MacOS X 10.2.1, DevTools August, 2002, installed on an empty partition, libdl from dlcompat-20010831. PowerMac G3/400, 50MHz bus, 128MB RAM, 2MB/200MHz backside cache. The gforth-0.5.0 builds used config.guess (timestamp='2002-07-23'), config.sub (timestamp='2002-07-03') and the corrected PPC _sync_cache_range.c. With each build, we did make clean, rm config.cache, rm config.status, make, make install, make bench, with all four combinations of the following: gcc_select 2 OR gcc_select 3 env ac_cv_sizeof_long_long=0 \ env "CC=cc -no-cpp-precomp -fno-gcse -Dunix -D_POSIX_SOURCE" \ configure --enable-force-reg --enable-direct-threaded --without-debug OR env ac_cv_sizeof_long_long=0 \ env "CC=cc -no-cpp-precomp -fno-gcse -Dunix -D_POSIX_SOURCE" \ configure --enable-force-reg --enable-indirect-threaded --without-debug Because there is so little difference between the results for gcc 3.1 with direct or indirect threading, we double checked those cases. Results for 2.95.2 with direct threading are very similar to (slightly faster than) old results for MacOS X Public Beta. The user slowdowns for gcc 2.95.2 indirect compared to direct threading are from 26% to 42%. The user slowdowns for gcc 3.1 compared to 2.95.2 with direct threading are from 41% to 55%. All times are user times; the sys times were in the noise. 2.95.2 2.95.2 3.1 3.1 direct indirect direct indirect siev 0m2.380s 0m3.200s 0m3.360s 0m3.650s bubble 0m2.720s 0m3.440s 0m3.870s 0m3.970s matrix 0m1.720s 0m2.440s 0m2.660s 0m2.800s fib 0m3.130s 0m4.030s 0m4.510s 0m4.510s dnw 25-Oct-02