Re: [patch] Move loop structures to gc memory

[Ian Lance Taylor <iant at google dot com>]

Michael Matz <matz@suse.de> writes:

> > > First, I don't really understand to what bug Ian refers to, so, I
> > > couldn't fix it, even if I wanted to.  Second, I'd predict at this
> > > point, that even if the bug were fixed, it would not be as fast as
> > > ggc_free.  If it isn't faster, I don't see what relevance it has to
> > > the two quoted above.  And lastly, I don't think this issue was
> > > contained in either of those two messages, just for the record.
> > 
> > The bug, if it is indeed a bug, is that the set of GC roots is static.
> > If you want to have a dynamically allocated pointer into GC memory,
> > you must put that pointer into GC memory itself.
> > 
> > Fixing that, and using it for the loop structures, would not be slower
> > than using ggc_free.
> 
> But it also wouldn't be different the slightest.  You just move the 
> problem of when to call ggc_free to when to remove that pointer from the 
> set of additional roots.  Unless of course you don't want to remove them 
> once allocated.  But that would be wasteful again, just like every memory 
> leak.

No, for example, we could enhance the GTY markings to say "this field
in cfun points to a linked list of structs, linked by this field.  For
each instance of the struct, these fields in the struct point into GC
memory."  Then a list of, e.g., loop structs would automatically
become GC roots with no additional work required.  That is simple and
safe.

Ian