session
Soundness passes: escape analysis, sound security annotations, verified constant-time
Between sessions 27 and 30 the compiler stopped trusting its own good intentions and started proving things: about user code and about itself.
References can no longer dangle
- Second-class references (
src/escape.rs,compiler/escape.fd): a&T/&mut Tcan no longer be stored in a struct, returned from a function’s locals, sent to a task, or captured by an escaping closure. The escape check runs in both compilers, so a reference that would outlive its referent is a compile error everywhere, not just on one backend. Vec<f64>reads andf64returns fixed on all three backends on the way, and the fixpoint was re-verified after the change.
Security annotations made sound
@taintedbecame real data-flow tracking: taint propagates through expressions until a sanitizer clears it or it hits a sink (SQL, shell, HTML, path), which is a compile error.@zeroizenow wipes heap data too, not just stack slots: PE output verified byte-for-byte on the affected tests.- Post-codegen constant-time verification: after the peephole stage,
verify_constant_time_asmscans the final x86-64 assembly of a@constant_timefunction for conditional/indirect jumps, memorycmov, indexed memory ops, variable-latency ops, and calls into non-CT functions. Violations are build errors: the check inspects what the CPU will actually run, not what the IR hoped for. (ELF path today; PE/ARM64/RISC-V coverage is planned.) cast_argABI repair in both C codegens heals&mut Vec→Vecforwards, andcl.exelinking was fixed on Windows (cmd /Cquote-stripping).
Branch consolidation
release-infra merged into main; 92 &mut call sites removed; str_equal replaced by plain ==; the bootstrap fixpoint re-verified byte-for-byte afterward.