Worse perf with NonZero

[hkBst]

In rust-lang/literal-escaper#15 I introduced NonZero into literal-escaper, but noticed perf regressions (on ARM).

On compiler explorer I noticed x86_64 asm differences that seem to indicate there may be a difference in inlining behavior: asm being generated without code differences in the surrounding code (does that not mean it is dead?). These can indeed be removed by uncommenting the commented #[inline] directives (top left pane). On arm, the changes are similar but more extensive and I can't tell whether extra function code is called or not.

Sprinkling the right #[inline] directives does remove the perf regressions on ARM. I haven't been able to do perf measurements on x86_64 yet.

[saethlin]

Adding #[inline] is probably relevant in a number of places for subtle reasons. Many of these functions are missing that attribute, but if they are trivial enough they get LocalCopy codegen anyway, which makes some LLVM optimizations more effective. When you add a call to a previously-trivial function, you need to add #[inline] to restore the same behavior.

This is probably what you are observing when you mention

asm being generated without code differences in the surrounding code (does that not mean it is dead?).

LocalCopy functions are only lowered in CGUs that reference them.