[SmallVector] Out-of-line the trivially-copyable push_back grow path (#206213)
In the approximately trivially-copyable specialization, push_back's grow
path does not early return. Both Clang and GCC likely keep `this` and
`Elt` live across the out-of-line `grow_pod` call, saving and restoring
them in the prologue/epilogue. Shrink wrapping can't sink it (the saved
values are used in the store block the fast path also reaches).
Move the grow-and-store into a noinline `growAndPushBack` helper and
tail call it. The fast path needs no callee-saved registers.
`push_back(int)` drops from 14 to 7 instructions on x86-64.
```
// void vec_pb_int(llvm::SmallVectorImpl<int>&v, int x){ v.push_back(x); }
mov eax, dword ptr [rdi + 8]
cmp eax, dword ptr [rdi + 12]
jae _ZN4llvm23SmallVectorTemplateBaseIiLb1EE15growAndPushBackEi # TAILCALL
mov rcx, qword ptr [rdi]
mov dword ptr [rcx + 4*rax], esi
inc dword ptr [rdi + 8]
ret
```
`noinline` keeps the fast path frame-free and the cold grow path out of
every call site (with a single COMDAT copy). It is load-bearing, as GCC
and Clang otherwise inline it back for some code. However, the
out-of-line call makes the element's address escape, which defeats
construct-in-place for large element types.
`T Tmp = Elt` copy preserves the internal-reference-during-grow case and
is elided for by-value element types.
llvm-project
fe1fc78c - [SmallVector] Out-of-line the trivially-copyable push_back grow path (#206213)
