[Mesa-dev] [PATCH 6/6] i965/fs: don't copy propagate if the instruction writes to more than two adjacent GRFs

[Francisco Jerez]

Samuel Iglesias Gonsálvez <siglesias at igalia.com> writes:

> This is not allowed by the HW and copy propagation can hide this issue to
> lower_simd_width pass, which is going to fix it.
>
> Signed-off-by: Samuel Iglesias Gonsálvez <siglesias at igalia.com>
> ---
>  src/mesa/drivers/dri/i965/brw_fs_copy_propagation.cpp | 1 +
>  1 file changed, 1 insertion(+)
>
> diff --git a/src/mesa/drivers/dri/i965/brw_fs_copy_propagation.cpp b/src/mesa/drivers/dri/i965/brw_fs_copy_propagation.cpp
> index 438f681..c7f7628 100644
> --- a/src/mesa/drivers/dri/i965/brw_fs_copy_propagation.cpp
> +++ b/src/mesa/drivers/dri/i965/brw_fs_copy_propagation.cpp
> @@ -752,6 +752,7 @@ can_propagate_from(fs_inst *inst)
>              inst->src[0].file == UNIFORM ||
>              inst->src[0].file == IMM) &&
>             inst->src[0].type == inst->dst.type &&
> +           inst->regs_written <= 2 &&

This doesn't look right to me, why should copy propagation care whether
the SIMD width of a MOV instruction it's going to propagate away is
allowed by the hardware or not?  The "illegal" copy instruction is
already there anyway, and preventing copy propagation from doing its job
in that case can only increase the likelihood that the unsupported
instruction will remain in the program which implies more work for the
SIMD lowering pass at a later point.

>             !inst->is_partial_write());
>  }
>  
> -- 
> 2.7.4
>
> _______________________________________________
> mesa-dev mailing list
> mesa-dev at lists.freedesktop.org
> https://lists.freedesktop.org/mailman/listinfo/mesa-dev
-------------- next part --------------
A non-text attachment was scrubbed...
Name: signature.asc
Type: application/pgp-signature
Size: 212 bytes
Desc: not available
URL: <https://lists.freedesktop.org/archives/mesa-dev/attachments/20160707/ba91c025/attachment.sig>