On 2012-07-03 19:36, Eric Niebler wrote:
On 7/3/2012 9:22 AM, Roland Bock wrote:
Given your test, I suggest you benchmark the preprocessed output rather than the source directly. OK, created preprocessed versions with clang -E and measured again, but
On 2012-07-03 16:52, Mathias Gaunard wrote: the results are similarly close (and way too much under the influence of the sheer amount of bytes that need to be processed).
I came up with a new test, see attachments:
Without optimization: Eric: real 0m11.457s user 0m11.080s sys 0m0.330s
Roland: real 0m11.228s user 0m10.790s sys 0m0.400s
With -O3: Eric: real 0m2.910s user 0m2.730s sys 0m0.150s
Roland: real 0m2.867s user 0m2.680s sys 0m0.160s
Interestingly enough, Eric's version takes a bit longer to compile on my machine: Too close to be very meaningful. I find that with TMP, the real costs don't become evident until you have a non-trivial program.
clang version 3.2 (trunk 155315) Target: x86_64-unknown-linux-gnu
And clang crashes when I add another row of parameters in Eric's version. No problems with my version... I hope you filed a bug. :-) Will do :-)