JIT: Optimize constant range tests

[EgorBo]

void Test(char c)
{
    if ((c >= 'a' && c <= 'z') || (c >= 'A' && c <= 'Z'))
        Console.WriteLine("a-zA-Z");

// also: "X < CNS1 || X > CNS2", etc
}

Current codegen:

; Assembly listing for method Prog:Test(ushort):this (FullOpts)
       movzx    rcx, dx
       cmp      ecx, 97
       jl       SHORT G_M56518_IG04
       cmp      ecx, 122
       jle      SHORT G_M56518_IG05
G_M56518_IG04:
       cmp      ecx, 65
       jl       SHORT G_M56518_IG07
       cmp      ecx, 90
       jg       SHORT G_M56518_IG07
G_M56518_IG05:
       mov      rcx, 0x13A00208D40      ; 'a-zA-Z'
       tail.jmp [System.Console:WriteLine(System.String)]
G_M56518_IG07:
       ret      
; Total bytes of code 40

New codegen:

; Assembly listing for method Prog:Test(ushort):this (FullOpts)
       movzx    rcx, dx
       mov      eax, ecx
       sub      eax, 97
       cmp      eax, 25
       jbe      SHORT G_M56518_IG04
       sub      ecx, 65
       cmp      ecx, 25
       ja       SHORT G_M56518_IG06
G_M56518_IG04:
       mov      rcx, 0x1E200008D40      ; 'a-zA-Z'
G_M56518_IG05:
       tail.jmp [System.Console:WriteLine(System.String)]
G_M56518_IG06:
       ret      
; Total bytes of code 38

Not a big size win (sometimes even a size regression actually), but less branches (takes PGO into account).

Alternative patterns:

• c < 'a' || c > 'z'
• c is >= 'a' and <= 'z' or >= 'A' and <= 'Z'

[ghost]

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

Issue Details

---

void Test(char c)
{
    if ((c >= 'a' && c <= 'z') || 
        (c >= 'A' && c <= 'Z'))
        Console.WriteLine("a-zA-Z");
}

Current codegen:

; Assembly listing for method Prog:Test(ushort):this (FullOpts)
       movzx    rcx, dx
       cmp      ecx, 97
       jl       SHORT G_M56518_IG04
       cmp      ecx, 122
       jle      SHORT G_M56518_IG05
G_M56518_IG04:
       cmp      ecx, 65
       jl       SHORT G_M56518_IG07
       cmp      ecx, 90
       jg       SHORT G_M56518_IG07
G_M56518_IG05:
       mov      rcx, 0x13A00208D40      ; 'a-zA-Z'
       tail.jmp [System.Console:WriteLine(System.String)]
G_M56518_IG07:
       ret      
; Total bytes of code 40

New codegen:

; Assembly listing for method Prog:Test(ushort):this (FullOpts)
       movzx    rcx, dx
       mov      eax, ecx
       sub      eax, 97
       cmp      eax, 25
       jbe      SHORT G_M56518_IG04
       sub      ecx, 65
       cmp      ecx, 25
       ja       SHORT G_M56518_IG06
G_M56518_IG04:
       mov      rcx, 0x1E200008D40      ; 'a-zA-Z'
G_M56518_IG05:
       tail.jmp [System.Console:WriteLine(System.String)]
G_M56518_IG06:
       ret      
; Total bytes of code 38

Not a big size win (sometimes even a size regression actually), but less branches (takes PGO into account).

Alternative patterns:

• c < 'a' || c > 'z'
• c is >= 'a' and <= 'z'

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

[EgorBo]

/azp list

[EgorBo]

/azp run runtime-coreclr outerloop, runtime-coreclr jitstress, Fuzzlyn

[azure-pipelines]

Azure Pipelines successfully started running 3 pipeline(s).

[EgorBo]

@AndyAyersMS @jakobbotsch @dotnet/jit-contrib PTAL, diffs are not too big because we only save 2 bytes on average, but we get rid of a branch, so, hopefully, it will be noticeable. Outerloop/Jitstress/fuzzlyn pass. Can be separately extended to:

• if (x >= 0 && x < arr.Length) -> if ((uint)x < arr.Length) (if arr is not null)
• if (x > 10 && x > 100) -> if (x > 100) (although, for this one it's better to be done in RedundantBranchOpts)

Motivated by https://github.com/ricomariani/wrath-othello benchmark.

[AndyAyersMS]

Nice.

I think it would be a bit more challenging to fit this sort of thing into RBO because it modifies the upper branch and not the lower one. But still worth doing someday.

[EgorBo]

Failure is #93527

[rhuijben]

I assume it isn't worth the effort to detect+apply the bit fiddling to map lower to upper case (masking 1 bit) and then check a single range, like in the new ascii helpers?
That would also remove the branch, but at another level.

[EgorBo]

I assume it isn't worth the effort to detect+apply the bit fiddling to map lower to upper case (masking 1 bit) and then check a single range, like in the new ascii helpers? That would also remove the branch, but at another level.

I am not sure I understood, can you explaind it via code? 🙂