 CCs[RTLIB::UNE_F32] = ISD::SETNE; 500 CCs[RTLIB::UNE_F64] = ISD::SETNE; 501 CCs[RTLIB::OGE_F32] = ISD::SETGE; 502 CCs[RTLIB::OGE_F64] = ISD::SETGE; 503 CCs[RTLIB::OLT_F32] = ISD::SETLT; 504 CCs[RTLIB::OLT_F64] = ISD::SETLT; 505 CCs[RTLIB::OLE_F32] = ISD::SETLE; 506 CCs[RTLIB::OLE_F64] = ISD::SETLE; 507 CCs[RTLIB::OGT_F32] = ISD::SETGT; 508 CCs[RTLIB::OGT_F64] = ISD::SETGT; 509 CCs[RTLIB::UO_F32] = ISD::SETNE; 510 CCs[RTLIB::UO_F64] = ISD::SETNE; 511 CCs[RTLIB::O_F32] = ISD::SETEQ; 512 CCs[RTLIB::O_F64] = ISD::SETEQ; 513} 514 515/// NOTE: The constructor takes ownership of TLOF. 516TargetLowering::TargetLowering(const TargetMachine &tm, 517 const TargetLoweringObjectFile *tlof) 518 : TM(tm), TD(TM.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 EVT &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 EVT 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(EVT VT) const { 716 const TargetRegisterInfo *TRI = getTargetMachine().getRegisterInfo(); 717 const TargetRegisterClass *RC = RegClassForVT[VT.getSimpleVT().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 EVT 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 EVT 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 EVT 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 EVT 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 EVT RegisterVT; 869 unsigned NumIntermediates; 870 NumRegistersForVT[i] = 871 getVectorTypeBreakdownMVT(VT, IntermediateVT, NumIntermediates, 872 RegisterVT, this); 873 RegisterTypeForVT[i] = RegisterVT; 874 875 EVT 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 EVT &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 RegisterVT = getTypeToTransformTo(Context, VT); 938 if (isTypeLegal(RegisterVT)) { 939 IntermediateVT = RegisterVT; 940 NumIntermediates = 1; 941 return 1; 942 } 943 } 944 945 // Figure out the right, legal destination reg to copy into. 946 EVT EltTy = VT.getVectorElementType(); 947 948 unsigned NumVectorRegs = 1; 949 950 // FIXME: We don't support non-power-of-2-sized vectors for now. Ideally we 951 // could break down into LHS/RHS like LegalizeDAG does. 952 if (!isPowerOf2_32(NumElts)) { 953 NumVectorRegs = NumElts; 954 NumElts = 1; 955 } 956 957 // Divide the input until we get to a supported size. This will always 958 // end with a scalar if the target doesn't support vectors. 959 while (NumElts > 1 && !isTypeLegal( 960 EVT::getVectorVT(Context, EltTy, NumElts))) { 961 NumElts >>= 1; 962 NumVectorRegs <<= 1; 963 } 964 965 NumIntermediates = NumVectorRegs; 966 967 EVT NewVT = EVT::getVectorVT(Context, EltTy, NumElts); 968 if (!isTypeLegal(NewVT)) 969 NewVT = EltTy; 970 IntermediateVT = NewVT; 971 972 EVT DestVT = getRegisterType(Context, NewVT); 973 RegisterVT = DestVT; 974 unsigned NewVTSize = NewVT.getSizeInBits(); 975 976 // Convert sizes such as i33 to i64. 977 if (!isPowerOf2_32(NewVTSize)) 978 NewVTSize = NextPowerOf2(NewVTSize); 979 980 if (DestVT.bitsLT(NewVT)) // Value is expanded, e.g. i64 -> i16. 981 return NumVectorRegs*(NewVTSize/DestVT.getSizeInBits()); 982 983 // Otherwise, promotion or legal types use the same number of registers as 984 // the vector decimated to the appropriate level. 985 return NumVectorRegs; 986} 987 988/// Get the EVTs and ArgFlags collections that represent the legalized return 989/// type of the given function. This does not require a DAG or a return value, 990/// and is suitable for use before any DAGs for the function are constructed. 991/// TODO: Move this out of TargetLowering.cpp. 992void llvm::GetReturnInfo(Type* ReturnType, Attributes attr, 993 SmallVectorImpl<ISD::OutputArg> &Outs, 994 const TargetLowering &TLI) { 995 SmallVector<EVT, 4> ValueVTs; 996 ComputeValueVTs(TLI, ReturnType, ValueVTs); 997 unsigned NumValues = ValueVTs.size(); 998 if (NumValues == 0) return; 999 1000 for (unsigned j = 0, f = NumValues; j != f; ++j) { 1001 EVT VT = ValueVTs[j]; 1002 ISD::NodeType ExtendKind = ISD::ANY_EXTEND; 1003 1004 if (attr.hasAttribute(Attributes::SExt)) 1005 ExtendKind = ISD::SIGN_EXTEND; 1006 else if (attr.hasAttribute(Attributes::ZExt)) 1007 ExtendKind = ISD::ZERO_EXTEND; 1008 1009 // FIXME: C calling convention requires the return type to be promoted to 1010 // at least 32-bit. But this is not necessary for non-C calling 1011 // conventions. The frontend should mark functions whose return values 1012 // require promoting with signext or zeroext attributes. 1013 if (ExtendKind != ISD::ANY_EXTEND && VT.isInteger()) { 1014 EVT MinVT = TLI.getRegisterType(ReturnType->getContext(), MVT::i32); 1015 if (VT.bitsLT(MinVT)) 1016 VT = MinVT; 1017 } 1018 1019 unsigned NumParts = TLI.getNumRegisters(ReturnType->getContext(), VT); 1020 EVT PartVT = TLI.getRegisterType(ReturnType->getContext(), VT); 1021 1022 // 'inreg' on function refers to return value 1023 ISD::ArgFlagsTy Flags = ISD::ArgFlagsTy(); 1024 if (attr.hasAttribute(Attributes::InReg)) 1025 Flags.setInReg(); 1026 1027 // Propagate extension type if any 1028 if (attr.hasAttribute(Attributes::SExt)) 1029 Flags.setSExt(); 1030 else if (attr.hasAttribute(Attributes::ZExt)) 1031 Flags.setZExt(); 1032 1033 for (unsigned i = 0; i < NumParts; ++i) 1034 Outs.push_back(ISD::OutputArg(Flags, PartVT, /*isFixed=*/true, 0, 0)); 1035 } 1036} 1037 1038/// getByValTypeAlignment - Return the desired alignment for ByVal aggregate 1039/// function arguments in the caller parameter area. This is the actual 1040/// alignment, not its logarithm. 1041unsigned TargetLowering::getByValTypeAlignment(Type *Ty) const { 1042 return TD->getCallFrameTypeAlignment(Ty); 1043} 1044 1045/// getJumpTableEncoding - Return the entry encoding for a jump table in the 1046/// current function. The returned value is a member of the 1047/// MachineJumpTableInfo::JTEntryKind enum. 1048unsigned TargetLowering::getJumpTableEncoding() const { 1049 // In non-pic modes, just use the address of a block. 1050 if (getTargetMachine().getRelocationModel() != Reloc::PIC_) 1051 return MachineJumpTableInfo::EK_BlockAddress; 1052 1053 // In PIC mode, if the target supports a GPRel32 directive, use it. 1054 if (getTargetMachine().getMCAsmInfo()->getGPRel32Directive() != 0) 1055 return MachineJumpTableInfo::EK_GPRel32BlockAddress; 1056 1057 // Otherwise, use a label difference. 1058 return MachineJumpTableInfo::EK_LabelDifference32; 1059} 1060 1061SDValue TargetLowering::getPICJumpTableRelocBase(SDValue Table, 1062 SelectionDAG &DAG) const { 1063 // If our PIC model is GP relative, use the global offset table as the base. 1064 unsigned JTEncoding = getJumpTableEncoding(); 1065 1066 if ((JTEncoding == MachineJumpTableInfo::EK_GPRel64BlockAddress) || 1067 (JTEncoding == MachineJumpTableInfo::EK_GPRel32BlockAddress)) 1068 return DAG.getGLOBAL_OFFSET_TABLE(getPointerTy(0)); 1069 1070 return Table; 1071} 1072 1073/// getPICJumpTableRelocBaseExpr - This returns the relocation base for the 1074/// given PIC jumptable, the same as getPICJumpTableRelocBase, but as an 1075/// MCExpr. 1076const MCExpr * 1077TargetLowering::getPICJumpTableRelocBaseExpr(const MachineFunction *MF, 1078 unsigned JTI,MCContext &Ctx) const{ 1079 // The normal PIC reloc base is the label at the start of the jump table. 1080 return MCSymbolRefExpr::Create(MF->getJTISymbol(JTI, Ctx), Ctx); 1081} 1082 1083bool 1084TargetLowering::isOffsetFoldingLegal(const GlobalAddressSDNode *GA) const { 1085 // Assume that everything is safe in static mode. 1086 if (getTargetMachine().getRelocationModel() == Reloc::Static) 1087 return true; 1088 1089 // In dynamic-no-pic mode, assume that known defined values are safe. 1090 if (getTargetMachine().getRelocationModel() == Reloc::DynamicNoPIC && 1091 GA && 1092 !GA->getGlobal()->isDeclaration() && 1093 !GA->getGlobal()->isWeakForLinker()) 1094 return true; 1095 1096 // Otherwise assume nothing is safe. 1097 return false; 1098} 1099 1100//===----------------------------------------------------------------------===// 1101// Optimization Methods 1102//===----------------------------------------------------------------------===// 1103 1104/// ShrinkDemandedConstant - Check to see if the specified operand of the 1105/// specified instruction is a constant integer. If so, check to see if there 1106/// are any bits set in the constant that are not demanded. If so, shrink the 1107/// constant and return true. 1108bool TargetLowering::TargetLoweringOpt::ShrinkDemandedConstant(SDValue Op, 1109 const APInt &Demanded) { 1110 DebugLoc dl = Op.getDebugLoc(); 1111 1112 // FIXME: ISD::SELECT, ISD::SELECT_CC 1113 switch (Op.getOpcode()) { 1114 default: break; 1115 case ISD::XOR: 1116 case ISD::AND: 1117 case ISD::OR: { 1118 ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op.getOperand(1)); 1119 if (!C) return false; 1120 1121 if (Op.getOpcode() == ISD::XOR && 1122 (C->getAPIntValue() | (~Demanded)).isAllOnesValue()) 1123 return false; 1124 1125 // if we can expand it to have all bits set, do it 1126 if (C->getAPIntValue().intersects(~Demanded)) { 1127 EVT VT = Op.getValueType(); 1128 SDValue New = DAG.getNode(Op.getOpcode(), dl, VT, Op.getOperand(0), 1129 DAG.getConstant(Demanded & 1130 C->getAPIntValue(), 1131 VT)); 1132 return CombineTo(Op, New); 1133 } 1134 1135 break; 1136 } 1137 } 1138 1139 return false; 1140} 1141 1142/// ShrinkDemandedOp - Convert x+y to (VT)((SmallVT)x+(SmallVT)y) if the 1143/// casts are free. This uses isZExtFree and ZERO_EXTEND for the widening 1144/// cast, but it could be generalized for targets with other types of 1145/// implicit widening casts. 1146bool 1147TargetLowering::TargetLoweringOpt::ShrinkDemandedOp(SDValue Op, 1148 unsigned BitWidth, 1149 const APInt &Demanded, 1150 DebugLoc dl) { 1151 assert(Op.getNumOperands() == 2 && 1152 "ShrinkDemandedOp only supports binary operators!"); 1153 assert(Op.getNode()->getNumValues() == 1 && 1154 "ShrinkDemandedOp only supports nodes with one result!"); 1155 1156 // Don't do this if the node has another user, which may require the 1157 // full value. 1158 if (!Op.getNode()->hasOneUse()) 1159 return false; 1160 1161 // Search for the smallest integer type with free casts to and from 1162 // Op's type. For expedience, just check power-of-2 integer types. 1163 const TargetLowering &TLI = DAG.getTargetLoweringInfo(); 1164 unsigned SmallVTBits = BitWidth - Demanded.countLeadingZeros(); 1165 if (!isPowerOf2_32(SmallVTBits)) 1166 SmallVTBits = NextPowerOf2(SmallVTBits); 1167 for (; SmallVTBits < BitWidth; SmallVTBits = NextPowerOf2(SmallVTBits)) { 1168 EVT SmallVT = EVT::getIntegerVT(*DAG.getContext(), SmallVTBits); 1169 if (TLI.isTruncateFree(Op.getValueType(), SmallVT) && 1170 TLI.isZExtFree(SmallVT, Op.getValueType())) { 1171 // We found a type with free casts. 1172 SDValue X = DAG.getNode(Op.getOpcode(), dl, SmallVT, 1173 DAG.getNode(ISD::TRUNCATE, dl, SmallVT, 1174 Op.getNode()->getOperand(0)), 1175 DAG.getNode(ISD::TRUNCATE, dl, SmallVT, 1176 Op.getNode()->getOperand(1))); 1177 SDValue Z = DAG.getNode(ISD::ZERO_EXTEND, dl, Op.getValueType(), X); 1178 return CombineTo(Op, Z); 1179 } 1180 } 1181 return false; 1182} 1183 1184/// SimplifyDemandedBits - Look at Op. At this point, we know that only the 1185/// DemandedMask bits of the result of Op are ever used downstream. If we can 1186/// use this information to simplify Op, create a new simplified DAG node and 1187/// return true, returning the original and new nodes in Old and New. Otherwise, 1188/// analyze the expression and return a mask of KnownOne and KnownZero bits for 1189/// the expression (used to simplify the caller). The KnownZero/One bits may 1190/// only be accurate for those bits in the DemandedMask. 1191bool TargetLowering::SimplifyDemandedBits(SDValue Op, 1192 const APInt &DemandedMask, 1193 APInt &KnownZero, 1194 APInt &KnownOne, 1195 TargetLoweringOpt &TLO, 1196 unsigned Depth) const { 1197 unsigned BitWidth = DemandedMask.getBitWidth(); 1198 assert(Op.getValueType().getScalarType().getSizeInBits() == BitWidth && 1199 "Mask size mismatches value type size!"); 1200 APInt NewMask = DemandedMask; 1201 DebugLoc dl = Op.getDebugLoc(); 1202 1203 // Don't know anything. 1204 KnownZero = KnownOne = APInt(BitWidth, 0); 1205 1206 // Other users may use these bits. 1207 if (!Op.getNode()->hasOneUse()) { 1208 if (Depth != 0) { 1209 // If not at the root, Just compute the KnownZero/KnownOne bits to 1210 // simplify things downstream. 