gcc/config/i386/x86-tune.def

   1 /* Definitions of x86 tunable features.
   2    Copyright (C) 2013-2023 Free Software Foundation, Inc.
   3
   4 This file is part of GCC.
   5
   6 GCC is free software; you can redistribute it and/or modify
   7 it under the terms of the GNU General Public License as published by
   8 the Free Software Foundation; either version 3, or (at your option)
   9 any later version.
  10
  11 GCC is distributed in the hope that it will be useful,
  12 but WITHOUT ANY WARRANTY; without even the implied warranty of
  13 MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
  14 GNU General Public License for more details.
  15
  16 You should have received a copy of the GNU General Public License and
  17 a copy of the GCC Runtime Library Exception along with this program;
  18 see the files COPYING3 and COPYING.RUNTIME respectively.  If not, see
  19 <http://www.gnu.org/licenses/>.  */
  20
  21 /* Tuning for a given CPU XXXX consists of:
  22     - adding new CPU into:
  23         - adding PROCESSOR_XXX to processor_type (in i386.h)
  24         - possibly adding XXX into CPU attribute in i386.md
  25         - adding XXX to processor_alias_table (in i386.cc)
  26     - introducing ix86_XXX_cost in i386.cc
  27         - Stringop generation table can be build based on test_stringop
  28         - script (once rest of tuning is complete)
  29     - designing a scheduler model in
  30         - XXXX.md file
  31         - Updating ix86_issue_rate and ix86_adjust_cost in i386.md
  32         - possibly updating ia32_multipass_dfa_lookahead, ix86_sched_reorder
  33           and ix86_sched_init_global if those tricks are needed.
  34     - Tunning the flags bellow. Those are split into sections and each
  35       section is very roughly ordered by importance.  */
  36
  37 /*****************************************************************************/
  38 /* Scheduling flags.                                                         */
  39 /*****************************************************************************/
  40
  41 /* X86_TUNE_SCHEDULE: Enable scheduling.  */
  42 DEF_TUNE (X86_TUNE_SCHEDULE, "schedule",
  43           m_PENT | m_LAKEMONT | m_PPRO | m_CORE_ALL | m_BONNELL | m_SILVERMONT
  44           | m_INTEL | m_KNL | m_KNM | m_K6_GEODE | m_AMD_MULTIPLE | m_LUJIAZUI
  45           | m_GOLDMONT | m_GOLDMONT_PLUS | m_TREMONT | m_ALDERLAKE | m_CORE_ATOM
  46           | m_GENERIC)
  47
  48 /* X86_TUNE_PARTIAL_REG_DEPENDENCY: Enable more register renaming
  49    on modern chips.  Prefer stores affecting whole integer register
  50    over partial stores.  For example prefer MOVZBL or MOVQ to load 8bit
  51    value over movb.  */
  52 DEF_TUNE (X86_TUNE_PARTIAL_REG_DEPENDENCY, "partial_reg_dependency",
  53           m_P4_NOCONA | m_CORE2 | m_NEHALEM  | m_SANDYBRIDGE | m_CORE_AVX2
  54           | m_BONNELL | m_SILVERMONT | m_GOLDMONT | m_GOLDMONT_PLUS | m_INTEL
  55           | m_KNL | m_KNM | m_AMD_MULTIPLE | m_LUJIAZUI | m_TREMONT
  56           | m_ALDERLAKE | m_CORE_ATOM | m_GENERIC)
  57
  58 /* X86_TUNE_SSE_PARTIAL_REG_DEPENDENCY: This knob promotes all store
  59    destinations to be 128bit to allow register renaming on 128bit SSE units,
  60    but usually results in one extra microop on 64bit SSE units.
  61    Experimental results shows that disabling this option on P4 brings over 20%
  62    SPECfp regression, while enabling it on K8 brings roughly 2.4% regression
  63    that can be partly masked by careful scheduling of moves.  */
  64 DEF_TUNE (X86_TUNE_SSE_PARTIAL_REG_DEPENDENCY, "sse_partial_reg_dependency",
  65           m_PPRO | m_P4_NOCONA | m_CORE_ALL | m_BONNELL | m_AMDFAM10
  66           | m_BDVER | m_ZNVER | m_LUJIAZUI | m_TREMONT | m_ALDERLAKE
  67           | m_CORE_ATOM | m_GENERIC)
  68
  69 /* X86_TUNE_SSE_PARTIAL_REG_FP_CONVERTS_DEPENDENCY: This knob avoids
  70    partial write to the destination in scalar SSE conversion from FP
  71    to FP.  */
  72 DEF_TUNE (X86_TUNE_SSE_PARTIAL_REG_FP_CONVERTS_DEPENDENCY,
  73           "sse_partial_reg_fp_converts_dependency",
  74           m_PPRO | m_P4_NOCONA | m_CORE_ALL | m_BONNELL | m_AMDFAM10
  75           | m_BDVER | m_ZNVER | m_LUJIAZUI | m_ALDERLAKE | m_CORE_ATOM
  76           | m_GENERIC)
  77
  78 /* X86_TUNE_SSE_PARTIAL_REG_CONVERTS_DEPENDENCY: This knob avoids partial
  79    write to the destination in scalar SSE conversion from integer to FP.  */
  80 DEF_TUNE (X86_TUNE_SSE_PARTIAL_REG_CONVERTS_DEPENDENCY,
  81           "sse_partial_reg_converts_dependency",
  82           m_PPRO | m_P4_NOCONA | m_CORE_ALL | m_BONNELL | m_AMDFAM10
  83           | m_BDVER | m_ZNVER | m_LUJIAZUI | m_ALDERLAKE | m_CORE_ATOM
  84           | m_GENERIC)
  85
  86 /* X86_TUNE_DEST_FALSE_DEP_FOR_GLC: This knob inserts zero-idiom before
  87    several insns to break false dependency on the dest register for GLC
  88    micro-architecture.  */
  89 DEF_TUNE (X86_TUNE_DEST_FALSE_DEP_FOR_GLC,
  90           "dest_false_dep_for_glc", m_SAPPHIRERAPIDS | m_ALDERLAKE
  91           | m_CORE_ATOM)
  92
  93 /* X86_TUNE_SSE_SPLIT_REGS: Set for machines where the type and dependencies
  94    are resolved on SSE register parts instead of whole registers, so we may
  95    maintain just lower part of scalar values in proper format leaving the
  96    upper part undefined.  */
  97 DEF_TUNE (X86_TUNE_SSE_SPLIT_REGS, "sse_split_regs", m_ATHLON_K8)
  98
  99 /* X86_TUNE_PARTIAL_FLAG_REG_STALL: this flag disables use of flags
 100    set by instructions affecting just some flags (in particular shifts).
