session
One integer meaning, sound builds, a real allocator
Session 36 was a semantics-and-soundness session: fewer ways for the three backends to disagree, and a real memory story underneath them.
Integers mean one thing now
add/sub/mul/negare checked at the declared width on the interpreter, C, and native paths:i8overflows ati8, not ati64.@wrappingwraps at the declared width, division by zero always panics, andu64is genuinely unsigned. (Lowering used to carryi64for every integer op, which was the root cause of most backend disagreements.)
Memory, actually managed
p_freeis honest: double-free and freeing a pointer that is not a live heap block panic on all backends. It already caught a real leak:rc.fdcounters allocated through plainmallocwere never freed.- Size-class allocator behind
malloc/free. - Automatic
Vecreclamation: an IR escape pass frees localVecs that provably never escape: the dropped-Vec benchmark went from 544 MB peak to 6 MB.
Faster self-hosting
Self-hosted type checking of the compiler went from about 205 s to 15 s, with Vec_get/Vec_len inlined natively.
Bugs that were invisible
p_system crashed every native program on return; a pert_str_hash edge case could lose HashMap keys near page ends; hex literals above i64::MAX read as 0 on fidesc; native panics now actually print (via a vectored exception handler for stack overflow and invalid access). The Rust suite sits at interp 286/286, C 285/285, PE 287/287, with 66/66 expected-fail tests correctly rejected, and every stdlib module now compiles standalone in CI (scripts/check-stdlib.sh).