Date: 2026-06-25
Area: bytecode
Pull Request: #802
Removed the per-call stack-trace string cost from the bytecode VM call path (#798). SetupNewFrame previously pushed (ATemplate.Name, ExecutionSourcePath, 0, 0) onto the shared TGocciaCallStack on every call, copying two reference-counted strings that are only ever read when an Error materialises .stack. It now pushes a deferred frame carrying the function-template pointer plus a module-path fallback (used only when the template has no own source file); a class-level resolver registered by the VM reproduces the name and source path at capture time, so an ordinary call performs no stack-trace string work. The shared call stack stays decoupled from the bytecode units: the template is held as an opaque Pointer and resolved through TGocciaTemplateTraceResolver, and the interpreter and native callers keep using the string-based push unchanged.
Fully removing the per-call push and reconstructing traces from the VM frame stack (FFrameStack + FCurrentClosure + saved-state roots) was considered and rejected: CaptureStackTrace's ASkipTop is calibrated to native frames (e.g. the Error constructor frame) that are absent from FFrameStack, and constructor frames are pushed by InvokeConstructor/Construct rather than SetupNewFrame, so a frame-stack reconstruction would drop or mis-skip frames. Keeping the push (cheap, pointer-only) preserves frame ordering, native/constructor interleaving, and ASkipTop exactly, so Error.stack output stays byte-identical to the previous behaviour. This is a deliberate trade-off of a one-time capture-side resolve for zero per-call cost. bytecode-vm.md. errors.md.