On Thu, Apr 24, 2025 at 12:50 AM Jan Hubicka <hubi...@ucw.cz> wrote:
>
> > In some benchmark, I notice stv failed due to cost unprofitable, but the
> > igain
> > is inside the loop, but sse<->integer conversion is outside the loop,
> > current cost
> > model doesn't consider the frequency of those gain/cost.
> > The patch weights those cost with frequency just like LRA does.
> >
> > Bootstrapped and regtested on x86_64-pc-linux-gnu{-m32,}.
> > Ok for GCC16?
> >
> > gcc/ChangeLog:
> >
> > * config/i386/i386-features.cc (scalar_chain::mark_dual_mode_def):
> > (general_scalar_chain::compute_convert_gain):
> > ---
> > gcc/config/i386/i386-features.cc | 9 +++++++--
> > 1 file changed, 7 insertions(+), 2 deletions(-)
> >
> > diff --git a/gcc/config/i386/i386-features.cc
> > b/gcc/config/i386/i386-features.cc
> > index c35ac24fd8a..ae0844a70c2 100644
> > --- a/gcc/config/i386/i386-features.cc
> > +++ b/gcc/config/i386/i386-features.cc
> > @@ -337,18 +337,20 @@ scalar_chain::mark_dual_mode_def (df_ref def)
> > /* Record the def/insn pair so we can later efficiently iterate over
> > the defs to convert on insns not in the chain. */
> > bool reg_new = bitmap_set_bit (defs_conv, DF_REF_REGNO (def));
> > + unsigned frequency
> > + = REG_FREQ_FROM_BB (BLOCK_FOR_INSN (DF_REF_INSN (def)));
>
> I am generally trying to get rid of remaing uses of REG_FREQ since the
> 10000 based fixed point arithmetics iot always working that well.
>
> You can do the sums in profile_count type (doing something reasonable
> when count is uninitialized) and then convert it to sreal for the final
> heuristics.
Thanks for the suggestion, let me try.
>
> Typically such code also wants skip scaling by count when optimizing for
> size (since in this case we want to count statically). Not sure how
> important it is for vector code but I suppose it can happen.
>
> Honza
> > if (!bitmap_bit_p (insns, DF_REF_INSN_UID (def)))
> > {
> > if (!bitmap_set_bit (insns_conv, DF_REF_INSN_UID (def))
> > && !reg_new)
> > return;
> > - n_integer_to_sse++;
> > + n_integer_to_sse += frequency;
> > }
> > else
> > {
> > if (!reg_new)
> > return;
> > - n_sse_to_integer++;
> > + n_sse_to_integer += frequency;
> > }
> >
> > if (dump_file)
> > @@ -556,6 +558,8 @@ general_scalar_chain::compute_convert_gain ()
> > rtx src = SET_SRC (def_set);
> > rtx dst = SET_DEST (def_set);
> > int igain = 0;
> > + unsigned frequency
> > + = REG_FREQ_FROM_BB (BLOCK_FOR_INSN (insn));
> >
> > if (REG_P (src) && REG_P (dst))
> > igain += 2 * m - ix86_cost->xmm_move;
> > @@ -755,6 +759,7 @@ general_scalar_chain::compute_convert_gain ()
> > }
> > }
> >
> > + igain *= frequency;
> > if (igain != 0 && dump_file)
> > {
> > fprintf (dump_file, " Instruction gain %d for ", igain);
> > --
> > 2.34.1
> >
--
BR,
Hongtao
