On Wed, Sep 15, 2021 at 10:09 AM <lili....@intel.com> wrote:
>
> From: "H.J. Lu" <hjl.to...@gmail.com>
>
> Initial -mtune=tremont update
>
> 1. Use Haswell scheduling model.
> 2. Assume that stack engine allows to execute push&pop instructions in
> parall.
> 3. Prepare for scheduling pass as -mtune=generic.
> 4. Use the same issue rate as -mtune=generic.
> 5. Enable partial_reg_dependency.
> 6. Disable accumulate_outgoing_args
> 7. Enable use_leave
> 8. Enable push_memory
> 9. Disable four_jump_limit
> 10. Disable opt_agu
> 11. Disable avoid_lea_for_addr
> 12. Disable avoid_mem_opnd_for_cmove
> 13. Enable misaligned_move_string_pro_epilogues
> 14. Enable use_cltd
> 16. Enable avoid_false_dep_for_bmi
> 17. Enable avoid_mfence
> 18. Disable expand_abs
> 19. Enable sse_typeless_stores
> 20. Enable sse_load0_by_pxor
> 21. Disable split_mem_opnd_for_fp_converts
> 22. Disable slow_pshufb
> 23. Enable partial_reg_dependency
>
> This is the first patch to tune for Tremont. With all patches applied,
> performance impacts on SPEC CPU 2017 are:
>
> 500.perlbench_r 1.81%
> 502.gcc_r 0.57%
> 505.mcf_r 1.16%
> 520.omnetpp_r 0.00%
> 523.xalancbmk_r 0.00%
> 525.x264_r 4.55%
> 531.deepsjeng_r 0.00%
> 541.leela_r 0.39%
> 548.exchange2_r 1.13%
> 557.xz_r 0.00%
> geomean for intrate 0.95%
> 503.bwaves_r 0.00%
> 507.cactuBSSN_r 6.94%
> 508.namd_r 12.37%
> 510.parest_r 1.01%
> 511.povray_r 3.70%
> 519.lbm_r 36.61%
> 521.wrf_r 8.79%
> 526.blender_r 2.91%
> 527.cam4_r 6.23%
> 538.imagick_r 0.28%
> 544.nab_r 21.99%
> 549.fotonik3d_r 3.63%
> 554.roms_r -1.20%
> geomean for fprate 7.50%
>
> gcc/ChangeLog
>
> * common/config/i386/i386-common.c: Use Haswell scheduling model
> for Tremont.
> * config/i386/i386.c (ix86_sched_init_global): Prepare for Tremont
> scheduling pass.
> * config/i386/x86-tune-sched.c (ix86_issue_rate): Change Tremont
> issue rate to 4.
> (ix86_adjust_cost): Handle Tremont.
> * config/i386/x86-tune.def (X86_TUNE_SSE_PARTIAL_REG_DEPENDENCY):
> Enable for Tremont.
> (X86_TUNE_USE_LEAVE): Likewise.
> (X86_TUNE_PUSH_MEMORY): Likewise.
> (X86_TUNE_MISALIGNED_MOVE_STRING_PRO_EPILOGUES): Likewise.
> (X86_TUNE_USE_CLTD): Likewise.
> (X86_TUNE_AVOID_FALSE_DEP_FOR_BMI): Likewise.
> (X86_TUNE_AVOID_MFENCE): Likewise.
> (X86_TUNE_SSE_TYPELESS_STORES): Likewise.
> (X86_TUNE_SSE_LOAD0_BY_PXOR): Likewise.
> (X86_TUNE_ACCUMULATE_OUTGOING_ARGS): Disable for Tremont.
> (X86_TUNE_FOUR_JUMP_LIMIT): Likewise.
> (X86_TUNE_OPT_AGU): Likewise.
> (X86_TUNE_AVOID_LEA_FOR_ADDR): Likewise.
> (X86_TUNE_AVOID_MEM_OPND_FOR_CMOVE): Likewise.
> (X86_TUNE_EXPAND_ABS): Likewise.
> (X86_TUNE_SPLIT_MEM_OPND_FOR_FP_CONVERTS): Likewise.
