https://gcc.gnu.org/bugzilla/show_bug.cgi?id=68892
--- Comment #7 from Richard Biener <rguenth at gcc dot gnu.org> ---
(In reply to Bill Seurer from comment #4)
> That fix broke the test case pr60203.c.
>
> PASS: gcc.target/powerpc/pr60203.c (test for excess errors)
> PASS: gcc.target/powerpc/pr60203.c scan-assembler-not stfd
> FAIL: gcc.target/powerpc/pr60203.c scan-assembler-not lfd
> PASS: gcc.target/powerpc/pr60203.c scan-assembler-not lxsdx
> PASS: gcc.target/powerpc/pr60203.c scan-assembler-not stxsdx
> PASS: gcc.target/powerpc/pr60203.c scan-assembler-not mfvsrd
> PASS: gcc.target/powerpc/pr60203.c scan-assembler-not mtvsrd
>
> There are two lfds and lots of other ops in the generated code with the fix
>
> pack:
> xxpermdi 0,2,1,0
> addi 9,1,-16
> xxpermdi 0,0,0,2
> stxvd2x 0,0,9
> lfd 1,-16(1)
> lfd 2,-8(1)
> blr
>
> and none without
>
> pack:
> blr
>
> Compiled with -mcpu=power8 -O3.
So this is very likely vectorization of
long double
pack (double a, double aa)
{
union u_ld u;
u.d[0] = a;
u.d[1] = aa;
return u.ld;
}
being no longer profitable. I'll investigate why, this shouldn't have changed
at all by the rev.
I get
# of expected passes 7
with a cross btw.
pack:
blr
it's already optimized at GIMPLE level to
pack (double a, double aa)
{
<bb 2>:
return a_2(D);
}
even before vectorization. So certainly unrelated to this change.
