[GR-56599] Create type units for types during debuginfo generation for referencing across multiple debug info files.

[graalvmbot]

Each type is currently generated within a compilation unit, with no way to reference it from outside the .debug_info section. During debuginfo generation at image runtime, we have to create a separate debuginfo file that references types generated during image build. This would require to reproduce debuginfo for types that was already created at image build.

One way to reference types in different objectfiles is to put them into so-called type units and create type signatures that can be used as references. Each type is placed in its own type unit and type references will be changed to refer to the type units. During the runtime debuginfo generation, we can then reference existing types by the type signature.

Type units were introduced in DWARF 4 in their own debugging section (.debug_types) and moved to the .debug_info section in DWARF 5 to reduce overhead. As a change to DWARF 5 requires little changes and makes creating type units more convenient, we will also update the DWARF version of the .debug_info section to DWARF 5.

[olpaw]

Hi @dominikmascherbauer

the PR is in good shape already. I only have a few questions before we can merge this:

• Did you check if this causes any regressions in debuginfo generation time of the image builder?
• Is it guaranteed that the classloader contributes to the type signatures for everything except primitive types?

[olpaw]

Looks good overall.

Please process the comments and we should be able to merge this soon.

[dominikmascherbauer]

Hi @dominikmascherbauer

the PR is in good shape already. I only have a few questions before we can merge this:

* Did you check if this causes any regressions in debuginfo generation time of the image builder?

The time for debug info generation stays about the same, I checked that with debug info files with sizes of up to a few hundred MB.

* Is it guaranteed that the classloader contributes to the type signatures for everything except primitive types?

Yes, the classloader is looked up for all types except primitives (and the single instance of the header type which represents the object header).
The type signature then is the hash of the loader name and the type name, except for built-in classloaders (those are ignored in the hash).

* How does this change to dwarf5 + type signatures affect debug-info size?

Overall, this achieves a reduction of around 30% in debug info size, mostly due to the dwarf5 format changes of the .debug_loc and .debug_range sections.
Also, the loading time of debug info in gdb is reduced by about 15%.