 101    This is because Core2 resolves dependencies on whole flags register
 102    and such sequences introduce false dependency on previous instruction
 103    setting full flags.
 104
 105    The flags does not affect generation of INC and DEC that is controlled
 106    by X86_TUNE_USE_INCDEC.  */
 107
 108 DEF_TUNE (X86_TUNE_PARTIAL_FLAG_REG_STALL, "partial_flag_reg_stall",
 109           m_CORE2)
 110
 111 /* X86_TUNE_MOVX: Enable to zero extend integer registers to avoid
 112    partial dependencies.  */
 113 DEF_TUNE (X86_TUNE_MOVX, "movx",
 114           m_PPRO | m_P4_NOCONA | m_CORE2 | m_NEHALEM  | m_SANDYBRIDGE
 115           | m_BONNELL | m_SILVERMONT | m_GOLDMONT | m_KNL | m_KNM | m_INTEL
 116           | m_GOLDMONT_PLUS | m_GEODE | m_AMD_MULTIPLE | m_LUJIAZUI
 117           | m_CORE_AVX2 | m_TREMONT | m_ALDERLAKE | m_CORE_ATOM | m_GENERIC)
 118
 119 /* X86_TUNE_MEMORY_MISMATCH_STALL: Avoid partial stores that are followed by
 120    full sized loads.  */
 121 DEF_TUNE (X86_TUNE_MEMORY_MISMATCH_STALL, "memory_mismatch_stall",
 122           m_P4_NOCONA | m_CORE_ALL | m_BONNELL | m_SILVERMONT | m_INTEL
 123           | m_KNL | m_KNM | m_GOLDMONT | m_GOLDMONT_PLUS | m_AMD_MULTIPLE
 124           | m_LUJIAZUI | m_TREMONT | m_ALDERLAKE | m_CORE_ATOM | m_GENERIC)
 125
 126 /* X86_TUNE_FUSE_CMP_AND_BRANCH_32: Fuse compare with a subsequent
 127    conditional jump instruction for 32 bit TARGET.  */
 128 DEF_TUNE (X86_TUNE_FUSE_CMP_AND_BRANCH_32, "fuse_cmp_and_branch_32",
 129           m_CORE_ALL | m_BDVER | m_ZNVER | m_LUJIAZUI | m_GENERIC)
 130
 131 /* X86_TUNE_FUSE_CMP_AND_BRANCH_64: Fuse compare with a subsequent
 132    conditional jump instruction for TARGET_64BIT.  */
 133 DEF_TUNE (X86_TUNE_FUSE_CMP_AND_BRANCH_64, "fuse_cmp_and_branch_64",
 134           m_NEHALEM | m_SANDYBRIDGE | m_CORE_AVX2 | m_BDVER
 135           | m_ZNVER | m_LUJIAZUI | m_GENERIC)
 136
 137 /* X86_TUNE_FUSE_CMP_AND_BRANCH_SOFLAGS: Fuse compare with a
 138    subsequent conditional jump instruction when the condition jump
 139    check sign flag (SF) or overflow flag (OF).  */
 140 DEF_TUNE (X86_TUNE_FUSE_CMP_AND_BRANCH_SOFLAGS, "fuse_cmp_and_branch_soflags",
 141           m_NEHALEM | m_SANDYBRIDGE | m_CORE_AVX2 | m_BDVER
 142           | m_ZNVER | m_LUJIAZUI | m_GENERIC)
 143
 144 /* X86_TUNE_FUSE_ALU_AND_BRANCH: Fuse alu with a subsequent conditional
 145    jump instruction when the alu instruction produces the CCFLAG consumed by
 146    the conditional jump instruction. */
 147 DEF_TUNE (X86_TUNE_FUSE_ALU_AND_BRANCH, "fuse_alu_and_branch",
 148                   m_SANDYBRIDGE | m_CORE_AVX2 | m_LUJIAZUI | m_GENERIC)
 149
 150
 151 /*****************************************************************************/
 152 /* Function prologue, epilogue and function calling sequences.               */
 153 /*****************************************************************************/
 154
 155 /* X86_TUNE_ACCUMULATE_OUTGOING_ARGS: Allocate stack space for outgoing
 156    arguments in prologue/epilogue instead of separately for each call
 157    by push/pop instructions.
 158    This increase code size by about 5% in 32bit mode, less so in 64bit mode
 159    because parameters are passed in registers.  It is considerable
 160    win for targets without stack engine that prevents multple push operations
 161    to happen in parallel.  */
 162
 163 DEF_TUNE (X86_TUNE_ACCUMULATE_OUTGOING_ARGS, "accumulate_outgoing_args",
 164           m_PPRO | m_P4_NOCONA | m_BONNELL | m_SILVERMONT | m_KNL | m_KNM | m_INTEL
 165           | m_GOLDMONT | m_GOLDMONT_PLUS | m_ATHLON_K8 | m_LUJIAZUI)
 166
 167 /* X86_TUNE_PROLOGUE_USING_MOVE: Do not use push/pop in prologues that are
 168    considered on critical path.  */
 169 DEF_TUNE (X86_TUNE_PROLOGUE_USING_MOVE, "prologue_using_move",
 170           m_PPRO | m_ATHLON_K8)
 171
 172 /* X86_TUNE_PROLOGUE_USING_MOVE: Do not use push/pop in epilogues that are
 173    considered on critical path.  */
 174 DEF_TUNE (X86_TUNE_EPILOGUE_USING_MOVE, "epilogue_using_move",
 175           m_PPRO | m_ATHLON_K8)
 176
 177 /* X86_TUNE_USE_LEAVE: Use "leave" instruction in epilogues where it fits.  */
 178 DEF_TUNE (X86_TUNE_USE_LEAVE, "use_leave",
 179           m_386 | m_CORE_ALL | m_K6_GEODE | m_AMD_MULTIPLE | m_LUJIAZUI
 180           | m_TREMONT | m_ALDERLAKE | m_CORE_ATOM | m_GENERIC)
 181
 182 /* X86_TUNE_PUSH_MEMORY: Enable generation of "push mem" instructions.
