[Jit] Skip `test` instruction after `add`?

[benaadams]

The following C#

long lengthToExamine = length - Vector256<byte>.Count;
if (lengthToExamine != 0)
{

Produces asm:

add      r8, -32
test     r8, r8
je       SHORT G_M64667_IG11

However add sets flags so could this be?

add      r8, -32
jz       SHORT G_M64667_IG11

category:cq
theme:optimization
skill-level:expert
cost:medium

[GrabYourPitchforks]

Related: #10742, #6794

[AndyAyersMS]

@benaadams do you have a repro? I see this getting optimized in

    [MethodImpl(MethodImplOptions.NoInlining)]
    static int Foo(long x)
    {
   	long lengthToExamine = x - Vector256<byte>.Count;
	if (lengthToExamine != 0)
	{    
            return 1;
	}
	else
        {
	    return 0;
        }
    }

as

       4883C1E0             add      rcx, -32
       7406                 je       SHORT G_M34204_IG05

[benaadams]

gist outputs for SequenceEqualThreshold:SequenceEqualPR(byref,byref,long):bool

G_M14281_IG09:
       cmp      r8, 32
       jl       SHORT G_M14281_IG12
       xor      r9, r9
       add      r8, -32
       test     r8, r8
       je       SHORT G_M14281_IG11

and

G_M14281_IG12:
       cmp      r8, 16
       jl       SHORT G_M14281_IG15
       xor      r9, r9
       add      r8, -16
       test     r8, r8
       je       SHORT G_M14281_IG14

[AndyAyersMS]

Ah, in your case we have split trees, so OptimizeConstCompare can't see the add. Roslyn is the one that inspires this since there are multiple uses of lengthToExamine ...

***** BB14
STMT00038 (IL 0x0A2...0x0AA)
N005 (  3,  3) [000154] -A------R---              *  ASG       long   $202
N004 (  1,  1) [000153] D------N----              +--*  LCL_VAR   long   V06 loc3         d:2 $202
N003 (  3,  3) [000152] ------------              \--*  ADD       long   $202
N001 (  1,  1) [000149] ------------                 +--*  LCL_VAR   long   V02 arg2         u:1 (last use) $c0
N002 (  1,  1) [000151] ------------                 \--*  CNS_INT   long   -32 $106

***** BB14
STMT00039 (IL 0x0AB...0x0AC)
N004 (  5,  5) [000158] ------------              *  JTRUE     void  
N003 (  3,  3) [000157] J------N----              \--*  EQ        int    $186
N001 (  1,  1) [000155] ------------                 +--*  LCL_VAR   long   V06 loc3         u:2 $202
N002 (  1,  1) [000156] ------------                 \--*  CNS_INT   long   0 $100

If I add a second use in my simple example we produce

       488D41E0             lea      rax, [rcx-32]
       4885C0               test     rax, rax
       7401                 je       SHORT G_M42597_IG04

which would be even harder to clean up later on.

[AndyAyersMS]

Seems like another case for forward substitution -- seems like the tracking issue for this is #4655.

[AndyAyersMS]

Have looked into implementing forward sub and am not yet happy with any of the approaches. So am going to mark this as future.

[benaadams]

@AndyAyersMS is the instruction stream still mutable post emit?

e.g. lea is emitted; next is test could the test emit then undo the lea and change it to an add? (Since when the lea is being emitted it doesn't know next will be a test; but the test can look back)

[AndyAyersMS]

Perhaps? We are streaming other things along with emission, in particular gc liveness tracking, so "undoing" a previously emitted instruction may not be so easy.

[AndyAyersMS]

cc @kunalspathak as this just came up after a presentation Kunal gave that included the new arm64 peepholes...