On January 18, 2019 2:38:44 PM GMT+01:00, Michael Ploujnikov <michael.ploujni...@oracle.com> wrote: >On 2019-01-18 3:45 a.m., Richard Biener wrote: >> On Thu, Jan 17, 2019 at 8:16 PM Michael Ploujnikov >> <michael.ploujni...@oracle.com> wrote: >>> >>> Hi, >>> >>> I've been doing some investigations that required using a bitmap to >keep track of decl IDs and I ran into segmentation fault after my >bitmap has been loaded from a PCH. Attached is a short patch that can >reliably reproduce my problem. Looks like head->current is set to point >to uninitialized memory for some reason. Could someone with a strong >understanding of GGC take a look and let me know what I might be doing >wrong and how to fix this? >> >> If you're just doing investigation why do you care about PCH? > >My idea is to write a patch that does something similar to this, but >not exactly (obviously not storing all of the object IDs for example). >It's very preliminary and I don't know about the feasibility of what >I'm trying to do, but I do know that I need a bitmap/hash_set for the >full duration of toplev::main which inevitably involves PCH. However >because of this segfault I haven't been able to actually try my idea >yet! > >> >> In any event I guess that bimaps are not supported by PCH because of >> GTY(skip)ing the bitmap_head current member (so that's not saved). >> I guess PCH writers/readers need to ignore skipping. > >Why is the current member not saved? I know the documentation says that >another pointer in the chain already points to the same thing, but >what's the harm in not skipping it?
It disturbs optimal chaining of the list elements during the GC walk. As said we probably need a different thing than Skip for this purpose. For your experiments simply remove the skipping. Richard. > >- Michael
