dwarf4 Separate Type Unit Entries

[Mark Wielaard <mjw at redhat dot com>]

Hi,

I was talking to Tom Tromey about the general idea of dwarf compression
through identification of similar DIE trees and replacing them with a
single instance and then referencing those similar trees in the main
tree through the use of DW_TAG_imported_unit DIEs that have a
DW_AT_import attribute pointing to the DW_TAG_partial_unit DIE tree that
we separated out.

Now dwarf4 adds another unit variant, the DW_TAG_type_unit (section
3.1.3) which is like a partial unit, but only for types. They cannot be
imported normally, but only through a DW_AT_signature attribute that
references the DW_TAG_type_unit through a 64bit signature/hash instead
of a normal reference, instead of the DIE having an DW_AT_declaration.

This seems a special case variant of the way we do dwarf compression.
But it may be indicative of the gains we might be seeing. Tom said that
for a gdb build on F14 -g -O2 -vs- -gdwarf-4 -O2, gives a size reduction
from 16M to 11M. The .debug_info went from 70ffa6 -> 17a63c, the abbrev
shrunk and the .debug_types section (where the type units are stored) is
smaller than the difference in debug_info.

One advantage of the type unit compression is that gdb already has some
support for it, while the support for imported partial units is still on
some branch.

So now we probably should figure out how to cash in on this alternative
compression variant. Does it make sense to create a dwarf writer that
outputs DW_TAG_type_units instead of DW_TAG_partial_units if they only
contain type information DIEs (I am not completely clear on what DIEs
are and which are not allowed to be included in the type units). But the
references seem somewhat bloated (don't know if that really adds up). Or
does it make more sense to transform the type units into partial units
when we read them in? But what would be needed to get the libdw/C++
dwarf branch to properly parse type units?

Cheers,

Mark