The bug which I reported occurring in MATMUL() at -O2 when egcs-19980906 is used as a back end for Pacific-Sierra f90, has been fixed in egcs-19980914. However, -O2 no longer runs faster than -Os in any of my test cases on Pentium 2.