It appears that rust-lang/rust#153343 introduced some mild regressions in hypot, which uses min and max. aarch64:
icount::icount_bench_math_group::icount_bench_hypot logspace:(setup_hypot())
Baselines: arch_enabled|arch_enabled (old)
Instructions: 19914|18946 (+5.10926%) [+1.05109x]
L1 Hits: 21329|20360 (+4.75933%) [+1.04759x]
LL Hits: 0|1 (-100.000%) [---inf---]
RAM Hits: 15|15 (No change)
Total read+write: 21344|20376 (+4.75069%) [+1.04751x]
Estimated Cycles: 21854|20890 (+4.61465%) [+1.04615x]
x86-64:
icount::icount_bench_math_group::icount_bench_hypot logspace:(setup_hypot())
Baselines: arch_enabled|arch_enabled (old)
Instructions: 40337|38446 (+4.91859%) [+1.04919x]
L1 Hits: 53116|51224 (+3.69358%) [+1.03694x]
LL Hits: 2|3 (-33.3333%) [-1.50000x]
RAM Hits: 35|35 (No change)
Total read+write: 53153|51262 (+3.68889%) [+1.03689x]
Estimated Cycles: 54351|52464 (+3.59675%) [+1.03597x]
From the CI runs at https://github.com/rust-lang/compiler-builtins/actions/runs/31156146463?pr=1248. I didn't bisect down to the specific commit but 2026-03-15 is fine and 2026-03-16 shows the regression. The lowering changed https://rust.godbolt.org/z/bbvY9jEPT
min:
fminnm d0, d0, d1
retmin:
fminnm d1, d1, d1
fminnm d0, d0, d0
fminnm d0, d0, d1
retThat change makes sense for rustc but we already know we're not dealing with NaN so we could use something better.
It appears that rust-lang/rust#153343 introduced some mild regressions in
hypot, which usesminandmax. aarch64:x86-64:
From the CI runs at https://github.com/rust-lang/compiler-builtins/actions/runs/31156146463?pr=1248. I didn't bisect down to the specific commit but 2026-03-15 is fine and 2026-03-16 shows the regression. The lowering changed https://rust.godbolt.org/z/bbvY9jEPT
That change makes sense for rustc but we already know we're not dealing with NaN so we could use something better.