JIT fails to merge trivial blocks after jump threading

[jakobbotsch]

[MethodImpl(MethodImplOptions.NoInlining)]
private static int Foo(bool cond)
{
    int x;
    if (cond)
    {
        x = 2;
    }
    else
    {
        x = 23;
    }

    int y;
    if (cond)
    {
        y = 2;
    }
    else
    {
        y = 23;
    }

    return x + y;
}

Output:

G_M7860_IG02:  ;; offset=0000H
       test     cl, cl
       je       SHORT G_M7860_IG04
                                                ;; size=4 bbWeight=1 PerfScore 1.25
G_M7860_IG03:  ;; offset=0004H
       mov      eax, 2
       jmp      SHORT G_M7860_IG05
                                                ;; size=7 bbWeight=0.50 PerfScore 1.12
G_M7860_IG04:  ;; offset=000BH
       mov      eax, 23
       jmp      SHORT G_M7860_IG06
                                                ;; size=7 bbWeight=1 PerfScore 2.25
G_M7860_IG05:  ;; offset=0012H
       mov      edx, 2
       jmp      SHORT G_M7860_IG07
                                                ;; size=7 bbWeight=0.50 PerfScore 1.12
G_M7860_IG06:  ;; offset=0019H
       mov      edx, 23
                                                ;; size=5 bbWeight=0.50 PerfScore 0.12
G_M7860_IG07:  ;; offset=001EH
       add      eax, edx
                                                ;; size=2 bbWeight=1 PerfScore 0.25
G_M7860_IG08:  ;; offset=0020H
       ret

The blocks resulting in IG03/IG05 and IG04/IG06 should have been merged at some point to avoid those unnecessary unconditional jumps.

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

Issue Details

---

[MethodImpl(MethodImplOptions.NoInlining)]
private static int Foo(bool cond)
{
    int x;
    if (cond)
    {
        x = 2;
    }
    else
    {
        x = 23;
    }

    int y;
    if (cond)
    {
        y = 2;
    }
    else
    {
        y = 23;
    }

    return x + y;
}

Output:

G_M7860_IG02:  ;; offset=0000H
       test     cl, cl
       je       SHORT G_M7860_IG04
                                                ;; size=4 bbWeight=1 PerfScore 1.25
G_M7860_IG03:  ;; offset=0004H
       mov      eax, 2
       jmp      SHORT G_M7860_IG05
                                                ;; size=7 bbWeight=0.50 PerfScore 1.12
G_M7860_IG04:  ;; offset=000BH
       mov      eax, 23
       jmp      SHORT G_M7860_IG06
                                                ;; size=7 bbWeight=1 PerfScore 2.25
G_M7860_IG05:  ;; offset=0012H
       mov      edx, 2
       jmp      SHORT G_M7860_IG07
                                                ;; size=7 bbWeight=0.50 PerfScore 1.12
G_M7860_IG06:  ;; offset=0019H
       mov      edx, 23
                                                ;; size=5 bbWeight=0.50 PerfScore 0.12
G_M7860_IG07:  ;; offset=001EH
       add      eax, edx
                                                ;; size=2 bbWeight=1 PerfScore 0.25
G_M7860_IG08:  ;; offset=0020H
       ret

IG03/IG05 and IG04/IG06 should have been merged at some point to avoid those unnecessary unconditional jumps.

Author:	jakobbotsch
Assignees:	-
Labels:	area-CodeGen-coreclr
Milestone:	-

[jl0pd]

Not sure if this is same issue, but Jit fails to optimize jump in following case:

.method static int32 IndexOf(int32[], int32)
    {
        .locals init (int32)
        
    Condition:
        ldloc.0
        ldarg.0
        ldlen        
        conv.i4
        bge NotFound
        
    Body:
        ldarg.0
        ldloc.0
        ldelem.i4
        ldarg.1
        beq Found
        
    Increment:
        ldloc.0
        ldc.i4.1
        add
        stloc.0
        br Condition

