------- Comment #27 from dnovillo at google dot com 2007-06-19 17:39 ------- Subject: Re: [4.2 Regression] Incorrect stack sharing causing removal of live code
On 6/19/07 1:26 PM, rth at gcc dot gnu dot org wrote: > ------- Comment #26 from rth at gcc dot gnu dot org 2007-06-19 17:26 ------- > (In reply to comment #10) >> Talked to Dan Berlin and Diego Novillo here at Google. They told me >> that all locals are promoted to function scope. > > That *only* applies to register variables, not stack variables. Yes, but our GIMPLE optimizers don't know that and we do not have a way of enforcing that in GIMPLE. There just are no scope boundaries in the SSA web. So, the end result is that in SSA we only have function scope. > We very very much want to preserve scope of stack variables, because > we very very much want to share stack space between stack variables > of different scopes. Failure to do so causes bad interactions with > inlining, and causes stack space consumtion to grow to unacceptable > levels. I.e. you can't build kernels anymore. Agreed. We have to find a way of either tying the hands of code motion transformations by making them use the SSA web *and* the TREE_BLOCKs, or make stack slot sharing aware of live ranges. Right now we have the unfortunate situation that an optimizer is making a valid transformation which breaks the scope assumption that stack sharing uses. I am not sure if it would be practical to force code motion optimizations to use anything other than the SSA web to do their analysis. Perhaps we could force scopes by building barriers on the SSA web itself (say, by putting PHI nodes at scope boundaries, or something). Alternatively, we could incorporate live-range analysis to stack slot sharing. That way, if code motion creates undue stack slot pressure, we would merely emit suboptimal code, instead of wrong code. -- http://gcc.gnu.org/bugzilla/show_bug.cgi?id=32327
