PhaseB spawned three child processes per module (mlir-opt, mlir-translate, llc), and the
process startup dominated the cost: a 0.3 KB module took ~170 ms, of which ~150 ms was
just starting the tools; the 59-module VKernel set cost 18.4 s CPU (3.3 s wall on 16
workers).
ircore.dll (IRCore) does MLIR parse + verify + LLVM IR translation + codegen in a single
call. PhaseB now uses it from the worker threads:
- loaded at runtime with LoadLibraryA/GetProcAddress, not through the import table, so the
compiler still starts when the DLL is missing; one retry points the loader at clang64's
bin directory because ircore.dll needs that exact LLVM build
- falls back to the legacy flow automatically (with a warning) when the DLL cannot load
- `--phaseb-legacy` forces the legacy backend; the old command path is kept intact
(compile_ll_to_obj / init_llc_task / _llc_task_worker / run_llc_parallel)
- ASan instrumentation, raw-IR merging (llvm-link/opt), --traceback frame pointers and
--phase2command keep using the external tools
- new failure stage LLC_STAGE_IRCORE so a failure is attributed to the FFI call
Measured: 59 VKernel modules 18.4 s -> 4.8 s CPU, 3.3 s -> 1.6 s wall; AstTest
--rebuild-includes 13.6 s -> 11.3 s; BuddyTest 4.8 s -> 2.9 s; ViperC self-host 125 s ->
119 s (PhaseA dominates it); 41 Test projects pass, plus the 5 pre-existing failures.
TermiNexus
ef4582dcfc