push_dc_driver() places drivers based on their priorities, so the newly created driver is not
necessarily on top. Thus in windrv_CreateDC(), find_dc_driver() should be used to find the DIB
driver instead of assuming the DIB driver is the top driver, which could be the path driver because
it has a higher priority.

The exact wrong code path was:
1. A path driver with priority 400 is created for a DC.
2. windrv_CreateDC() is called to create a window driver for the DC.
3. Then in dibdrv_CreateDC(), push_dc_driver() is called with 'dev' pointing to the top driver, which is the path driver.
4. push_dc_driver() updates 'dev' to point to the address of the next driver because DIB driver has a lower 300 priority.
5. The DIB driver is assigned to 'dev', which is not the original parameter passed into push_dc_driver().
6. In windrv_CreateDC(), get_dibdrv_pdev(*dev) is called, assuming the top driver is the DIB driver. But actually the top
   driver that '*dev' points to is still the path driver.

The added tests can demonstrate the memory corruption before this fix is applied.

Wine-Bug: https://bugs.winehq.org/show_bug.cgi?id=55185

Signed-off-by: Nikolay Sivov <nsivov@codeweavers.com>

LLVM no longer allows non-private labels to appear between
.cfi_startproc/endproc when targeting Mach-O.

Based on a patch by Jacek Caban.

Wine-Bug: https://bugs.winehq.org/show_bug.cgi?id=55863

Wine-Bug: https://bugs.winehq.org/show_bug.cgi?id=55920Signed-off-by: Myah Caron <qsniyg@protonmail.com>

In rare cases we get an event during the second iteration of the loop,
changing the first element. This would lead to an assertion later.

Wine-Bug: https://bugs.winehq.org/show_bug.cgi?id=55054

It was confusing before since it made it seem like it might use the outer
window, while in fact the document is unchanged on native. Now the "new"
doc is used for navigating, since it's already checked to be the same as
the iframes_doc (but that test fails in wine and is todo_wine).

The idea is to reuse the existing code to handle SRVs, which simplifies the GL
code and essentially allows the Vulkan code to work "for free" (which is to say,
by writing this patch, rather than by adding support for flat textures to the
Vulkan renderer.)

This is a large patch; it consists the following parts:

* Create identity SRVs for d3d 1-9 textures. Store those in
  state->shader_resource_view instead of in state->texture.

* (Re)use wined3d_context_gl_bind_shader_resources() instead of state_sampler()
  to bind them.

  - Introduce code to that function to handle FFP textures.

  - Bind the sRGB texture if necessary in wined3d_shader_resource_view_gl_bind.

* (Re)use context_gl_load_shader_resources() instead of
  context_preload_textures() to load them.

  - Introduce code to that function to handle FFP textures.

  - Load the sRGB texture if necessary.

  - Port the SRV/RTV feedback loop check from context_preload_textures().

* Invalidate STATE_GRAPHICS_SHADER_RESOURCE_BINDING in places that now need to
  account for texture binding being guarded by that state instead of
  STATE_SAMPLER.

  Transitioning the remaining users of STATE_SAMPLER to
  STATE_GRAPHICS_SHADER_RESOURCE_BINDING, and removing STATE_SAMPLER, is left
  for future patches.

Allow us to avoid grabbing a temporary reference. This becomes a problem with
the next patch, where we would otherwise grab a reference while a texture is
being destroyed, and hence destroy it twice.

Currently we invalidate STATE_SAMPLER whenever an SRV is bound, and hence rely
on sampler() to do this for us, which is a bit obscure and won't work with the
next patch.