> (X86_TUNE_SLOW_PSHUFB): Likewise.
OK. (Tuning patches are kind of obvious).
Thanks,
Uros.
> ---
> gcc/common/config/i386/i386-common.c | 2 +-
> gcc/config/i386/i386.c | 1 +
> gcc/config/i386/x86-tune-sched.c | 2 ++
> gcc/config/i386/x86-tune.def | 37 ++++++++++++++--------------
> 4 files changed, 23 insertions(+), 19 deletions(-)
>
> diff --git a/gcc/common/config/i386/i386-common.c
> b/gcc/common/config/i386/i386-common.c
> index 00c65ba15ab..2c9e1ccbc6e 100644
> --- a/gcc/common/config/i386/i386-common.c
> +++ b/gcc/common/config/i386/i386-common.c
> @@ -1935,7 +1935,7 @@ const pta processor_alias_table[] =
> M_CPU_TYPE (INTEL_GOLDMONT), P_PROC_SSE4_2},
> {"goldmont-plus", PROCESSOR_GOLDMONT_PLUS, CPU_GLM, PTA_GOLDMONT_PLUS,
> M_CPU_TYPE (INTEL_GOLDMONT_PLUS), P_PROC_SSE4_2},
> - {"tremont", PROCESSOR_TREMONT, CPU_GLM, PTA_TREMONT,
> + {"tremont", PROCESSOR_TREMONT, CPU_HASWELL, PTA_TREMONT,
> M_CPU_TYPE (INTEL_TREMONT), P_PROC_SSE4_2},
> {"knl", PROCESSOR_KNL, CPU_SLM, PTA_KNL,
> M_CPU_TYPE (INTEL_KNL), P_PROC_AVX512F},
> diff --git a/gcc/config/i386/i386.c b/gcc/config/i386/i386.c
> index 7b173bc0beb..2927e2884c9 100644
> --- a/gcc/config/i386/i386.c
> +++ b/gcc/config/i386/i386.c
> @@ -16976,6 +16976,7 @@ ix86_sched_init_global (FILE *, int, int)
> case PROCESSOR_NEHALEM:
> case PROCESSOR_SANDYBRIDGE:
> case PROCESSOR_HASWELL:
> + case PROCESSOR_TREMONT:
> case PROCESSOR_GENERIC:
> /* Do not perform multipass scheduling for pre-reload schedule
> to save compile time. */
> diff --git a/gcc/config/i386/x86-tune-sched.c
> b/gcc/config/i386/x86-tune-sched.c
> index 2e5ee4e4444..56ada99a450 100644
> --- a/gcc/config/i386/x86-tune-sched.c
> +++ b/gcc/config/i386/x86-tune-sched.c
> @@ -71,6 +71,7 @@ ix86_issue_rate (void)
> case PROCESSOR_NEHALEM:
> case PROCESSOR_SANDYBRIDGE:
> case PROCESSOR_HASWELL:
> + case PROCESSOR_TREMONT:
> case PROCESSOR_GENERIC:
> return 4;
>
> @@ -429,6 +430,7 @@ ix86_adjust_cost (rtx_insn *insn, int dep_type, rtx_insn
> *dep_insn, int cost,
> case PROCESSOR_NEHALEM:
> case PROCESSOR_SANDYBRIDGE:
> case PROCESSOR_HASWELL:
> + case PROCESSOR_TREMONT:
> case PROCESSOR_GENERIC:
> /* Stack engine allows to execute push&pop instructions in parall. */
> if ((insn_type == TYPE_PUSH || insn_type == TYPE_POP)
> diff --git a/gcc/config/i386/x86-tune.def b/gcc/config/i386/x86-tune.def
> index 2f221b1f8c9..385e275bbd9 100644
> --- a/gcc/config/i386/x86-tune.def
> +++ b/gcc/config/i386/x86-tune.def
> @@ -62,7 +62,7 @@ DEF_TUNE (X86_TUNE_PARTIAL_REG_DEPENDENCY,