1211 TLO.DAG.ComputeMaskedBits(Op, KnownZero, KnownOne, Depth); 1212 return false; 1213 } 1214 // If this is the root being simplified, allow it to have multiple uses, 1215 // just set the NewMask to all bits. 1216 NewMask = APInt::getAllOnesValue(BitWidth); 1217 } else if (DemandedMask == 0) { 1218 // Not demanding any bits from Op. 1219 if (Op.getOpcode() != ISD::UNDEF) 1220 return TLO.CombineTo(Op, TLO.DAG.getUNDEF(Op.getValueType())); 1221 return false; 1222 } else if (Depth == 6) { // Limit search depth. 1223 return false; 1224 } 1225 1226 APInt KnownZero2, KnownOne2, KnownZeroOut, KnownOneOut; 1227 switch (Op.getOpcode()) { 1228 case ISD::Constant: 1229 // We know all of the bits for a constant! 1230 KnownOne = cast<ConstantSDNode>(Op)->getAPIntValue(); 1231 KnownZero = ~KnownOne; 1232 return false; // Don't fall through, will infinitely loop. 1233 case ISD::AND: 1234 // If the RHS is a constant, check to see if the LHS would be zero without 1235 // using the bits from the RHS. Below, we use knowledge about the RHS to 1236 // simplify the LHS, here we're using information from the LHS to simplify 1237 // the RHS. 1238 if (ConstantSDNode *RHSC = dyn_cast<ConstantSDNode>(Op.getOperand(1))) { 1239 APInt LHSZero, LHSOne; 1240 // Do not increment Depth here; that can cause an infinite loop. 1241 TLO.DAG.ComputeMaskedBits(Op.getOperand(0), LHSZero, LHSOne, Depth); 1242 // If the LHS already has zeros where RHSC does, this and is dead. 1243 if ((LHSZero & NewMask) == (~RHSC->getAPIntValue() & NewMask)) 1244 return TLO.CombineTo(Op, Op.getOperand(0)); 1245 // If any of the set bits in the RHS are known zero on the LHS, shrink 1246 // the constant. 1247 if (TLO.ShrinkDemandedConstant(Op, ~LHSZero & NewMask)) 1248 return true; 1249 } 1250 1251 if (SimplifyDemandedBits(Op.getOperand(1), NewMask, KnownZero, 1252 KnownOne, TLO, Depth+1)) 1253 return true; 1254 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?"); 1255 if (SimplifyDemandedBits(Op.getOperand(0), ~KnownZero & NewMask, 1256 KnownZero2, KnownOne2, TLO, Depth+1)) 1257 return true; 1258 assert((KnownZero2 & KnownOne2) == 0 && "Bits known to be one AND zero?"); 1259 1260 // If all of the demanded bits are known one on one side, return the other. 1261 // These bits cannot contribute to the result of the 'and'. 1262 if ((NewMask & ~KnownZero2 & KnownOne) == (~KnownZero2 & NewMask)) 1263 return TLO.CombineTo(Op, Op.getOperand(0)); 1264 if ((NewMask & ~KnownZero & KnownOne2) == (~KnownZero & NewMask)) 1265 return TLO.CombineTo(Op, Op.getOperand(1)); 1266 // If all of the demanded bits in the inputs are known zeros, return zero. 1267 if ((NewMask & (KnownZero|KnownZero2)) == NewMask) 1268 return TLO.CombineTo(Op, TLO.DAG.getConstant(0, Op.getValueType())); 1269 // If the RHS is a constant, see if we can simplify it. 1270 if (TLO.ShrinkDemandedConstant(Op, ~KnownZero2 & NewMask)) 1271 return true; 1272 // If the operation can be done in a smaller type, do so. 1273 if (TLO.ShrinkDemandedOp(Op, BitWidth, NewMask, dl)) 1274 return true; 1275 1276 // Output known-1 bits are only known if set in both the LHS & RHS. 1277 KnownOne &= KnownOne2; 1278 // Output known-0 are known to be clear if zero in either the LHS | RHS. 1279 KnownZero |= KnownZero2; 1280 break; 1281 case ISD::OR: 1282 if (SimplifyDemandedBits(Op.getOperand(1), NewMask, KnownZero, 1283 KnownOne, TLO, Depth+1)) 1284 return true; 1285 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?"); 1286 if (SimplifyDemandedBits(Op.getOperand(0), ~KnownOne & NewMask, 1287 KnownZero2, KnownOne2, TLO, Depth+1)) 1288 return true; 1289 assert((KnownZero2 & KnownOne2) == 0 && "Bits known to be one AND zero?"); 1290 1291 // If all of the demanded bits are known zero on one side, return the other. 1292 // These bits cannot contribute to the result of the 'or'. 1293 if ((NewMask & ~KnownOne2 & KnownZero) == (~KnownOne2 & NewMask)) 1294 return TLO.CombineTo(Op, Op.getOperand(0)); 1295 if ((NewMask & ~KnownOne & KnownZero2) == (~KnownOne & NewMask)) 1296 return TLO.CombineTo(Op, Op.getOperand(1)); 1297 // If all of the potentially set bits on one side are known to be set on 1298 // the other side, just use the 'other' side. 1299 if ((NewMask & ~KnownZero & KnownOne2) == (~KnownZero & NewMask)) 1300 return TLO.CombineTo(Op, Op.getOperand(0)); 1301 if ((NewMask & ~KnownZero2 & KnownOne) == (~KnownZero2 & NewMask)) 1302 return TLO.CombineTo(Op, Op.getOperand(1)); 1303 // If the RHS is a constant, see if we can simplify it. 1304 if (TLO.ShrinkDemandedConstant(Op, NewMask)) 1305 return true; 1306 // If the operation can be done in a smaller type, do so. 1307 if (TLO.ShrinkDemandedOp(Op, BitWidth, NewMask, dl)) 1308 return true; 1309 1310 // Output known-0 bits are only known if clear in both the LHS & RHS. 1311 KnownZero &= KnownZero2; 1312 // Output known-1 are known to be set if set in either the LHS | RHS. 1313 KnownOne |= KnownOne2; 1314 break; 1315 case ISD::XOR: 1316 if (SimplifyDemandedBits(Op.getOperand(1), NewMask, KnownZero, 1317 KnownOne, TLO, Depth+1)) 1318 return true; 1319 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?"); 1320 if (SimplifyDemandedBits(Op.getOperand(0), NewMask, KnownZero2, 1321 KnownOne2, TLO, Depth+1)) 1322 return true; 1323 assert((KnownZero2 & KnownOne2) == 0 && "Bits known to be one AND zero?"); 1324 1325 // If all of the demanded bits are known zero on one side, return the other. 1326 // These bits cannot contribute to the result of the 'xor'. 1327 if ((KnownZero & NewMask) == NewMask) 1328 return TLO.CombineTo(Op, Op.getOperand(0)); 1329 if ((KnownZero2 & NewMask) == NewMask) 1330 return TLO.CombineTo(Op, Op.getOperand(1)); 1331 // If the operation can be done in a smaller type, do so. 1332 if (TLO.ShrinkDemandedOp(Op, BitWidth, NewMask, dl)) 1333 return true; 1334 1335 // If all of the unknown bits are known to be zero on one side or the other 1336 // (but not both) turn this into an *inclusive* or. 1337 // e.g. (A & C1)^(B & C2) -> (A & C1)|(B & C2) iff C1&C2 == 0 1338 if ((NewMask & ~KnownZero & ~KnownZero2) == 0) 1339 return TLO.CombineTo(Op, TLO.DAG.getNode(ISD::OR, dl, Op.getValueType(), 1340 Op.getOperand(0), 1341 Op.getOperand(1))); 1342 1343 // Output known-0 bits are known if clear or set in both the LHS & RHS. 1344 KnownZeroOut = (KnownZero & KnownZero2) | (KnownOne & KnownOne2); 1345 // Output known-1 are known to be set if set in only one of the LHS, RHS. 1346 KnownOneOut = (KnownZero & KnownOne2) | (KnownOne & KnownZero2); 1347 1348 // If all of the demanded bits on one side are known, and all of the set 1349 // bits on that side are also known to be set on the other side, turn this 1350 // into an AND, as we know the bits will be cleared. 1351 // e.g. (X | C1) ^ C2 --> (X | C1) & ~C2 iff (C1&C2) == C2 1352 // NB: it is okay if more bits are known than are requested 1353 if ((NewMask & (KnownZero|KnownOne)) == NewMask) { // all known on one side 1354 if (KnownOne == KnownOne2) { // set bits are the same on both sides 1355 EVT VT = Op.getValueType(); 1356 SDValue ANDC = TLO.DAG.getConstant(~KnownOne & NewMask, VT); 1357 return TLO.CombineTo(Op, TLO.DAG.getNode(ISD::AND, dl, VT, 1358 Op.getOperand(0), ANDC)); 1359 } 1360 } 1361 1362 // If the RHS is a constant, see if we can simplify it. 1363 // for XOR, we prefer to force bits to 1 if they will make a -1. 1364 // if we can't force bits, try to shrink constant 1365 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op.getOperand(1))) { 1366 APInt Expanded = C->getAPIntValue() | (~NewMask); 1367 // if we can expand it to have all bits set, do it 1368 if (Expanded.isAllOnesValue()) { 1369 if (Expanded != C->getAPIntValue()) { 1370 EVT VT = Op.getValueType(); 1371 SDValue New = TLO.DAG.getNode(Op.getOpcode(), dl,VT, Op.getOperand(0), 1372 TLO.DAG.getConstant(Expanded, VT)); 1373 return TLO.CombineTo(Op, New); 1374 } 1375 // if it already has all the bits set, nothing to change 1376 // but don't shrink either! 1377 } else if (TLO.ShrinkDemandedConstant(Op, NewMask)) { 1378 return true; 1379 } 1380 } 1381 1382 KnownZero = KnownZeroOut; 1383 KnownOne = KnownOneOut; 1384 break; 1385 case ISD::SELECT: 1386 if (SimplifyDemandedBits(Op.getOperand(2), NewMask, KnownZero, 1387 KnownOne, TLO, Depth+1)) 1388 return true; 1389 if (SimplifyDemandedBits(Op.getOperand(1), NewMask, KnownZero2, 1390 KnownOne2, TLO, Depth+1)) 1391 return true; 1392 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?"); 1393 assert((KnownZero2 & KnownOne2) == 0 && "Bits known to be one AND zero?"); 1394 1395 // If the operands are constants, see if we can simplify them. 1396 if (TLO.ShrinkDemandedConstant(Op, NewMask)) 1397 return true; 1398 1399 // Only known if known in both the LHS and RHS. 1400 KnownOne &= KnownOne2; 1401 KnownZero &= KnownZero2; 1402 break; 1403 case ISD::SELECT_CC: 1404 if (SimplifyDemandedBits(Op.getOperand(3), NewMask, KnownZero, 1405 KnownOne, TLO, Depth+1)) 1406 return true; 1407 if (SimplifyDemandedBits(Op.getOperand(2), NewMask, KnownZero2, 1408 KnownOne2, TLO, Depth+1)) 1409 return true; 1410 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?"); 1411 assert((KnownZero2 & KnownOne2) == 0 && "Bits known to be one AND zero?"); 1412 1413 // If the operands are constants, see if we can simplify them. 1414 if (TLO.ShrinkDemandedConstant(Op, NewMask)) 1415 return true; 1416 1417 // Only known if known in both the LHS and RHS. 1418 KnownOne &= KnownOne2; 1419 KnownZero &= KnownZero2; 1420 break; 1421 case ISD::SHL: 1422 if (ConstantSDNode *SA = dyn_cast<ConstantSDNode>(Op.getOperand(1))) { 1423 unsigned ShAmt = SA->getZExtValue(); 1424 SDValue InOp = Op.getOperand(0); 1425 1426 // If the shift count is an invalid immediate, don't do anything. 1427 if (ShAmt >= BitWidth) 1428 break; 1429 1430 // If this is ((X >>u C1) << ShAmt), see if we can simplify this into a 1431 // single shift. We can do this if the bottom bits (which are shifted 1432 // out) are never demanded. 1433 if (InOp.getOpcode() == ISD::SRL && 1434 isa<ConstantSDNode>(InOp.getOperand(1))) { 1435 if (ShAmt && (NewMask & APInt::getLowBitsSet(BitWidth, ShAmt)) == 0) { 1436 unsigned C1= cast<ConstantSDNode>(InOp.getOperand(1))->getZExtValue(); 1437 unsigned Opc = ISD::SHL; 1438 int Diff = ShAmt-C1; 1439 if (Diff < 0) { 1440 Diff = -Diff; 1441 Opc = ISD::SRL; 1442 } 1443 1444 SDValue NewSA = 1445 TLO.DAG.getConstant(Diff, Op.getOperand(1).getValueType()); 1446 EVT VT = Op.getValueType(); 1447 return TLO.CombineTo(Op, TLO.DAG.getNode(Opc, dl, VT, 1448 InOp.getOperand(0), NewSA)); 1449 } 1450 } 1451 1452 if (SimplifyDemandedBits(InOp, NewMask.lshr(ShAmt), 1453 KnownZero, KnownOne, TLO, Depth+1)) 1454 return true; 1455 1456 // Convert (shl (anyext x, c)) to (anyext (shl x, c)) if the high bits 1457 // are not demanded. This will likely allow the anyext to be folded away. 1458 if (InOp.getNode()->getOpcode() == ISD::ANY_EXTEND) { 1459 SDValue InnerOp = InOp.getNode()->getOperand(0); 1460 EVT InnerVT = InnerOp.getValueType(); 1461 unsigned InnerBits = InnerVT.getSizeInBits(); 1462 if (ShAmt < InnerBits && NewMask.lshr(InnerBits) == 0 && 1463 isTypeDesirableForOp(ISD::SHL, InnerVT)) { 1464 EVT ShTy = getShiftAmountTy(InnerVT); 1465 if (!APInt(BitWidth, ShAmt).isIntN(ShTy.getSizeInBits())) 1466 ShTy = InnerVT; 1467 SDValue NarrowShl = 1468 TLO.DAG.getNode(ISD::SHL, dl, InnerVT, InnerOp, 1469 TLO.DAG.getConstant(ShAmt, ShTy)); 1470 return 1471 TLO.CombineTo(Op, 1472 TLO.DAG.getNode(ISD::ANY_EXTEND, dl, Op.getValueType(), 1473 NarrowShl)); 1474 } 1475 } 1476 1477 KnownZero <<= SA->getZExtValue(); 1478 KnownOne <<= SA->getZExtValue(); 1479 // low bits known zero. 1480 KnownZero |= APInt::getLowBitsSet(BitWidth, SA->getZExtValue()); 1481 } 1482 break; 1483 case ISD::SRL: 1484 if (ConstantSDNode *SA = dyn_cast<ConstantSDNode>(Op.getOperand(1))) { 1485 EVT VT = Op.getValueType(); 1486 unsigned ShAmt = SA->getZExtValue(); 1487 unsigned VTSize = VT.getSizeInBits(); 1488 SDValue InOp = Op.getOperand(0); 1489 1490 // If the shift count is an invalid immediate, don't do anything. 1491 if (ShAmt >= BitWidth) 1492 break; 1493 1494 // If this is ((X << C1) >>u ShAmt), see if we can simplify this into a 1495 // single shift. We can do this if the top bits (which are shifted out) 1496 // are never demanded. 