[PATCH v2 0/1] MR5070: win32u: Cache is_virtual_desktop.

[Gabriel Ivăncescu (＠insn)]

It can be substantially faster in some cases. Notepad++ (with several plugins) start up goes down from around ~1.03 sec to ~0.61 sec on my machine (with wineprefix active).

9b7669592d6f8b40976b571b70f8543777d35167 is what introduced this performance regression. I don't know why exactly, but it made my Notepad++ noticeably slower when launched from a file manager (so it's not just at wineprefix startup overhead), which is very annoying.

--
  v2: win32u: Cache is_virtual_desktop.

https://gitlab.winehq.org/wine/wine/-/merge_requests/5070

[PATCH v2 1/1] win32u: Cache is_virtual_desktop.

[Gabriel Ivăncescu]

From: Gabriel Ivăncescu

It can be substantially faster in some cases. Notepad++ (with several plugins)
start up goes down from around ~1.03 sec to ~0.61 sec on my machine (with
wineprefix active).

Signed-off-by: Gabriel Ivăncescu 
---
 dlls/win32u/ntuser_private.h |  1 +
 dlls/win32u/winstation.c     | 14 ++++++++++++--
 2 files changed, 13 insertions(+), 2 deletions(-)

diff --git a/dlls/win32u/ntuser_private.h b/dlls/win32u/ntuser_private.h
index 3b6cab5bdc9..0cffa8b2208 100644
--- a/dlls/win32u/ntuser_private.h
+++ b/dlls/win32u/ntuser_private.h
@@ -127,6 +127,7 @@ struct user_thread_info
     UINT                          spy_indent;             /* Current spy indent */
     BOOL                          clipping_cursor;        /* thread is currently clipping */
     DWORD                         clipping_reset;         /* time when clipping was last reset */
+    signed char                   cached_is_virtual_desktop;  /* ternary value: zero = unknown, negative = FALSE, positive = TRUE */
 };

C_ASSERT( sizeof(struct user_thread_info) <= sizeof(((TEB *)0)->Win32ClientInfo) );
diff --git a/dlls/win32u/winstation.c b/dlls/win32u/winstation.c
index 9df27517a43..70179d3c8f6 100644
--- a/dlls/win32u/winstation.c
+++ b/dlls/win32u/winstation.c
@@ -42,12 +42,21 @@ WINE_DECLARE_DEBUG_CHANNEL(win);

BOOL is_virtual_desktop(void)
 {
-    HANDLE desktop = NtUserGetThreadDesktop( GetCurrentThreadId() );
+    struct user_thread_info *thread_info = get_user_thread_info();
     USEROBJECTFLAGS flags = {0};
+    HANDLE desktop;
     DWORD len;
+    BOOL ret;
+
+    if (thread_info->cached_is_virtual_desktop)
+        return thread_info->cached_is_virtual_desktop > 0;

+    desktop = NtUserGetThreadDesktop( GetCurrentThreadId() );
     if (!NtUserGetObjectInformation( desktop, UOI_FLAGS, &flags, sizeof(flags), &len )) return FALSE;
-    return !!(flags.dwFlags & DF_WINE_VIRTUAL_DESKTOP);
+    ret = !!(flags.dwFlags & DF_WINE_VIRTUAL_DESKTOP);
+
+    thread_info->cached_is_virtual_desktop = ret ? 1 : -1;
+    return ret;
 }

/***********************************************************************
@@ -261,6 +270,7 @@ BOOL WINAPI NtUserSetThreadDesktop( HDESK handle )
     {
         struct user_thread_info *thread_info = get_user_thread_info();
         struct user_key_state_info *key_state_info = thread_info->key_state;
+        thread_info->cached_is_virtual_desktop = 0;
         thread_info->client_info.top_window = 0;
         thread_info->client_info.msg_window = 0;
         if (key_state_info) key_state_info->time = 0;

[Gabriel Ivăncescu (＠insn)]

Well I rebased it, although I know it's unlikely to go in without further investigation.

Can you reproduce similar numbers/improvements with an active prefix at least?

[Rémi Bernon (＠rbernon)]

Of course starting notepad in an active prefix is faster but I still don't reproduce any difference.

Do you see `X11DRV_UpdateDisplayDevices` being called (ie: the TRACE after `if (!force && !force_display_devices_refresh) return TRUE;`) when you run notepad in an active prefix?

[Gabriel Ivăncescu (＠insn)]

On Sun Feb 18 20:42:12 2024 +0000, Rémi Bernon wrote:
> Of course starting notepad in an active prefix is faster but I still
> don't reproduce any difference.
> Do you see `X11DRV_UpdateDisplayDevices` being called (ie: the TRACE
> after `if (!force && !force_display_devices_refresh) return TRUE;`) when
> you run notepad in an active prefix?

Nope. It's not called. But I'm talking about notepad++ (plus plus, not
the built in notepad), with a bunch of plugins (not sure if that
matters). Built in notepad launches fast enough that I can't see a
difference.

FWIW, I don't see it called when launching notepad++ in an active prefix
either, but it's still slow, so that can't be it.

[Rémi Bernon (＠rbernon)]

Recent changes in virtual desktop mode have skipped a couple of unnecessary cache invalidation, is this still making a difference?

As an additional prefix startup optimization I think we could write the display device modes to the registry all at once, as done in https://gitlab.winehq.org/wine/wine/-/merge_requests/5579.

[Gabriel Ivăncescu (＠insn)]

On Sat May  4 18:56:44 2024 +0000, Rémi Bernon wrote:
> Recent changes in virtual desktop mode have skipped a couple of
> unnecessary cache invalidation, is this still making a difference?
> As an additional prefix startup optimization I think we could write the
> display device modes to the registry all at once, as done in https://gitlab.winehq.org/wine/wine/-/merge_requests/5579.

Just tested now on an active prefix (so, no prefix startup):

Upstream: ~630 ms
With my patch on top: ~590 ms

Of note is that it's actually faster than before even with my patch
(great!), but I think the difference now is too small to warrant this,
so I'll close this MR. (well unless someone changes their mind to reopen
it, but I don't think it's worth it anymore)

Thanks!

[PATCH v2 0/1] MR5070: win32u: Cache is_virtual_desktop. - closed

[Gabriel Ivăncescu (＠insn)]

This merge request was closed by Gabriel Ivăncescu.