runtime: introduce G/M/P pthread backend

[cpunion]

Problem

The compiler currently lowers every go statement directly to pthread setup. That couples goroutine code generation to one OS backend and provides no runtime-owned G/M/P boundary for a future scheduler.

Change

• extend the pthread-key g merged by runtime: store goroutine state behind pthread-key getg #2090 into minimal Go-compatible G, M, and P objects;
• retain the current 1:1 model: one detached pthread, one G, one M, and one P per goroutine;
• make the compiler emit only a startup record and call runtime.NewProc;
• move pthread creation, optional stack-size setup, startup, and cleanup behind newproc1, newm, mstart, and mexit;
• route normal return and runtime.Goexit through the same cleanup path;
• keep the G/M/P context in one GC root, released by mexit for runtime-created goroutines and by the pthread destructor for bootstrap/external threads;
• publish P status atomically on threaded targets while retaining direct bare-metal access without compiler-rt atomic dependencies.

This establishes the replaceable native backend boundary proposed in #2152. It does not add M:N scheduling, preemption, growable stacks, LLVM coroutines, or runtime assembly reuse.

The independent diff is +622/-181 across 15 files. It contains no #2079 locality compiler/runtime files.

Validation

Tests were serialized with GOMAXPROCS=2, GOMEMLIMIT=8GiB, and -p=1.

• macOS arm64, Go 1.26.5: complete ./ssa, runtime compilation, and real-LLGo G/M/P tests pass.
• Ubuntu arm64, Go 1.26.5, LLVM 19.1.7: compiler/runtime and real-LLGo tests pass in an 8 GiB, 2 CPU, no-swap, 512-PID container.
• Ownership tests cover parent IDs, unique G/M/P IDs, running states, bidirectional links, getg isolation, and defer isolation.
• Existing goroutine and select IR fixtures pass.
• No test is skipped or ignored.

Reusing pinned GOROOT runtime and assembly code remains tracked separately in #2165.

[fennoai]

Review: runtime G/M/P pthread backend

Overall this is a solid, well-structured change. The runtime side is clean and, if anything, a small improvement: go lowering now emits a single runtime.NewProc call, and newproc1 does one combined AllocRoot for the G/M/P context. The runtimeContext lifecycle (alloc → mexit → FreeRoot), the mutually-exclusive arg-free paths, and the EnterLocalContext/LeaveLocalContext pairing are internally consistent. The switch of the FuncForPC/symtab caches to atomic pointer publication is the correct call now that goroutines run on real threads. No blocking memory-safety defect found.

The actionable items below are all compile-time performance and maintainability concerns — none block correctness. Inline comments carry the specifics; a couple of non-inline notes follow.

Non-inline notes (low severity / nits):

• runtime/internal/runtime/os_pthread.go header still reads Copyright (c) 2024 (carried over from the z_thread.go rename); every other new/rewritten file in this PR uses 2026.
• runtime/internal/runtime/os_pthread.go newosproc ignores the return of attr.Destroy(). Low risk and conventional (the create-critical returns — Init/SetDetached/SetStackSize/Create — are all checked), noting only for completeness.
• internal/build/main_module.go defineLibraryRuntimeInit: the doc block describes only the global_ctors registration but the function now also conditionally emits the enterLibraryLocalContext call when Prog.NeedsLocalContext() is true. Worth extending the comment, since that call is what lets exported Go calls on a C constructor thread inherit a context.

[cpunion]

Addressed the non-inline review notes as well:

• updated os_pthread.go copyright to 2026;
• made every pthread attribute Destroy result explicitly ignored, with a comment explaining why a post-Create destroy failure cannot be surfaced as thread-creation failure;
• extended defineLibraryRuntimeInit documentation in parent runtime: store goroutine state behind pthread-key getg #2090 to include the conditional process-lifetime locality context.

Validation on the final head:

• macOS/arm64 and Linux/amd64 runtime execution tests pass, including a 2 MB pthread stack;
• WASI, JS/Wasm, and ESP32-C3 cross-builds pass;
• full repository coverage run passes at 75.61% (16,539 / 21,873 statements; ssa 92.0%, cl 97.0%).

[codecov]

Codecov Report

✅ All modified and coverable lines are covered by tests.

📢 Thoughts on this report? Let us know!