On 09/28/2016 02:48 AM, Kyrill Tkachov wrote:
Hi all,
This patch tries to avoid materialising an immediate
when comparison has shown that it is already loaded.
This is performed during ifcvt and only when the target has the
conditional negate
or inverse optabs, which for now is only aarch64.
Thus for the code:
int
foo (int a, int b)
{
return a == 5 ? -5 : b;
}
we currently emit on aarch64:
foo:
mov w2, -5
cmp w0, 5
csel w0, w1, w2, ne
ret
but with this patch we emit:
foo:
cmp w0, 5
csneg w0, w1, w0, ne
ret
Bootstrapped and tested on arm, aarch64, x86_64.
Ok for trunk?
Thanks,
Kyrill
P.S. The simpler form of the transformation: X == CST ? CST : Y --> X ==
CST ? X : Y
could also be beneficial IMO but I found that it can cause bad codegen
in combine
due to extending the live range of X. I'm still working on a way to work
around that,
but that is a separate piece of work anyway.
FWIW, the simpler form of the transformation is already done just prior
to leaving SSA form in tree-ssa-uncprop.c. So there may not be a ton of
opportunities for the simpler form in the RTL optimizers.
My only concern here is this transformation isn't significantly
different than the simpler form, which is apparently causing some kind
of combine problem.
I'd like to understand the combine problem better before giving final
approval to this patch.
Jeff