 183    Some chips, like 486 and Pentium works faster with separate load
 184    and push instructions.  */
 185 DEF_TUNE (X86_TUNE_PUSH_MEMORY, "push_memory",
 186           m_386 | m_P4_NOCONA | m_CORE_ALL | m_K6_GEODE | m_AMD_MULTIPLE
 187           | m_LUJIAZUI | m_TREMONT | m_ALDERLAKE | m_CORE_ATOM | m_GENERIC)
 188
 189 /* X86_TUNE_SINGLE_PUSH: Enable if single push insn is preferred
 190    over esp subtraction.  */
 191 DEF_TUNE (X86_TUNE_SINGLE_PUSH, "single_push", m_386 | m_486 | m_PENT
 192           | m_LAKEMONT | m_K6_GEODE)
 193
 194 /* X86_TUNE_DOUBLE_PUSH. Enable if double push insn is preferred
 195    over esp subtraction.  */
 196 DEF_TUNE (X86_TUNE_DOUBLE_PUSH, "double_push", m_PENT | m_LAKEMONT
 197           | m_K6_GEODE)
 198
 199 /* X86_TUNE_SINGLE_POP: Enable if single pop insn is preferred
 200    over esp addition.  */
 201 DEF_TUNE (X86_TUNE_SINGLE_POP, "single_pop", m_386 | m_486 | m_PENT
 202           | m_LAKEMONT | m_PPRO)
 203
 204 /* X86_TUNE_DOUBLE_POP: Enable if double pop insn is preferred
 205    over esp addition.  */
 206 DEF_TUNE (X86_TUNE_DOUBLE_POP, "double_pop", m_PENT | m_LAKEMONT)
 207
 208 /*****************************************************************************/
 209 /* Branch predictor tuning                                                   */
 210 /*****************************************************************************/
 211
 212 /* X86_TUNE_PAD_SHORT_FUNCTION: Make every function to be at least 4
 213    instructions long.  */
 214 DEF_TUNE (X86_TUNE_PAD_SHORT_FUNCTION, "pad_short_function", m_BONNELL)
 215
 216 /* X86_TUNE_PAD_RETURNS: Place NOP before every RET that is a destination
 217    of conditional jump or directly preceded by other jump instruction.
 218    This is important for AND K8-AMDFAM10 because the branch prediction
 219    architecture expect at most one jump per 2 byte window.  Failing to
 220    pad returns leads to misaligned return stack.  */
 221 DEF_TUNE (X86_TUNE_PAD_RETURNS, "pad_returns",
 222           m_ATHLON_K8 | m_AMDFAM10)
 223
 224 /* X86_TUNE_FOUR_JUMP_LIMIT: Some CPU cores are not able to predict more
 225    than 4 branch instructions in the 16 byte window.  */
 226 DEF_TUNE (X86_TUNE_FOUR_JUMP_LIMIT, "four_jump_limit",
 227           m_PPRO | m_P4_NOCONA | m_BONNELL | m_SILVERMONT | m_KNL | m_KNM
 228           | m_GOLDMONT | m_GOLDMONT_PLUS | m_INTEL | m_ATHLON_K8 | m_AMDFAM10)
 229
 230 /*****************************************************************************/
 231 /* Integer instruction selection tuning                                      */
 232 /*****************************************************************************/
 233
 234 /* X86_TUNE_SOFTWARE_PREFETCHING_BENEFICIAL: Enable software prefetching
 235    at -O3.  For the moment, the prefetching seems badly tuned for Intel
 236    chips.  */
 237 DEF_TUNE (X86_TUNE_SOFTWARE_PREFETCHING_BENEFICIAL, "software_prefetching_beneficial",
 238           m_K6_GEODE | m_ATHLON_K8 | m_AMDFAM10 | m_BDVER | m_BTVER)
 239
 240 /* X86_TUNE_LCP_STALL: Avoid an expensive length-changing prefix stall
 241    on 16-bit immediate moves into memory on Core2 and Corei7.  */
 242 DEF_TUNE (X86_TUNE_LCP_STALL, "lcp_stall", m_CORE_ALL | m_LUJIAZUI | m_GENERIC)
 243
 244 /* X86_TUNE_READ_MODIFY: Enable use of read-modify instructions such
 245    as "add mem, reg".  */
 246 DEF_TUNE (X86_TUNE_READ_MODIFY, "read_modify", ~(m_PENT | m_LAKEMONT | m_PPRO))
 247
 248 /* X86_TUNE_USE_INCDEC: Enable use of inc/dec instructions.
 249
 250    Core2 and nehalem has stall of 7 cycles for partial flag register stalls.
 251    Sandy bridge and Ivy bridge generate extra uop.  On Haswell this extra uop
 252    is output only when the values needs to be really merged, which is not
 253    done by GCC generated code.  */
 254 DEF_TUNE (X86_TUNE_USE_INCDEC, "use_incdec",
 255           ~(m_P4_NOCONA | m_CORE2 | m_NEHALEM  | m_SANDYBRIDGE
 256             | m_BONNELL | m_SILVERMONT | m_INTEL |  m_KNL | m_KNM | m_GOLDMONT
 257             | m_GOLDMONT_PLUS | m_TREMONT | m_ALDERLAKE | m_CORE_ATOM
 258             | m_LUJIAZUI | m_GENERIC))
 259
 260 /* X86_TUNE_INTEGER_DFMODE_MOVES: Enable if integer moves are preferred
 261    for DFmode copies */
 262 DEF_TUNE (X86_TUNE_INTEGER_DFMODE_MOVES, "integer_dfmode_moves",
 263           ~(m_PPRO | m_P4_NOCONA | m_CORE_ALL | m_BONNELL | m_SILVERMONT
 264             | m_KNL | m_KNM | m_INTEL | m_GEODE | m_AMD_MULTIPLE | m_LUJIAZUI
 265             | m_GOLDMONT | m_GOLDMONT_PLUS | m_TREMONT | m_ALDERLAKE
 266             | m_CORE_ATOM | m_GENERIC))
 267
 268 /* X86_TUNE_OPT_AGU: Optimize for Address Generation Unit. This flag
 269    will impact LEA instruction selection. */
 270 DEF_TUNE (X86_TUNE_OPT_AGU, "opt_agu", m_BONNELL | m_SILVERMONT | m_KNL
 271          | m_KNM | m_GOLDMONT | m_GOLDMONT_PLUS | m_INTEL | m_LUJIAZUI)
 272
 273 /* X86_TUNE_AVOID_LEA_FOR_ADDR: Avoid lea for address computation.  */
 274 DEF_TUNE (X86_TUNE_AVOID_LEA_FOR_ADDR, "avoid_lea_for_addr",
 275           m_BONNELL | m_SILVERMONT | m_GOLDMONT | m_GOLDMONT_PLUS
 276           | m_KNL | m_KNM)
 277
 278 /* X86_TUNE_SLOW_IMUL_IMM32_MEM: Imul of 32-bit constant and memory is
 279    vector path on AMD machines.
