[Mesa-dev] [PATCH v2 49/53] intel/compiler: fix cmod propagation for non 32-bit types

[Iago Toral]

On Wed, 2019-01-02 at 12:42 +0200, Pohjolainen, Topi wrote:
> On Wed, Dec 19, 2018 at 12:51:17PM +0100, Iago Toral Quiroga wrote:
> > ---
> >  src/intel/compiler/brw_fs_cmod_propagation.cpp | 8 +++-----
> >  1 file changed, 3 insertions(+), 5 deletions(-)
> > 
> > diff --git a/src/intel/compiler/brw_fs_cmod_propagation.cpp
> > b/src/intel/compiler/brw_fs_cmod_propagation.cpp
> > index 7bb5c9afbc9..dfef9d720a2 100644
> > --- a/src/intel/compiler/brw_fs_cmod_propagation.cpp
> > +++ b/src/intel/compiler/brw_fs_cmod_propagation.cpp
> > @@ -244,8 +244,7 @@ opt_cmod_propagation_local(const
> > gen_device_info *devinfo,
> >              /* CMP's result is the same regardless of dest type.
> > */
> >              if (inst->conditional_mod == BRW_CONDITIONAL_NZ &&
> >                  scan_inst->opcode == BRW_OPCODE_CMP &&
> > -                (inst->dst.type == BRW_REGISTER_TYPE_D ||
> > -                 inst->dst.type == BRW_REGISTER_TYPE_UD)) {
> > +                brw_reg_type_is_integer(inst->dst.type)) {
> >                 inst->remove(block);
> >                 progress = true;
> >                 break;
> > @@ -258,9 +257,8 @@ opt_cmod_propagation_local(const
> > gen_device_info *devinfo,
> >                 break;
> >  
> >              /* Comparisons operate differently for ints and floats
> > */
> > -            if (scan_inst->dst.type != inst->dst.type &&
> 
> This wouldn't let, for example, (DF, F) pair thru while the new
> version does.
> Should we keep this line?

This is about not turning a floating point comparison into an integer
comparison and viceversa, but F and DF are both floating point and
follow the same rules, so I think this sould be okay in that aspect.

With that being said though, I wonder if we shold prevent propagation
when the types don't have the same bit-size, since even if the
underlying comparison semantics are the same, we could have different
behaviors for out-of-range values... I wonder if that can actually ever
happen though, since that would mean that scan_inst writes, for example
a DF, and then the CMP reads it as F... but anyway, I guess being extra
careful here doesn't hurt.

> > -                (scan_inst->dst.type == BRW_REGISTER_TYPE_F ||
> > -                 inst->dst.type == BRW_REGISTER_TYPE_F))
> > +            if (brw_reg_type_is_floating_point(scan_inst-
> > >dst.type) !=
> > +                brw_reg_type_is_floating_point(inst->dst.type))
> >                 break;
> >  
> >              /* If the instruction generating inst's source also
> > wrote the
> > -- 
> > 2.17.1
> > 
> > _______________________________________________
> > mesa-dev mailing list
> > mesa-dev at lists.freedesktop.org
> > https://lists.freedesktop.org/mailman/listinfo/mesa-dev
> 
>