Expand String.EndsWith/MemoryExtensions.EndsWith in JIT

[EgorBo]

It turns out it's not difficult to extend the existing *.StartsWith expansion to cover *.EndsWith as well. Example:

const string corelib = "System.Private.CoreLib.dll";

bool Test() => str.EndsWith(corelib, StringComparison.OrdinalIgnoreCase);

// Same for Span<char>/ReadOnlySpan<char> APIs

Old codegen:

; Assembly listing for method Program:Test
       mov      rdx, 0x20D80109480 ; 'System.Private.CoreLib.dll'
       mov      r8d, 5
       cmp      dword ptr [rcx], ecx
       tail.jmp [System.String:EndsWith(System.String,int):ubyte:this]

New codegen:

; Assembly listing for method Program:Test
       cmp      dword ptr [rcx+0x08], 26
       jl       SHORT G_M64019_IG04
       mov      eax, dword ptr [rcx+0x08]
       lea      rax, bword ptr [rcx+2*rax-0x34]
       vmovups  ymm0, ymmword ptr [rax+0x0C]
       vpor     ymm0, ymm0, ymmword ptr [reloc @RWD00]
       vpxor    ymm0, ymm0, ymmword ptr [reloc @RWD32]
       vmovups  ymm1, ymmword ptr [rax+0x20]
       vpor     ymm1, ymm1, ymmword ptr [reloc @RWD64]
       vpxor    ymm1, ymm1, ymmword ptr [reloc @RWD96]
       vpor     ymm0, ymm1, ymm0
       vptest   ymm0, ymm0
       sete     al
       movzx    rax, al
       jmp      SHORT G_M64019_IG05
G_M64019_IG04:
       xor      eax, eax
G_M64019_IG05:
       movzx    rax, al
       vzeroupper 
       ret      
RWD00  	dq	0020002000200020h, 0020000000200020h, ...
RWD32  	dq	0074007300790073h, 0070002E006D0065h, ...
RWD64  	dq	0020002000200020h, 0020002000200000h, ...
RWD96  	dq	0065007400610076h, 0072006F0063002Eh, ...

[ghost]

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

Issue Details

---

It turns out it's not difficult to extend the existing *.StartsWith expansion to cover *.EndsWith as well. Affected APIs:

const string corelib = "System.Private.CoreLib.dll";

bool Test() => str.EndsWith(corelib, StringComparison.OrdinalIgnoreCase);

// Same for Span<char>/ReadOnlySpan<char> APIs

Old codegen:

; Assembly listing for method Program:Test
       mov      rdx, 0x20D80109480 ; 'System.Private.CoreLib.dll'
       mov      r8d, 5
       cmp      dword ptr [rcx], ecx
       tail.jmp [System.String:EndsWith(System.String,int):ubyte:this]

New codegen:

; Assembly listing for method Program:Test
       cmp      dword ptr [rcx+0x08], 26
       jl       SHORT G_M64019_IG04
       mov      eax, dword ptr [rcx+0x08]
       lea      rax, bword ptr [rcx+2*rax-0x34]
       vmovups  ymm0, ymmword ptr [rax+0x0C]
       vpor     ymm0, ymm0, ymmword ptr [reloc @RWD00]
       vpxor    ymm0, ymm0, ymmword ptr [reloc @RWD32]
       vmovups  ymm1, ymmword ptr [rax+0x20]
       vpor     ymm1, ymm1, ymmword ptr [reloc @RWD64]
       vpxor    ymm1, ymm1, ymmword ptr [reloc @RWD96]
       vpor     ymm0, ymm1, ymm0
       vptest   ymm0, ymm0
       sete     al
       movzx    rax, al
       jmp      SHORT G_M64019_IG05
G_M64019_IG04:
       xor      eax, eax
G_M64019_IG05:
       movzx    rax, al
       vzeroupper 
       ret      
RWD00  	dq	0020002000200020h, 0020000000200020h, ...
RWD32  	dq	0074007300790073h, 0070002E006D0065h, ...
RWD64  	dq	0020002000200020h, 0020002000200000h, ...
RWD96  	dq	0065007400610076h, 0072006F0063002Eh, ...

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

[EgorBo]

@MihuBot

[EgorBo]

@AndyAyersMS @jakobbotsch @dotnet/jit-contrib PTAL. Simple change to extend StartsWith expansion with EndsWith. Basically, EndsWith is SequenceEquals where we shift the offset to dataPtr + (stringLength * 2 - constValueLength * 2)

[EgorBo]

Actually, some of the unrolling can probably be removed entirely since we already have a generic SequenceEquals unrolling in a late phase. But it's still needed for a bit complicated OrdinalIgnoreCase

[jakobbotsch]

LGTM otherwise

[EgorBo]

CI Failure is #98406