[ ... ] > -xregs=%frameptr > > would seem to be the magic. > > Although it does apparently slow down performance.
Has this been quantified ? I don't mean just function throughput benchmarks, but rather like the effect it has on something like specweb if I use HTTPS done via OpenSSL with/without framepointers ? How much does it _really_ gain for non-HPC code to ditch %ebp/%rbp ? > > Is it really so important to have dtrace have extra visibility > into the openssl shared libs? Do these libs do systemcalls ? Do these libs call other things from libc ? If so, you don't only loose visibility into the openssl libs. You also loose visibility of the callers. I.e. you don't "just" make openssl invisible to DTrace's ustack(), but all codepaths that call into openssl. I mean, for completely self-consistent code (pure arithmetics, aka things like BLAS/LAPACK) that doesn't "call out", not maintaining framepointers is definitely ok. That may apply to some of the openssl crypto code, but really the whole library ? (I don't know it well enough - is is "pure math" ?) Btw, register use: I'd suggest _not_ to use -fpic/-Kpic on 32bit x86, since that steals another register from you. Before ditching framepointers from openssl, I'd ditch pic (on 32bit) and accept a possiblty slightly higher time to start. For 64bit/amd64, that doesn't matter. FrankH. _______________________________________________ opensolaris-discuss mailing list opensolaris-discuss@opensolaris.org
