[Bug 57061] New: SRWLock is slower than CriticalSection on wine. Why?

[WineHQ Bugzilla]

https://bugs.winehq.org/show_bug.cgi?id=57061

Bug ID: 57061
           Summary: SRWLock is slower than CriticalSection on wine. Why?
           Product: Wine
           Version: unspecified
          Hardware: x86-64
                OS: Linux
            Status: UNCONFIRMED
          Severity: normal
          Priority: P2
         Component: ntdll
          Assignee: [e-mail]
          Reporter: [e-mail]
      Distribution: ---

https://github.com/cppfastio/fast_io/blob/next/benchmark/0002.file_io_tsc/ti...

[cqwrteur@DESKTOP-9B705LH time]$ wine
./iobuf_file_lockable_critical_section.exe 
output:0.3784567s
input:0.1407622s
[cqwrteur@DESKTOP-9B705LH time]$ rm *.txt
[cqwrteur@DESKTOP-9B705LH time]$ wine ./iobuf_file_lockable_srwlock.exe 
output:0.5857926s
input:0.3284965s
[cqwrteur@DESKTOP-9B705LH time]$ rm *.txt

This is very strange tbh. Because on windows SRWLock is faster than
CriticalSection. I have checked wine's implementation and I don't see why
SRWLock is slower than CriticalSection on wine either since the implementation
is very simple.

[Bug 57061] SRWLock is slower than CriticalSection on wine. Why?

[WineHQ Bugzilla]

https://bugs.winehq.org/show_bug.cgi?id=57061

cqwrteur  changed:

What    |Removed                     |Added
----------------------------------------------------------------------------
                 CC|                            |[e-mail]

[Bug 57061] SRWLock is slower than CriticalSection on wine. Why?

[WineHQ Bugzilla]

https://bugs.winehq.org/show_bug.cgi?id=57061

Zeb Figura  changed:

What    |Removed                     |Added
----------------------------------------------------------------------------
                 CC|                            |[e-mail]

--- Comment #1 from Zeb Figura  ---
What kind of numbers do you get on Windows (ideally on the same hardware?)

[Bug 57061] SRWLock is slower than CriticalSection on wine. Why?

[WineHQ Bugzilla]

https://bugs.winehq.org/show_bug.cgi?id=57061

--- Comment #2 from cqwrteur  ---
(In reply to Zeb Figura from comment #1)
> What kind of numbers do you get on Windows (ideally on the same hardware?)

D:\msys64\home\unlvs\projects\fast_io\benchmark\0002.file_io_tsc\time>clang++
-o iobuf_file_lockable_srwlock.exe iobuf_file_lockable.cc -O3 -std=c++26
-I../../../include --target=x86_64-windows-msvc
--sysroot=D:\toolchains\windows-msvc-sysroot -D_DLL=1 -lmsvcrt -fuse-ld=lld
-flto=thin

D:\msys64\home\unlvs\projects\fast_io\benchmark\0002.file_io_tsc\time>clang++
-o iobuf_file_lockable_critical_section.exe iobuf_file_lockable.cc -O3
-std=c++26 -I../../../include --target=x86_64-windows-msvc
--sysroot=D:\toolchains\windows-msvc-sysroot -D_DLL=1 -lmsvcrt -fuse-ld=lld
-flto=thin

D:\msys64\home\unlvs\projects\fast_io\benchmark\0002.file_io_tsc\time>del *.txt
Could Not Find
D:\msys64\home\unlvs\projects\fast_io\benchmark\0002.file_io_tsc\time*.txt

D:\msys64\home\unlvs\projects\fast_io\benchmark\0002.file_io_tsc\time>iobuf_file_lockable_srwlock.exe
output:0.1471754s
input:0.1363283s

D:\msys64\home\unlvs\projects\fast_io\benchmark\0002.file_io_tsc\time>del *.txt

D:\msys64\home\unlvs\projects\fast_io\benchmark\0002.file_io_tsc\time>iobuf_file_lockable_critical_section.exe
output:0.1584142s
input:0.1377649s

SRWLock is faster than CriticalSection.

I'll send you the binary

[Bug 57061] SRWLock is slower than CriticalSection on wine. Why?

[WineHQ Bugzilla]

https://bugs.winehq.org/show_bug.cgi?id=57061

--- Comment #3 from cqwrteur  ---
Created attachment 76931
  --> https://bugs.winehq.org/attachment.cgi?id=76931
CriticalSection vs SRWLock binary

Try yourself with both windows and wine.

Wine is much slower while windows is faster for SRWLock.

[Bug 57061] SRWLock is slower than CriticalSection on wine. Why?

[WineHQ Bugzilla]

https://bugs.winehq.org/show_bug.cgi?id=57061

--- Comment #4 from cqwrteur  ---
(In reply to Zeb Figura from comment #1)
> What kind of numbers do you get on Windows (ideally on the same hardware?)

I tried it on my android phone and I got the same result. SRWLock is very slow
for wine not just on x86_64 but aarch64 too.

[Bug 57061] SRWLock is slower than CriticalSection on wine. Why?

[WineHQ Bugzilla]

https://bugs.winehq.org/show_bug.cgi?id=57061

--- Comment #5 from cqwrteur  ---
The wine version is the git upstream, and I just compiled it from the source
yesterday.

[Bug 57061] SRWLock is slower than CriticalSection on wine. Why?

[WineHQ Bugzilla]

https://bugs.winehq.org/show_bug.cgi?id=57061

--- Comment #6 from cqwrteur  ---
Created attachment 76932
  --> https://bugs.winehq.org/attachment.cgi?id=76932
Another benchmark with leb128 instead of text

https://github.com/cppfastio/fast_io/blob/next/benchmark/0002.file_io_tsc/ti...

windows:
iobuf_file_lockable_leb128_critical_section.exe
output:0.1355566s
input:0.1343125s

iobuf_file_lockable_leb128_srwlock.exe
output:0.1290951s
input:0.1271444s

wine:
$ wine ./iobuf_file_lockable_leb128_critical_section.exe
output:0.1806937s
input:0.1704513s
$ wine ./iobuf_file_lockable_leb128_srwlock.exe
output:0.3658482s
input:0.3381468s

Same binary. SRWLock is nearly 2x slower.

[Bug 57061] SRWLock is slower than CriticalSection on wine. Why?

[WineHQ Bugzilla]

https://bugs.winehq.org/show_bug.cgi?id=57061

Fabian Maurer  changed:

What    |Removed                     |Added
----------------------------------------------------------------------------
                 CC|                            |[e-mail]

[Bug 57061] SRWLock is slower than CriticalSection on wine. Why?

[WineHQ Bugzilla]

https://bugs.winehq.org/show_bug.cgi?id=57061

William Horvath  changed:

What    |Removed                     |Added
----------------------------------------------------------------------------
                 CC|                            |[e-mail]

--- Comment #7 from William Horvath  ---
https://gitlab.winehq.org/wine/wine/-/merge_requests/3504

Just to add to the discussion, here are some results running these benchmarks
with and without this patchset applied:

Before:

CriticalSection
output:0.0854515s
input:0.0727398s

SRWLock
output:0.1495588s
input:0.1295463s

After:

CriticalSection
output:0.0900452s
input:0.0716042s

SRWLock
output:0.0934296s
input:0.0835009s

This was after reverting the patchset that superseded it (which doesn't help
this benchmark), and applying it to Wine 9.17.