[Bug middle-end/56548] [4.8 Regression] ICE in emit_move_insn, at expr.c:3486 with -march=pentium{pro,2,3} -O3

[sje at gcc dot gnu.org]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=56548

--- Comment #5 from Steve Ellcey <sje at gcc dot gnu.org> 2013-03-06 18:04:30 UTC ---
For the record, this does undo the performance regression fix on MIPS.  I will
submit a new bug for that issue.  The problem (on mips) is that we want to do a
conditional move of a short to a destination that already contains a short.
Since we don't have a short conditional move we use an int conditional move,
but we shouldn't need to 'reconvert' the moved integer back to a short by
masking off the upper bits.  They are already zero because the source and dest
of the conditional move were both shorts.