1497 if (InOp.getOpcode() == ISD::SHL && 1498 isa<ConstantSDNode>(InOp.getOperand(1))) { 1499 if (ShAmt && (NewMask & APInt::getHighBitsSet(VTSize, ShAmt)) == 0) { 1500 unsigned C1= cast<ConstantSDNode>(InOp.getOperand(1))->getZExtValue(); 1501 unsigned Opc = ISD::SRL; 1502 int Diff = ShAmt-C1; 1503 if (Diff < 0) { 1504 Diff = -Diff; 1505 Opc = ISD::SHL; 1506 } 1507 1508 SDValue NewSA = 1509 TLO.DAG.getConstant(Diff, Op.getOperand(1).getValueType()); 1510 return TLO.CombineTo(Op, TLO.DAG.getNode(Opc, dl, VT, 1511 InOp.getOperand(0), NewSA)); 1512 } 1513 } 1514 1515 // Compute the new bits that are at the top now. 1516 if (SimplifyDemandedBits(InOp, (NewMask << ShAmt), 1517 KnownZero, KnownOne, TLO, Depth+1)) 1518 return true; 1519 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?"); 1520 KnownZero = KnownZero.lshr(ShAmt); 1521 KnownOne = KnownOne.lshr(ShAmt); 1522 1523 APInt HighBits = APInt::getHighBitsSet(BitWidth, ShAmt); 1524 KnownZero |= HighBits; // High bits known zero. 1525 } 1526 break; 1527 case ISD::SRA: 1528 // If this is an arithmetic shift right and only the low-bit is set, we can 1529 // always convert this into a logical shr, even if the shift amount is 1530 // variable. The low bit of the shift cannot be an input sign bit unless 1531 // the shift amount is >= the size of the datatype, which is undefined. 1532 if (NewMask == 1) 1533 return TLO.CombineTo(Op, 1534 TLO.DAG.getNode(ISD::SRL, dl, Op.getValueType(), 1535 Op.getOperand(0), Op.getOperand(1))); 1536 1537 if (ConstantSDNode *SA = dyn_cast<ConstantSDNode>(Op.getOperand(1))) { 1538 EVT VT = Op.getValueType(); 1539 unsigned ShAmt = SA->getZExtValue(); 1540 1541 // If the shift count is an invalid immediate, don't do anything. 1542 if (ShAmt >= BitWidth) 1543 break; 1544 1545 APInt InDemandedMask = (NewMask << ShAmt); 1546 1547 // If any of the demanded bits are produced by the sign extension, we also 1548 // demand the input sign bit. 1549 APInt HighBits = APInt::getHighBitsSet(BitWidth, ShAmt); 1550 if (HighBits.intersects(NewMask)) 1551 InDemandedMask |= APInt::getSignBit(VT.getScalarType().getSizeInBits()); 1552 1553 if (SimplifyDemandedBits(Op.getOperand(0), InDemandedMask, 1554 KnownZero, KnownOne, TLO, Depth+1)) 1555 return true; 1556 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?"); 1557 KnownZero = KnownZero.lshr(ShAmt); 1558 KnownOne = KnownOne.lshr(ShAmt); 1559 1560 // Handle the sign bit, adjusted to where it is now in the mask. 1561 APInt SignBit = APInt::getSignBit(BitWidth).lshr(ShAmt); 1562 1563 // If the input sign bit is known to be zero, or if none of the top bits 1564 // are demanded, turn this into an unsigned shift right. 1565 if (KnownZero.intersects(SignBit) || (HighBits & ~NewMask) == HighBits) { 1566 return TLO.CombineTo(Op, TLO.DAG.getNode(ISD::SRL, dl, VT, 1567 Op.getOperand(0), 1568 Op.getOperand(1))); 1569 } else if (KnownOne.intersects(SignBit)) { // New bits are known one. 1570 KnownOne |= HighBits; 1571 } 1572 } 1573 break; 1574 case ISD::SIGN_EXTEND_INREG: { 1575 EVT ExVT = cast<VTSDNode>(Op.getOperand(1))->getVT(); 1576 1577 APInt MsbMask = APInt::getHighBitsSet(BitWidth, 1); 1578 // If we only care about the highest bit, don't bother shifting right. 1579 if (MsbMask == DemandedMask) { 1580 unsigned ShAmt = ExVT.getScalarType().getSizeInBits(); 1581 SDValue InOp = Op.getOperand(0); 1582 1583 // Compute the correct shift amount type, which must be getShiftAmountTy 1584 // for scalar types after legalization. 1585 EVT ShiftAmtTy = Op.getValueType(); 1586 if (TLO.LegalTypes() && !ShiftAmtTy.isVector()) 1587 ShiftAmtTy = getShiftAmountTy(ShiftAmtTy); 1588 1589 SDValue ShiftAmt = TLO.DAG.getConstant(BitWidth - ShAmt, ShiftAmtTy); 1590 return TLO.CombineTo(Op, TLO.DAG.getNode(ISD::SHL, dl, 1591 Op.getValueType(), InOp, ShiftAmt)); 1592 } 1593 1594 // Sign extension. Compute the demanded bits in the result that are not 1595 // present in the input. 1596 APInt NewBits = 1597 APInt::getHighBitsSet(BitWidth, 1598 BitWidth - ExVT.getScalarType().getSizeInBits()); 1599 1600 // If none of the extended bits are demanded, eliminate the sextinreg. 1601 if ((NewBits & NewMask) == 0) 1602 return TLO.CombineTo(Op, Op.getOperand(0)); 1603 1604 APInt InSignBit = 1605 APInt::getSignBit(ExVT.getScalarType().getSizeInBits()).zext(BitWidth); 1606 APInt InputDemandedBits = 1607 APInt::getLowBitsSet(BitWidth, 1608 ExVT.getScalarType().getSizeInBits()) & 1609 NewMask; 1610 1611 // Since the sign extended bits are demanded, we know that the sign 1612 // bit is demanded. 1613 InputDemandedBits |= InSignBit; 1614 1615 if (SimplifyDemandedBits(Op.getOperand(0), InputDemandedBits, 1616 KnownZero, KnownOne, TLO, Depth+1)) 1617 return true; 1618 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?"); 1619 1620 // If the sign bit of the input is known set or clear, then we know the 1621 // top bits of the result. 1622 1623 // If the input sign bit is known zero, convert this into a zero extension. 1624 if (KnownZero.intersects(InSignBit)) 1625 return TLO.CombineTo(Op, 1626 TLO.DAG.getZeroExtendInReg(Op.getOperand(0),dl,ExVT)); 1627 1628 if (KnownOne.intersects(InSignBit)) { // Input sign bit known set 1629 KnownOne |= NewBits; 1630 KnownZero &= ~NewBits; 1631 } else { // Input sign bit unknown 1632 KnownZero &= ~NewBits; 1633 KnownOne &= ~NewBits; 1634 } 1635 break; 1636 } 1637 case ISD::ZERO_EXTEND: { 1638 unsigned OperandBitWidth = 1639 Op.getOperand(0).getValueType().getScalarType().getSizeInBits(); 1640 APInt InMask = NewMask.trunc(OperandBitWidth); 1641 1642 // If none of the top bits are demanded, convert this into an any_extend. 1643 APInt NewBits = 1644 APInt::getHighBitsSet(BitWidth, BitWidth - OperandBitWidth) & NewMask; 1645 if (!NewBits.intersects(NewMask)) 1646 return TLO.CombineTo(Op, TLO.DAG.getNode(ISD::ANY_EXTEND, dl, 1647 Op.getValueType(), 1648 Op.getOperand(0))); 1649 1650 if (SimplifyDemandedBits(Op.getOperand(0), InMask, 1651 KnownZero, KnownOne, TLO, Depth+1)) 1652 return true; 1653 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?"); 1654 KnownZero = KnownZero.zext(BitWidth); 1655 KnownOne = KnownOne.zext(BitWidth); 1656 KnownZero |= NewBits; 1657 break; 1658 } 1659 case ISD::SIGN_EXTEND: { 1660 EVT InVT = Op.getOperand(0).getValueType(); 1661 unsigned InBits = InVT.getScalarType().getSizeInBits(); 1662 APInt InMask = APInt::getLowBitsSet(BitWidth, InBits); 1663 APInt InSignBit = APInt::getBitsSet(BitWidth, InBits - 1, InBits); 1664 APInt NewBits = ~InMask & NewMask; 1665 1666 // If none of the top bits are demanded, convert this into an any_extend. 1667 if (NewBits == 0) 1668 return TLO.CombineTo(Op,TLO.DAG.getNode(ISD::ANY_EXTEND, dl, 1669 Op.getValueType(), 1670 Op.getOperand(0))); 1671 1672 // Since some of the sign extended bits are demanded, we know that the sign 1673 // bit is demanded. 1674 APInt InDemandedBits = InMask & NewMask; 1675 InDemandedBits |= InSignBit; 1676 InDemandedBits = InDemandedBits.trunc(InBits); 1677 1678 if (SimplifyDemandedBits(Op.getOperand(0), InDemandedBits, KnownZero, 1679 KnownOne, TLO, Depth+1)) 1680 return true; 1681 KnownZero = KnownZero.zext(BitWidth); 1682 KnownOne = KnownOne.zext(BitWidth); 1683 1684 // If the sign bit is known zero, convert this to a zero extend. 1685 if (KnownZero.intersects(InSignBit)) 1686 return TLO.CombineTo(Op, TLO.DAG.getNode(ISD::ZERO_EXTEND, dl, 1687 Op.getValueType(), 1688 Op.getOperand(0))); 1689 1690 // If the sign bit is known one, the top bits match. 1691 if (KnownOne.intersects(InSignBit)) { 1692 KnownOne |= NewBits; 1693 assert((KnownZero & NewBits) == 0); 1694 } else { // Otherwise, top bits aren't known. 1695 assert((KnownOne & NewBits) == 0); 1696 assert((KnownZero & NewBits) == 0); 1697 } 1698 break; 1699 } 1700 case ISD::ANY_EXTEND: { 1701 unsigned OperandBitWidth = 1702 Op.getOperand(0).getValueType().getScalarType().getSizeInBits(); 1703 APInt InMask = NewMask.trunc(OperandBitWidth); 1704 if (SimplifyDemandedBits(Op.getOperand(0), InMask, 1705 KnownZero, KnownOne, TLO, Depth+1)) 1706 return true; 1707 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?"); 1708 KnownZero = KnownZero.zext(BitWidth); 1709 KnownOne = KnownOne.zext(BitWidth); 1710 break; 1711 } 1712 case ISD::TRUNCATE: { 1713 // Simplify the input, using demanded bit information, and compute the known 1714 // zero/one bits live out. 1715 unsigned OperandBitWidth = 1716 Op.getOperand(0).getValueType().getScalarType().getSizeInBits(); 1717 APInt TruncMask = NewMask.zext(OperandBitWidth); 1718 if (SimplifyDemandedBits(Op.getOperand(0), TruncMask, 1719 KnownZero, KnownOne, TLO, Depth+1)) 1720 return true; 1721 KnownZero = KnownZero.trunc(BitWidth); 1722 KnownOne = KnownOne.trunc(BitWidth); 1723 1724 // If the input is only used by this truncate, see if we can shrink it based 1725 // on the known demanded bits. 1726 if (Op.getOperand(0).getNode()->hasOneUse()) { 1727 SDValue In = Op.getOperand(0); 1728 switch (In.getOpcode()) { 1729 default: break; 1730 case ISD::SRL: 1731 // Shrink SRL by a constant if none of the high bits shifted in are 1732 // demanded. 1733 if (TLO.LegalTypes() && 1734 !isTypeDesirableForOp(ISD::SRL, Op.getValueType())) 1735 // Do not turn (vt1 truncate (vt2 srl)) into (vt1 srl) if vt1 is 1736 // undesirable. 1737 break; 1738 ConstantSDNode *ShAmt = dyn_cast<ConstantSDNode>(In.getOperand(1)); 1739 if (!ShAmt) 1740 break; 1741 SDValue Shift = In.getOperand(1); 1742 if (TLO.LegalTypes()) { 1743 uint64_t ShVal = ShAmt->getZExtValue(); 1744 Shift = 1745 TLO.DAG.getConstant(ShVal, getShiftAmountTy(Op.getValueType())); 1746 } 1747 1748 APInt HighBits = APInt::getHighBitsSet(OperandBitWidth, 1749 OperandBitWidth - BitWidth); 1750 HighBits = HighBits.lshr(ShAmt->getZExtValue()).trunc(BitWidth); 1751 1752 if (ShAmt->getZExtValue() < BitWidth && !(HighBits & NewMask)) { 1753 // None of the shifted in bits are needed. Add a truncate of the 1754 // shift input, then shift it. 1755 SDValue NewTrunc = TLO.DAG.getNode(ISD::TRUNCATE, dl, 1756 Op.getValueType(), 1757 In.getOperand(0)); 1758 return TLO.CombineTo(Op, TLO.DAG.getNode(ISD::SRL, dl, 1759 Op.getValueType(), 1760 NewTrunc, 1761 Shift)); 1762 } 1763 break; 1764 } 1765 } 1766 1767 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?"); 1768 break; 1769 } 1770 case ISD::AssertZext: { 1771 // AssertZext demands all of the high bits, plus any of the low bits 1772 // demanded by its users. 1773 EVT VT = cast<VTSDNode>(Op.getOperand(1))->getVT(); 1774 APInt InMask = APInt::getLowBitsSet(BitWidth, 1775 VT.getSizeInBits()); 1776 if (SimplifyDemandedBits(Op.getOperand(0), ~InMask | NewMask, 1777 KnownZero, KnownOne, TLO, Depth+1)) 1778 return true; 1779 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?"); 1780 1781 KnownZero |= ~InMask & NewMask; 1782 break; 1783 } 1784 case ISD::BITCAST: 1785 // If this is an FP->Int bitcast and if the sign bit is the only 1786 // thing demanded, turn this into a FGETSIGN. 1787 if (!TLO.LegalOperations() && 1788 !Op.getValueType().isVector() && 1789 !Op.getOperand(0).getValueType().isVector() && 1790 NewMask == APInt::getSignBit(Op.getValueType().getSizeInBits()) && 1791 Op.getOperand(0).getValueType().isFloatingPoint()) { 1792 bool OpVTLegal = isOperationLegalOrCustom(ISD::FGETSIGN, Op.getValueType()); 1793 bool i32Legal = isOperationLegalOrCustom(ISD::FGETSIGN, MVT::i32); 1794 if ((OpVTLegal || i32Legal) && Op.getValueType().isSimple()) { 1795 EVT Ty = OpVTLegal ? Op.getValueType() : MVT::i32; 1796 // Make a FGETSIGN + SHL to move the sign bit into the appropriate 1797 // place. We expect the SHL to be eliminated by other optimizations. 1798 SDValue Sign = TLO.DAG.getNode(ISD::FGETSIGN, dl, Ty, Op.getOperand(0)); 1799 unsigned OpVTSizeInBits = Op.getValueType().getSizeInBits(); 1800 if (!OpVTLegal && OpVTSizeInBits > 32) 1801 Sign = TLO.DAG.getNode(ISD::ZERO_EXTEND, dl, Op.getValueType(), Sign); 1802 unsigned ShVal = Op.getValueType().getSizeInBits()-1; 1803 SDValue ShAmt = TLO.DAG.getConstant(ShVal, Op.getValueType()); 1804 return TLO.CombineTo(Op, TLO.DAG.getNode(ISD::SHL, dl, 1805 Op.getValueType(), 1806 Sign, ShAmt)); 1807 } 1808 } 1809 break; 1810 case ISD::ADD: 1811 case ISD::MUL: 1812 case ISD::SUB: { 1813 // Add, Sub, and Mul don't demand any bits in positions beyond that 1814 // of the highest bit demanded of them. 1815 APInt LoMask = APInt::getLowBitsSet(BitWidth, 1816 BitWidth - NewMask.countLeadingZeros()); 1817 if (SimplifyDemandedBits(Op.getOperand(0), LoMask, KnownZero2, 1818 KnownOne2, TLO, Depth+1)) 1819 return true; 1820 if (SimplifyDemandedBits(Op.getOperand(1), LoMask, KnownZero2, 1821 KnownOne2, TLO, Depth+1)) 1822 return true; 1823 // See if the operation should be performed at a smaller bit width. 1824 if (TLO.ShrinkDemandedOp(Op, BitWidth, NewMask, dl)) 1825 return true; 1826 } 1827 // FALL THROUGH 1828 default: 1829 // Just use ComputeMaskedBits to compute output bits. 1830 TLO.DAG.ComputeMaskedBits(Op, KnownZero, KnownOne, Depth); 1831 break; 1832 } 1833 1834 // If we know the value of all of the demanded bits, return this as a 1835 // constant. 