On Tue, Apr 28, 2020 at 11:16:13AM +0200, Richard Biener wrote:
> I think we generally want to use :s, but OTOH even for your new case,
> since you restrict to nop-conversions, :s isn't strictly needed since
> VN should know how to materialize nop-converted variants. I think
> it's different for the 2nd case you add since there we deal with
> bitops in different sign.
>
> So maybe just remove the single_use () calls?
I've tried doing that, but that regresses e.g.
void bar (unsigned int, unsigned int);
unsigned int
foo (int x, int y)
{
unsigned int x2 = x;
unsigned int y2 = y;
bar (x2, y2);
return x2 | y2;
}
@@ -6,12 +6,14 @@ foo (int x, int y)
unsigned int y2;
unsigned int x2;
unsigned int _7;
+ int _8;
<bb 2> [local count: 1073741824]:
x2_2 = (unsigned int) x_1(D);
y2_4 = (unsigned int) y_3(D);
bar (x2_2, y2_4);
- _7 = x2_2 | y2_4;
+ _8 = x_1(D) | y_3(D);
+ _7 = (unsigned int) _8;
return _7;
}
As the patch was meant purely as an IL size optimization, that is not the
case here.
> > Or should I e.g. move this GIMPLE rule as a separate simplify (under the
> > same for) and then use :s in there?
>
> That wouldn't work, if a structural pattern matched others are not
> considered so it wouldn't ever be used. genmatch even warns for this
> I think.
Ah, ok.
Jakub
