https://gcc.gnu.org/bugzilla/show_bug.cgi?id=85406
Bug ID: 85406
Summary: Unnecessary blend when vectorizing short-cutted
calculations
Product: gcc
Version: unknown
Status: UNCONFIRMED
Severity: normal
Priority: P3
Component: tree-optimization
Assignee: unassigned at gcc dot gnu.org
Reporter: linux at carewolf dot com
Target Milestone: ---
If you have something like this:
inline unsigned qPremultiply(unsigned x)
{
const unsigned a = x >> 24;
if (a == 255)
return x;
unsigned t = (x & 0xff00ff) * a;
t = (t + ((t >> 8) & 0xff00ff) + 0x800080) >> 8;
t &= 0xff00ff;
x = ((x >> 8) & 0xff) * a;
x = (x + ((x >> 8) & 0xff) + 0x80);
x &= 0xff00;
return x | t | (a << 24);
}
Gcc will vectorize it so that the longer calculation is always performed and
with an added blend in the end to merge the two different return values. This
is however unnecessary as the calculation will give the same result, and thus
the blend can be saved.
Also in any case it is actually a bit unsafe to vectorize as the performance
difference between the two branches is substantial, and it happens that in this
case the short-cut is likely to be valid most of the time, so a nonvectorized
loop might be faster than a vectorized one by doing a lot less.
The latter can be fixed, if the short-cut was also vectorized, for instance
making the test for 4 values at a time and skip the long route if none of them
need it.