1836 if ((NewMask & (KnownZero|KnownOne)) == NewMask) 1837 return TLO.CombineTo(Op, TLO.DAG.getConstant(KnownOne, Op.getValueType())); 1838 1839 return false; 1840} 1841 1842/// computeMaskedBitsForTargetNode - Determine which of the bits specified 1843/// in Mask are known to be either zero or one and return them in the 1844/// KnownZero/KnownOne bitsets. 1845void TargetLowering::computeMaskedBitsForTargetNode(const SDValue Op, 1846 APInt &KnownZero, 1847 APInt &KnownOne, 1848 const SelectionDAG &DAG, 1849 unsigned Depth) const { 1850 assert((Op.getOpcode() >= ISD::BUILTIN_OP_END || 1851 Op.getOpcode() == ISD::INTRINSIC_WO_CHAIN || 1852 Op.getOpcode() == ISD::INTRINSIC_W_CHAIN || 1853 Op.getOpcode() == ISD::INTRINSIC_VOID) && 1854 "Should use MaskedValueIsZero if you don't know whether Op" 1855 " is a target node!"); 1856 KnownZero = KnownOne = APInt(KnownOne.getBitWidth(), 0); 1857} 1858 1859/// ComputeNumSignBitsForTargetNode - This method can be implemented by 1860/// targets that want to expose additional information about sign bits to the 1861/// DAG Combiner. 1862unsigned TargetLowering::ComputeNumSignBitsForTargetNode(SDValue Op, 1863 unsigned Depth) const { 1864 assert((Op.getOpcode() >= ISD::BUILTIN_OP_END || 1865 Op.getOpcode() == ISD::INTRINSIC_WO_CHAIN || 1866 Op.getOpcode() == ISD::INTRINSIC_W_CHAIN || 1867 Op.getOpcode() == ISD::INTRINSIC_VOID) && 1868 "Should use ComputeNumSignBits if you don't know whether Op" 1869 " is a target node!"); 1870 return 1; 1871} 1872 1873/// ValueHasExactlyOneBitSet - Test if the given value is known to have exactly 1874/// one bit set. This differs from ComputeMaskedBits in that it doesn't need to 1875/// determine which bit is set. 1876/// 1877static bool ValueHasExactlyOneBitSet(SDValue Val, const SelectionDAG &DAG) { 1878 // A left-shift of a constant one will have exactly one bit set, because 1879 // shifting the bit off the end is undefined. 1880 if (Val.getOpcode() == ISD::SHL) 1881 if (ConstantSDNode *C = 1882 dyn_cast<ConstantSDNode>(Val.getNode()->getOperand(0))) 1883 if (C->getAPIntValue() == 1) 1884 return true; 1885 1886 // Similarly, a right-shift of a constant sign-bit will have exactly 1887 // one bit set. 1888 if (Val.getOpcode() == ISD::SRL) 1889 if (ConstantSDNode *C = 1890 dyn_cast<ConstantSDNode>(Val.getNode()->getOperand(0))) 1891 if (C->getAPIntValue().isSignBit()) 1892 return true; 1893 1894 // More could be done here, though the above checks are enough 1895 // to handle some common cases. 1896 1897 // Fall back to ComputeMaskedBits to catch other known cases. 1898 EVT OpVT = Val.getValueType(); 1899 unsigned BitWidth = OpVT.getScalarType().getSizeInBits(); 1900 APInt KnownZero, KnownOne; 1901 DAG.ComputeMaskedBits(Val, KnownZero, KnownOne); 1902 return (KnownZero.countPopulation() == BitWidth - 1) && 1903 (KnownOne.countPopulation() == 1); 1904} 1905 1906/// SimplifySetCC - Try to simplify a setcc built with the specified operands 1907/// and cc. If it is unable to simplify it, return a null SDValue. 1908SDValue 1909TargetLowering::SimplifySetCC(EVT VT, SDValue N0, SDValue N1, 1910 ISD::CondCode Cond, bool foldBooleans, 1911 DAGCombinerInfo &DCI, DebugLoc dl) const { 1912 SelectionDAG &DAG = DCI.DAG; 1913 1914 // These setcc operations always fold. 1915 switch (Cond) { 1916 default: break; 1917 case ISD::SETFALSE: 1918 case ISD::SETFALSE2: return DAG.getConstant(0, VT); 1919 case ISD::SETTRUE: 1920 case ISD::SETTRUE2: return DAG.getConstant(1, VT); 1921 } 1922 1923 // Ensure that the constant occurs on the RHS, and fold constant 1924 // comparisons. 1925 if (isa<ConstantSDNode>(N0.getNode())) 1926 return DAG.getSetCC(dl, VT, N1, N0, ISD::getSetCCSwappedOperands(Cond)); 1927 1928 if (ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1.getNode())) { 1929 const APInt &C1 = N1C->getAPIntValue(); 1930 1931 // If the LHS is '(srl (ctlz x), 5)', the RHS is 0/1, and this is an 1932 // equality comparison, then we're just comparing whether X itself is 1933 // zero. 1934 if (N0.getOpcode() == ISD::SRL && (C1 == 0 || C1 == 1) && 1935 N0.getOperand(0).getOpcode() == ISD::CTLZ && 1936 N0.getOperand(1).getOpcode() == ISD::Constant) { 1937 const APInt &ShAmt 1938 = cast<ConstantSDNode>(N0.getOperand(1))->getAPIntValue(); 1939 if ((Cond == ISD::SETEQ || Cond == ISD::SETNE) && 1940 ShAmt == Log2_32(N0.getValueType().getSizeInBits())) { 1941 if ((C1 == 0) == (Cond == ISD::SETEQ)) { 1942 // (srl (ctlz x), 5) == 0 -> X != 0 1943 // (srl (ctlz x), 5) != 1 -> X != 0 1944 Cond = ISD::SETNE; 1945 } else { 1946 // (srl (ctlz x), 5) != 0 -> X == 0 1947 // (srl (ctlz x), 5) == 1 -> X == 0 1948 Cond = ISD::SETEQ; 1949 } 1950 SDValue Zero = DAG.getConstant(0, N0.getValueType()); 1951 return DAG.getSetCC(dl, VT, N0.getOperand(0).getOperand(0), 1952 Zero, Cond); 1953 } 1954 } 1955 1956 SDValue CTPOP = N0; 1957 // Look through truncs that don't change the value of a ctpop. 1958 if (N0.hasOneUse() && N0.getOpcode() == ISD::TRUNCATE) 1959 CTPOP = N0.getOperand(0); 1960 1961 if (CTPOP.hasOneUse() && CTPOP.getOpcode() == ISD::CTPOP && 1962 (N0 == CTPOP || N0.getValueType().getSizeInBits() > 1963 Log2_32_Ceil(CTPOP.getValueType().getSizeInBits()))) { 1964 EVT CTVT = CTPOP.getValueType(); 1965 SDValue CTOp = CTPOP.getOperand(0); 1966 1967 // (ctpop x) u< 2 -> (x & x-1) == 0 1968 // (ctpop x) u> 1 -> (x & x-1) != 0 1969 if ((Cond == ISD::SETULT && C1 == 2) || (Cond == ISD::SETUGT && C1 == 1)){ 1970 SDValue Sub = DAG.getNode(ISD::SUB, dl, CTVT, CTOp, 1971 DAG.getConstant(1, CTVT)); 1972 SDValue And = DAG.getNode(ISD::AND, dl, CTVT, CTOp, Sub); 1973 ISD::CondCode CC = Cond == ISD::SETULT ? ISD::SETEQ : ISD::SETNE; 1974 return DAG.getSetCC(dl, VT, And, DAG.getConstant(0, CTVT), CC); 1975 } 1976 1977 // TODO: (ctpop x) == 1 -> x && (x & x-1) == 0 iff ctpop is illegal. 1978 } 1979 1980 // (zext x) == C --> x == (trunc C) 1981 if (DCI.isBeforeLegalize() && N0->hasOneUse() && 1982 (Cond == ISD::SETEQ || Cond == ISD::SETNE)) { 1983 unsigned MinBits = N0.getValueSizeInBits(); 1984 SDValue PreZExt; 1985 if (N0->getOpcode() == ISD::ZERO_EXTEND) { 1986 // ZExt 1987 MinBits = N0->getOperand(0).getValueSizeInBits(); 1988 PreZExt = N0->getOperand(0); 1989 } else if (N0->getOpcode() == ISD::AND) { 1990 // DAGCombine turns costly ZExts into ANDs 1991 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(N0->getOperand(1))) 1992 if ((C->getAPIntValue()+1).isPowerOf2()) { 1993 MinBits = C->getAPIntValue().countTrailingOnes(); 1994 PreZExt = N0->getOperand(0); 1995 } 1996 } else if (LoadSDNode *LN0 = dyn_cast<LoadSDNode>(N0)) { 1997 // ZEXTLOAD 1998 if (LN0->getExtensionType() == ISD::ZEXTLOAD) { 1999 MinBits = LN0->getMemoryVT().getSizeInBits(); 2000 PreZExt = N0; 2001 } 2002 } 2003 2004 // Make sure we're not losing bits from the constant. 2005 if (MinBits < C1.getBitWidth() && MinBits > C1.getActiveBits()) { 2006 EVT MinVT = EVT::getIntegerVT(*DAG.getContext(), MinBits); 2007 if (isTypeDesirableForOp(ISD::SETCC, MinVT)) { 2008 // Will get folded away. 2009 SDValue Trunc = DAG.getNode(ISD::TRUNCATE, dl, MinVT, PreZExt); 2010 SDValue C = DAG.getConstant(C1.trunc(MinBits), MinVT); 2011 return DAG.getSetCC(dl, VT, Trunc, C, Cond); 2012 } 2013 } 2014 } 2015 2016 // If the LHS is '(and load, const)', the RHS is 0, 2017 // the test is for equality or unsigned, and all 1 bits of the const are 2018 // in the same partial word, see if we can shorten the load. 2019 if (DCI.isBeforeLegalize() && 2020 N0.getOpcode() == ISD::AND && C1 == 0 && 2021 N0.getNode()->hasOneUse() && 2022 isa<LoadSDNode>(N0.getOperand(0)) && 2023 N0.getOperand(0).getNode()->hasOneUse() && 2024 isa<ConstantSDNode>(N0.getOperand(1))) { 2025 LoadSDNode *Lod = cast<LoadSDNode>(N0.getOperand(0)); 2026 APInt bestMask; 2027 unsigned bestWidth = 0, bestOffset = 0; 2028 if (!Lod->isVolatile() && Lod->isUnindexed()) { 2029 unsigned origWidth = N0.getValueType().getSizeInBits(); 2030 unsigned maskWidth = origWidth; 2031 // We can narrow (e.g.) 16-bit extending loads on 32-bit target to 2032 // 8 bits, but have to be careful... 2033 if (Lod->getExtensionType() != ISD::NON_EXTLOAD) 2034 origWidth = Lod->getMemoryVT().getSizeInBits(); 2035 const APInt &Mask = 2036 cast<ConstantSDNode>(N0.getOperand(1))->getAPIntValue(); 2037 for (unsigned width = origWidth / 2; width>=8; width /= 2) { 2038 APInt newMask = APInt::getLowBitsSet(maskWidth, width); 2039 for (unsigned offset=0; offset<origWidth/width; offset++) { 2040 if ((newMask & Mask) == Mask) { 2041 if (!TD->isLittleEndian()) 2042 bestOffset = (origWidth/width - offset - 1) * (width/8); 2043 else 2044 bestOffset = (uint64_t)offset * (width/8); 2045 bestMask = Mask.lshr(offset * (width/8) * 8); 2046 bestWidth = width; 2047 break; 2048 } 2049 newMask = newMask << width; 2050 } 2051 } 2052 } 2053 if (bestWidth) { 2054 EVT newVT = EVT::getIntegerVT(*DAG.getContext(), bestWidth); 2055 if (newVT.isRound()) { 2056 EVT PtrType = Lod->getOperand(1).getValueType(); 2057 SDValue Ptr = Lod->getBasePtr(); 2058 if (bestOffset != 0) 2059 Ptr = DAG.getNode(ISD::ADD, dl, PtrType, Lod->getBasePtr(), 2060 DAG.getConstant(bestOffset, PtrType)); 2061 unsigned NewAlign = MinAlign(Lod->getAlignment(), bestOffset); 2062 SDValue NewLoad = DAG.getLoad(newVT, dl, Lod->getChain(), Ptr, 2063 Lod->getPointerInfo().getWithOffset(bestOffset), 2064 false, false, false, NewAlign); 2065 return DAG.getSetCC(dl, VT, 2066 DAG.getNode(ISD::AND, dl, newVT, NewLoad, 2067 DAG.getConstant(bestMask.trunc(bestWidth), 2068 newVT)), 2069 DAG.getConstant(0LL, newVT), Cond); 2070 } 2071 } 2072 } 2073 2074 // If the LHS is a ZERO_EXTEND, perform the comparison on the input. 2075 if (N0.getOpcode() == ISD::ZERO_EXTEND) { 2076 unsigned InSize = N0.getOperand(0).getValueType().getSizeInBits(); 2077 2078 // If the comparison constant has bits in the upper part, the 2079 // zero-extended value could never match. 2080 if (C1.intersects(APInt::getHighBitsSet(C1.getBitWidth(), 2081 C1.getBitWidth() - InSize))) { 2082 switch (Cond) { 2083 case ISD::SETUGT: 2084 case ISD::SETUGE: 2085 case ISD::SETEQ: return DAG.getConstant(0, VT); 2086 case ISD::SETULT: 2087 case ISD::SETULE: 2088 case ISD::SETNE: return DAG.getConstant(1, VT); 2089 case ISD::SETGT: 2090 case ISD::SETGE: 2091 // True if the sign bit of C1 is set. 2092 return DAG.getConstant(C1.isNegative(), VT); 2093 case ISD::SETLT: 2094 case ISD::SETLE: 2095 // True if the sign bit of C1 isn't set. 2096 return DAG.getConstant(C1.isNonNegative(), VT); 2097 default: 2098 break; 2099 } 2100 } 2101 2102 // Otherwise, we can perform the comparison with the low bits. 2103 switch (Cond) { 2104 case ISD::SETEQ: 2105 case ISD::SETNE: 2106 case ISD::SETUGT: 2107 case ISD::SETUGE: 2108 case ISD::SETULT: 2109 case ISD::SETULE: { 2110 EVT newVT = N0.getOperand(0).getValueType(); 2111 if (DCI.isBeforeLegalizeOps() || 2112 (isOperationLegal(ISD::SETCC, newVT) && 2113 getCondCodeAction(Cond, newVT)==Legal)) 2114 return DAG.getSetCC(dl, VT, N0.getOperand(0), 2115 DAG.getConstant(C1.trunc(InSize), newVT), 2116 Cond); 2117 break; 2118 } 2119 default: 2120 break; // todo, be more careful with signed comparisons 2121 } 2122 } else if (N0.getOpcode() == ISD::SIGN_EXTEND_INREG && 2123 (Cond == ISD::SETEQ || Cond == ISD::SETNE)) { 2124 EVT ExtSrcTy = cast<VTSDNode>(N0.getOperand(1))->getVT(); 2125 unsigned ExtSrcTyBits = ExtSrcTy.getSizeInBits(); 2126 EVT ExtDstTy = N0.getValueType(); 2127 unsigned ExtDstTyBits = ExtDstTy.getSizeInBits(); 2128 2129 // If the constant doesn't fit into the number of bits for the source of 2130 // the sign extension, it is impossible for both sides to be equal. 2131 if (C1.getMinSignedBits() > ExtSrcTyBits) 2132 return DAG.getConstant(Cond == ISD::SETNE, VT); 2133 2134 SDValue ZextOp; 2135 EVT Op0Ty = N0.getOperand(0).getValueType(); 2136 if (Op0Ty == ExtSrcTy) { 2137 ZextOp = N0.getOperand(0); 2138 } else { 2139 APInt Imm = APInt::getLowBitsSet(ExtDstTyBits, ExtSrcTyBits); 2140 ZextOp = DAG.getNode(ISD::AND, dl, Op0Ty, N0.getOperand(0), 2141 DAG.getConstant(Imm, Op0Ty)); 2142 } 2143 if (!DCI.isCalledByLegalizer()) 2144 DCI.AddToWorklist(ZextOp.getNode()); 2145 // Otherwise, make this a use of a zext. 2146 return DAG.getSetCC(dl, VT, ZextOp, 2147 DAG.getConstant(C1 & APInt::getLowBitsSet( 2148 ExtDstTyBits, 2149 ExtSrcTyBits), 2150 ExtDstTy), 2151 Cond); 2152 } else if ((N1C->isNullValue() || N1C->getAPIntValue() == 1) && 2153 (Cond == ISD::SETEQ || Cond == ISD::SETNE)) { 2154 // SETCC (SETCC), [0|1], [EQ|NE] -> SETCC 2155 if (N0.getOpcode() == ISD::SETCC && 2156 isTypeLegal(VT) && VT.bitsLE(N0.getValueType())) { 2157 bool TrueWhenTrue = (Cond == ISD::SETEQ) ^ (N1C->getAPIntValue() != 1); 2158 if (TrueWhenTrue) 2159 return DAG.getNode(ISD::TRUNCATE, dl, VT, N0); 2160 // Invert the condition. 