| Issue |
184762
|
| Summary |
Pessimistic code generation for `next_multiple_of` for powers of two
|
| Labels |
new issue
|
| Assignees |
|
| Reporter |
coastalwhite
|
LLVM is unable to optimize `a.div_ceil(64) * 64` to `(a + 63) & ~0x3F`. This seems to apply to all powers of two. The result is worse code.
This was originally noticed in Rust, but it also applies in `C`. Here is a Godbolt link to illustrate the problem.
<https://godbolt.org/z/vE6nqKqer>
```c
#include <stdint.h>
typedef uint32_t u32;
static inline u32 next_multiple_of(u32 a, u32 b) {
u32 r = a % b;
if (r == 0) {
return a;
} else {
return a + (b - r);
}
}
static inline u32 div_ceil(u32 a, u32 b) {
u32 d = a / b;
u32 r = a % b;
if (r > 0) {
return d + 1;
} else {
return d;
}
}
// Generates worse code.
u32 next_multiple_of_64_div_ceil(u32 v) {
return div_ceil(v, 64) * 64;
}
// Generates correct code.
u32 next_multiple_of_64(u32 v) {
return next_multiple_of(v, 64);
}
// Generates correct code.
u32 next_multiple_of_64_raw(u32 v) {
return (v + 63) & ~((u32)0x3F);
}
```
![]()
_______________________________________________
llvm-bugs mailing list
[email protected]
https://lists.llvm.org/cgi-bin/mailman/listinfo/llvm-bugs
