https://gcc.gnu.org/bugzilla/show_bug.cgi?id=126696
Bug ID: 126696
Summary: Fold a fused multiply-add by one into an addition
Product: gcc
Version: 17.0
Status: UNCONFIRMED
Keywords: easyhack, missed-optimization
Severity: normal
Priority: P3
Component: tree-optimization
Assignee: unassigned at gcc dot gnu.org
Reporter: ktkachov at gcc dot gnu.org
Target Milestone: ---
Multiplying by one is exact, so the single rounding of the fused
operation is the rounding of the sum, and the same holds for minus one
and a subtraction. Dropping the fma also drops the materialisation of
the constant.
When both operands are NaN the result is a NaN either way, but which one
is returned can change. IEEE 754 leaves that unspecified, and GCC
already folds x * 1.0 into x on the same basis.
double f (double x, double y) { return __builtin_fma (x, 1.0, y); }
Can generate just:
f:
fadd d0, d0, d1
ret
which Clang already does. GCC does:
f:
fmov d31, 1.0e+0
fmadd d0, d0, d31, d1
ret
