On Mon, Jun 17, 2013 at 6:54 PM, Joseph S. Myers <jos...@codesourcery.com> wrote: > On Mon, 17 Jun 2013, Jakub Jelinek wrote: > >> The following patch shows a performance regression caused by the C++ changes >> to evaluate side-effects in x += foo (); first and only then do the >> addition. Previously if x was say int and foo () returned unsigned long >> long, convert_to_integer would narrow the addition to unsigned int, but >> as we now pass to convert_to_* a COMPOUND_EXPR with the side-effects on the >> op0 and addition on op1 of the COMPOUND_EXPR, convert_to_integer doesn't >> perform this shortening and unfortunately we don't have any gimple >> optimization yet to do this later on. This patch fixes it by handling >> COMPOUND_EXPR in convert_to_* where we care about TREE_CODE of the expr. >> >> Ok for trunk? Ok for 4.8 as well after a while? > > I suppose it's OK to fix the regression, though really we should be > eliminating these early convert_to_* optimizations (optimizing later on > GIMPLE if possible) rather than adding to them.
I agree. The correct place for this is in tree-ssa-forwprop.c (for now). Richard.