 280    FIXME: Do we need to enable this for core? */
 281 DEF_TUNE (X86_TUNE_SLOW_IMUL_IMM32_MEM, "slow_imul_imm32_mem",
 282           m_K8 | m_AMDFAM10)
 283
 284 /* X86_TUNE_SLOW_IMUL_IMM8: Imul of 8-bit constant is vector path on AMD
 285    machines.
 286    FIXME: Do we need to enable this for core? */
 287 DEF_TUNE (X86_TUNE_SLOW_IMUL_IMM8, "slow_imul_imm8",
 288           m_K8 | m_AMDFAM10)
 289
 290 /* X86_TUNE_AVOID_MEM_OPND_FOR_CMOVE: Try to avoid memory operands for
 291    a conditional move.  */
 292 DEF_TUNE (X86_TUNE_AVOID_MEM_OPND_FOR_CMOVE, "avoid_mem_opnd_for_cmove",
 293           m_BONNELL | m_SILVERMONT | m_GOLDMONT | m_GOLDMONT_PLUS | m_KNL
 294           | m_KNM | m_INTEL)
 295
 296 /* X86_TUNE_SINGLE_STRINGOP: Enable use of single string operations, such
 297    as MOVS and STOS (without a REP prefix) to move/set sequences of bytes.  */
 298 DEF_TUNE (X86_TUNE_SINGLE_STRINGOP, "single_stringop", m_386 | m_P4_NOCONA)
 299
 300 /* X86_TUNE_PREFER_KNOWN_REP_MOVSB_STOSB: Enable use of REP MOVSB/STOSB to
 301    move/set sequences of bytes with known size.  */
 302 DEF_TUNE (X86_TUNE_PREFER_KNOWN_REP_MOVSB_STOSB,
 303           "prefer_known_rep_movsb_stosb",
 304           m_SKYLAKE | m_ALDERLAKE | m_CORE_ATOM | m_TREMONT | m_CORE_AVX512
 305           | m_LUJIAZUI)
 306
 307 /* X86_TUNE_MISALIGNED_MOVE_STRING_PRO_EPILOGUES: Enable generation of
 308    compact prologues and epilogues by issuing a misaligned moves.  This
 309    requires target to handle misaligned moves and partial memory stalls
 310    reasonably well.
 311    FIXME: This may actualy be a win on more targets than listed here.  */
 312 DEF_TUNE (X86_TUNE_MISALIGNED_MOVE_STRING_PRO_EPILOGUES,
 313           "misaligned_move_string_pro_epilogues",
 314           m_386 | m_486 | m_CORE_ALL | m_AMD_MULTIPLE | m_LUJIAZUI | m_TREMONT
 315           | m_ALDERLAKE | m_CORE_ATOM | m_GENERIC)
 316
 317 /* X86_TUNE_USE_SAHF: Controls use of SAHF.  */
 318 DEF_TUNE (X86_TUNE_USE_SAHF, "use_sahf",
 319           m_PPRO | m_P4_NOCONA | m_CORE_ALL | m_BONNELL | m_SILVERMONT
 320           | m_KNL | m_KNM | m_INTEL | m_K6_GEODE | m_K8 | m_AMDFAM10 | m_BDVER
 321           | m_BTVER | m_ZNVER | m_LUJIAZUI | m_GOLDMONT | m_GOLDMONT_PLUS
 322           | m_TREMONT | m_ALDERLAKE | m_CORE_ATOM | m_GENERIC)
 323
 324 /* X86_TUNE_USE_CLTD: Controls use of CLTD and CTQO instructions.  */
 325 DEF_TUNE (X86_TUNE_USE_CLTD, "use_cltd",
 326           ~(m_PENT | m_LAKEMONT | m_BONNELL | m_SILVERMONT | m_KNL | m_KNM | m_INTEL
 327             | m_K6 | m_GOLDMONT | m_GOLDMONT_PLUS))
 328
 329 /* X86_TUNE_USE_BT: Enable use of BT (bit test) instructions.  */
 330 DEF_TUNE (X86_TUNE_USE_BT, "use_bt",
 331           m_CORE_ALL | m_BONNELL | m_SILVERMONT | m_KNL | m_KNM | m_INTEL
 332           | m_LAKEMONT | m_AMD_MULTIPLE | m_LUJIAZUI | m_GOLDMONT
 333           | m_GOLDMONT_PLUS | m_TREMONT | m_ALDERLAKE | m_CORE_ATOM | m_GENERIC)
 334
 335 /* X86_TUNE_AVOID_FALSE_DEP_FOR_BMI: Avoid false dependency
 336    for bit-manipulation instructions.  */
 337 DEF_TUNE (X86_TUNE_AVOID_FALSE_DEP_FOR_BMI, "avoid_false_dep_for_bmi",
 338           m_SANDYBRIDGE | m_CORE_AVX2 | m_TREMONT | m_ALDERLAKE | m_CORE_ATOM
 339           | m_LUJIAZUI | m_GENERIC)
 340
 341 /* X86_TUNE_ADJUST_UNROLL: This enables adjusting the unroll factor based
 342    on hardware capabilities. Bdver3 hardware has a loop buffer which makes
 343    unrolling small loop less important. For, such architectures we adjust
 344    the unroll factor so that the unrolled loop fits the loop buffer.  */
 345 DEF_TUNE (X86_TUNE_ADJUST_UNROLL, "adjust_unroll_factor", m_BDVER3 | m_BDVER4)
 346
 347 /* X86_TUNE_ONE_IF_CONV_INSNS: Restrict a number of cmov insns in
 348    if-converted sequence to one.  */
 349 DEF_TUNE (X86_TUNE_ONE_IF_CONV_INSN, "one_if_conv_insn",
 350           m_SILVERMONT | m_KNL | m_KNM | m_INTEL | m_CORE_ALL | m_GOLDMONT
 351           | m_GOLDMONT_PLUS | m_TREMONT | m_ALDERLAKE | m_CORE_ATOM | m_LUJIAZUI
 352           | m_GENERIC)
 353
 354 /* X86_TUNE_AVOID_MFENCE: Use lock prefixed instructions instead of mfence.  */
 355 DEF_TUNE (X86_TUNE_AVOID_MFENCE, "avoid_mfence",
 356          m_CORE_ALL | m_BDVER | m_ZNVER | m_LUJIAZUI | m_TREMONT | m_ALDERLAKE
 357          | m_CORE_ATOM | m_GENERIC)
 358
 359 /* X86_TUNE_EXPAND_ABS: This enables a new abs pattern by
 360    generating instructions for abs (x) = (((signed) x >> (W-1) ^ x) -
 361    (signed) x >> (W-1)) instead of cmove or SSE max/abs instructions.  */
 362 DEF_TUNE (X86_TUNE_EXPAND_ABS, "expand_abs",
 363           m_CORE_ALL | m_SILVERMONT | m_KNL | m_KNM | m_GOLDMONT
 364           | m_GOLDMONT_PLUS | m_LUJIAZUI)
 365
 366 /*****************************************************************************/
 367 /* 387 instruction selection tuning                                          */
 368 /*****************************************************************************/
 369
 370 /* X86_TUNE_USE_HIMODE_FIOP: Enables use of x87 instructions with 16bit
 371    integer operand.
