JIT: simple range check is not branchless

[EgorBo]

bool IsInRange(int value, int lower, int upper) => 
    value >= lower && value <= upper;

Current arm64 codegen:

; Method Program:IsInRange(int,int,int):ubyte:this (FullOpts)
G_M32893_IG01:  ;; offset=0x0000
            stp     fp, lr, [sp, #-0x10]!
            mov     fp, sp
G_M32893_IG02:  ;; offset=0x0008
            cmp     w1, w2
            blt     G_M32893_IG05
G_M32893_IG03:  ;; offset=0x0010
            cmp     w1, w3
            cset    x0, le
G_M32893_IG04:  ;; offset=0x0018
            ldp     fp, lr, [sp], #0x10
            ret     lr
G_M32893_IG05:  ;; offset=0x0020
            mov     w0, wzr
G_M32893_IG06:  ;; offset=0x0024
            ldp     fp, lr, [sp], #0x10
            ret     lr
; Total bytes of code: 44

Expected codgen:

G_M32893_IG01:  ;; offset=0x0000
            stp     fp, lr, [sp, #-0x10]!
            mov     fp, sp
G_M32893_IG02:  ;; offset=0x0008
            cmp     w1, w2
            ccmp    w1, w3, #0, ge
            cset    w1, le
            ret
G_M32893_IG06:  ;; offset=0x0024
            ldp     fp, lr, [sp], #0x10
            ret     lr
; Total bytes of code: 32

(based on https://godbolt.org/z/oqjfMPosj)

cc @jakobbotsch @dotnet/jit-contrib

Tagging subscribers to this area: @JulieLeeMSFT, @jakobbotsch
See info in area-owners.md if you want to be subscribed.

Issue Details

---

bool IsInRange(int value, int lower, int upper) => 
    value >= lower && value <= upper;

Current arm64 codegen:

; Method Program:IsInRange(int,int,int):ubyte:this (FullOpts)
G_M32893_IG01:  ;; offset=0x0000
            stp     fp, lr, [sp, #-0x10]!
            mov     fp, sp
G_M32893_IG02:  ;; offset=0x0008
            cmp     w1, w2
            blt     G_M32893_IG05
G_M32893_IG03:  ;; offset=0x0010
            cmp     w1, w3
            cset    x0, le
G_M32893_IG04:  ;; offset=0x0018
            ldp     fp, lr, [sp], #0x10
            ret     lr
G_M32893_IG05:  ;; offset=0x0020
            mov     w0, wzr
G_M32893_IG06:  ;; offset=0x0024
            ldp     fp, lr, [sp], #0x10
            ret     lr
; Total bytes of code: 44

Expected codgen:

G_M32893_IG01:  ;; offset=0x0000
            stp     fp, lr, [sp, #-0x10]!
            mov     fp, sp
G_M32893_IG02:  ;; offset=0x0008
        cmp     w1, w2
        ccmp    w1, w3, #0, ge
        cset    w1, le
        ret
G_M32893_IG06:  ;; offset=0x0024
            ldp     fp, lr, [sp], #0x10
            ret     lr

(based on https://godbolt.org/z/oqjfMPosj)

cc @jakobbotsch @dotnet/jit-contrib

Author:	EgorBo
Assignees:	-
Labels:	area-CodeGen-coreclr, untriaged
Milestone:	-

[EgorBo]

Ah, I presume it's BBJ_RETURN <cond> messes things up.

void IsInRange(int value, int lower, int upper)
{
    if (value >= lower && value <= upper)
        Console.WriteLine();
}

works as expected

[jakobbotsch]

cc @a74nh

[jakobbotsch]

Note that the handling of conversion from short circuiting to bitwise ops is done inside "optimize bools" phase, in optOptimizeCompareChainCondBlock, so likely that's where there we could enhance treatment of GT_RETURN. Ideally these kind of conversions would be PGO based since it's not clear that it is always an optimization to make your example branchless. That would require us to have reliable PGO data at the point of "optimize bools" though, which we don't have yet. The example works fine if the user themselves make it branchless (=> value >= lower & value <= upper).

@a74nh are you interested in working on it?

[a74nh]

Note that the handling of conversion from short circuiting to bitwise ops is done inside "optimize bools" phase, in optOptimizeCompareChainCondBlock, so likely that's where there we could enhance treatment of GT_RETURN.

Right, should be similar to the if conversion code.

@a74nh are you interested in working on it?

Yes, but not in the immediate future as there is a lot of SVE work to do.

[AndyAyersMS]

Seems like we should move this one out of .NET 9?

[a74nh]

Seems like we should move this one out of .NET 9?

Agreed. We're not going to get to this in time.