2161 ISD::CondCode CC = cast<CondCodeSDNode>(N0.getOperand(2))->get(); 2162 CC = ISD::getSetCCInverse(CC, 2163 N0.getOperand(0).getValueType().isInteger()); 2164 return DAG.getSetCC(dl, VT, N0.getOperand(0), N0.getOperand(1), CC); 2165 } 2166 2167 if ((N0.getOpcode() == ISD::XOR || 2168 (N0.getOpcode() == ISD::AND && 2169 N0.getOperand(0).getOpcode() == ISD::XOR && 2170 N0.getOperand(1) == N0.getOperand(0).getOperand(1))) && 2171 isa<ConstantSDNode>(N0.getOperand(1)) && 2172 cast<ConstantSDNode>(N0.getOperand(1))->getAPIntValue() == 1) { 2173 // If this is (X^1) == 0/1, swap the RHS and eliminate the xor. We 2174 // can only do this if the top bits are known zero. 2175 unsigned BitWidth = N0.getValueSizeInBits(); 2176 if (DAG.MaskedValueIsZero(N0, 2177 APInt::getHighBitsSet(BitWidth, 2178 BitWidth-1))) { 2179 // Okay, get the un-inverted input value. 2180 SDValue Val; 2181 if (N0.getOpcode() == ISD::XOR) 2182 Val = N0.getOperand(0); 2183 else { 2184 assert(N0.getOpcode() == ISD::AND && 2185 N0.getOperand(0).getOpcode() == ISD::XOR); 2186 // ((X^1)&1)^1 -> X & 1 2187 Val = DAG.getNode(ISD::AND, dl, N0.getValueType(), 2188 N0.getOperand(0).getOperand(0), 2189 N0.getOperand(1)); 2190 } 2191 2192 return DAG.getSetCC(dl, VT, Val, N1, 2193 Cond == ISD::SETEQ ? ISD::SETNE : ISD::SETEQ); 2194 } 2195 } else if (N1C->getAPIntValue() == 1 && 2196 (VT == MVT::i1 || 2197 getBooleanContents(false) == ZeroOrOneBooleanContent)) { 2198 SDValue Op0 = N0; 2199 if (Op0.getOpcode() == ISD::TRUNCATE) 2200 Op0 = Op0.getOperand(0); 2201 2202 if ((Op0.getOpcode() == ISD::XOR) && 2203 Op0.getOperand(0).getOpcode() == ISD::SETCC && 2204 Op0.getOperand(1).getOpcode() == ISD::SETCC) { 2205 // (xor (setcc), (setcc)) == / != 1 -> (setcc) != / == (setcc) 2206 Cond = (Cond == ISD::SETEQ) ? ISD::SETNE : ISD::SETEQ; 2207 return DAG.getSetCC(dl, VT, Op0.getOperand(0), Op0.getOperand(1), 2208 Cond); 2209 } else if (Op0.getOpcode() == ISD::AND && 2210 isa<ConstantSDNode>(Op0.getOperand(1)) && 2211 cast<ConstantSDNode>(Op0.getOperand(1))->getAPIntValue() == 1) { 2212 // If this is (X&1) == / != 1, normalize it to (X&1) != / == 0. 2213 if (Op0.getValueType().bitsGT(VT)) 2214 Op0 = DAG.getNode(ISD::AND, dl, VT, 2215 DAG.getNode(ISD::TRUNCATE, dl, VT, Op0.getOperand(0)), 2216 DAG.getConstant(1, VT)); 2217 else if (Op0.getValueType().bitsLT(VT)) 2218 Op0 = DAG.getNode(ISD::AND, dl, VT, 2219 DAG.getNode(ISD::ANY_EXTEND, dl, VT, Op0.getOperand(0)), 2220 DAG.getConstant(1, VT)); 2221 2222 return DAG.getSetCC(dl, VT, Op0, 2223 DAG.getConstant(0, Op0.getValueType()), 2224 Cond == ISD::SETEQ ? ISD::SETNE : ISD::SETEQ); 2225 } 2226 } 2227 } 2228 2229 APInt MinVal, MaxVal; 2230 unsigned OperandBitSize = N1C->getValueType(0).getSizeInBits(); 2231 if (ISD::isSignedIntSetCC(Cond)) { 2232 MinVal = APInt::getSignedMinValue(OperandBitSize); 2233 MaxVal = APInt::getSignedMaxValue(OperandBitSize); 2234 } else { 2235 MinVal = APInt::getMinValue(OperandBitSize); 2236 MaxVal = APInt::getMaxValue(OperandBitSize); 2237 } 2238 2239 // Canonicalize GE/LE comparisons to use GT/LT comparisons. 2240 if (Cond == ISD::SETGE || Cond == ISD::SETUGE) { 2241 if (C1 == MinVal) return DAG.getConstant(1, VT); // X >= MIN --> true 2242 // X >= C0 --> X > (C0-1) 2243 return DAG.getSetCC(dl, VT, N0, 2244 DAG.getConstant(C1-1, N1.getValueType()), 2245 (Cond == ISD::SETGE) ? ISD::SETGT : ISD::SETUGT); 2246 } 2247 2248 if (Cond == ISD::SETLE || Cond == ISD::SETULE) { 2249 if (C1 == MaxVal) return DAG.getConstant(1, VT); // X <= MAX --> true 2250 // X <= C0 --> X < (C0+1) 2251 return DAG.getSetCC(dl, VT, N0, 2252 DAG.getConstant(C1+1, N1.getValueType()), 2253 (Cond == ISD::SETLE) ? ISD::SETLT : ISD::SETULT); 2254 } 2255 2256 if ((Cond == ISD::SETLT || Cond == ISD::SETULT) && C1 == MinVal) 2257 return DAG.getConstant(0, VT); // X < MIN --> false 2258 if ((Cond == ISD::SETGE || Cond == ISD::SETUGE) && C1 == MinVal) 2259 return DAG.getConstant(1, VT); // X >= MIN --> true 2260 if ((Cond == ISD::SETGT || Cond == ISD::SETUGT) && C1 == MaxVal) 2261 return DAG.getConstant(0, VT); // X > MAX --> false 2262 if ((Cond == ISD::SETLE || Cond == ISD::SETULE) && C1 == MaxVal) 2263 return DAG.getConstant(1, VT); // X <= MAX --> true 2264 2265 // Canonicalize setgt X, Min --> setne X, Min 2266 if ((Cond == ISD::SETGT || Cond == ISD::SETUGT) && C1 == MinVal) 2267 return DAG.getSetCC(dl, VT, N0, N1, ISD::SETNE); 2268 // Canonicalize setlt X, Max --> setne X, Max 2269 if ((Cond == ISD::SETLT || Cond == ISD::SETULT) && C1 == MaxVal) 2270 return DAG.getSetCC(dl, VT, N0, N1, ISD::SETNE); 2271 2272 // If we have setult X, 1, turn it into seteq X, 0 2273 if ((Cond == ISD::SETLT || Cond == ISD::SETULT) && C1 == MinVal+1) 2274 return DAG.getSetCC(dl, VT, N0, 2275 DAG.getConstant(MinVal, N0.getValueType()), 2276 ISD::SETEQ); 2277 // If we have setugt X, Max-1, turn it into seteq X, Max 2278 else if ((Cond == ISD::SETGT || Cond == ISD::SETUGT) && C1 == MaxVal-1) 2279 return DAG.getSetCC(dl, VT, N0, 2280 DAG.getConstant(MaxVal, N0.getValueType()), 2281 ISD::SETEQ); 2282 2283 // If we have "setcc X, C0", check to see if we can shrink the immediate 2284 // by changing cc. 2285 2286 // SETUGT X, SINTMAX -> SETLT X, 0 2287 if (Cond == ISD::SETUGT && 2288 C1 == APInt::getSignedMaxValue(OperandBitSize)) 2289 return DAG.getSetCC(dl, VT, N0, 2290 DAG.getConstant(0, N1.getValueType()), 2291 ISD::SETLT); 2292 2293 // SETULT X, SINTMIN -> SETGT X, -1 2294 if (Cond == ISD::SETULT && 2295 C1 == APInt::getSignedMinValue(OperandBitSize)) { 2296 SDValue ConstMinusOne = 2297 DAG.getConstant(APInt::getAllOnesValue(OperandBitSize), 2298 N1.getValueType()); 2299 return DAG.getSetCC(dl, VT, N0, ConstMinusOne, ISD::SETGT); 2300 } 2301 2302 // Fold bit comparisons when we can. 2303 if ((Cond == ISD::SETEQ || Cond == ISD::SETNE) && 2304 (VT == N0.getValueType() || 2305 (isTypeLegal(VT) && VT.bitsLE(N0.getValueType()))) && 2306 N0.getOpcode() == ISD::AND) 2307 if (ConstantSDNode *AndRHS = 2308 dyn_cast<ConstantSDNode>(N0.getOperand(1))) { 2309 EVT ShiftTy = DCI.isBeforeLegalizeOps() ? 2310 getPointerTy() : getShiftAmountTy(N0.getValueType()); 2311 if (Cond == ISD::SETNE && C1 == 0) {// (X & 8) != 0 --> (X & 8) >> 3 2312 // Perform the xform if the AND RHS is a single bit. 2313 if (AndRHS->getAPIntValue().isPowerOf2()) { 2314 return DAG.getNode(ISD::TRUNCATE, dl, VT, 2315 DAG.getNode(ISD::SRL, dl, N0.getValueType(), N0, 2316 DAG.getConstant(AndRHS->getAPIntValue().logBase2(), ShiftTy))); 2317 } 2318 } else if (Cond == ISD::SETEQ && C1 == AndRHS->getAPIntValue()) { 2319 // (X & 8) == 8 --> (X & 8) >> 3 2320 // Perform the xform if C1 is a single bit. 2321 if (C1.isPowerOf2()) { 2322 return DAG.getNode(ISD::TRUNCATE, dl, VT, 2323 DAG.getNode(ISD::SRL, dl, N0.getValueType(), N0, 2324 DAG.getConstant(C1.logBase2(), ShiftTy))); 2325 } 2326 } 2327 } 2328 2329 if (C1.getMinSignedBits() <= 64 && 2330 !isLegalICmpImmediate(C1.getSExtValue())) { 2331 // (X & -256) == 256 -> (X >> 8) == 1 2332 if ((Cond == ISD::SETEQ || Cond == ISD::SETNE) && 2333 N0.getOpcode() == ISD::AND && N0.hasOneUse()) { 2334 if (ConstantSDNode *AndRHS = 2335 dyn_cast<ConstantSDNode>(N0.getOperand(1))) { 2336 const APInt &AndRHSC = AndRHS->getAPIntValue(); 2337 if ((-AndRHSC).isPowerOf2() && (AndRHSC & C1) == C1) { 2338 unsigned ShiftBits = AndRHSC.countTrailingZeros(); 2339 EVT ShiftTy = DCI.isBeforeLegalizeOps() ? 2340 getPointerTy() : getShiftAmountTy(N0.getValueType()); 2341 EVT CmpTy = N0.getValueType(); 2342 SDValue Shift = DAG.getNode(ISD::SRL, dl, CmpTy, N0.getOperand(0), 2343 DAG.getConstant(ShiftBits, ShiftTy)); 2344 SDValue CmpRHS = DAG.getConstant(C1.lshr(ShiftBits), CmpTy); 2345 return DAG.getSetCC(dl, VT, Shift, CmpRHS, Cond); 2346 } 2347 } 2348 } else if (Cond == ISD::SETULT || Cond == ISD::SETUGE || 2349 Cond == ISD::SETULE || Cond == ISD::SETUGT) { 2350 bool AdjOne = (Cond == ISD::SETULE || Cond == ISD::SETUGT); 2351 // X < 0x100000000 -> (X >> 32) < 1 2352 // X >= 0x100000000 -> (X >> 32) >= 1 2353 // X <= 0x0ffffffff -> (X >> 32) < 1 2354 // X > 0x0ffffffff -> (X >> 32) >= 1 2355 unsigned ShiftBits; 2356 APInt NewC = C1; 2357 ISD::CondCode NewCond = Cond; 2358 if (AdjOne) { 2359 ShiftBits = C1.countTrailingOnes(); 2360 NewC = NewC + 1; 2361 NewCond = (Cond == ISD::SETULE) ? ISD::SETULT : ISD::SETUGE; 2362 } else { 2363 ShiftBits = C1.countTrailingZeros(); 2364 } 2365 NewC = NewC.lshr(ShiftBits); 2366 if (ShiftBits && isLegalICmpImmediate(NewC.getSExtValue())) { 2367 EVT ShiftTy = DCI.isBeforeLegalizeOps() ? 2368 getPointerTy() : getShiftAmountTy(N0.getValueType()); 2369 EVT CmpTy = N0.getValueType(); 2370 SDValue Shift = DAG.getNode(ISD::SRL, dl, CmpTy, N0, 2371 DAG.getConstant(ShiftBits, ShiftTy)); 2372 SDValue CmpRHS = DAG.getConstant(NewC, CmpTy); 2373 return DAG.getSetCC(dl, VT, Shift, CmpRHS, NewCond); 2374 } 2375 } 2376 } 2377 } 2378 2379 if (isa<ConstantFPSDNode>(N0.getNode())) { 2380 // Constant fold or commute setcc. 2381 SDValue O = DAG.FoldSetCC(VT, N0, N1, Cond, dl); 2382 if (O.getNode()) return O; 2383 } else if (ConstantFPSDNode *CFP = dyn_cast<ConstantFPSDNode>(N1.getNode())) { 2384 // If the RHS of an FP comparison is a constant, simplify it away in 2385 // some cases. 2386 if (CFP->getValueAPF().isNaN()) { 2387 // If an operand is known to be a nan, we can fold it. 2388 switch (ISD::getUnorderedFlavor(Cond)) { 2389 default: llvm_unreachable("Unknown flavor!"); 2390 case 0: // Known false. 2391 return DAG.getConstant(0, VT); 2392 case 1: // Known true. 2393 return DAG.getConstant(1, VT); 2394 case 2: // Undefined. 2395 return DAG.getUNDEF(VT); 2396 } 2397 } 2398 2399 // Otherwise, we know the RHS is not a NaN. Simplify the node to drop the 2400 // constant if knowing that the operand is non-nan is enough. We prefer to 2401 // have SETO(x,x) instead of SETO(x, 0.0) because this avoids having to 2402 // materialize 0.0. 2403 if (Cond == ISD::SETO || Cond == ISD::SETUO) 2404 return DAG.getSetCC(dl, VT, N0, N0, Cond); 2405 2406 // If the condition is not legal, see if we can find an equivalent one 2407 // which is legal. 2408 if (!isCondCodeLegal(Cond, N0.getValueType())) { 2409 // If the comparison was an awkward floating-point == or != and one of 2410 // the comparison operands is infinity or negative infinity, convert the 2411 // condition to a less-awkward <= or >=. 2412 if (CFP->getValueAPF().isInfinity()) { 2413 if (CFP->getValueAPF().isNegative()) { 2414 if (Cond == ISD::SETOEQ && 2415 isCondCodeLegal(ISD::SETOLE, N0.getValueType())) 2416 return DAG.getSetCC(dl, VT, N0, N1, ISD::SETOLE); 2417 if (Cond == ISD::SETUEQ && 2418 isCondCodeLegal(ISD::SETOLE, N0.getValueType())) 2419 return DAG.getSetCC(dl, VT, N0, N1, ISD::SETULE); 2420 if (Cond == ISD::SETUNE && 2421 isCondCodeLegal(ISD::SETUGT, N0.getValueType())) 2422 return DAG.getSetCC(dl, VT, N0, N1, ISD::SETUGT); 2423 if (Cond == ISD::SETONE && 2424 isCondCodeLegal(ISD::SETUGT, N0.getValueType())) 2425 return DAG.getSetCC(dl, VT, N0, N1, ISD::SETOGT); 2426 } else { 2427 if (Cond == ISD::SETOEQ && 2428 isCondCodeLegal(ISD::SETOGE, N0.getValueType())) 2429 return DAG.getSetCC(dl, VT, N0, N1, ISD::SETOGE); 2430 if (Cond == ISD::SETUEQ && 2431 isCondCodeLegal(ISD::SETOGE, N0.getValueType())) 2432 return DAG.getSetCC(dl, VT, N0, N1, ISD::SETUGE); 2433 if (Cond == ISD::SETUNE && 2434 isCondCodeLegal(ISD::SETULT, N0.getValueType())) 2435 return DAG.getSetCC(dl, VT, N0, N1, ISD::SETULT); 2436 if (Cond == ISD::SETONE && 2437 isCondCodeLegal(ISD::SETULT, N0.getValueType())) 2438 return DAG.getSetCC(dl, VT, N0, N1, ISD::SETOLT); 2439 } 2440 } 2441 } 2442 } 2443 2444 if (N0 == N1) { 2445 // The sext(setcc()) => setcc() optimization relies on the appropriate 2446 // constant being emitted. 2447 uint64_t EqVal = 0; 2448 switch (getBooleanContents(N0.getValueType().isVector())) { 2449 case UndefinedBooleanContent: 2450 case ZeroOrOneBooleanContent: 2451 EqVal = ISD::isTrueWhenEqual(Cond); 2452 break; 2453 case ZeroOrNegativeOneBooleanContent: 2454 EqVal = ISD::isTrueWhenEqual(Cond) ? -1 : 0; 2455 break; 2456 } 2457 2458 // We can always fold X == X for integer setcc's. 2459 if (N0.getValueType().isInteger()) { 2460 return DAG.getConstant(EqVal, VT); 2461 } 2462 unsigned UOF = ISD::getUnorderedFlavor(Cond); 2463 if (UOF == 2) // FP operators that are undefined on NaNs. 2464 return DAG.getConstant(EqVal, VT); 2465 if (UOF == unsigned(ISD::isTrueWhenEqual(Cond))) 2466 return DAG.getConstant(EqVal, VT); 2467 // Otherwise, we can't fold it. However, we can simplify it to SETUO/SETO 2468 // if it is not already. 