[Mesa-dev] [PATCH 5/5] nir: Propagate negates up multiplication chains.

[Ian Romanick]

On 02/22/2016 04:13 PM, Matt Turner wrote:
> total instructions in shared programs: 7127270 -> 7103195 (-0.34%)
> instructions in affected programs: 1376832 -> 1352757 (-1.75%)
> helped: 7394
> HURT: 622
> 
> GAINED: 4
> LOST:   2
> ---
>  src/compiler/nir/nir_opt_algebraic.py | 4 ++++
>  1 file changed, 4 insertions(+)
> 
> diff --git a/src/compiler/nir/nir_opt_algebraic.py b/src/compiler/nir/nir_opt_algebraic.py
> index cc2c229..1863033 100644
> --- a/src/compiler/nir/nir_opt_algebraic.py
> +++ b/src/compiler/nir/nir_opt_algebraic.py
> @@ -227,6 +227,10 @@ optimizations = [
>     (('fabs', ('fsub', 0.0, a)), ('fabs', a)),
>     (('iabs', ('isub', 0, a)), ('iabs', a)),
>  
> +   # Propagate negation up multiplication chains
> +   (('fmul', ('fneg', a), b), ('fneg', ('fmul', a, b))),
> +   (('imul', ('ineg', a), b), ('ineg', ('fmul', a, b))),
> +

We had tried things like this before, and it wasn't clearly a win.  Is
it a win now because of your early change to try to match a larger
expression?  My recollection is that doing this thwarted some attempts
at generating MAD.

Should we also add rules like

   (('ffma', ('fneg', a), b, c), ('fadd', ('fneg', ('fmul', a, b)), c),
'options->lower_ffma'),

Finally, have you looked at the hurt cases?  I also seem to recall that
cases where one of the operands of the fmul was an immediate were
usually better off moving the negation from the

>     # Misc. lowering
>     (('fmod', a, b), ('fsub', a, ('fmul', b, ('ffloor', ('fdiv', a, b)))), 'options->lower_fmod'),
>     (('uadd_carry', a, b), ('b2i', ('ult', ('iadd', a, b), a)), 'options->lower_uadd_carry'),
>