 372    FIXME: Why this is disabled for modern chips?  */
 373 DEF_TUNE (X86_TUNE_USE_HIMODE_FIOP, "use_himode_fiop",
 374           m_386 | m_486 | m_K6_GEODE)
 375
 376 /* X86_TUNE_USE_SIMODE_FIOP: Enables use of x87 instructions with 32bit
 377    integer operand.  */
 378 DEF_TUNE (X86_TUNE_USE_SIMODE_FIOP, "use_simode_fiop",
 379           ~(m_PENT | m_LAKEMONT | m_PPRO | m_CORE_ALL | m_BONNELL
 380             | m_SILVERMONT | m_KNL | m_KNM | m_INTEL | m_AMD_MULTIPLE
 381             | m_LUJIAZUI | m_GOLDMONT | m_GOLDMONT_PLUS | m_TREMONT
 382             | m_ALDERLAKE | m_CORE_ATOM | m_GENERIC))
 383
 384 /* X86_TUNE_USE_FFREEP: Use freep instruction instead of fstp.  */
 385 DEF_TUNE (X86_TUNE_USE_FFREEP, "use_ffreep", m_AMD_MULTIPLE | m_LUJIAZUI)
 386
 387 /* X86_TUNE_EXT_80387_CONSTANTS: Use fancy 80387 constants, such as PI.  */
 388 DEF_TUNE (X86_TUNE_EXT_80387_CONSTANTS, "ext_80387_constants",
 389           m_PPRO | m_P4_NOCONA | m_CORE_ALL | m_BONNELL | m_SILVERMONT
 390           | m_KNL | m_KNM | m_INTEL | m_K6_GEODE | m_ATHLON_K8 | m_LUJIAZUI
 391           | m_GOLDMONT | m_GOLDMONT_PLUS | m_TREMONT | m_ALDERLAKE | m_CORE_ATOM
 392           | m_GENERIC)
 393
 394 /*****************************************************************************/
 395 /* SSE instruction selection tuning                                          */
 396 /*****************************************************************************/
 397
 398 /* X86_TUNE_GENERAL_REGS_SSE_SPILL: Try to spill general regs to SSE
 399    regs instead of memory.  */
 400 DEF_TUNE (X86_TUNE_GENERAL_REGS_SSE_SPILL, "general_regs_sse_spill",
 401           m_CORE_ALL)
 402
 403 /* X86_TUNE_SSE_UNALIGNED_LOAD_OPTIMAL: Use movups for misaligned loads instead
 404    of a sequence loading registers by parts.  */
 405 DEF_TUNE (X86_TUNE_SSE_UNALIGNED_LOAD_OPTIMAL, "sse_unaligned_load_optimal",
 406           m_NEHALEM | m_SANDYBRIDGE | m_CORE_AVX2 | m_SILVERMONT | m_KNL | m_KNM
 407           | m_INTEL | m_GOLDMONT | m_GOLDMONT_PLUS | m_TREMONT | m_ALDERLAKE
 408           | m_CORE_ATOM | m_AMDFAM10 | m_BDVER | m_BTVER | m_ZNVER | m_LUJIAZUI
 409           | m_GENERIC)
 410
 411 /* X86_TUNE_SSE_UNALIGNED_STORE_OPTIMAL: Use movups for misaligned stores
 412    instead of a sequence loading registers by parts.  */
 413 DEF_TUNE (X86_TUNE_SSE_UNALIGNED_STORE_OPTIMAL, "sse_unaligned_store_optimal",
 414           m_NEHALEM | m_SANDYBRIDGE | m_CORE_AVX2 | m_SILVERMONT | m_KNL | m_KNM
 415           | m_INTEL | m_GOLDMONT | m_GOLDMONT_PLUS | m_TREMONT | m_ALDERLAKE
 416           | m_CORE_ATOM | m_BDVER | m_ZNVER | m_LUJIAZUI | m_GENERIC)
 417
 418 /* X86_TUNE_SSE_PACKED_SINGLE_INSN_OPTIMAL: Use packed single
 419    precision 128bit instructions instead of double where possible.   */
 420 DEF_TUNE (X86_TUNE_SSE_PACKED_SINGLE_INSN_OPTIMAL, "sse_packed_single_insn_optimal",
 421           m_BDVER | m_ZNVER)
 422
 423 /* X86_TUNE_SSE_TYPELESS_STORES: Always movaps/movups for 128bit stores.   */
 424 DEF_TUNE (X86_TUNE_SSE_TYPELESS_STORES, "sse_typeless_stores",
 425           m_AMD_MULTIPLE | m_LUJIAZUI | m_CORE_ALL | m_TREMONT | m_ALDERLAKE
 426           | m_CORE_ATOM | m_GENERIC)
 427
 428 /* X86_TUNE_SSE_LOAD0_BY_PXOR: Always use pxor to load0 as opposed to
 429    xorps/xorpd and other variants.  */
 430 DEF_TUNE (X86_TUNE_SSE_LOAD0_BY_PXOR, "sse_load0_by_pxor",
 431           m_PPRO | m_P4_NOCONA | m_CORE_ALL | m_BDVER | m_BTVER | m_ZNVER
 432           | m_LUJIAZUI | m_TREMONT | m_ALDERLAKE | m_CORE_ATOM | m_GENERIC)
 433
 434 /* X86_TUNE_INTER_UNIT_MOVES_TO_VEC: Enable moves in from integer
 435    to SSE registers.  If disabled, the moves will be done by storing
 436    the value to memory and reloading.