> "partial_reg_dependency",
> that can be partly masked by careful scheduling of moves. */
> DEF_TUNE (X86_TUNE_SSE_PARTIAL_REG_DEPENDENCY, "sse_partial_reg_dependency",
> m_PPRO | m_P4_NOCONA | m_CORE_ALL | m_BONNELL | m_AMDFAM10
> - | m_BDVER | m_ZNVER | m_GENERIC)
> + | m_BDVER | m_ZNVER | m_TREMONT | m_GENERIC)
>
> /* X86_TUNE_SSE_SPLIT_REGS: Set for machines where the type and dependencies
> are resolved on SSE register parts instead of whole registers, so we may
> @@ -136,7 +136,7 @@ DEF_TUNE (X86_TUNE_FUSE_ALU_AND_BRANCH,
> "fuse_alu_and_branch",
>
> DEF_TUNE (X86_TUNE_ACCUMULATE_OUTGOING_ARGS, "accumulate_outgoing_args",
> m_PPRO | m_P4_NOCONA | m_BONNELL | m_SILVERMONT | m_KNL | m_KNM |
> m_INTEL
> - | m_GOLDMONT | m_GOLDMONT_PLUS | m_TREMONT | m_ATHLON_K8)
> + | m_GOLDMONT | m_GOLDMONT_PLUS | m_ATHLON_K8)
>
> /* X86_TUNE_PROLOGUE_USING_MOVE: Do not use push/pop in prologues that are
> considered on critical path. */
> @@ -150,14 +150,15 @@ DEF_TUNE (X86_TUNE_EPILOGUE_USING_MOVE,
> "epilogue_using_move",
>
> /* X86_TUNE_USE_LEAVE: Use "leave" instruction in epilogues where it fits.
> */
> DEF_TUNE (X86_TUNE_USE_LEAVE, "use_leave",
> - m_386 | m_CORE_ALL | m_K6_GEODE | m_AMD_MULTIPLE | m_GENERIC)
> + m_386 | m_CORE_ALL | m_K6_GEODE | m_AMD_MULTIPLE | m_TREMONT
> + | m_GENERIC)
>
> /* X86_TUNE_PUSH_MEMORY: Enable generation of "push mem" instructions.
> Some chips, like 486 and Pentium works faster with separate load
> and push instructions. */
> DEF_TUNE (X86_TUNE_PUSH_MEMORY, "push_memory",
> m_386 | m_P4_NOCONA | m_CORE_ALL | m_K6_GEODE | m_AMD_MULTIPLE
> - | m_GENERIC)
> + | m_TREMONT | m_GENERIC)
>
> /* X86_TUNE_SINGLE_PUSH: Enable if single push insn is preferred
> over esp subtraction. */
> @@ -198,8 +199,7 @@ DEF_TUNE (X86_TUNE_PAD_RETURNS, "pad_returns",
> than 4 branch instructions in the 16 byte window. */
> DEF_TUNE (X86_TUNE_FOUR_JUMP_LIMIT, "four_jump_limit",
> m_PPRO | m_P4_NOCONA | m_BONNELL | m_SILVERMONT | m_KNL | m_KNM
> - | m_GOLDMONT | m_GOLDMONT_PLUS | m_TREMONT | m_INTEL | m_ATHLON_K8
> - | m_AMDFAM10)
> + | m_GOLDMONT | m_GOLDMONT_PLUS | m_INTEL | m_ATHLON_K8 | m_AMDFAM10)
>
>
> /*****************************************************************************/
> /* Integer instruction selection tuning
> */
> @@ -240,11 +240,11 @@ DEF_TUNE (X86_TUNE_INTEGER_DFMODE_MOVES,
> "integer_dfmode_moves",
> /* X86_TUNE_OPT_AGU: Optimize for Address Generation Unit. This flag
> will impact LEA instruction selection. */
> DEF_TUNE (X86_TUNE_OPT_AGU, "opt_agu", m_BONNELL | m_SILVERMONT | m_KNL
