| Issue |
149685
|
| Summary |
[AArch64] Fold `cmp a, b; sub a, c; ccmp a, c` into `subs a, c; ccmp a, b`
|
| Labels |
backend:AArch64,
missed-optimization
|
| Assignees |
|
| Reporter |
Kmeakin
|
Consider this rust function for indexing into a slice:
https://godbolt.org/z/fsfdxPs8e
```rust
#[inline(never)]
pub fn get(slice: &[u8], range: std::ops::Range<usize>) -> Option<&[u8]> {
slice.get(range)
}
```
on AArch64, this produces
```asm
;; parameters:
;; x0: slice.ptr
;; x1: slice.len
;; x2: range.start
;; x3: range.end
;;
;; return:
;; x0: ret.ptr
;; x1: ret.len
example::get::h40fd170020d218c4:
cmp x3, x1 ;; range.end <=> slice.len
add x8, x0, x2 ;; new_ptr = slice.ptr + range.start
sub x1, x3, x2 ;; ret.len = range.end - range.start
ccmp x3, x2, #0, ls ;; range.end <=> range.start
csel x0, xzr, x8, lo ;; ret.ptr = (range.start > range.end || range.end > slice.len) ? null : new_ptr
ret
```
Notice that the comparison between `range.end` and `range.start` is performed twice:
* In `sub x1, x3, x2`, we calculate `range.end - range.start` (without setting flags)
* In `ccmp x3, x2, #0, ls`, we calculate `range.end - range.start` again, this time only to modify the flags
We could avoid the redundant calculation by performing the subtraction and setting the flags in the same instruction (`subs`) and using `ccmp` for comparing `range.end` with `slice.len`:
```asm
;; x0: slice.ptr
;; x1: slice.len
;; x2: range.start
;; x3: range.end
;;
;; return:
;; x0: ret.ptr
;; x1: ret.len
add x8, x0, x2 ;; new_ptr = slice.ptr + range.start
subs x9, x3, x2 ;; new_len = range.end - range.start and set flags
ccmp x3, x1, #0, lo ;; range.end <=> range.start
csel x0, xzr, x8, ls ;; ret.ptr = (range.start > range.end || range.end > slice.len) ? null : new_ptr
mov x1, x9 ;; ret.len = new_len
ret
```
![]()
_______________________________________________
llvm-bugs mailing list
llvm-bugs@lists.llvm.org
https://lists.llvm.org/cgi-bin/mailman/listinfo/llvm-bugs
