https://gcc.gnu.org/bugzilla/show_bug.cgi?id=109892
Richard Biener <rguenth at gcc dot gnu.org> changed:
What |Removed |Added
----------------------------------------------------------------------------
Assignee|unassigned at gcc dot gnu.org |rguenth at gcc dot
gnu.org
Status|NEW |ASSIGNED
Summary|SLP failure with explicit |SLP reduction failure with
|fma |explicit fma
--- Comment #4 from Richard Biener <rguenth at gcc dot gnu.org> ---
Vectorization of the following works with -O2 -mfma
void f(double x[restrict], double *y, double *z)
{
x[0] = __builtin_fma (x[0], y[0], z[0]);
x[1] = __builtin_fma (x[1], y[1], z[1]);
}
so SLP of fma() works. The issue is that the FMA is not supported as
reduction operation:
t3.c:4:12: note: vect_is_simple_use: operand r1_23 = PHI <r1_12(7), 0.0(6)>,
type of def: reduction
t3.c:4:12: missed: no reduc code for scalar code.
Same for g() with -ffp-contract=on (and .FMA):
t3.c:8:12: note: vect_is_simple_use: operand r1_23 = PHI <_12(7), 0.0(6)>,
type of def: reduction
t3.c:8:12: missed: no reduc code for scalar code.
t3.c:5:8: missed: not vectorized: relevant stmt not supported: r1_23 = PHI
<_12(7), 0.0(6)>
t3.c:8:12: note: unsupported SLP instance starting from: _12 = .FMA (_2, _2,
r1_23);
We currently only support .FMAX and .FMIN, I have a pach to add .FMA.
