https://gcc.gnu.org/bugzilla/show_bug.cgi?id=104547
--- Comment #4 from Jonathan Wakely <redi at gcc dot gnu.org> ---
(In reply to Jonathan Wakely from comment #0)
> This is better:
>
> void shrink_pop(std::vector<int>& v, unsigned n) {
> while (n--)
> v.pop_back();
> }
>
> _Z10shrink_popRSt6vectorIiSaIiEEj:
> .LFB866:
> .cfi_startproc
> testl %esi, %esi
> je .L10
> movq 8(%rdi), %rax
> movl %esi, %esi
> negq %rsi
> leaq (%rax,%rsi,4), %rdx
> .p2align 4,,10
> .p2align 3
> .L12:
> subq $4, %rax
> movq %rax, 8(%rdi)
> cmpq %rax, %rdx
> jne .L12
> .L10:
> ret
> .cfi_endproc
This should be A LOT better, but see PR 96717 comment 6 and 7. It's still
better than the resize(v.size() - n) code even with that regression.
Clang's code for shrink_pop is:
_Z10shrink_popRSt6vectorIiSaIiEEm: # @_Z10shrink_popRSt6vectorIiSaIiEEm
.cfi_startproc
# %bb.0:
testq %rsi, %rsi
je .LBB0_2
# %bb.1:
shlq $2, %rsi
subq %rsi, 8(%rdi)
.LBB0_2:
retq
I was expecting something like that for v.resize(v.size() - n) too.
