Reduce v2 self-host memory

[medvednikov]

Summary

• store v2 expression types sparsely and release the cache after ARM64 SSA build
• avoid large checker fallback-var map clones and string-heavy type name comparisons
• prune unused compiler backend modules on ARM64 before transform/SSA and release SSA/MIR/ARM64 codegen arenas once consumed

Impact

ARM64 v2 self-host peak RSS drops from ~9.25 GB to ~4.12 GB on:
./v2 --no-parallel -nocache -gc none -backend arm64 -O0 -o /tmp/v2_mem_final v2.v.

Cleanc remains heavier on this branch. Current measurement:
./v2 --no-parallel -nocache -gc none -backend cleanc -O0 -o /tmp/v2_mem_final_cleanc v2.v

• max RSS: ~6.62 GB
• peak memory footprint: ~12.24 GB
• total time: ~40s, with C generation taking ~29.7s

This does not reach the <1 GB target; remaining memory is dominated by transformed AST + backend-specific generated representation overlap and should be handled by the flat-AST/direct-lowering work.

Validation

• ../../v -gc none -cc cc -o v2 v2.v
• ./v -gc none vlib/v2/types/checker_test.v
• ./v -gc none vlib/v2/transformer/propagate_types_from_flat_test.v
• ./v -gc none vlib/v2/transformer/transformer_v2_darwin_test.v
• ./v -gc none vlib/v2/abi/abi_test.v
• ./v -gc none vlib/v2/builder/target_os_test.v
• cd cmd/v2 && sh test_all.sh

[chatgpt-codex-connector]

💡 Codex Review

Here are some automated review suggestions for this pull request.

Reviewed commit: 0c229f3f46

ℹ️ About Codex in GitHub

Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you

• Open a pull request for review
• Mark a draft as ready
• Comment "@codex review".

If Codex has suggestions, it will comment; otherwise it will react with 👍.

Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".

[chatgpt-codex-connector]

Limit ARM64 backend pruning to compiler backend files

This makes every ARM64 native build populate skip_modules for module basenames like x64, c, v, eval, and cleanc. The SSA builder’s filter only checks b.cur_module in b.skip_modules (not the file path), so any reachable user/imported module with one of those names is dropped entirely on ARM64 even though it is not an unused compiler backend. This was previously gated by --single-backend; the new arch == .arm64 default can produce missing functions or broken output for ordinary ARM64 builds that happen to use one of these module names.

Useful? React with 👍 / 👎.