https://gcc.gnu.org/bugzilla/show_bug.cgi?id=103771
Richard Biener <rguenth at gcc dot gnu.org> changed:
What |Removed |Added
----------------------------------------------------------------------------
Ever confirmed|0 |1
Last reconfirmed| |2022-01-13
CC| |rguenth at gcc dot gnu.org,
| |rsandifo at gcc dot gnu.org
Status|UNCONFIRMED |NEW
--- Comment #6 from Richard Biener <rguenth at gcc dot gnu.org> ---
(In reply to Tamar Christina from comment #1)
> Looks like the change causes the simpler conditional to be detected by the
> vectorizer as a masked operation, which in principle makes sense:
>
> note: vect_recog_mask_conversion_pattern: detected: iftmp.0_21 = x.1_14 >
> 255 ? iftmp.0_19 : iftmp.0_20;
> note: mask_conversion pattern recognized: patt_43 = patt_42 ? iftmp.0_19 :
> iftmp.0_20;
> note: extra pattern stmt: patt_40 = x.1_14 > 255;
> note: extra pattern stmt: patt_42 = (<signed-boolean:8>) patt_40;
if we look at the ifcvt result we see
iftmp.0_19 = (unsigned char) _18;
iftmp.0_20 = (unsigned char) _5;
iftmp.0_21 = x.1_14 > 255 ? iftmp.0_19 : iftmp.0_20;
*_6 = iftmp.0_21;
I suspect we intended to carry the fact that x.1_14 > 255 will produce
a mask from a SImode vector element compare but we need a mask suitable
for a QImode vector element select (iftmp.0_19 and iftmp.0_20). That's
not something we can express in scalar code and thus a pattern I think.
So the pattern as generated doesn't make very much sense to me. I suppose
it might try to convert a AVX512 mask to a AVX2 style mask but that needs
to be done with
patt_42 = patt_40 ? (<signed-boolean:8>)-1 : 0;
not with a conversion. But it's also somewhat pointless since it will
simply cause the same issue again - the COND_EXPR vectorization will need
to cobble up 4 AVX512 masks to produce the desired result.
