gdb: extend maint print address-spaces to show all dbgapi available info

[czidev-amd]

Motivation

The maint print address-spaces command previously printed only a flat list of address-space names. Rewrite address_spaces_dump to query dbgapi for five properties per address space and emit them as a structured ui_out table: name, DWARF number, address size in bits, null address, and access class.

Technical Details

A new helper, address_space_access_to_string, maps the amd_dbgapi_address_space_access_t enum to a human-readable string (all, program-constant, dispatch-constant). The command help text and the maint print address-spaces section in gdb.texinfo are updated to document the new columns.

Test Plan

A new test, gdb.rocm/maint-print-address-spaces.exp, runs the command, and checks that a header row and at least one data row are printed without error.

Submission Checklist

• Look over the contributing guidelines at https://github.com/ROCm/ROCm/blob/develop/CONTRIBUTING.md#pull-requests.

[czidev-amd]

Pushed a new version with the next changes:

• address_space_access_to_string: change return type to std::string and print the raw value in hex for the unknown case, e.g. "unknown (0x42)",to aid debugging.

• address_spaces_dump: collect all dbgapi properties in a single pass into an aspace_info vector, so each property is queried exactly once. Width calculation and printing both iterate the cached vector. This eliminates the double-query-per-property pattern and the single-statement braced if blocks in the old first-pass loop.

• Width sentinels: use strlen(("Name")) / strlen(("Access")) instead of the bare string literals so the gettext result is measured, not the key.

• Numeric fields: switch dwarf and addr columns from field_fmt("%lu") to field_unsigned(), which is portable (no LLP64 truncation on Windows) and preserves correct right-alignment.

• Test: anchor the header regex so -lbl matching fires on a full header line, not a substring buried elsewhere in the output.

[czidev-amd]

Thank you @amd-shahab for your input, I have fix that and each error message will contain the returned error status too (as Pedro requested in #225)

[lumachad]

Just nits. Take it or not, your choice.

[palves]

What does "setting a different architecture" mean?

Did you mean "testing"?

We do "set architecture amdgcn:gfx90a" above, though, so the architecture is fixed. So how can the premise that the output varies be true?

[czidev-amd]

Trailing message from previous refactoring, removing.

[palves]

Well, that still leaves the whole premise that lead to not parsing actual specific address spaces incorrect -- if we're hard coding one specific architecture, then we could make the testcase look closer to the address spaces that came out. OTOH, maybe we want instead to loop over all architectures and issue the command, to confirm that we (gdb, dbgapi, etc.) don't do anything bad in any.

[czidev-amd]

That (having a strict testcase for a specific architecture) was my initial draft, however, after speaking with Lancelot, I've simplified the test case to what we have right now. However, I can loop over a list of known architectures, and check if we crash or something of a sort.

[lancesix]

The issue with testing the values is: aperture configuration are not part of the ABI. Those are not expected to change often, but they could, and are different on Windows and Linux.

We could have "probe" kernels which we use to figure out the configuration and deduce valid conversions we want to test for if we really wanted to test conversion.

[palves]

You are thinking of the address alias PR instead? This PR is about listing this:

Name              DWARF  Addr Bits    Null Address  Access
global                0         64             0x0     all
generic               1         64             0x0     all
local                 3         32      0xffffffff     all
private_lane          5         32      0xffffffff     all
private_wave          6         32      0xffffffff     all

[lancesix]

ah, sorry for the confusion. Yes, I was thinking about the alias PR.

The list of address space supported is a per architecture, but we could teach GDB what to expect for here, and match the actual output.

[czidev-amd]

I have updated the test, now it loops over the gdb's known amdgpu arches and executes for each one the maintenance command. Please let me know if you want to strength more up the test. For example, null address can be either 0x0+ or 0xf+

[lancesix]

Thanks, this LGTM.

Maybe give some time to other people who contributed reviews before merging see if they are happy to.

[palves]

This still has the same problem:

Display address-space properties supported by the current architecture.\n\

The set of properties is fixed! What varies is the address spaces.

For each address space, print its name, DWARF id, address size,\n
null address, and access class."),

So this would be more correct:

Display properties of address spaces supported by the current architecture.\n
For each address space, print its name, DWARF id, address size,\n
null address, and access class."),