| Issue |
171969
|
| Summary |
Popcount loop doesn't get folded on some targets
|
| Labels |
missed-optimization
|
| Assignees |
|
| Reporter |
koachan
|
Using the example adapted from [here](https://xania.org/202512/11-pop-goes-the-weasel-er-count), it seems that the loop fails to optimize on some targets, even when I tell clang that the target has a hardware instruction for it:
```c
// SPARC64: clang -O2 -mpopc
// MIPS64: clang -O2 -march=octeon
// X86-64: clang -O2 -mpopcnt
unsigned population_count_loop(long long value) {
unsigned result = 0;
while (value) {
value &= value - 1;
++result;
}
return result;
}
```
At least sparc64 and mips64 is affected (but I haven't checked the other targets, more could be affected):
```
! SPARC64
population_count_loop:
brz %o0, .LBB1_2
mov %g0, %o1
.LBB1_1:
add %o0, -1, %o2
and %o2, %o0, %o0
brnz %o0, .LBB1_1
add %o1, 1, %o1
.LBB1_2:
retl
srl %o1, 0, %o0
```
```
! MIPS64
population_count_loop:
.Lfunc_begin1 = .Ltmp3
daddiu $sp, $sp, -16
sd $ra, 8($sp)
sd $fp, 0($sp)
move $fp, $sp
beqz $4, .LBB1_2
addiu $2, $zero, 0
.LBB1_1:
daddiu $1, $4, -1
and $4, $1, $4
bnez $4, .LBB1_1
addiu $2, $2, 1
.LBB1_2:
sll $2, $2, 0
move $sp, $fp
ld $fp, 0($sp)
ld $ra, 8($sp)
jr $ra
daddiu $sp, $sp, 16
```
On the other hand, the builtin is compiled down to the hardware instruction as expected:
```c
unsigned population_count_builtin(long long value) {
return __builtin_popcountll(value);
}
```
```
! SPARC64
population_count_builtin:
retl
popc %o0, %o0
```
```
! MIPS64
population_count_builtin:
.Lfunc_begin0 = .Ltmp0
daddiu $sp, $sp, -16
sd $ra, 8($sp)
sd $fp, 0($sp)
move $fp, $sp
dpop $2, $4
move $sp, $fp
ld $fp, 0($sp)
ld $ra, 8($sp)
jr $ra
daddiu $sp, $sp, 16
```
For reference when targeting x86-64 both functions compile down to the popcount instruction:
```
! X86-64
population_count_builtin:
popcnt rax, rdi
ret
population_count_loop:
popcnt rax, rdi
ret
```
Generally I'd expect that such optimizations are performed in a target-independent manner.
[Godbolt link](https://godbolt.org/z/d66zdzoMz).
![]()
_______________________________________________
llvm-bugs mailing list
[email protected]
https://lists.llvm.org/cgi-bin/mailman/listinfo/llvm-bugs
