https://gcc.gnu.org/bugzilla/show_bug.cgi?id=81904
Bug ID: 81904
Summary: FMA and addsub instructions
Product: gcc
Version: 8.0
Status: UNCONFIRMED
Keywords: missed-optimization
Severity: enhancement
Priority: P3
Component: target
Assignee: unassigned at gcc dot gnu.org
Reporter: glisse at gcc dot gnu.org
Target Milestone: ---
Target: x86_64-*-*
(asked in
https://stackoverflow.com/questions/45298855/how-to-write-portable-simd-code-for-complex-multiplicative-reduction/45401182#comment77780455_45401182
)
Intel has instructions like vfmaddsubps. Gcc manages, under certain
circumstances, to merge mult and plus or mult and minus into FMA, but not mult
and this strange addsub mix.
#include <x86intrin.h>
__m128d f(__m128d x, __m128d y, __m128d z){
return _mm_addsub_pd(_mm_mul_pd(x,y),z);
}
__m128d g(__m128d x, __m128d y, __m128d z){
return _mm_fmaddsub_pd(x,y,z);
}
(the order of the arguments is probably not right)
My first guess as to how this could be implemented without too much trouble is
in ix86_gimple_fold_builtin: for IX86_BUILTIN_ADDSUBPD and others, check that
we are late enough in the optimization pipeline (roughly where "widening_mul"
is), that contractions are enabled, and that the first (?) argument is a
single-use MULT_EXPR.
I didn't check what the situation is with the vectorizer (which IIRC can now
generate code that ends up as addsub).
