On Wed, Sep 19, 2012 at 10:48 AM, Jan Hubicka <hubi...@ucw.cz> wrote: >> Hi, >> >> I've integrated all the reviews from this thread (Thank you guys for >> helping refine this patch). >> >> Now the patch can pass all gcc testsuite as well as all spec2006 >> benchmarks (with LTO). Concerning memory consumption, for extreme >> benchmarks like tramp3d, this patch incurs around 2% peak memory >> overhead (mostly from the extra blocks that have been set NULL in the >> original implementation.) >> >> Attached is the new patch. Honza, could you help me try this on >> Mozzila lto to see if the error is gone? > Hi, > I tested the last version you posed and it works fine. (i.e. no ICE) > I also observed no real differences in memory use. > linemap lookup seems to be bottleneck on streaming out stage of WPA. > I wonder if we can't stream location better into LTO objects, by perhaps > using same encoding as we do in memory (i.e. streaming out locators and > separate table)
Yes, I think we should consider streaming locations separately so that we can read them in in one chunk and sort them to be able to assign most compact location_t's to them. Dehao, the patch is ok for trunk - please be on the watch to address possible fallout. Thanks, Richard. > Honza
