Re: [PATCH][GCC] Simplification of 1U << (31 - x)

[Wilco Dijkstra <Wilco.Dijkstra@arm.com>]

>On Wed, Apr 12, 2017 at 09:29:55AM +0000, Sudi Das wrote:
> > Hi all
> > 
> > This is a fix for PR 80131 
> > Currently the code A << (B - C) is not simplified.
>> However at least a more specific case of 1U << (C -x) where C = precision(type) - 1 can be simplified to (1 << C) >> x.
>
> Is that always a win though?

Yes assuming left and right shift have the same performance.

> Some constants have higher costs than others on various targets, some
> significantly higher.  This change might be beneficial only
> if if C is as expensive as 1, then you get rid of a one (typically cheap)
> operation.

Most targets can create the constant cheaply. Targets that can't would need 2
instructions (move+shift) or a literal load. That's not worse compared to the
original sequence (3 operations). The constant can be shared or lifted out of
a loop, so we're saving 1 subtract per use of the sequence.

> Which makes me wonder whether this should be done at GIMPLE time and not
> at RTL time (expansion or combine etc.) when one can compare the RTX costs.
> Or do this at match.pd as canonicalization and then have RTL transformation
> to rewrite such (1U << C) >> X as 1U << (C - X) if the latter is faster (or
> shorter).

I can't see why that would be useful for just this pattern. There are lots more useful
cases where GCC should simplify immediates to fit the target but it doesn't currently.
For example it changes x < 0x100000 to x <= 0xfffff which is worse on many targets.

Wilco