2469 ISD::CondCode NewCond = UOF == 0 ? ISD::SETO : ISD::SETUO; 2470 if (NewCond != Cond && (DCI.isBeforeLegalizeOps() || 2471 getCondCodeAction(NewCond, N0.getValueType()) == Legal)) 2472 return DAG.getSetCC(dl, VT, N0, N1, NewCond); 2473 } 2474 2475 if ((Cond == ISD::SETEQ || Cond == ISD::SETNE) && 2476 N0.getValueType().isInteger()) { 2477 if (N0.getOpcode() == ISD::ADD || N0.getOpcode() == ISD::SUB || 2478 N0.getOpcode() == ISD::XOR) { 2479 // Simplify (X+Y) == (X+Z) --> Y == Z 2480 if (N0.getOpcode() == N1.getOpcode()) { 2481 if (N0.getOperand(0) == N1.getOperand(0)) 2482 return DAG.getSetCC(dl, VT, N0.getOperand(1), N1.getOperand(1), Cond); 2483 if (N0.getOperand(1) == N1.getOperand(1)) 2484 return DAG.getSetCC(dl, VT, N0.getOperand(0), N1.getOperand(0), Cond); 2485 if (DAG.isCommutativeBinOp(N0.getOpcode())) { 2486 // If X op Y == Y op X, try other combinations. 2487 if (N0.getOperand(0) == N1.getOperand(1)) 2488 return DAG.getSetCC(dl, VT, N0.getOperand(1), N1.getOperand(0), 2489 Cond); 2490 if (N0.getOperand(1) == N1.getOperand(0)) 2491 return DAG.getSetCC(dl, VT, N0.getOperand(0), N1.getOperand(1), 2492 Cond); 2493 } 2494 } 2495 2496 // If RHS is a legal immediate value for a compare instruction, we need 2497 // to be careful about increasing register pressure needlessly. 2498 bool LegalRHSImm = false; 2499 2500 if (ConstantSDNode *RHSC = dyn_cast<ConstantSDNode>(N1)) { 2501 if (ConstantSDNode *LHSR = dyn_cast<ConstantSDNode>(N0.getOperand(1))) { 2502 // Turn (X+C1) == C2 --> X == C2-C1 2503 if (N0.getOpcode() == ISD::ADD && N0.getNode()->hasOneUse()) { 2504 return DAG.getSetCC(dl, VT, N0.getOperand(0), 2505 DAG.getConstant(RHSC->getAPIntValue()- 2506 LHSR->getAPIntValue(), 2507 N0.getValueType()), Cond); 2508 } 2509 2510 // Turn (X^C1) == C2 into X == C1^C2 iff X&~C1 = 0. 2511 if (N0.getOpcode() == ISD::XOR) 2512 // If we know that all of the inverted bits are zero, don't bother 2513 // performing the inversion. 2514 if (DAG.MaskedValueIsZero(N0.getOperand(0), ~LHSR->getAPIntValue())) 2515 return 2516 DAG.getSetCC(dl, VT, N0.getOperand(0), 2517 DAG.getConstant(LHSR->getAPIntValue() ^ 2518 RHSC->getAPIntValue(), 2519 N0.getValueType()), 2520 Cond); 2521 } 2522 2523 // Turn (C1-X) == C2 --> X == C1-C2 2524 if (ConstantSDNode *SUBC = dyn_cast<ConstantSDNode>(N0.getOperand(0))) { 2525 if (N0.getOpcode() == ISD::SUB && N0.getNode()->hasOneUse()) { 2526 return 2527 DAG.getSetCC(dl, VT, N0.getOperand(1), 2528 DAG.getConstant(SUBC->getAPIntValue() - 2529 RHSC->getAPIntValue(), 2530 N0.getValueType()), 2531 Cond); 2532 } 2533 } 2534 2535 // Could RHSC fold directly into a compare? 2536 if (RHSC->getValueType(0).getSizeInBits() <= 64) 2537 LegalRHSImm = isLegalICmpImmediate(RHSC->getSExtValue()); 2538 } 2539 2540 // Simplify (X+Z) == X --> Z == 0 2541 // Don't do this if X is an immediate that can fold into a cmp 2542 // instruction and X+Z has other uses. It could be an induction variable 2543 // chain, and the transform would increase register pressure. 2544 if (!LegalRHSImm || N0.getNode()->hasOneUse()) { 2545 if (N0.getOperand(0) == N1) 2546 return DAG.getSetCC(dl, VT, N0.getOperand(1), 2547 DAG.getConstant(0, N0.getValueType()), Cond); 2548 if (N0.getOperand(1) == N1) { 2549 if (DAG.isCommutativeBinOp(N0.getOpcode())) 2550 return DAG.getSetCC(dl, VT, N0.getOperand(0), 2551 DAG.getConstant(0, N0.getValueType()), Cond); 2552 else if (N0.getNode()->hasOneUse()) { 2553 assert(N0.getOpcode() == ISD::SUB && "Unexpected operation!"); 2554 // (Z-X) == X --> Z == X<<1 2555 SDValue SH = DAG.getNode(ISD::SHL, dl, N1.getValueType(), N1, 2556 DAG.getConstant(1, getShiftAmountTy(N1.getValueType()))); 2557 if (!DCI.isCalledByLegalizer()) 2558 DCI.AddToWorklist(SH.getNode()); 2559 return DAG.getSetCC(dl, VT, N0.getOperand(0), SH, Cond); 2560 } 2561 } 2562 } 2563 } 2564 2565 if (N1.getOpcode() == ISD::ADD || N1.getOpcode() == ISD::SUB || 2566 N1.getOpcode() == ISD::XOR) { 2567 // Simplify X == (X+Z) --> Z == 0 2568 if (N1.getOperand(0) == N0) { 2569 return DAG.getSetCC(dl, VT, N1.getOperand(1), 2570 DAG.getConstant(0, N1.getValueType()), Cond); 2571 } else if (N1.getOperand(1) == N0) { 2572 if (DAG.isCommutativeBinOp(N1.getOpcode())) { 2573 return DAG.getSetCC(dl, VT, N1.getOperand(0), 2574 DAG.getConstant(0, N1.getValueType()), Cond); 2575 } else if (N1.getNode()->hasOneUse()) { 2576 assert(N1.getOpcode() == ISD::SUB && "Unexpected operation!"); 2577 // X == (Z-X) --> X<<1 == Z 2578 SDValue SH = DAG.getNode(ISD::SHL, dl, N1.getValueType(), N0, 2579 DAG.getConstant(1, getShiftAmountTy(N0.getValueType()))); 2580 if (!DCI.isCalledByLegalizer()) 2581 DCI.AddToWorklist(SH.getNode()); 2582 return DAG.getSetCC(dl, VT, SH, N1.getOperand(0), Cond); 2583 } 2584 } 2585 } 2586 2587 // Simplify x&y == y to x&y != 0 if y has exactly one bit set. 2588 // Note that where y is variable and is known to have at most 2589 // one bit set (for example, if it is z&1) we cannot do this; 2590 // the expressions are not equivalent when y==0. 2591 if (N0.getOpcode() == ISD::AND) 2592 if (N0.getOperand(0) == N1 || N0.getOperand(1) == N1) { 2593 if (ValueHasExactlyOneBitSet(N1, DAG)) { 2594 Cond = ISD::getSetCCInverse(Cond, /*isInteger=*/true); 2595 SDValue Zero = DAG.getConstant(0, N1.getValueType()); 2596 return DAG.getSetCC(dl, VT, N0, Zero, Cond); 2597 } 2598 } 2599 if (N1.getOpcode() == ISD::AND) 2600 if (N1.getOperand(0) == N0 || N1.getOperand(1) == N0) { 2601 if (ValueHasExactlyOneBitSet(N0, DAG)) { 2602 Cond = ISD::getSetCCInverse(Cond, /*isInteger=*/true); 2603 SDValue Zero = DAG.getConstant(0, N0.getValueType()); 2604 return DAG.getSetCC(dl, VT, N1, Zero, Cond); 2605 } 2606 } 2607 } 2608 2609 // Fold away ALL boolean setcc's. 2610 SDValue Temp; 2611 if (N0.getValueType() == MVT::i1 && foldBooleans) { 2612 switch (Cond) { 2613 default: llvm_unreachable("Unknown integer setcc!"); 2614 case ISD::SETEQ: // X == Y -> ~(X^Y) 2615 Temp = DAG.getNode(ISD::XOR, dl, MVT::i1, N0, N1); 2616 N0 = DAG.getNOT(dl, Temp, MVT::i1); 2617 if (!DCI.isCalledByLegalizer()) 2618 DCI.AddToWorklist(Temp.getNode()); 2619 break; 2620 case ISD::SETNE: // X != Y --> (X^Y) 2621 N0 = DAG.getNode(ISD::XOR, dl, MVT::i1, N0, N1); 2622 break; 2623 case ISD::SETGT: // X >s Y --> X == 0 & Y == 1 --> ~X & Y 2624 case ISD::SETULT: // X <u Y --> X == 0 & Y == 1 --> ~X & Y 2625 Temp = DAG.getNOT(dl, N0, MVT::i1); 2626 N0 = DAG.getNode(ISD::AND, dl, MVT::i1, N1, Temp); 2627 if (!DCI.isCalledByLegalizer()) 2628 DCI.AddToWorklist(Temp.getNode()); 2629 break; 2630 case ISD::SETLT: // X <s Y --> X == 1 & Y == 0 --> ~Y & X 2631 case ISD::SETUGT: // X >u Y --> X == 1 & Y == 0 --> ~Y & X 2632 Temp = DAG.getNOT(dl, N1, MVT::i1); 2633 N0 = DAG.getNode(ISD::AND, dl, MVT::i1, N0, Temp); 2634 if (!DCI.isCalledByLegalizer()) 2635 DCI.AddToWorklist(Temp.getNode()); 2636 break; 2637 case ISD::SETULE: // X <=u Y --> X == 0 | Y == 1 --> ~X | Y 2638 case ISD::SETGE: // X >=s Y --> X == 0 | Y == 1 --> ~X | Y 2639 Temp = DAG.getNOT(dl, N0, MVT::i1); 2640 N0 = DAG.getNode(ISD::OR, dl, MVT::i1, N1, Temp); 2641 if (!DCI.isCalledByLegalizer()) 2642 DCI.AddToWorklist(Temp.getNode()); 2643 break; 2644 case ISD::SETUGE: // X >=u Y --> X == 1 | Y == 0 --> ~Y | X 2645 case ISD::SETLE: // X <=s Y --> X == 1 | Y == 0 --> ~Y | X 2646 Temp = DAG.getNOT(dl, N1, MVT::i1); 2647 N0 = DAG.getNode(ISD::OR, dl, MVT::i1, N0, Temp); 2648 break; 2649 } 2650 if (VT != MVT::i1) { 2651 if (!DCI.isCalledByLegalizer()) 2652 DCI.AddToWorklist(N0.getNode()); 2653 // FIXME: If running after legalize, we probably can't do this. 2654 N0 = DAG.getNode(ISD::ZERO_EXTEND, dl, VT, N0); 2655 } 2656 return N0; 2657 } 2658 2659 // Could not fold it. 2660 return SDValue(); 2661} 2662 2663/// isGAPlusOffset - Returns true (and the GlobalValue and the offset) if the 2664/// node is a GlobalAddress + offset. 2665bool TargetLowering::isGAPlusOffset(SDNode *N, const GlobalValue *&GA, 2666 int64_t &Offset) const { 2667 if (isa<GlobalAddressSDNode>(N)) { 2668 GlobalAddressSDNode *GASD = cast<GlobalAddressSDNode>(N); 2669 GA = GASD->getGlobal(); 2670 Offset += GASD->getOffset(); 2671 return true; 2672 } 2673 2674 if (N->getOpcode() == ISD::ADD) { 2675 SDValue N1 = N->getOperand(0); 2676 SDValue N2 = N->getOperand(1); 2677 if (isGAPlusOffset(N1.getNode(), GA, Offset)) { 2678 ConstantSDNode *V = dyn_cast<ConstantSDNode>(N2); 2679 if (V) { 2680 Offset += V->getSExtValue(); 2681 return true; 2682 } 2683 } else if (isGAPlusOffset(N2.getNode(), GA, Offset)) { 2684 ConstantSDNode *V = dyn_cast<ConstantSDNode>(N1); 2685 if (V) { 2686 Offset += V->getSExtValue(); 2687 return true; 2688 } 2689 } 2690 } 2691 2692 return false; 2693} 2694 2695 2696SDValue TargetLowering:: 2697PerformDAGCombine(SDNode *N, DAGCombinerInfo &DCI) const { 2698 // Default implementation: no optimization. 2699 return SDValue(); 2700} 2701 2702//===----------------------------------------------------------------------===// 2703// Inline Assembler Implementation Methods 2704//===----------------------------------------------------------------------===// 2705 2706 2707TargetLowering::ConstraintType 2708TargetLowering::getConstraintType(const std::string &Constraint) const { 2709 if (Constraint.size() == 1) { 2710 switch (Constraint[0]) { 2711 default: break; 2712 case 'r': return C_RegisterClass; 2713 case 'm': // memory 2714 case 'o': // offsetable 2715 case 'V': // not offsetable 2716 return C_Memory; 2717 case 'i': // Simple Integer or Relocatable Constant 2718 case 'n': // Simple Integer 2719 case 'E': // Floating Point Constant 2720 case 'F': // Floating Point Constant 2721 case 's': // Relocatable Constant 2722 case 'p': // Address. 2723 case 'X': // Allow ANY value. 2724 case 'I': // Target registers. 2725 case 'J': 2726 case 'K': 2727 case 'L': 2728 case 'M': 2729 case 'N': 2730 case 'O': 2731 case 'P': 2732 case '<': 2733 case '>': 2734 return C_Other; 2735 } 2736 } 2737 2738 if (Constraint.size() > 1 && Constraint[0] == '{' && 2739 Constraint[Constraint.size()-1] == '}') 2740 return C_Register; 2741 return C_Unknown; 2742} 2743 2744/// LowerXConstraint - try to replace an X constraint, which matches anything, 2745/// with another that has more specific requirements based on the type of the 2746/// corresponding operand. 2747const char *TargetLowering::LowerXConstraint(EVT ConstraintVT) const{ 2748 if (ConstraintVT.isInteger()) 2749 return "r"; 2750 if (ConstraintVT.isFloatingPoint()) 2751 return "f"; // works for many targets 2752 return 0; 2753} 2754 2755/// LowerAsmOperandForConstraint - Lower the specified operand into the Ops 2756/// vector. If it is invalid, don't add anything to Ops. 2757void TargetLowering::LowerAsmOperandForConstraint(SDValue Op, 2758 std::string &Constraint, 2759 std::vector<SDValue> &Ops, 2760 SelectionDAG &DAG) const { 2761 2762 if (Constraint.length() > 1) return; 2763 2764 char ConstraintLetter = Constraint[0]; 2765 switch (ConstraintLetter) { 2766 default: break; 2767 case 'X': // Allows any operand; labels (basic block) use this. 2768 if (Op.getOpcode() == ISD::BasicBlock) { 2769 Ops.push_back(Op); 2770 return; 2771 } 2772 // fall through 2773 case 'i': // Simple Integer or Relocatable Constant 2774 case 'n': // Simple Integer 2775 case 's': { // Relocatable Constant 2776 // These operands are interested in values of the form (GV+C), where C may 2777 // be folded in as an offset of GV, or it may be explicitly added. Also, it 2778 // is possible and fine if either GV or C are missing. 2779 ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op); 2780 GlobalAddressSDNode *GA = dyn_cast<GlobalAddressSDNode>(Op); 2781 2782 // If we have "(add GV, C)", pull out GV/C 2783 if (Op.getOpcode() == ISD::ADD) { 2784 C = dyn_cast<ConstantSDNode>(Op.getOperand(1)); 2785 GA = dyn_cast<GlobalAddressSDNode>(Op.getOperand(0)); 2786 if (C == 0 || GA == 0) { 2787 C = dyn_cast<ConstantSDNode>(Op.getOperand(0)); 2788 GA = dyn_cast<GlobalAddressSDNode>(Op.getOperand(1)); 2789 } 2790 if (C == 0 || GA == 0) 2791 C = 0, GA = 0; 2792 } 2793 2794 // If we find a valid operand, map to the TargetXXX version so that the 2795 // value itself doesn't get selected. 