 437    Enable this flag for generic - the only relevant architecture preferring
 438    no inter-unit moves is Buldozer. While this makes small regression on SPECfp
 439    scores (sub 0.3%), disabling inter-unit moves penalizes noticeably hand
 440    written vectorized code which use i.e. _mm_set_epi16.  */
 441 DEF_TUNE (X86_TUNE_INTER_UNIT_MOVES_TO_VEC, "inter_unit_moves_to_vec",
 442           ~(m_ATHLON_K8 | m_AMDFAM10 | m_BDVER | m_BTVER))
 443
 444 /* X86_TUNE_INTER_UNIT_MOVES_TO_VEC: Enable moves in from SSE
 445    to integer registers.  If disabled, the moves will be done by storing
 446    the value to memory and reloading.  */
 447 DEF_TUNE (X86_TUNE_INTER_UNIT_MOVES_FROM_VEC, "inter_unit_moves_from_vec",
 448           ~m_ATHLON_K8)
 449
 450 /* X86_TUNE_INTER_UNIT_CONVERSIONS: Enable float<->integer conversions
 451    to use both SSE and integer registers at a same time.  */
 452 DEF_TUNE (X86_TUNE_INTER_UNIT_CONVERSIONS, "inter_unit_conversions",
 453           ~(m_AMDFAM10 | m_BDVER))
 454
 455 /* X86_TUNE_SPLIT_MEM_OPND_FOR_FP_CONVERTS: Try to split memory operand for
 456    fp converts to destination register.  */
 457 DEF_TUNE (X86_TUNE_SPLIT_MEM_OPND_FOR_FP_CONVERTS, "split_mem_opnd_for_fp_converts",
 458           m_SILVERMONT | m_KNL | m_KNM | m_GOLDMONT | m_GOLDMONT_PLUS
 459           | m_INTEL)
 460
 461 /* X86_TUNE_USE_VECTOR_FP_CONVERTS: Prefer vector packed SSE conversion
 462    from FP to FP.  This form of instructions avoids partial write to the
 463    destination.  */
 464 DEF_TUNE (X86_TUNE_USE_VECTOR_FP_CONVERTS, "use_vector_fp_converts",
 465           m_AMDFAM10)
 466
 467 /* X86_TUNE_USE_VECTOR_CONVERTS: Prefer vector packed SSE conversion
 468    from integer to FP. */
 469 DEF_TUNE (X86_TUNE_USE_VECTOR_CONVERTS, "use_vector_converts", m_AMDFAM10)
 470
 471 /* X86_TUNE_SLOW_SHUFB: Indicates tunings with slow pshufb instruction.  */
 472 DEF_TUNE (X86_TUNE_SLOW_PSHUFB, "slow_pshufb",
 473           m_BONNELL | m_SILVERMONT | m_KNL | m_KNM | m_GOLDMONT
 474           | m_GOLDMONT_PLUS | m_INTEL)
 475
 476 /* X86_TUNE_AVOID_4BYTE_PREFIXES: Avoid instructions requiring 4+ bytes of prefixes.  */
 477 DEF_TUNE (X86_TUNE_AVOID_4BYTE_PREFIXES, "avoid_4byte_prefixes",
 478           m_SILVERMONT | m_GOLDMONT | m_GOLDMONT_PLUS | m_TREMONT | m_ALDERLAKE
 479           | m_CORE_ATOM | m_INTEL)
 480
 481 /* X86_TUNE_USE_GATHER_2PARTS: Use gather instructions for vectors with 2
 482    elements.  */
 483 DEF_TUNE (X86_TUNE_USE_GATHER_2PARTS, "use_gather_2parts",
 484           ~(m_ZNVER1 | m_ZNVER2 | m_ZNVER3 | m_ZNVER4 | m_ALDERLAKE | m_CORE_ATOM | m_GENERIC))
 485
 486 /* X86_TUNE_USE_GATHER_4PARTS: Use gather instructions for vectors with 4
 487    elements.  */
 488 DEF_TUNE (X86_TUNE_USE_GATHER_4PARTS, "use_gather_4parts",
 489           ~(m_ZNVER1 | m_ZNVER2 | m_ZNVER3 | m_ZNVER4 | m_ALDERLAKE | m_CORE_ATOM | m_GENERIC))
 490
 491 /* X86_TUNE_USE_GATHER: Use gather instructions for vectors with 8 or more
 492    elements.  */
 493 DEF_TUNE (X86_TUNE_USE_GATHER, "use_gather",
 494           ~(m_ZNVER1 | m_ZNVER2 | m_ALDERLAKE | m_CORE_ATOM | m_GENERIC))
 495
 496 /* X86_TUNE_AVOID_128FMA_CHAINS: Avoid creating loops with tight 128bit or
 497    smaller FMA chain.  */
 498 DEF_TUNE (X86_TUNE_AVOID_128FMA_CHAINS, "avoid_fma_chains", m_ZNVER)
 499
 500 /* X86_TUNE_AVOID_256FMA_CHAINS: Avoid creating loops with tight 256bit or
 501    smaller FMA chain.  */
 502 DEF_TUNE (X86_TUNE_AVOID_256FMA_CHAINS, "avoid_fma256_chains", m_ZNVER2 | m_ZNVER3 | m_ZNVER4
 503           | m_ALDERLAKE | m_SAPPHIRERAPIDS | m_CORE_ATOM)
 504
