On 4/18/06, Aldy Hernandez <[EMAIL PROTECTED]> wrote: > > losing a slight missed optimization on the tree level. > > Yay, exactly what I'm trying to fix. Glad you agree.
You can try disabling the write_only heuristic completely on mem-ssa branch and see if you get huge regressions on the compile time of fortran polyhedron benchmark. If you do, you might want to explore other ways of "fixing" this problem. On mainline I cannot see a way around this hack other than maybe imposing some per-function number-of-SFTs limit or so. But then, structures only written to (in a sane way) are usually not worth decomposing (which was my observation with the fortran testcase). Richard.