[perf regression] recursive self-calls slower on 0.5.1265

[dlee]

What happened

Functions that recurse (specific to self-recursion) run ~1.4–1.6x slower on 0.5.1265 than 0.5.1260.

Measurements (min ms, from-source builds, same host)

workload	0.5.1260	0.5.1265	ratio
fib(30) ×3 (~8M calls)	179	280	1.57
sumrec(1000) ×8000 (single self-call, ~8M)	236	322	1.36
flat 8M indirect non-recursive calls (control)	245	112	0.46 (faster)
iterative fib(30) ×3	~0	~0	—

Minimal reproduction

function fib(n) { return n < 2 ? n : (fib(n - 1) + fib(n - 2)) | 0; }
function work() { var a = 0; for (var k = 0; k < 3; k++) a = (a + fib(30)) | 0; return a | 0; }
work();
var best = 1e9, c;
for (var r = 0; r < 3; r++) { var t = Date.now(); c = work(); var d = Date.now() - t; if (d < best) best = d; }
console.log(best + " ms");   // ~179 ms on 0.5.1260, ~280 ms on 0.5.1265

perry compile m.js -o m && ./m.

Environment

Perry 0.5.1260 (2e105ca) vs 0.5.1265 (a3b31c0d), cargo build --release -p perry
from tagged source, each linked to its own full runtime. macOS 27.0 arm64.

[proggeramlug]

Audited this against current Perry and the later codegen history.

The exact JavaScript reproducer is still a recursive-numeric performance case (about 367-396 ms here on Perry 0.5.1482, Windows x86_64, with the expected 2496120 result), but this report is now superseded by #7244, which tracks the narrower actionable work: a sound recursive-numeric specialization.

The old speed came from emit_i64_specializations, which ran eligible number-typed self-recursive functions as i64. #7238 showed that path was unsound in two independent ways: it truncated fractional arguments on entry and diverged from JavaScript once intermediates crossed the exact-double range (and later i64 range). #7242 therefore removed it, explicitly measured the resulting Fibonacci cost, and opened #7244 for a guarded/deoptimizing replacement. That follow-up's acceptance criteria cover both restoring the performance and preserving the #7238 exactness tests.

I also A/B'd the current Phase-2 specialized ABI on the issue's verbatim repro; the best timings were 367 ms enabled and 370 ms disabled, so there is no separate modern ABI regression to fix here.

Closing this broad historical regression as superseded by #7244. No code change or version bump is appropriate on this issue.