 505 /* X86_TUNE_AVOID_512FMA_CHAINS: Avoid creating loops with tight 512bit or
 506    smaller FMA chain.  */
 507 DEF_TUNE (X86_TUNE_AVOID_512FMA_CHAINS, "avoid_fma512_chains", m_ZNVER4)
 508
 509 /* X86_TUNE_V2DF_REDUCTION_PREFER_PHADDPD: Prefer haddpd
 510    for v2df vector reduction.  */
 511 DEF_TUNE (X86_TUNE_V2DF_REDUCTION_PREFER_HADDPD,
 512           "v2df_reduction_prefer_haddpd", m_NONE)
 513
 514 /*****************************************************************************/
 515 /* AVX instruction selection tuning (some of SSE flags affects AVX, too)     */
 516 /*****************************************************************************/
 517
 518 /* X86_TUNE_AVX256_UNALIGNED_LOAD_OPTIMAL: if false, unaligned loads are
 519    split.  */
 520 DEF_TUNE (X86_TUNE_AVX256_UNALIGNED_LOAD_OPTIMAL, "256_unaligned_load_optimal",
 521           ~(m_NEHALEM | m_SANDYBRIDGE))
 522
 523 /* X86_TUNE_AVX256_UNALIGNED_STORE_OPTIMAL: if false, unaligned stores are
 524    split.  */
 525 DEF_TUNE (X86_TUNE_AVX256_UNALIGNED_STORE_OPTIMAL, "256_unaligned_store_optimal",
 526           ~(m_NEHALEM | m_SANDYBRIDGE | m_BDVER | m_ZNVER1))
 527
 528 /* X86_TUNE_AVX256_SPLIT_REGS: if true, AVX256 ops are split into two AVX128 ops.  */
 529 DEF_TUNE (X86_TUNE_AVX256_SPLIT_REGS, "avx256_split_regs",m_BDVER | m_BTVER2
 530           | m_ZNVER1)
 531
 532 /* X86_TUNE_AVX128_OPTIMAL: Enable 128-bit AVX instruction generation for
 533    the auto-vectorizer.  */
 534 DEF_TUNE (X86_TUNE_AVX128_OPTIMAL, "avx128_optimal", m_BDVER | m_BTVER2
 535           | m_ZNVER1)
 536
 537 /* X86_TUNE_AVX256_OPTIMAL: Use 256-bit AVX instructions instead of 512-bit AVX
 538    instructions in the auto-vectorizer.  */
 539 DEF_TUNE (X86_TUNE_AVX256_OPTIMAL, "avx256_optimal", m_CORE_AVX512 | m_ZNVER4)
 540
 541 /* X86_TUNE_AVX256_SPLIT_REGS: if true, AVX512 ops are split into two AVX256 ops.  */
 542 DEF_TUNE (X86_TUNE_AVX512_SPLIT_REGS, "avx512_split_regs", m_ZNVER4)
 543
 544 /* X86_TUNE_AVX256_MOVE_BY_PIECES: Optimize move_by_pieces with 256-bit
 545    AVX instructions.  */
 546 DEF_TUNE (X86_TUNE_AVX256_MOVE_BY_PIECES, "avx256_move_by_pieces",
 547           m_ALDERLAKE | m_CORE_AVX2 | m_ZNVER1 | m_ZNVER2 | m_ZNVER3)
 548
 549 /* X86_TUNE_AVX256_STORE_BY_PIECES: Optimize store_by_pieces with 256-bit
 550    AVX instructions.  */
 551 DEF_TUNE (X86_TUNE_AVX256_STORE_BY_PIECES, "avx256_store_by_pieces",
 552           m_ALDERLAKE | m_CORE_AVX2 | m_ZNVER1 | m_ZNVER2 | m_ZNVER3)
 553
 554 /* X86_TUNE_AVX512_MOVE_BY_PIECES: Optimize move_by_pieces with 512-bit
 555    AVX instructions.  */
 556 DEF_TUNE (X86_TUNE_AVX512_MOVE_BY_PIECES, "avx512_move_by_pieces",
 557           m_SAPPHIRERAPIDS | m_ZNVER4)
 558
 559 /* X86_TUNE_AVX512_STORE_BY_PIECES: Optimize store_by_pieces with 512-bit
 560    AVX instructions.  */
 561 DEF_TUNE (X86_TUNE_AVX512_STORE_BY_PIECES, "avx512_store_by_pieces",
 562           m_SAPPHIRERAPIDS | m_ZNVER4)
 563
 564 /*****************************************************************************/
 565 /*****************************************************************************/
 566 /* Historical relics: tuning flags that helps a specific old CPU designs     */
 567 /*****************************************************************************/
 568
 569 /* X86_TUNE_DOUBLE_WITH_ADD: Use add instead of sal to double value in
 570    an integer register.  */
 571 DEF_TUNE (X86_TUNE_DOUBLE_WITH_ADD, "double_with_add", ~m_386)
 572
 573 /* X86_TUNE_ALWAYS_FANCY_MATH_387: controls use of fancy 387 operations,
 574    such as fsqrt, fprem, fsin, fcos, fsincos etc.