> - | m_KNM | m_GOLDMONT | m_GOLDMONT_PLUS | m_TREMONT | m_INTEL)
> + | m_KNM | m_GOLDMONT | m_GOLDMONT_PLUS | m_INTEL)
>
> /* X86_TUNE_AVOID_LEA_FOR_ADDR: Avoid lea for address computation. */
> DEF_TUNE (X86_TUNE_AVOID_LEA_FOR_ADDR, "avoid_lea_for_addr",
> - m_BONNELL | m_SILVERMONT | m_GOLDMONT | m_GOLDMONT_PLUS | m_TREMONT
> + m_BONNELL | m_SILVERMONT | m_GOLDMONT | m_GOLDMONT_PLUS
> | m_KNL | m_KNM)
>
> /* X86_TUNE_SLOW_IMUL_IMM32_MEM: Imul of 32-bit constant and memory is
> @@ -263,7 +263,7 @@ DEF_TUNE (X86_TUNE_SLOW_IMUL_IMM8, "slow_imul_imm8",
> a conditional move. */
> DEF_TUNE (X86_TUNE_AVOID_MEM_OPND_FOR_CMOVE, "avoid_mem_opnd_for_cmove",
> m_BONNELL | m_SILVERMONT | m_GOLDMONT | m_GOLDMONT_PLUS | m_KNL
> - | m_KNM | m_TREMONT | m_INTEL)
> + | m_KNM | m_INTEL)
>
> /* X86_TUNE_SINGLE_STRINGOP: Enable use of single string operations, such
> as MOVS and STOS (without a REP prefix) to move/set sequences of bytes.
> */
> @@ -282,7 +282,8 @@ DEF_TUNE (X86_TUNE_PREFER_KNOWN_REP_MOVSB_STOSB,
> FIXME: This may actualy be a win on more targets than listed here. */
> DEF_TUNE (X86_TUNE_MISALIGNED_MOVE_STRING_PRO_EPILOGUES,
> "misaligned_move_string_pro_epilogues",
> - m_386 | m_486 | m_CORE_ALL | m_AMD_MULTIPLE | m_GENERIC)
> + m_386 | m_486 | m_CORE_ALL | m_AMD_MULTIPLE | m_TREMONT
> + | m_GENERIC)
>
> /* X86_TUNE_USE_SAHF: Controls use of SAHF. */
> DEF_TUNE (X86_TUNE_USE_SAHF, "use_sahf",
> @@ -294,7 +295,7 @@ DEF_TUNE (X86_TUNE_USE_SAHF, "use_sahf",
> /* X86_TUNE_USE_CLTD: Controls use of CLTD and CTQO instructions. */
> DEF_TUNE (X86_TUNE_USE_CLTD, "use_cltd",
> ~(m_PENT | m_LAKEMONT | m_BONNELL | m_SILVERMONT | m_KNL | m_KNM |
> m_INTEL
> - | m_K6 | m_GOLDMONT | m_GOLDMONT_PLUS | m_TREMONT))
> + | m_K6 | m_GOLDMONT | m_GOLDMONT_PLUS))
>
> /* X86_TUNE_USE_BT: Enable use of BT (bit test) instructions. */
> DEF_TUNE (X86_TUNE_USE_BT, "use_bt",
> @@ -305,7 +306,7 @@ DEF_TUNE (X86_TUNE_USE_BT, "use_bt",
> /* X86_TUNE_AVOID_FALSE_DEP_FOR_BMI: Avoid false dependency
> for bit-manipulation instructions. */
> DEF_TUNE (X86_TUNE_AVOID_FALSE_DEP_FOR_BMI, "avoid_false_dep_for_bmi",
> - m_SANDYBRIDGE | m_CORE_AVX2 | m_GENERIC)
> + m_SANDYBRIDGE | m_CORE_AVX2 | m_TREMONT | m_GENERIC)
>
> /* X86_TUNE_ADJUST_UNROLL: This enables adjusting the unroll factor based
> on hardware capabilities. Bdver3 hardware has a loop buffer which makes
> @@ -321,14 +322,14 @@ DEF_TUNE (X86_TUNE_ONE_IF_CONV_INSN, "one_if_conv_insn",
>
> /* X86_TUNE_AVOID_MFENCE: Use lock prefixed instructions instead of mfence.
