| Issue |
174621
|
| Summary |
[[musttaill]] generates extra unnecessary instructions at the end
|
| Labels |
new issue
|
| Assignees |
|
| Reporter |
brevzin
|
Consider this example (courtesy of Matt Godbolt in his Advent of Compiler Optimization series, https://www.youtube.com/watch?v=J1vtP0QDLLU):
```cpp
#ifdef WITH_TAIL
#define MUSTTAIL [[clang::musttail]]
#else
#define MUSTTAIL
#endif
unsigned gcd(unsigned a, unsigned b) {
if (b == 0) return a;
MUSTTAIL return gcd(b, a % b);
}
```
Even without the `musttail` attribute, clang x86-64 `-O2` performs tail call optimization here. Except that with `musttail`, it adds two more instructions at the end, that do nothing?
<table>
<tr><th>without [[musttail]]</th><th>with [[musttail]]</th></tr>
<tr>
<td>
```asm
gcd(unsigned int, unsigned int):
mov eax, edi
test esi, esi
je .LBB0_4
mov edx, esi
.LBB0_2:
mov ecx, edx
xor edx, edx
div ecx
mov eax, ecx
test edx, edx
jne .LBB0_2
mov eax, ecx
.LBB0_4:
ret
```
</td>
<td>
```asm
gcd(unsigned int, unsigned int):
mov eax, edi
test esi, esi
je .LBB0_1
mov edx, esi
.LBB0_4:
mov ecx, edx
xor edx, edx
div ecx
mov eax, ecx
test edx, edx
jne .LBB0_4
mov eax, ecx
ret
.LBB0_1:
mov ecx, eax
mov eax, ecx
ret
```
</td>
</tr>
</table>
You can see the extra instructions generated with `[[musttaill]]`, even though there is already a tail call. [On compiler explorer](https://compiler-explorer.com/z/KMvsGbdPe). This is only true on `-O2` and `-O3`, on `-O1` they generate identical code.
![]()
_______________________________________________
llvm-bugs mailing list
[email protected]
https://lists.llvm.org/cgi-bin/mailman/listinfo/llvm-bugs