2796 if (GA) { // Either &GV or &GV+C 2797 if (ConstraintLetter != 'n') { 2798 int64_t Offs = GA->getOffset(); 2799 if (C) Offs += C->getZExtValue(); 2800 Ops.push_back(DAG.getTargetGlobalAddress(GA->getGlobal(), 2801 C ? C->getDebugLoc() : DebugLoc(), 2802 Op.getValueType(), Offs)); 2803 return; 2804 } 2805 } 2806 if (C) { // just C, no GV. 2807 // Simple constants are not allowed for 's'. 2808 if (ConstraintLetter != 's') { 2809 // gcc prints these as sign extended. Sign extend value to 64 bits 2810 // now; without this it would get ZExt'd later in 2811 // ScheduleDAGSDNodes::EmitNode, which is very generic. 2812 Ops.push_back(DAG.getTargetConstant(C->getAPIntValue().getSExtValue(), 2813 MVT::i64)); 2814 return; 2815 } 2816 } 2817 break; 2818 } 2819 } 2820} 2821 2822std::pair<unsigned, const TargetRegisterClass*> TargetLowering:: 2823getRegForInlineAsmConstraint(const std::string &Constraint, 2824 EVT VT) const { 2825 if (Constraint[0] != '{') 2826 return std::make_pair(0u, static_cast<TargetRegisterClass*>(0)); 2827 assert(*(Constraint.end()-1) == '}' && "Not a brace enclosed constraint?"); 2828 2829 // Remove the braces from around the name. 2830 StringRef RegName(Constraint.data()+1, Constraint.size()-2); 2831 2832 // Figure out which register class contains this reg. 2833 const TargetRegisterInfo *RI = TM.getRegisterInfo(); 2834 for (TargetRegisterInfo::regclass_iterator RCI = RI->regclass_begin(), 2835 E = RI->regclass_end(); RCI != E; ++RCI) { 2836 const TargetRegisterClass *RC = *RCI; 2837 2838 // If none of the value types for this register class are valid, we 2839 // can't use it. For example, 64-bit reg classes on 32-bit targets. 2840 if (!isLegalRC(RC)) 2841 continue; 2842 2843 for (TargetRegisterClass::iterator I = RC->begin(), E = RC->end(); 2844 I != E; ++I) { 2845 if (RegName.equals_lower(RI->getName(*I))) 2846 return std::make_pair(*I, RC); 2847 } 2848 } 2849 2850 return std::make_pair(0u, static_cast<const TargetRegisterClass*>(0)); 2851} 2852 2853//===----------------------------------------------------------------------===// 2854// Constraint Selection. 2855 2856/// isMatchingInputConstraint - Return true of this is an input operand that is 2857/// a matching constraint like "4". 2858bool TargetLowering::AsmOperandInfo::isMatchingInputConstraint() const { 2859 assert(!ConstraintCode.empty() && "No known constraint!"); 2860 return isdigit(ConstraintCode[0]); 2861} 2862 2863/// getMatchedOperand - If this is an input matching constraint, this method 2864/// returns the output operand it matches. 2865unsigned TargetLowering::AsmOperandInfo::getMatchedOperand() const { 2866 assert(!ConstraintCode.empty() && "No known constraint!"); 2867 return atoi(ConstraintCode.c_str()); 2868} 2869 2870 2871/// ParseConstraints - Split up the constraint string from the inline 2872/// assembly value into the specific constraints and their prefixes, 2873/// and also tie in the associated operand values. 2874/// If this returns an empty vector, and if the constraint string itself 2875/// isn't empty, there was an error parsing. 2876TargetLowering::AsmOperandInfoVector TargetLowering::ParseConstraints( 2877 ImmutableCallSite CS) const { 2878 /// ConstraintOperands - Information about all of the constraints. 2879 AsmOperandInfoVector ConstraintOperands; 2880 const InlineAsm *IA = cast<InlineAsm>(CS.getCalledValue()); 2881 unsigned maCount = 0; // Largest number of multiple alternative constraints. 2882 2883 // Do a prepass over the constraints, canonicalizing them, and building up the 2884 // ConstraintOperands list. 2885 InlineAsm::ConstraintInfoVector 2886 ConstraintInfos = IA->ParseConstraints(); 2887 2888 unsigned ArgNo = 0; // ArgNo - The argument of the CallInst. 2889 unsigned ResNo = 0; // ResNo - The result number of the next output. 2890 2891 for (unsigned i = 0, e = ConstraintInfos.size(); i != e; ++i) { 2892 ConstraintOperands.push_back(AsmOperandInfo(ConstraintInfos[i])); 2893 AsmOperandInfo &OpInfo = ConstraintOperands.back(); 2894 2895 // Update multiple alternative constraint count. 2896 if (OpInfo.multipleAlternatives.size() > maCount) 2897 maCount = OpInfo.multipleAlternatives.size(); 2898 2899 OpInfo.ConstraintVT = MVT::Other; 2900 2901 // Compute the value type for each operand. 2902 switch (OpInfo.Type) { 2903 case InlineAsm::isOutput: 2904 // Indirect outputs just consume an argument. 2905 if (OpInfo.isIndirect) { 2906 OpInfo.CallOperandVal = const_cast<Value *>(CS.getArgument(ArgNo++)); 2907 break; 2908 } 2909 2910 // The return value of the call is this value. As such, there is no 2911 // corresponding argument. 2912 assert(!CS.getType()->isVoidTy() && 2913 "Bad inline asm!"); 2914 if (StructType *STy = dyn_cast<StructType>(CS.getType())) { 2915 OpInfo.ConstraintVT = getValueType(STy->getElementType(ResNo)); 2916 } else { 2917 assert(ResNo == 0 && "Asm only has one result!"); 2918 OpInfo.ConstraintVT = getValueType(CS.getType()); 2919 } 2920 ++ResNo; 2921 break; 2922 case InlineAsm::isInput: 2923 OpInfo.CallOperandVal = const_cast<Value *>(CS.getArgument(ArgNo++)); 2924 break; 2925 case InlineAsm::isClobber: 2926 // Nothing to do. 2927 break; 2928 } 2929 2930 if (OpInfo.CallOperandVal) { 2931 llvm::Type *OpTy = OpInfo.CallOperandVal->getType(); 2932 if (OpInfo.isIndirect) { 2933 llvm::PointerType *PtrTy = dyn_cast<PointerType>(OpTy); 2934 if (!PtrTy) 2935 report_fatal_error("Indirect operand for inline asm not a pointer!"); 2936 OpTy = PtrTy->getElementType(); 2937 } 2938 2939 // Look for vector wrapped in a struct. e.g. { <16 x i8> }. 2940 if (StructType *STy = dyn_cast<StructType>(OpTy)) 2941 if (STy->getNumElements() == 1) 2942 OpTy = STy->getElementType(0); 2943 2944 // If OpTy is not a single value, it may be a struct/union that we 2945 // can tile with integers. 2946 if (!OpTy->isSingleValueType() && OpTy->isSized()) { 2947 unsigned BitSize = TD->getTypeSizeInBits(OpTy); 2948 switch (BitSize) { 2949 default: break; 2950 case 1: 2951 case 8: 2952 case 16: 2953 case 32: 2954 case 64: 2955 case 128: 2956 OpInfo.ConstraintVT = 2957 EVT::getEVT(IntegerType::get(OpTy->getContext(), BitSize), true); 2958 break; 2959 } 2960 } else if (PointerType *PT = dyn_cast<PointerType>(OpTy)) { 2961 OpInfo.ConstraintVT = MVT::getIntegerVT( 2962 8*TD->getPointerSize(PT->getAddressSpace())); 2963 } else { 2964 OpInfo.ConstraintVT = EVT::getEVT(OpTy, true); 2965 } 2966 } 2967 } 2968 2969 // If we have multiple alternative constraints, select the best alternative. 2970 if (ConstraintInfos.size()) { 2971 if (maCount) { 2972 unsigned bestMAIndex = 0; 2973 int bestWeight = -1; 2974 // weight: -1 = invalid match, and 0 = so-so match to 5 = good match. 2975 int weight = -1; 2976 unsigned maIndex; 2977 // Compute the sums of the weights for each alternative, keeping track 2978 // of the best (highest weight) one so far. 2979 for (maIndex = 0; maIndex < maCount; ++maIndex) { 2980 int weightSum = 0; 2981 for (unsigned cIndex = 0, eIndex = ConstraintOperands.size(); 2982 cIndex != eIndex; ++cIndex) { 2983 AsmOperandInfo& OpInfo = ConstraintOperands[cIndex]; 2984 if (OpInfo.Type == InlineAsm::isClobber) 2985 continue; 2986 2987 // If this is an output operand with a matching input operand, 2988 // look up the matching input. If their types mismatch, e.g. one 2989 // is an integer, the other is floating point, or their sizes are 2990 // different, flag it as an maCantMatch. 2991 if (OpInfo.hasMatchingInput()) { 2992 AsmOperandInfo &Input = ConstraintOperands[OpInfo.MatchingInput]; 2993 if (OpInfo.ConstraintVT != Input.ConstraintVT) { 2994 if ((OpInfo.ConstraintVT.isInteger() != 2995 Input.ConstraintVT.isInteger()) || 2996 (OpInfo.ConstraintVT.getSizeInBits() != 2997 Input.ConstraintVT.getSizeInBits())) { 2998 weightSum = -1; // Can't match. 2999 break; 3000 } 3001 } 3002 } 3003 weight = getMultipleConstraintMatchWeight(OpInfo, maIndex); 3004 if (weight == -1) { 3005 weightSum = -1; 3006 break; 3007 } 3008 weightSum += weight; 3009 } 3010 // Update best. 3011 if (weightSum > bestWeight) { 3012 bestWeight = weightSum; 3013 bestMAIndex = maIndex; 3014 } 3015 } 3016 3017 // Now select chosen alternative in each constraint. 3018 for (unsigned cIndex = 0, eIndex = ConstraintOperands.size(); 3019 cIndex != eIndex; ++cIndex) { 3020 AsmOperandInfo& cInfo = ConstraintOperands[cIndex]; 3021 if (cInfo.Type == InlineAsm::isClobber) 3022 continue; 3023 cInfo.selectAlternative(bestMAIndex); 3024 } 3025 } 3026 } 3027 3028 // Check and hook up tied operands, choose constraint code to use. 3029 for (unsigned cIndex = 0, eIndex = ConstraintOperands.size(); 3030 cIndex != eIndex; ++cIndex) { 3031 AsmOperandInfo& OpInfo = ConstraintOperands[cIndex]; 3032 3033 // If this is an output operand with a matching input operand, look up the 3034 // matching input. If their types mismatch, e.g. one is an integer, the 3035 // other is floating point, or their sizes are different, flag it as an 3036 // error. 3037 if (OpInfo.hasMatchingInput()) { 3038 AsmOperandInfo &Input = ConstraintOperands[OpInfo.MatchingInput]; 3039 3040 if (OpInfo.ConstraintVT != Input.ConstraintVT) { 3041 std::pair<unsigned, const TargetRegisterClass*> MatchRC = 3042 getRegForInlineAsmConstraint(OpInfo.ConstraintCode, 3043 OpInfo.ConstraintVT); 3044 std::pair<unsigned, const TargetRegisterClass*> InputRC = 3045 getRegForInlineAsmConstraint(Input.ConstraintCode, 3046 Input.ConstraintVT); 3047 if ((OpInfo.ConstraintVT.isInteger() != 3048 Input.ConstraintVT.isInteger()) || 3049 (MatchRC.second != InputRC.second)) { 3050 report_fatal_error("Unsupported asm: input constraint" 3051 " with a matching output constraint of" 3052 " incompatible type!"); 3053 } 3054 } 3055 3056 } 3057 } 3058 3059 return ConstraintOperands; 3060} 3061 3062 3063/// getConstraintGenerality - Return an integer indicating how general CT 3064/// is. 3065static unsigned getConstraintGenerality(TargetLowering::ConstraintType CT) { 3066 switch (CT) { 3067 case TargetLowering::C_Other: 3068 case TargetLowering::C_Unknown: 3069 return 0; 3070 case TargetLowering::C_Register: 3071 return 1; 3072 case TargetLowering::C_RegisterClass: 3073 return 2; 3074 case TargetLowering::C_Memory: 3075 return 3; 3076 } 3077 llvm_unreachable("Invalid constraint type"); 3078} 3079 3080/// Examine constraint type and operand type and determine a weight value. 3081/// This object must already have been set up with the operand type 3082/// and the current alternative constraint selected. 3083TargetLowering::ConstraintWeight 3084 TargetLowering::getMultipleConstraintMatchWeight( 3085 AsmOperandInfo &info, int maIndex) const { 3086 InlineAsm::ConstraintCodeVector *rCodes; 3087 if (maIndex >= (int)info.multipleAlternatives.size()) 3088 rCodes = &info.Codes; 3089 else 3090 rCodes = &info.multipleAlternatives[maIndex].Codes; 3091 ConstraintWeight BestWeight = CW_Invalid; 3092 3093 // Loop over the options, keeping track of the most general one. 3094 for (unsigned i = 0, e = rCodes->size(); i != e; ++i) { 3095 ConstraintWeight weight = 3096 getSingleConstraintMatchWeight(info, (*rCodes)[i].c_str()); 3097 if (weight > BestWeight) 3098 BestWeight = weight; 3099 } 3100 3101 return BestWeight; 3102} 3103 3104/// Examine constraint type and operand type and determine a weight value. 3105/// This object must already have been set up with the operand type 3106/// and the current alternative constraint selected. 3107TargetLowering::ConstraintWeight 3108 TargetLowering::getSingleConstraintMatchWeight( 3109 AsmOperandInfo &info, const char *constraint) const { 3110 ConstraintWeight weight = CW_Invalid; 3111 Value *CallOperandVal = info.CallOperandVal; 3112 // If we don't have a value, we can't do a match, 3113 // but allow it at the lowest weight. 3114 if (CallOperandVal == NULL) 3115 return CW_Default; 3116 // Look at the constraint type. 3117 switch (*constraint) { 3118 case 'i': // immediate integer. 3119 case 'n': // immediate integer with a known value. 3120 if (isa<ConstantInt>(CallOperandVal)) 3121 weight = CW_Constant; 3122 break; 3123 case 's': // non-explicit intregal immediate. 3124 if (isa<GlobalValue>(CallOperandVal)) 3125 weight = CW_Constant; 3126 break; 3127 case 'E': // immediate float if host format. 3128 case 'F': // immediate float. 3129 if (isa<ConstantFP>(CallOperandVal)) 3130 weight = CW_Constant; 3131 break; 3132 case '<': // memory operand with autodecrement. 3133 case '>': // memory operand with autoincrement. 3134 case 'm': // memory operand. 3135 case 'o': // offsettable memory operand 3136 case 'V': // non-offsettable memory operand 3137 weight = CW_Memory; 3138 break; 3139 case 'r': // general register. 