[Mesa-dev] MR: NIR: Partial redundancy elimination for compares

[Ian Romanick]

I was really hoping to avoid having review discussion for a patch series
on both the mailing list and the MR, but I guess that ship has sailed.

On 12/17/18 2:39 PM, Roland Scheidegger wrote:
> Am 17.12.18 um 23:27 schrieb Roland Scheidegger:
>> Am 17.12.18 um 23:07 schrieb Ilia Mirkin:
>>> On Mon, Dec 17, 2018 at 5:05 PM Ian Romanick <idr at freedesktop.org> wrote:
>>>>
>>>> https://na01.safelinks.protection.outlook.com/?url=https%3A%2F%2Fgitlab.freedesktop.org%2Fmesa%2Fmesa%2Fmerge_requests%2F22&data=02%7C01%7Csroland%40vmware.com%7C5773f37aa397417e6beb08d6646c24c6%7Cb39138ca3cee4b4aa4d6cd83d9dd62f0%7C0%7C0%7C636806812986635916&sdata=z2ev4sVJEW8Kw2hsoNWfYHh6FkSwSy%2B5CdTItxoh%2FPE%3D&reserved=0
>>>>
>>>> This series adds a new optimization pass that tries to replace code
>>>> sequences like
>>>>
>>>> if (x < y) {
>>>>     z = y - x;
>>>>     ...
>>>> }
>>>>
>>>> with a sequence like
>>>>
>>>> t = x - y;
>>>> if (t < 0) {
>>>>     z = -t;
>>>>     ...
>>>> }
>>>
>>> Is it worth worrying about infinities? e.g. if x = -Infinity, y =
>>> Infinity, "x < y" will be true, but "x - y < 0" will not be (pretty
>>> sure it'll be a NaN, which is not < 0).
>>
>> I was wondering the same, but I think this should still work.
>> -Inf - Inf = -Inf, so no problem there.
> 
> Although it looks like the optimization might be a bit problematic? For
> it to work you really need not just be able to use flags generated by
> sub, but also you need to be able to eliminate the negation one way or
> another (e.g. free input negate going into another operation, turning
> subsequent adds into subs, ...). But well maybe that's often possible.

The optimization handles various combinations of 'x < y', 'x > y', 'x <=
y', and 'x >= y' combined with various combinations of 'x - y' and 'y -
x'.  Not all of the cases require a negation in the replacement.

Right now this is only enabled in the i965 driver, and negation is free
on that architecture.  If this is enabled on other architectures,
someone might want to add a flag that limits the kind of replacements to
avoid the negation... even that may be overly conservative.  In the
example case, it's possible that the only use of z is in an expression
like 'w - z'.

> Roland