On Thu, Jul 03, 2025 at 08:12:22PM +0200, Richard Biener wrote:
> > So, instead at least for now, the following patch keeps doing the
> > optimization, just doesn't perform it in pointer arithmetics.
> > pointer_int_sum itself actually adds the multiplication by size_exp,
> > so ptr + expr is turned into ptr p+ expr * size_exp,
> > so this patch will try to optimize
> > ptr + (expr +- cst)
> > into
> > ptr p+ ((sizetype)expr * size_exp +- (sizetype)cst * size_exp)
> > and
> > ptr - (expr +- cst)
> > into
> > ptr p+ -((sizetype)expr * size_exp +- (sizetype)cst * size_exp)
>
> So the important part is the distribution of the multiplication, not the
> (invalid) association?
For the regressions yes.
Consider even trivial cases like
int a[1024];
int *
foo (int x)
{
return a + (x - 1);
}
Vanilla trunk as well as trunk with this patch compile this on x86_64 -O2
to
movslq %edi, %rdi
leaq a-4(,%rdi,4), %rax
ret
With just the c-common.cc opt removed, I get
subl $1, %edi
movslq %edi, %rdi
leaq a(,%rdi,4), %rax
ret
instead. I admit I haven't checked other targets, just ia32 and there is
no change whatsoever (expectedly, sizetype is 32-bit, so RTL opts can fold
it into leal just fine. But my hope is that
this patch keeps similar code like before more often, even when clearly it
will not be always the case.
E.g. for the (x ? a : b)[y - 1] case when a and b are arrays,
we used to do (x ? &a - 4 : &b - 4) p+ y * 4 so the subtraction done
on each of the branches, now it will be just added later on. But at least
on sizetype and not on the narrower type.
Jakub
