| Message ID | 20141114173347.GB5026@tucnak.redhat.com |
|---|---|
| State | New |
| Headers | show |
On 11/14/14 10:33, Jakub Jelinek wrote: > Hi! > > When looking at PR59984, I've noticed that in certain cases we leave > around GOMP_SIMD_LANE calls without lhs. > This internal call is intentionally not ECF_CONST, we don't want it moved > before the loop, and the argument is magic. But, if nothing uses it > anymore, it isn't needed. > > Fixed thusly, bootstrapped/regtested on x86_64-linux and i686-linux, ok for > trunk? > > 2014-11-14 Jakub Jelinek <jakub@redhat.com> > > * tree-ssa.dce.c (eliminate_unnecessary_stmts): Eliminate > IFN_GOMP_SIMD_LANE without lhs as useless. OK. I presume we don't have other functions with similar characteristics? If we did, then we might want a more general interface or some way to tag them rather than checking for them by name in DCE. jeff
--- gcc/tree-ssa-dce.c.jj 2014-11-12 13:28:47.000000000 +0100 +++ gcc/tree-ssa-dce.c 2014-11-14 16:15:10.356124955 +0100 @@ -1402,6 +1402,11 @@ eliminate_unnecessary_stmts (void) maybe_clean_or_replace_eh_stmt (stmt, stmt); update_stmt (stmt); release_ssa_name (name); + + /* GOMP_SIMD_LANE without lhs is not needed. */ + if (gimple_call_internal_p (stmt) + && gimple_call_internal_fn (stmt) == IFN_GOMP_SIMD_LANE) + remove_dead_stmt (&gsi, bb); } else if (gimple_call_internal_p (stmt)) switch (gimple_call_internal_fn (stmt))