[cecetaca]

The gap

math::fxp exports le, and no lt, ge or gt. The fixpoint deftype has all four as methods. So working with the type gives you every comparison; working with a raw int<A+B> gives you one.

That split shows up at a channel boundary, where values arrive as int<W> and the port type is fixed. Porting C to ACT, every signed comparison then has to go through le with the operands swapped and an (x != y) term bolted on — which is precisely what the deftype methods already do internally:

function lt(...) { self := (x != rhs.x) & fxp::le<A,B>(x, rhs.x) }
function ge(...) { self := fxp::le<A,B>(rhs.x, x) }
function gt(...) { self := (x != rhs.x) & fxp::le<A,B>(rhs.x, x) }

The logic exists; it just isn't reachable from the function API, so callers write it a third time.

The change

Adds lt, ge and gt to the fxp namespace in le's style, and then has the three deftype methods call them instead of keeping a second copy of the expression. le, mults, divs and everything else are untouched, and the methods compute exactly what they did before. Header comment updated to list the four.

Net +34 / −4 in one file.

Verification

No test harness in this repo, so this was checked under actsim.

Exhaustive. All 256 ordered pairs of Q(4,0) — every 4-bit two's-complement value against every other — with each of the four comparisons checked against an oracle that flips the sign bit and compares unsigned. That is deliberately a different method from the one under test, which compares sign bits and falls through to an unsigned compare:

PASS 1024 comparisons, 0 wrong

Negative control. Dropping the (x != y) strictness term from gt makes the same bench report FAIL 16 of 1024 wrong, so it can fail.

The refactor. The three fixpoint methods were separately checked to agree with the functions they now delegate to, across a spread of Q(8,8) values.

🤖 Generated with Claude Code