[AArch64] Need to adjust comparison with -1 to comparison with 0 when it is beneficial

[AZero13]

When I corrected an issue with getAArch64Cmp: the following changed:

; CHECK-SD-NEXT: ands w8, w8, w8, asr #31
; CHECK-SD-NEXT: csinv w8, w8, wzr, ge

to

; CHECK-SD-NEXT: and w8, w8, w8, asr #31
; CHECK-SD-NEXT: cmn w8, #1
; CHECK-SD-NEXT: csinv w8, w8, wzr, gt

Now, this is only one regression; the rest of the changes are improvements. However, it made me realize an issue: the compiler is peephole optimizing a compare with -1 but not because it is better, but because it thought -1 was not a legal immediate for a compare, which my patch corrects the compiler into realizing -1 is a valid cmp, so it changes it to a comparison with 0, which is beneficial in a tiny number of cases, like this one, but does more harm than good.

So the issue is not with this patch, but rather, we ought to peephole optimize a comparison with -1 to be a comparison with 0, if we know cse will do the job.

[llvmbot]

@llvm/issue-subscribers-backend-aarch64

Author: AZero13 (AZero13)

When I corrected an issue with getAArch64Cmp: the following changed:

; CHECK-SD-NEXT: ands w8, w8, w8, asr #31
; CHECK-SD-NEXT: csinv w8, w8, wzr, ge

to

; CHECK-SD-NEXT: and w8, w8, w8, asr #31
; CHECK-SD-NEXT: cmn w8, #1
; CHECK-SD-NEXT: csinv w8, w8, wzr, gt

Now, this is only one regression; the rest of the changes are improvements. However, it made me realize an issue: the compiler is peephole optimizing a compare with -1 but not because it is better, but because it thought -1 was not a legal immediate for a compare, which my patch corrects the compiler into realizing -1 is a valid cmp, so it changes it to a comparison with 0, which is beneficial in a tiny number of cases, like this one, but does more harm than good.

So the issue is not with this patch, but rather, we ought to peephole optimize a comparison with -1 to be a comparison with 0, if we know cse will do the job.