> */
> DEF_TUNE (X86_TUNE_AVOID_MFENCE, "avoid_mfence",
> - m_CORE_ALL | m_BDVER | m_ZNVER | m_GENERIC)
> + m_CORE_ALL | m_BDVER | m_ZNVER | m_TREMONT | m_GENERIC)
>
> /* X86_TUNE_EXPAND_ABS: This enables a new abs pattern by
> generating instructions for abs (x) = (((signed) x >> (W-1) ^ x) -
> (signed) x >> (W-1)) instead of cmove or SSE max/abs instructions. */
> DEF_TUNE (X86_TUNE_EXPAND_ABS, "expand_abs",
> m_CORE_ALL | m_SILVERMONT | m_KNL | m_KNM | m_GOLDMONT
> - | m_GOLDMONT_PLUS | m_TREMONT )
> + | m_GOLDMONT_PLUS)
>
>
> /*****************************************************************************/
> /* 387 instruction selection tuning
> */
> @@ -386,13 +387,13 @@ DEF_TUNE (X86_TUNE_SSE_PACKED_SINGLE_INSN_OPTIMAL,
> "sse_packed_single_insn_optim
>
> /* X86_TUNE_SSE_TYPELESS_STORES: Always movaps/movups for 128bit stores. */
> DEF_TUNE (X86_TUNE_SSE_TYPELESS_STORES, "sse_typeless_stores",
> - m_AMD_MULTIPLE | m_CORE_ALL | m_GENERIC)
> + m_AMD_MULTIPLE | m_CORE_ALL | m_TREMONT | m_GENERIC)
>
> /* X86_TUNE_SSE_LOAD0_BY_PXOR: Always use pxor to load0 as opposed to
> xorps/xorpd and other variants. */
> DEF_TUNE (X86_TUNE_SSE_LOAD0_BY_PXOR, "sse_load0_by_pxor",
> m_PPRO | m_P4_NOCONA | m_CORE_ALL | m_BDVER | m_BTVER | m_ZNVER
> - | m_GENERIC)
> + | m_TREMONT | m_GENERIC)
>
> /* X86_TUNE_INTER_UNIT_MOVES_TO_VEC: Enable moves in from integer
> to SSE registers. If disabled, the moves will be done by storing
> @@ -419,7 +420,7 @@ DEF_TUNE (X86_TUNE_INTER_UNIT_CONVERSIONS,
> "inter_unit_conversions",
> fp converts to destination register. */
> DEF_TUNE (X86_TUNE_SPLIT_MEM_OPND_FOR_FP_CONVERTS,
> "split_mem_opnd_for_fp_converts",
> m_SILVERMONT | m_KNL | m_KNM | m_GOLDMONT | m_GOLDMONT_PLUS
> - | m_TREMONT | m_INTEL)
> + | m_INTEL)
>
> /* X86_TUNE_USE_VECTOR_FP_CONVERTS: Prefer vector packed SSE conversion
> from FP to FP. This form of instructions avoids partial write to the
> @@ -434,7 +435,7 @@ DEF_TUNE (X86_TUNE_USE_VECTOR_CONVERTS,
> "use_vector_converts", m_AMDFAM10)
> /* X86_TUNE_SLOW_SHUFB: Indicates tunings with slow pshufb instruction. */
> DEF_TUNE (X86_TUNE_SLOW_PSHUFB, "slow_pshufb",
> m_BONNELL | m_SILVERMONT | m_KNL | m_KNM | m_GOLDMONT
> - | m_GOLDMONT_PLUS | m_TREMONT | m_INTEL)
> + | m_GOLDMONT_PLUS | m_INTEL)
>
> /* X86_TUNE_AVOID_4BYTE_PREFIXES: Avoid instructions requiring 4+ bytes of
> prefixes. */
> DEF_TUNE (X86_TUNE_AVOID_4BYTE_PREFIXES, "avoid_4byte_prefixes",
> --
> 2.17.1
>
