https://gcc.gnu.org/bugzilla/show_bug.cgi?id=108471
Bug ID: 108471
Summary: Suboptimal codegen for __int128 subtraction on x86_64
Product: gcc
Version: 13.0
Status: UNCONFIRMED
Severity: normal
Priority: P3
Component: other
Assignee: unassigned at gcc dot gnu.org
Reporter: rl.alt.accnt at gmail dot com
Target Milestone: ---
On x86_64, GCC generates an excessive amount of redundant `mov` instructions
for `__int128` subtraction in C/C++. Clicking through versions on godbolt shows
that this started getting worse in GCC 9.1 and later versions. See also
https://godbolt.org/z/86v6ar457
The code:
```
__int128 sub(__int128 a, __int128 b) { return a - b; }
```
At -O3 or -O2, GCC (trunk) generates:
```
sub:
mov r8, rdi
mov rax, rsi
mov rsi, r8
mov rdi, rax
mov r8, rdx
mov rax, rsi
mov rdx, rdi
sub rax, r8
sbb rdx, rcx
ret
```
Interestingly, the use of `r8` in the first three instructions disappears when
compiling w/ -O1, and those instructions are folded into two `mov`s instead.
By contrast, Clang (also at -O3) generates:
```
sub:
mov rax, rdi
sub rax, rdx
sbb rsi, rcx
mov rdx, rsi
ret
```
This is probably not a high-priority bug; I just wanted to bring attention to
the fact that this is happening.
