| Issue |
137514
|
| Summary |
Missed optimization: Suboptimal div / rem for 128-bits int by constant lowering
|
| Labels |
new issue
|
| Assignees |
|
| Reporter |
MaxGraey
|
For most constant divisors, the 128-bit type is generated a libcall instead of using magic constant + inlined mul_high_128 as GCC does.
Let's take `(__uint128_t)x / 19` for instance:
```c++
uint64_t div_by_19(__uint128_t x) {
return x / 19;
}
```
clang (trunk):
```asm
div_by_19(unsigned __int128):
push rax
mov edx, 19
xor ecx, ecx
call __udivti3@PLT
pop rcx
ret
```
Expected / GCC (trunk):
```asm
div_by_19(unsigned __int128):
movabs rcx, 18014398509481983
mov r8, rdi
mov rax, r8
and rdi, rcx
shrd rax, rsi, 54
shr rsi, 44
and rax, rcx
add rdi, rax
movabs rax, -2912643801112034465
add rsi, rdi
mul rsi
shr rdx, 4
lea rax, [rdx+rdx*8]
lea rax, [rdx+rax*2]
sub rsi, rax
movabs rax, -8737931403336103397
sub r8, rsi
imul rax, r8
ret
```
Link to more examples: https://godbolt.org/z/qhM6975b4
![]()
_______________________________________________
llvm-bugs mailing list
llvm-bugs@lists.llvm.org
https://lists.llvm.org/cgi-bin/mailman/listinfo/llvm-bugs
