On Mon, Feb 17, 2020 at 09:01:13AM +0100, Richard Biener wrote:
> On Mon, 17 Feb 2020, Li Jia He wrote:
> > This patch wants to fix PR66552 on gimple and optimizes (x shift (n mod C))
> > to (x shift (n bit_and (C - 1))) when C is a constant and power of two as
> > discussed in PR66552.
> >
> > The regression testing for the patch was done on GCC mainline on
> >
> > powerpc64le-unknown-linux-gnu (Power 9 LE)
> >
> > with no regressions. Is it OK for GCC11 ?
>
> I fail to see the connection with a shift operation, can you explain?
As mentioned in the PR, the reason why we can only optimize unsigned
mod C into bit_and C-1 (where C is pow2p) is that signed modulo can be negative
for
negative values (is non-positive). For shifts negative values would be UB
and so we can assume it will not be negative (i.e. x will be either positive
or a negative multiple of C) and can use bit_and then.
Jakub
