http://gcc.gnu.org/bugzilla/show_bug.cgi?id=58721
--- Comment #15 from Jan Hubicka <hubicka at ucw dot cz> --- > The threshold is ~6000 (exactly 5941), i.e. more than twice the default value > 2700. Thanks for working it out. This may be only testcase I know of where large-function-insns brings significant regression. Did you try to experiment with large-stack-frame limits for fortran? Those hit quite commonly, perhaps we miss useful inlining here. I never did really serious tunning on large-function-insns except for quick check that increasing it doesn't affect SPEC scores. But lets fix the underlying branch prediction problem and see if we find another testcase supporting the change. Honza
