m6w6
commented
4 years ago - Comment by: lp:~tstarling:
- Created at: 2013-11-15T02:35:47Z
Note that according to http://gcc.gnu.org/wiki/Graphite/Parallelization and the linked document
...auto-parallelization is extremely stupid, in that it will add a synchronisation overhead to every execution of every inner loop, even if the loop would complete on a single core in much less time than it takes to do the synchronisation. So there is a good chance that enabling it would actually slow down the application.
Imported from Launchpad using lp2gh.
One might think, reading the description of -floop-parallelize-all in the gcc manual, that -floop-parallelize-all will cause automatic parallelization of all loops by enabling a rarely-used, alpha-quality automatic parallelization feature called Graphite. However, this scary prospect is averted because -floop-parallelize-all doesn't actually generate parallel code unless you also specify -ftree-parallelize-loops=N where N is the thread count (that's right, it's fixed at compile time). This is explained at http://gcc.gnu.org/wiki/Graphite/Parallelization
So all -floop-parallelize-all does is cause the build to fail if libcloog is missing (i.e. bug #1232551), and increase compile time somewhat. I suggest that it be removed.