Repository navigation
v2.8.32 — signedness fix for 32-bit backends + 1.90x xtensa perf
A correctness release with a large 32-bit performance improvement.
Anyone targeting riscv32 or xtensa should upgrade — v2.8.31 and earlier
silently miscompile unsigned variables there.
Fixed
A variable's signedness came from its value, not its declaration. Two places
let the signed flag leak from an initialiser or an assigned value onto an
explicitly unsigned variable:
u32 ux = <i32 expr> // vardecl: copied the RHS's flag
u32 ux = 7; ux = <i32> // assign: passed the rvalue's flag through
ux and everything derived from it then used SIGNED operations: ux >> 1
lowered to an arithmetic shift and comparisons became signed. Invisible on
x86_64 and arm64, where a u32 occupies a 64-bit slot and never reaches the
sign bit — but wrong on the 32-bit backends, where 0x80000000 >> 31 produced
0xFFFFFFFF instead of 1. A declared type is now authoritative in both
directions: a store cannot retype a variable.
Also fixes a latent hole where copy coalescing could approve a merge its own
result could not satisfy, judging it against the whole register file rather than
the merged node's colour ceiling. The consequence was a spill, not wrong code.
Performance
Measured end to end on real ESP32 silicon (YuNet face detection, 96px input),
with every output value byte-identical to the previous release at each step:
| change | cumulative |
|---|---|
| xtensa: 9 register colours instead of 4, in leaf functions | 1.44x |
| xtensa: extend that to functions that make calls | 1.58x |
| xtensa/riscv32: ADD/SUB immediate fusion (was x86-only) | 1.85x |
| all targets: stop staging loop back-edge copies through temps | 1.90x |
1.90x overall — 4.58s to 2.44s per inference. The last change is in shared
IR and shrinks every target, including the compiler itself: krc self-compiles
to 1,334,832 bytes against 1,351,600.
Verification
699/699 tests, bootstrap reaches a fixed point (stage3 == stage4).