v2: self-compilation with `-v2` uses excessive memory

[GGRei]

Describe the bug

When trying to compile the V2 compiler by invoking V with -v2 directly on cmd/v2/v2.v, memory usage grows very high.

Reproduction Steps

From a fresh V checkout, run:

./v -v2 -o /tmp/v2_bootstrap cmd/v2/v2.v

If needed, also try the low-parallelism variants:

VJOBS=1 ./v -v2 -o /tmp/v2_bootstrap cmd/v2/v2.v
./v -v2 -no-parallel -o /tmp/v2_bootstrap cmd/v2/v2.v
VJOBS=1 ./v -v2 -no-parallel -o /tmp/v2_bootstrap cmd/v2/v2.v

For comparison, this command works and stays bounded:

./v -o /tmp/v2_bootstrap cmd/v2/v2.v

Expected Behavior

The command should either compile successfully with bounded memory usage or fail cleanly with an explicit compiler error.

It should not consume excessive memory, swap heavily or make the OS session unstable.

Current Behavior

The direct -v2 self-compilation path consumes excessive RAM.

Possible Solution

No response

Additional Information/Context

The memory blow-up happens inside the delegated V2 compiler during Transform and Mark Used and CleanC generation.

V version

V 0.5.1 10a27a3

Environment details (OS name and version, etc.)

Ubuntu 26.04 LTS

Note

You can use the 👍 reaction to increase the issue's priority for developers.

Please note that only the 👍 reaction to the issue itself counts as a vote.
Other reactions and those to comments will not be taken into account.

[programmingkidx]

On Mac OS 12.3 when I ran ./v -v2 -o /tmp/v2_bootstrap cmd/v2/v2.v, the memory usage went past 15GB at the 'Mark' and 'C Gen' steps.

[GGRei]

The main issue is that direct V2 self-compilation uses too much memory because the delegated V2 compiler is currently built with -gc none.

When that same delegated compiler is built with -gc boehm, memory stays low and stable.

Once the memory issue is avoided, there are still a few smaller follow-up points to check around generated C output.

[programmingkidx]

@GGRei Is there a noticeable difference in compile time when using boehm?

[GGRei]

@programmingkidx I still need to collect clean before/after timings, but the memory difference is clear

@medvednikov Would it be acceptable to build only the cached delegated V2 compiler with -gc boehm, while keeping the user target flags unchanged? Temporary solution while the v2 autofree.

[medvednikov]

Flat ast will resolve this.

Right now there's both traditional ast and flat.

Also V has always been compiled with -gc none. It's much faster this way.

[medvednikov]

15 GB is not normal though, I'll fix it today.

[GGRei]

I can confirm the same behavior on my side: the direct -v2 self-compilation also climbs to around 15 GB of RAM here. ( Ubuntu ).

[medvednikov]

I'm still on it

[GGRei]

@medvednikov Yup, I saw that PR #27344 is ready.

I have another PR coming for the x64 backend, but since your PR and mine touch some of the same V2 areas.

I’ll wait for yours to be merged first. After that I’ll rebase/update my branch.

[medvednikov]

Had to fix generic logic in transformer first, now that that's done, I can merge the flat ast PR.

[medvednikov]

Merged.

[GGRei]

Part 2 of the native x64 backend work is now ready.

Do you want me to also help with the Flat AST refactor work, or should I continue with the next native x64 backend step first?

[medvednikov]

flat ast is tough, it's not as isolated as the x64 backend, so it's probably not possible to work on it together