>> One bug might be that Guile 1.8.6 runs every benchmark (some of them >> are real-world applications, some simple synthetic benchmarks) faster >> than the current version with this new VM. > > Hah! Yes that sounds like a bug to me :) Are you compiling the > benchmarks? I'd like to look into this. Let me know if you can get an > isolated test case.
Hi, I have tried both compiling the benchmarks ahead-of-time and simply running them with guile -s, compiling is naturally little faster but still a little slow. From C stack trace I noticed that debugging VM was always used but I don't know how to change the VM. Trunk version was compiled with '-O2 -march=athlon' and 1.8.6 is a binary installed from Fedora Core 10 repository. I attached the worst offender (ctak) which takes 1.6GB of core and about one minute to run on the trunk, but insignificant amount of memory and about two seconds on 1.8.6. I also attached a very simple fibonacci benchmark that is somewhat slower on trunk. Other, bigger benchmarks are slower as well but this was easiest way to make an isolated test case :) BR, Juhani
ctak.scm
Description: Binary data
fib.scm
Description: Binary data
