https://gcc.gnu.org/bugzilla/show_bug.cgi?id=116979
Richard Biener <rguenth at gcc dot gnu.org> changed:
What |Removed |Added
----------------------------------------------------------------------------
Keywords| |deferred
--- Comment #23 from Richard Biener <rguenth at gcc dot gnu.org> ---
So in particular, vect_bb_slp_scalar_cost only ever considers scalar stmts
recorded in SLP_TREE_SCALAR_STMTS which for the case of the FMADDSUB
SLP pattern does not get to see the multiplication. There's a similar
issue for scalar patterns where we only consider the scalar pattern root
but not all original covered stmts (or alternatively all stmts covered
in the pattern def sequence).
A reasonable fix would look at this (scalar stmt coverage) and the problem
of determining out-of-SLP-subgraph uses (live lanes) at the same time since
both are very much related. Which means a rewrite - I don't see any easy
workaround suitable for stage4 or even backporting.