 575    Should be enabled for all targets that always has coprocesor.  */
 576 DEF_TUNE (X86_TUNE_ALWAYS_FANCY_MATH_387, "always_fancy_math_387",
 577           ~(m_386 | m_486 | m_LAKEMONT))
 578
 579 /* X86_TUNE_UNROLL_STRLEN: Produce (quite lame) unrolled sequence for
 580    inline strlen.  This affects only -minline-all-stringops mode. By
 581    default we always dispatch to a library since our internal strlen
 582    is bad.  */
 583 DEF_TUNE (X86_TUNE_UNROLL_STRLEN, "unroll_strlen", ~m_386)
 584
 585 /* X86_TUNE_SHIFT1: Enables use of short encoding of "sal reg" instead of
 586    longer "sal $1, reg".  */
 587 DEF_TUNE (X86_TUNE_SHIFT1, "shift1", ~m_486)
 588
 589 /* X86_TUNE_ZERO_EXTEND_WITH_AND: Use AND instruction instead
 590    of mozbl/movwl.  */
 591 DEF_TUNE (X86_TUNE_ZERO_EXTEND_WITH_AND, "zero_extend_with_and",
 592           m_486 | m_PENT)
 593
 594 /* X86_TUNE_PROMOTE_HIMODE_IMUL: Modern CPUs have same latency for HImode
 595    and SImode multiply, but 386 and 486 do HImode multiply faster.  */
 596 DEF_TUNE (X86_TUNE_PROMOTE_HIMODE_IMUL, "promote_himode_imul",
 597           ~(m_386 | m_486))
 598
 599 /* X86_TUNE_FAST_PREFIX: Enable demoting some 32bit or 64bit arithmetic
 600    into 16bit/8bit when resulting sequence is shorter.  For example
 601    for "and $-65536, reg" to 16bit store of 0.  */
 602 DEF_TUNE (X86_TUNE_FAST_PREFIX, "fast_prefix",
 603           ~(m_386 | m_486 | m_PENT | m_LAKEMONT))
 604
 605 /* X86_TUNE_READ_MODIFY_WRITE: Enable use of read modify write instructions
 606    such as "add $1, mem".  */
 607 DEF_TUNE (X86_TUNE_READ_MODIFY_WRITE, "read_modify_write",
 608           ~(m_PENT | m_LAKEMONT))
 609
 610 /* X86_TUNE_MOVE_M1_VIA_OR: On pentiums, it is faster to load -1 via OR
 611    than a MOV.  */
 612 DEF_TUNE (X86_TUNE_MOVE_M1_VIA_OR, "move_m1_via_or", m_PENT | m_LAKEMONT)
 613
 614 /* X86_TUNE_NOT_UNPAIRABLE: NOT is not pairable on Pentium, while XOR is,
 615    but one byte longer.  */
 616 DEF_TUNE (X86_TUNE_NOT_UNPAIRABLE, "not_unpairable", m_PENT | m_LAKEMONT)
 617
 618 /* X86_TUNE_PARTIAL_REG_STALL: Pentium pro, unlike later chips, handled
 619    use of partial registers by renaming.  This improved performance of 16bit
 620    code where upper halves of registers are not used.  It also leads to
 621    an penalty whenever a 16bit store is followed by 32bit use.  This flag
 622    disables production of such sequences in common cases.
 623    See also X86_TUNE_HIMODE_MATH.
 624
 625    In current implementation the partial register stalls are not eliminated
 626    very well - they can be introduced via subregs synthesized by combine
 627    and can happen in caller/callee saving sequences.  */
 628 DEF_TUNE (X86_TUNE_PARTIAL_REG_STALL, "partial_reg_stall", m_PPRO)
 629
 630 /* X86_TUNE_PROMOTE_QIMODE: When it is cheap, turn 8bit arithmetic to
 631    corresponding 32bit arithmetic.  */
 632 DEF_TUNE (X86_TUNE_PROMOTE_QIMODE, "promote_qimode",
 633           ~m_PPRO)
 634
 635 /* X86_TUNE_PROMOTE_HI_REGS: Same, but for 16bit artihmetic.  Again we avoid
 636    partial register stalls on PentiumPro targets. */
 637 DEF_TUNE (X86_TUNE_PROMOTE_HI_REGS, "promote_hi_regs", m_PPRO)
 638
 639 /* X86_TUNE_HIMODE_MATH: Enable use of 16bit arithmetic.
 640    On PPro this flag is meant to avoid partial register stalls.  */
 641 DEF_TUNE (X86_TUNE_HIMODE_MATH, "himode_math", ~m_PPRO)
 642
 643 /* X86_TUNE_SPLIT_LONG_MOVES: Avoid instructions moving immediates
 644    directly to memory.  */
 645 DEF_TUNE (X86_TUNE_SPLIT_LONG_MOVES, "split_long_moves", m_PPRO)
 646
 647 /* X86_TUNE_USE_XCHGB: Use xchgb %rh,%rl instead of rolw/rorw $8,rx.  */
 648 DEF_TUNE (X86_TUNE_USE_XCHGB, "use_xchgb", m_PENT4)
 649
 650 /* X86_TUNE_USE_MOV0: Use "mov $0, reg" instead of "xor reg, reg" to clear
 651    integer register.  */
 652 DEF_TUNE (X86_TUNE_USE_MOV0, "use_mov0", m_K6)
 653
 654 /* X86_TUNE_NOT_VECTORMODE: On AMD K6, NOT is vector decoded with memory
 655    operand that cannot be represented using a modRM byte.  The XOR
 656    replacement is long decoded, so this split helps here as well.  */
 657 DEF_TUNE (X86_TUNE_NOT_VECTORMODE, "not_vectormode", m_K6)
 658
 659 /* X86_TUNE_AVOID_VECTOR_DECODE: Enable splitters that avoid vector decoded
 660    forms of instructions on K8 targets.  */
 661 DEF_TUNE (X86_TUNE_AVOID_VECTOR_DECODE, "avoid_vector_decode",
 662           m_K8)
 663
 664 /*****************************************************************************/
 665 /* This never worked well before.                                            */
 666 /*****************************************************************************/
 667
 668 /* X86_TUNE_BRANCH_PREDICTION_HINTS: Branch hints were put in P4 based
 669    on simulation result. But after P4 was made, no performance benefit
 670    was observed with branch hints.  It also increases the code size.
 671    As a result, icc never generates branch hints.  */
 672 DEF_TUNE (X86_TUNE_BRANCH_PREDICTION_HINTS, "branch_prediction_hints", m_NONE)
 673
 674 /* X86_TUNE_QIMODE_MATH: Enable use of 8bit arithmetic.  */
 675 DEF_TUNE (X86_TUNE_QIMODE_MATH, "qimode_math", m_ALL)
 676
 677 /* X86_TUNE_PROMOTE_QI_REGS: This enables generic code that promotes all 8bit
 678    arithmetic to 32bit via PROMOTE_MODE macro.  This code generation scheme
 679    is usually used for RISC targets.  */
 680 DEF_TUNE (X86_TUNE_PROMOTE_QI_REGS, "promote_qi_regs", m_NONE)
 681
 682 /* X86_TUNE_EMIT_VZEROUPPER: This enables vzeroupper instruction insertion
 683    before a transfer of control flow out of the function.  */
 684 DEF_TUNE (X86_TUNE_EMIT_VZEROUPPER, "emit_vzeroupper", ~m_KNL)