    Found:
        ldloc.0
        br Ret
        
    NotFound:
        ldc.i4.m1

    Ret:
        ret
    }

Produces

Program.IndexOf(Int32[], Int32)
    L0000: push ebp
    L0001: mov ebp, esp
    L0003: push esi
    L0004: xor eax, eax
    L0006: mov esi, [ecx+4]
    L0009: cmp esi, eax
    L000b: jle short L0018
    L000d: cmp [ecx+eax*4+8], edx
    L0011: je short L0016
    L0013: inc eax
    L0014: jmp short L0006
    L0016: jmp short L001d
    L0018: mov eax, 0xffffffff
    L001d: pop esi
    L001e: pop ebp
    L001f: ret

Where L0011: je short L0016 can be simplified to L0011: je short L001d, which will remove L0016 and save 1 unconditional jump

[AndyAyersMS]

@jakobbotsch 's example is now a bit cleaner. One might consider pushing the add up (eg if we're at a join and all phi values for some local operations are known in each pred (but different) consider tail duplicating the operations.

; Method Program:<<Main>$>g__Foo|0_0(ubyte):int (FullOpts)
G_M49102_IG01:  ;; offset=0x0000
						;; size=0 bbWeight=1 PerfScore 0.00

G_M49102_IG02:  ;; offset=0x0000
       movzx    rax, cl
       test     eax, eax
       jne      SHORT G_M49102_IG04
						;; size=7 bbWeight=1 PerfScore 1.50

G_M49102_IG03:  ;; offset=0x0007
       mov      eax, 23
       mov      ecx, 23
       jmp      SHORT G_M49102_IG05
						;; size=12 bbWeight=0.50 PerfScore 1.25

G_M49102_IG04:  ;; offset=0x0013
       mov      eax, 2
       mov      ecx, 2
						;; size=10 bbWeight=0.50 PerfScore 0.25

G_M49102_IG05:  ;; offset=0x001D
       add      eax, ecx
						;; size=2 bbWeight=1 PerfScore 0.25

G_M49102_IG06:  ;; offset=0x001F
       ret      
						;; size=1 bbWeight=1 PerfScore 1.00
; Total bytes of code: 32

@jl0pd 's example is now

G_M11661_IG01:  ;; offset=0x0000
       push     ebp
       mov      ebp, esp
       push     esi
       xor      eax, eax
                                                ;; size=6 bbWeight=1 PerfScore 2.50
G_M11661_IG02:  ;; offset=0x0006
       mov      esi, dword ptr [ecx+0x04]
       jmp      SHORT G_M11661_IG04
       align    [0 bytes for IG03]
                                                ;; size=5 bbWeight=1 PerfScore 4.00
G_M11661_IG03:  ;; offset=0x000B
       inc      eax
                                                ;; size=1 bbWeight=4 PerfScore 1.00
G_M11661_IG04:  ;; offset=0x000C
       cmp      esi, eax
       jle      SHORT G_M11661_IG07
                                                ;; size=4 bbWeight=8 PerfScore 10.00
G_M11661_IG05:  ;; offset=0x0010
       cmp      dword ptr [ecx+4*eax+0x08], edx
       jne      SHORT G_M11661_IG03
                                                ;; size=6 bbWeight=4 PerfScore 16.00
G_M11661_IG06:  ;; offset=0x0016
       pop      esi
       pop      ebp
       ret
                                                ;; size=3 bbWeight=1 PerfScore 2.00
G_M11661_IG07:  ;; offset=0x0019
       mov      eax, -1
       jmp      SHORT G_M11661_IG06
                                                ;; size=7 bbWeight=0.50 PerfScore 1.12

; Total bytes of code 32, prolog size 6, PerfScore 36.62, instruction count 17, allocated bytes for code 32 (MethodHash=3333d272) for method X:IndexOf(int[],int):int (FullOpts)

where there are no in-loop unconditional branches. Here one might like to see the loop be rotated and perhaps strength reduce the array access / down-count?

I think we can close this one.