--
v11: vkd3d-shader/tpf: Add support for writing 'resinfo' instruction.
vkd3d-shader/tpf: Add support for writing 'sampleinfo' instruction.
vkd3d-shader/hlsl: Parse GetDimensions() method.
tests: Add some tests for GetDimensions().
https://gitlab.winehq.org/wine/vkd3d/-/merge_requests/218
DirectX's FFP normalizes the zero vector to the zero vector, but GLSL
normalizes them to all NaN.
This patch creates `ffp_normalize`, which normalizes vectors but
has DirectX's behavior on 0 vectors. Further patches in this set switch
other calls from GLSL's `normalize` to the 0-safe version.
This fixes [36564](https://bugs.winehq.org/show_bug.cgi?id=36564).
--
v2: wined3d: Use ffp_normalize in shader_glsl_ffp_vertex_lighting_footer.
wined3d: Use ffp_normalize in shader_glsl_ffp_vertex_lighting.
wined3d: Implement a zero-safe normalize function for FFP.
https://gitlab.winehq.org/wine/wine/-/merge_requests/3419
When the sample size is too small, we were releasing it from the memory,
but kept the next_sample pointer set. Later, when the transform removes
the sample from the allocator, its refcount was decremented one more
time, effectively leaking the sample.
--
https://gitlab.winehq.org/wine/wine/-/merge_requests/3427
On Thu Jul 27 20:38:34 2023 +0000, Zebediah Figura wrote:
> > Beyond the test included in the patches, I did roughly the following:
> ptr = VirtualAlloc(NULL 4096 \* 2, PAGE_NOACCESS); VirtualProtect(ptr,
> 4096, PAGE_READWRITE); params = ptr + 4096 - sizeof(\*params). And used
> that pointer. So far it doesn't access it beyond the end of accessible
> memory (that succeeds).
> Ah yes, that definitely sounds convincing to me :-)
> > It might happen that with some other values for unknown parameter it
> is using more data from the structure, but IMO exploring more details
> about those parameters, in the absence of any known app which uses those
> functions, looks like a waste of time to me. It is relatively easy here
> with nsi and was already explored for other functions so I didn't want
> to break this tradition in nsi, but in general it seems impractical to
> me to spend huge amount of time reversing the exact parameters of the
> undocumented functions which nothing is using directly. Not feasible for
> implementing larger APIs and the benefit is not obvious. And once
> something is using directly in a way we don't support we can rather
> easily look how and test with those parameters.
> Sure. I don't think we necessarily need to figure out the unknown
> parameter, or spend too much effort on reverse-engineering undocumented
> things in general, or closely matching undocumented APIs.
> I only get nervous about the prospect of implementing undocumented APIs
> that are incorrect in ways that would be subtle or hard to debug—wrong
> number of parameters, wrong struct sizes, and so on. To a degree, the
> fact that an application even uses such APIs means that it's a first
> place to look when debugging, but that's not exactly the kind of thing
> that's trivial to notice.
> > I could possibly add a fixme for nonzero unk parameter?
> That sounds like a unilaterally good idea to me, at least.
Now when this patchset is split in two this first part has only a stub FIXME there anyway; I will update the second part with the specific fixme for unknown parameter.
--
https://gitlab.winehq.org/wine/wine/-/merge_requests/3423#note_40573
> Beyond the test included in the patches, I did roughly the following: ptr = VirtualAlloc(NULL 4096 \* 2, PAGE_NOACCESS); VirtualProtect(ptr, 4096, PAGE_READWRITE); params = ptr + 4096 - sizeof(\*params). And used that pointer. So far it doesn't access it beyond the end of accessible memory (that succeeds).
Ah yes, that definitely sounds convincing to me :-)
> It might happen that with some other values for unknown parameter it is using more data from the structure, but IMO exploring more details about those parameters, in the absence of any known app which uses those functions, looks like a waste of time to me. It is relatively easy here with nsi and was already explored for other functions so I didn't want to break this tradition in nsi, but in general it seems impractical to me to spend huge amount of time reversing the exact parameters of the undocumented functions which nothing is using directly. Not feasible for implementing larger APIs and the benefit is not obvious. And once something is using directly in a way we don't support we can rather easily look how and test with those parameters.
Sure. I don't think we necessarily need to figure out the unknown parameter, or spend too much effort on reverse-engineering undocumented things in general, or closely matching undocumented APIs.
I only get nervous about the prospect of implementing undocumented APIs that are incorrect in ways that would be subtle or hard to debug—wrong number of parameters, wrong struct sizes, and so on. To a degree, the fact that an application even uses such APIs means that it's a first place to look when debugging, but that's not exactly the kind of thing that's trivial to notice.
> I could possibly add a fixme for nonzero unk parameter?
That sounds like a unilaterally good idea to me, at least.
--
https://gitlab.winehq.org/wine/wine/-/merge_requests/3423#note_40571
Beyond the test included in the patches, I did roughly the following: ptr = VirtualAlloc(NULL 4096 * 2, PAGE_NOACCESS); VirtualProtect(ptr, 4096, PAGE_READWRITE); params = ptr + 4096 - sizeof(*params). And used that pointer. So far it doesn't access it beyond the end of accessible memory (that succeeds). It might happen that with some other values for unknown parameter it is using more data from the structure, but IMO exploring more details about those parameters, in the absence of any known app which uses those functions, looks like a waste of time to me. It is relatively easy here with nsi and was already explored for other functions so I didn't want to break this tradition in nsi, but in general it seems impractical to me to spend huge amount of time reversing the exact parameters of the undocumented functions which nothing is using directly. Not feasible for implementing larger APIs and the benefit is not obvious. And once something is using directly in a way we don't support we
can rather easily look how and test with those parameters. I could possibly add a fixme for nonzero unk parameter?
--
https://gitlab.winehq.org/wine/wine/-/merge_requests/3423#note_40570