On Wed, 2018-10-31 at 13:22 +1100, Timothy Arceri wrote:
> On 31/10/18 1:23 am, Jason Ekstrand wrote:
> > Weird. I didn't expect this patch to have any impact whatsoever.
> > I
> > thought it was just moving around the way we emit stuff.
>
> I think I've spotted the problem. Iago does patch 1
On 31/10/18 1:23 am, Jason Ekstrand wrote:
Weird. I didn't expect this patch to have any impact whatsoever. I
thought it was just moving around the way we emit stuff.
I think I've spotted the problem. Iago does patch 1 help with the
regressions you are seeing.
Weird. I didn't expect this patch to have any impact whatsoever. I thought
it was just moving around the way we emit stuff.
On October 30, 2018 08:40:01 Iago Toral wrote:
Jason, JFYI, I have been looking into the cases where the boolean
bitsize lowering pass was producing worse instruction
Jason, JFYI, I have been looking into the cases where the boolean
bitsize lowering pass was producing worse instruction counts that the
default 32-bit pass and I have tracked it down to this patch. Reverting
this makes the instruction count much better for some tests, I'll check
why this happens
On 22/10/18 23:13, Jason Ekstrand wrote:
> Instead of doing our own constant folding, we just emit instructions and
> let constant folding happen. This is substantially simpler and lets us
> use the nir_imm_bool helper instead of dealing with the const_value's
> ourselves.
> ---
>
Looks ok to me.
Reviewed-by: Timothy Arceri
On 23/10/18 9:13 am, Jason Ekstrand wrote:
Instead of doing our own constant folding, we just emit instructions and
let constant folding happen. This is substantially simpler and lets us
use the nir_imm_bool helper instead of dealing with the
Instead of doing our own constant folding, we just emit instructions and
let constant folding happen. This is substantially simpler and lets us
use the nir_imm_bool helper instead of dealing with the const_value's
ourselves.
---
src/compiler/nir/nir_opt_if.c | 91