[AArch64] Fold `cmp a, b; sub a, c; ccmp a, c` into `subs a, c; ccmp a, b`

[Kmeakin]

Consider this rust function for indexing into a slice:

https://godbolt.org/z/fsfdxPs8e

#[inline(never)]
pub fn get(slice: &[u8], range: std::ops::Range<usize>) -> Option<&[u8]> {
    slice.get(range)
}

on AArch64, this produces

;; 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:

;; 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

[llvmbot]

@llvm/issue-subscribers-backend-aarch64

Author: Karl Meakin (Kmeakin)

Consider this rust function for indexing into a slice:

https://godbolt.org/z/fsfdxPs8e

#[inline(never)]
pub fn get(slice: &[u8], range: std::ops::Range<usize>) -> Option<&[u8]> {
    slice.get(range)
}

on AArch64, this produces

;; 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 &lt;=&gt; range.start
        csel    x0, xzr, x8, lo  ;; ret.ptr = (range.start &gt; range.end || range.end &gt; 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:

;; 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 &lt;=&gt; range.start
        csel    x0, xzr, x8, ls    ;; ret.ptr = (range.start &gt; range.end || range.end &gt; slice.len) ? null : new_ptr
        mov     x1, x9             ;; ret.len = new_len
        ret