[WTF] Reduce stack use and code size of Vector append slow paths - #76313

22f00b3

[WTF] Reduce stack use and code size of Vector append slow paths
<https://bugs.webkit.org/show_bug.cgi?id=326804>
<rdar://189511255>

Reviewed by Marcus Plutowski.

Pass small trivially copyable arguments by value to the out-of-line slow
paths of `Vector::append()` and `Vector::constructAndAppend()`.  The
`append()` slow path takes the argument's address so `expandCapacity()`
can rebase it if it points into the buffer being reallocated, and taking
that address forces callers to keep the value in a stack slot, even for
an `int` or a pointer.  Copying such a value into a parameter before the
buffer grows makes the rebase unnecessary.

Also move the `constructAndAppend()` slow path out of line, as
201368@main did for the `append()` reallocation in `expandCapacity()`.
Without `NEVER_INLINE`, clang inlines its reallocation sequence into
many call sites, which grows code at each `constructAndAppend()` caller.

Make `constructAndAppendSlowCase()` `ALWAYS_INLINE` into the by-value
wrapper rather than `NEVER_INLINE` itself, since its by-reference
parameters would make each caller copy a small argument from a register
to a stack slot just to pass the slot's address.

In a macOS release build, this shrinks `__text` by about 104 KB (0.1%)
across JavaScriptCore, WebCore, WebKit, WebKitLegacy, and WebGPU, and
the sum of their stack frame sizes by about 9.7 KB, mostly in
JavaScriptCore and WebCore.  Speedometer 3 is neutral.

No new tests since no change in behavior.

* Source/WTF/wtf/Vector.h:
(WTF::Vector::append):
(WTF::Vector::constructAndAppend):
(WTF::Vector::appendSlowCaseByValue): Add.
(WTF::Vector::constructAndAppendSlowCaseOutOfLine): Add.
(WTF::Vector::constructAndAppendSlowCase):

Canonical link: https://commits.webkit.org/323213@main

4e3fbfe