On Tue, 29 Jul 2025, Richard Biener wrote:
> On Tue, 29 Jul 2025, Robin Dapp wrote:
>
> > > - if (len >= 3
> > > + if (!reassoc_insert_powi_p
> > > + && len >= 3
> > > && (!has_fma
> > > /* width > 1 means ranking ops results in better
> > > parallelism. Check current value to avoid
> >
> > Last time I tried this I saw x86 regressions in bswap test cases (hence we
> > decided not to go forward with it). I'm assuming they didn't show up in
> > your
> > regtest but I wonder what's different now.
> >
> > > FAIL: gcc.target/i386/pr115102.c scan-assembler bswaphisi2_lowpart
> > > FAIL: gcc.target/i386/xchg-4.c scan-assembler rolw
> > > FAIL: gcc.target/i386/xchg-4.c scan-assembler-not mov
>
> They did, but somehow I thought they must come from an unrelated
> change. They could be fixed by allowing == 3 ops re-assocs but
> in the end improving rotate detection in forwprop and bswap
> in the bswap pass would be better. It's also that == 3 will
> still break reduction chains, just for 2 elements:
>
> double
> frd (double *p, double *lastone)
> {
> double sum = 0;
> for (; p <= lastone; p += 2)
> sum += p[0] + p[1];
> return sum;
> }
>
> I'll add a testcase for that case as well.
I have opened PR121298.
Richard.
