On Sun, Nov 22, 2020 at 11:57 AM Luc Van Oostenryck
<luc.vanoostenryck@xxxxxxxxx> wrote:
>
> In the case here with your example, the if-conversion doesn't happen
> because the phi-sources is not defined in the top block because of
> the OR:
Ahh. Yes, it's more obvious with a more realistic test-case that
actually translates one set of bits to another set of bits (which is
something we do in the kernel for various reasons - different bit
"namespaces", for example user interfaces etc):
#define BIT1 4
#define BIT1x 16
#define BIT2 8
#define BIT2x 32
int translate_bits(int x)
{
int y = 0;
if (x & BIT1)
y |= BIT1x;
if (x & BIT2)
y |= BIT2x;
return y;
}
and the first one gets nicely translated as
and.32 %r2 <- %arg1, $4
select.32 %r14(y) <- %r2, $16, $0
but then the second one doesn't for the reason you mention.
Honestly, particularly in the conditional form, the OP_SEL
optimization might not even be the right thing. It adds register
pressure.
So maybe a better model would be to not try to do jump-conversion, but
have some kind of general "can we simplify phi nodes", where jump
conversion to OP_SEL is just one of the options.
Linus