3140 case 'g': // general register, memory operand or immediate integer. 3141 // note: Clang converts "g" to "imr". 3142 if (CallOperandVal->getType()->isIntegerTy()) 3143 weight = CW_Register; 3144 break; 3145 case 'X': // any operand. 3146 default: 3147 weight = CW_Default; 3148 break; 3149 } 3150 return weight; 3151} 3152 3153/// ChooseConstraint - If there are multiple different constraints that we 3154/// could pick for this operand (e.g. "imr") try to pick the 'best' one. 3155/// This is somewhat tricky: constraints fall into four classes: 3156/// Other -> immediates and magic values 3157/// Register -> one specific register 3158/// RegisterClass -> a group of regs 3159/// Memory -> memory 3160/// Ideally, we would pick the most specific constraint possible: if we have 3161/// something that fits into a register, we would pick it. The problem here 3162/// is that if we have something that could either be in a register or in 3163/// memory that use of the register could cause selection of *other* 3164/// operands to fail: they might only succeed if we pick memory. Because of 3165/// this the heuristic we use is: 3166/// 3167/// 1) If there is an 'other' constraint, and if the operand is valid for 3168/// that constraint, use it. This makes us take advantage of 'i' 3169/// constraints when available. 3170/// 2) Otherwise, pick the most general constraint present. This prefers 3171/// 'm' over 'r', for example. 3172/// 3173static void ChooseConstraint(TargetLowering::AsmOperandInfo &OpInfo, 3174 const TargetLowering &TLI, 3175 SDValue Op, SelectionDAG *DAG) { 3176 assert(OpInfo.Codes.size() > 1 && "Doesn't have multiple constraint options"); 3177 unsigned BestIdx = 0; 3178 TargetLowering::ConstraintType BestType = TargetLowering::C_Unknown; 3179 int BestGenerality = -1; 3180 3181 // Loop over the options, keeping track of the most general one. 3182 for (unsigned i = 0, e = OpInfo.Codes.size(); i != e; ++i) { 3183 TargetLowering::ConstraintType CType = 3184 TLI.getConstraintType(OpInfo.Codes[i]); 3185 3186 // If this is an 'other' constraint, see if the operand is valid for it. 3187 // For example, on X86 we might have an 'rI' constraint. If the operand 3188 // is an integer in the range [0..31] we want to use I (saving a load 3189 // of a register), otherwise we must use 'r'. 3190 if (CType == TargetLowering::C_Other && Op.getNode()) { 3191 assert(OpInfo.Codes[i].size() == 1 && 3192 "Unhandled multi-letter 'other' constraint"); 3193 std::vector<SDValue> ResultOps; 3194 TLI.LowerAsmOperandForConstraint(Op, OpInfo.Codes[i], 3195 ResultOps, *DAG); 3196 if (!ResultOps.empty()) { 3197 BestType = CType; 3198 BestIdx = i; 3199 break; 3200 } 3201 } 3202 3203 // Things with matching constraints can only be registers, per gcc 3204 // documentation. This mainly affects "g" constraints. 3205 if (CType == TargetLowering::C_Memory && OpInfo.hasMatchingInput()) 3206 continue; 3207 3208 // This constraint letter is more general than the previous one, use it. 3209 int Generality = getConstraintGenerality(CType); 3210 if (Generality > BestGenerality) { 3211 BestType = CType; 3212 BestIdx = i; 3213 BestGenerality = Generality; 3214 } 3215 } 3216 3217 OpInfo.ConstraintCode = OpInfo.Codes[BestIdx]; 3218 OpInfo.ConstraintType = BestType; 3219} 3220 3221/// ComputeConstraintToUse - Determines the constraint code and constraint 3222/// type to use for the specific AsmOperandInfo, setting 3223/// OpInfo.ConstraintCode and OpInfo.ConstraintType. 3224void TargetLowering::ComputeConstraintToUse(AsmOperandInfo &OpInfo, 3225 SDValue Op, 3226 SelectionDAG *DAG) const { 3227 assert(!OpInfo.Codes.empty() && "Must have at least one constraint"); 3228 3229 // Single-letter constraints ('r') are very common. 3230 if (OpInfo.Codes.size() == 1) { 3231 OpInfo.ConstraintCode = OpInfo.Codes[0]; 3232 OpInfo.ConstraintType = getConstraintType(OpInfo.ConstraintCode); 3233 } else { 3234 ChooseConstraint(OpInfo, *this, Op, DAG); 3235 } 3236 3237 // 'X' matches anything. 3238 if (OpInfo.ConstraintCode == "X" && OpInfo.CallOperandVal) { 3239 // Labels and constants are handled elsewhere ('X' is the only thing 3240 // that matches labels). For Functions, the type here is the type of 3241 // the result, which is not what we want to look at; leave them alone. 3242 Value *v = OpInfo.CallOperandVal; 3243 if (isa<BasicBlock>(v) || isa<ConstantInt>(v) || isa<Function>(v)) { 3244 OpInfo.CallOperandVal = v; 3245 return; 3246 } 3247 3248 // Otherwise, try to resolve it to something we know about by looking at 3249 // the actual operand type. 3250 if (const char *Repl = LowerXConstraint(OpInfo.ConstraintVT)) { 3251 OpInfo.ConstraintCode = Repl; 3252 OpInfo.ConstraintType = getConstraintType(OpInfo.ConstraintCode); 3253 } 3254 } 3255} 3256 3257//===----------------------------------------------------------------------===// 3258// Loop Strength Reduction hooks 3259//===----------------------------------------------------------------------===// 3260 3261/// isLegalAddressingMode - Return true if the addressing mode represented 3262/// by AM is legal for this target, for a load/store of the specified type. 3263bool TargetLowering::isLegalAddressingMode(const AddrMode &AM, 3264 Type *Ty) const { 3265 // The default implementation of this implements a conservative RISCy, r+r and 3266 // r+i addr mode. 3267 3268 // Allows a sign-extended 16-bit immediate field. 3269 if (AM.BaseOffs <= -(1LL << 16) || AM.BaseOffs >= (1LL << 16)-1) 3270 return false; 3271 3272 // No global is ever allowed as a base. 3273 if (AM.BaseGV) 3274 return false; 3275 3276 // Only support r+r, 3277 switch (AM.Scale) { 3278 case 0: // "r+i" or just "i", depending on HasBaseReg. 3279 break; 3280 case 1: 3281 if (AM.HasBaseReg && AM.BaseOffs) // "r+r+i" is not allowed. 3282 return false; 3283 // Otherwise we have r+r or r+i. 3284 break; 3285 case 2: 3286 if (AM.HasBaseReg || AM.BaseOffs) // 2*r+r or 2*r+i is not allowed. 3287 return false; 3288 // Allow 2*r as r+r. 3289 break; 3290 } 3291 3292 return true; 3293} 3294 3295/// BuildExactDiv - Given an exact SDIV by a constant, create a multiplication 3296/// with the multiplicative inverse of the constant. 3297SDValue TargetLowering::BuildExactSDIV(SDValue Op1, SDValue Op2, DebugLoc dl, 3298 SelectionDAG &DAG) const { 3299 ConstantSDNode *C = cast<ConstantSDNode>(Op2); 3300 APInt d = C->getAPIntValue(); 3301 assert(d != 0 && "Division by zero!"); 3302 3303 // Shift the value upfront if it is even, so the LSB is one. 3304 unsigned ShAmt = d.countTrailingZeros(); 3305 if (ShAmt) { 3306 // TODO: For UDIV use SRL instead of SRA. 3307 SDValue Amt = DAG.getConstant(ShAmt, getShiftAmountTy(Op1.getValueType())); 3308 Op1 = DAG.getNode(ISD::SRA, dl, Op1.getValueType(), Op1, Amt); 3309 d = d.ashr(ShAmt); 3310 } 3311 3312 // Calculate the multiplicative inverse, using Newton's method. 3313 APInt t, xn = d; 3314 while ((t = d*xn) != 1) 3315 xn *= APInt(d.getBitWidth(), 2) - t; 3316 3317 Op2 = DAG.getConstant(xn, Op1.getValueType()); 3318 return DAG.getNode(ISD::MUL, dl, Op1.getValueType(), Op1, Op2); 3319} 3320 3321/// BuildSDIVSequence - Given an ISD::SDIV node expressing a divide by constant, 3322/// return a DAG expression to select that will generate the same value by 3323/// multiplying by a magic number. See: 3324/// <http://the.wall.riscom.net/books/proc/ppc/cwg/code2.html> 3325SDValue TargetLowering:: 3326BuildSDIV(SDNode *N, SelectionDAG &DAG, bool IsAfterLegalization, 3327 std::vector<SDNode*> *Created) const { 3328 EVT VT = N->getValueType(0); 3329 DebugLoc dl= N->getDebugLoc(); 3330 3331 // Check to see if we can do this. 3332 // FIXME: We should be more aggressive here. 3333 if (!isTypeLegal(VT)) 3334 return SDValue(); 3335 3336 APInt d = cast<ConstantSDNode>(N->getOperand(1))->getAPIntValue(); 3337 APInt::ms magics = d.magic(); 3338 3339 // Multiply the numerator (operand 0) by the magic value 3340 // FIXME: We should support doing a MUL in a wider type 3341 SDValue Q; 3342 if (IsAfterLegalization ? isOperationLegal(ISD::MULHS, VT) : 3343 isOperationLegalOrCustom(ISD::MULHS, VT)) 3344 Q = DAG.getNode(ISD::MULHS, dl, VT, N->getOperand(0), 3345 DAG.getConstant(magics.m, VT)); 3346 else if (IsAfterLegalization ? isOperationLegal(ISD::SMUL_LOHI, VT) : 3347 isOperationLegalOrCustom(ISD::SMUL_LOHI, VT)) 3348 Q = SDValue(DAG.getNode(ISD::SMUL_LOHI, dl, DAG.getVTList(VT, VT), 3349 N->getOperand(0), 3350 DAG.getConstant(magics.m, VT)).getNode(), 1); 3351 else 3352 return SDValue(); // No mulhs or equvialent 3353 // If d > 0 and m < 0, add the numerator 3354 if (d.isStrictlyPositive() && magics.m.isNegative()) { 3355 Q = DAG.getNode(ISD::ADD, dl, VT, Q, N->getOperand(0)); 3356 if (Created) 3357 Created->push_back(Q.getNode()); 3358 } 3359 // If d < 0 and m > 0, subtract the numerator. 3360 if (d.isNegative() && magics.m.isStrictlyPositive()) { 3361 Q = DAG.getNode(ISD::SUB, dl, VT, Q, N->getOperand(0)); 3362 if (Created) 3363 Created->push_back(Q.getNode()); 3364 } 3365 // Shift right algebraic if shift value is nonzero 3366 if (magics.s > 0) { 3367 Q = DAG.getNode(ISD::SRA, dl, VT, Q, 3368 DAG.getConstant(magics.s, getShiftAmountTy(Q.getValueType()))); 3369 if (Created) 3370 Created->push_back(Q.getNode()); 3371 } 3372 // Extract the sign bit and add it to the quotient 3373 SDValue T = 3374 DAG.getNode(ISD::SRL, dl, VT, Q, DAG.getConstant(VT.getSizeInBits()-1, 3375 getShiftAmountTy(Q.getValueType()))); 3376 if (Created) 3377 Created->push_back(T.getNode()); 3378 return DAG.getNode(ISD::ADD, dl, VT, Q, T); 3379} 3380 3381/// BuildUDIVSequence - Given an ISD::UDIV node expressing a divide by constant, 3382/// return a DAG expression to select that will generate the same value by 3383/// multiplying by a magic number. See: 3384/// <http://the.wall.riscom.net/books/proc/ppc/cwg/code2.html> 3385SDValue TargetLowering:: 3386BuildUDIV(SDNode *N, SelectionDAG &DAG, bool IsAfterLegalization, 3387 std::vector<SDNode*> *Created) const { 3388 EVT VT = N->getValueType(0); 3389 DebugLoc dl = N->getDebugLoc(); 3390 3391 // Check to see if we can do this. 3392 // FIXME: We should be more aggressive here. 3393 if (!isTypeLegal(VT)) 3394 return SDValue(); 3395 3396 // FIXME: We should use a narrower constant when the upper 3397 // bits are known to be zero. 3398 const APInt &N1C = cast<ConstantSDNode>(N->getOperand(1))->getAPIntValue(); 3399 APInt::mu magics = N1C.magicu(); 3400 3401 SDValue Q = N->getOperand(0); 3402 3403 // If the divisor is even, we can avoid using the expensive fixup by shifting 3404 // the divided value upfront. 3405 if (magics.a != 0 && !N1C[0]) { 3406 unsigned Shift = N1C.countTrailingZeros(); 3407 Q = DAG.getNode(ISD::SRL, dl, VT, Q, 3408 DAG.getConstant(Shift, getShiftAmountTy(Q.getValueType()))); 3409 if (Created) 3410 Created->push_back(Q.getNode()); 3411 3412 // Get magic number for the shifted divisor. 3413 magics = N1C.lshr(Shift).magicu(Shift); 3414 assert(magics.a == 0 && "Should use cheap fixup now"); 3415 } 3416 3417 // Multiply the numerator (operand 0) by the magic value 3418 // FIXME: We should support doing a MUL in a wider type 3419 if (IsAfterLegalization ? isOperationLegal(ISD::MULHU, VT) : 3420 isOperationLegalOrCustom(ISD::MULHU, VT)) 3421 Q = DAG.getNode(ISD::MULHU, dl, VT, Q, DAG.getConstant(magics.m, VT)); 3422 else if (IsAfterLegalization ? isOperationLegal(ISD::UMUL_LOHI, VT) : 3423 isOperationLegalOrCustom(ISD::UMUL_LOHI, VT)) 3424 Q = SDValue(DAG.getNode(ISD::UMUL_LOHI, dl, DAG.getVTList(VT, VT), Q, 3425 DAG.getConstant(magics.m, VT)).getNode(), 1); 3426 else 3427 return SDValue(); // No mulhu or equvialent 3428 if (Created) 3429 Created->push_back(Q.getNode()); 3430 3431 if (magics.a == 0) { 3432 assert(magics.s < N1C.getBitWidth() && 3433 "We shouldn't generate an undefined shift!"); 3434 return DAG.getNode(ISD::SRL, dl, VT, Q, 3435 DAG.getConstant(magics.s, getShiftAmountTy(Q.getValueType()))); 3436 } else { 3437 SDValue NPQ = DAG.getNode(ISD::SUB, dl, VT, N->getOperand(0), Q); 3438 if (Created) 3439 Created->push_back(NPQ.getNode()); 3440 NPQ = DAG.getNode(ISD::SRL, dl, VT, NPQ, 3441 DAG.getConstant(1, getShiftAmountTy(NPQ.getValueType()))); 3442 if (Created) 3443 Created->push_back(NPQ.getNode()); 3444 NPQ = DAG.getNode(ISD::ADD, dl, VT, NPQ, Q); 3445 if (Created) 3446 Created->push_back(NPQ.getNode()); 3447 return DAG.getNode(ISD::SRL, dl, VT, NPQ, 3448 DAG.getConstant(magics.s-1, getShiftAmountTy(NPQ.getValueType()))); 3449 } 3450} 3451