Suboptimal code generated for even a trivial ValueConverter

[hez2010]

Description

I'm seeing suboptimal code generated for ValueConverters even when what the converter does is trivial.

For example,

class Program
{
    private static readonly Foo foo = new();
    private static readonly CultureInfo culture = CultureInfo.CurrentCulture;

    public int Test()
    {
        var x = 3;
        var y = (int)foo.Convert(x, typeof(int), null, culture);
        return y;
    }
}

interface IValueConverter
{
    object Convert(object value, Type targetType, object? parameter, CultureInfo? culture);
}

sealed class Foo : IValueConverter
{
    public object Convert(object value, Type targetType, object? parameter, CultureInfo? culture)
    {
        return (int)value + 1;
    }
}

The codegen:

; Assembly listing for method Program:Test():int:this
; Emitting BLENDED_CODE for X64 CPU with AVX - Windows
; optimized code
; rsp based frame
; partially interruptible
; No PGO data
; 0 inlinees with PGO data; 1 single block inlinees; 0 inlinees without PGO data

G_M000_IG01:                ;; offset=0000H
       push     rdi
       push     rsi
       sub      rsp, 40
 
G_M000_IG02:                ;; offset=0006H
       mov      rsi, 0x7FF7BD400858
       mov      rcx, rsi
       call     CORINFO_HELP_NEWSFAST
       mov      rdi, rax
       test     byte  ptr [(reloc 0x7ff7bd6f9344)], 1
       je       SHORT G_M000_IG05
 
G_M000_IG03:                ;; offset=0024H
       mov      rcx, 0x23591401DA8
       mov      rcx, gword ptr [rcx]
       mov      dword ptr [rdi+08H], 3
       cmp      byte  ptr [rcx], cl
       mov      rcx, rsi
       call     CORINFO_HELP_NEWSFAST
       mov      ecx, dword ptr [rdi+08H]
       inc      ecx
       mov      dword ptr [rax+08H], ecx
       mov      eax, dword ptr [rax+08H]
 
G_M000_IG04:                ;; offset=004DH
       add      rsp, 40
       pop      rsi
       pop      rdi
       ret      
 
G_M000_IG05:                ;; offset=0054H
       mov      rcx, 0x7FF7BD6F9310
       mov      edx, 4
       call     CORINFO_HELP_GETSHARED_NONGCSTATIC_BASE
       jmp      SHORT G_M000_IG03
 
; Total bytes of code 106

This is important because ValueConverters are heavily used in Desktop apps that use XAML (WPF, UWP, MAUI and etc.), and we want to minimize the overhead caused by ValueConverters.

Expected codegen:

mov eax, 4
ret

Configuration

.NET 8 preview 3

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

Issue Details

---

Description

I'm seeing suboptimal code generated for ValueConverters even when the what the converter does is trivial.

For example,

class Program
{
    private static readonly Foo foo = new();
    private static readonly CultureInfo culture = CultureInfo.CurrentCulture;

    public int Test()
    {
        var x = 3;
        var y = (int)foo.Convert(x, typeof(int), null, culture);
        return y;
    }
}

interface IValueConverter
{
    object Convert(object value, Type targetType, object? parameter, CultureInfo? culture);
}

sealed class Foo : IValueConverter
{
    public object Convert(object value, Type targetType, object? parameter, CultureInfo? culture)
    {
        return (int)value + 1;
    }
}

The codegen:

; Assembly listing for method Program:Test():int:this
; Emitting BLENDED_CODE for X64 CPU with AVX - Windows
; optimized code
; rsp based frame
; partially interruptible
; No PGO data
; 0 inlinees with PGO data; 1 single block inlinees; 0 inlinees without PGO data

G_M000_IG01:                ;; offset=0000H
       push     rdi
       push     rsi
       sub      rsp, 40
 
G_M000_IG02:                ;; offset=0006H
       mov      rsi, 0x7FF7BD400858
       mov      rcx, rsi
       call     CORINFO_HELP_NEWSFAST
       mov      rdi, rax
       test     byte  ptr [(reloc 0x7ff7bd6f9344)], 1
       je       SHORT G_M000_IG05
 
G_M000_IG03:                ;; offset=0024H
       mov      rcx, 0x23591401DA8
       mov      rcx, gword ptr [rcx]
       mov      dword ptr [rdi+08H], 3
       cmp      byte  ptr [rcx], cl
       mov      rcx, rsi
       call     CORINFO_HELP_NEWSFAST
       mov      ecx, dword ptr [rdi+08H]
       inc      ecx
       mov      dword ptr [rax+08H], ecx
       mov      eax, dword ptr [rax+08H]
 
G_M000_IG04:                ;; offset=004DH
       add      rsp, 40
       pop      rsi
       pop      rdi
       ret      
 
G_M000_IG05:                ;; offset=0054H
       mov      rcx, 0x7FF7BD6F9310
       mov      edx, 4
       call     CORINFO_HELP_GETSHARED_NONGCSTATIC_BASE
       jmp      SHORT G_M000_IG03
 
; Total bytes of code 106

This is important because ValueConverters are heavily used in Desktop apps that uses XAML (WPF, UWP, MAUI and etc.), and we want to minimize the overhead caused by ValueConverters.

Expected codegen:

mov eax, 4
ret

Configuration

.NET 8 preview 3

Author:	hez2010
Assignees:	-
Labels:	tenet-performance, area-CodeGen-coreclr
Milestone:	-

[EgorBo]

We call it "multi-use boxing" problem, #9118

[EgorBo]

cc @markples

[hez2010]

Note that the code in the original post is a trivial case. In real-world cases, there'll be some type tests in the converter to avoid casting exceptions.

For example,

if (value is int d) return d + 1;
else return default(int);

The return value is boxed.

[markples]

Thank you for the example.

• The order of the first allocation and type initialization check/call require the allocated object to be spilled. Of course, this is moot if the allocation is removed.
• The input allocation is unnecessary.
• The null check on foo (I think?) remains. I suppose readonly+nonnull in cctor could help here.
• The output allocation is unnecessary, and in the more complicated example there are two allocations, one on each path, making it a bit more complicated. (Well, I assume so.. either the source or JIT could merge them.)

I'll include this in #9118.