[Bug 52393] New: Sacred 2 Gold: Textures largely missing since 5.0-rc3

[WineHQ Bugzilla]

https://bugs.winehq.org/show_bug.cgi?id=52393

Bug ID: 52393
           Summary: Sacred 2 Gold: Textures largely missing since 5.0-rc3
           Product: Wine
           Version: 6.23
          Hardware: x86-64
                OS: Linux
            Status: UNCONFIRMED
          Severity: normal
          Priority: P2
         Component: -unknown
          Assignee: [e-mail]
          Reporter: [e-mail]
      Distribution: ---

Created attachment 71619
  --> https://bugs.winehq.org/attachment.cgi?id=71619
main menu without background textures

I found a driver and gpu independent regression, which apparently affects just
Sacred 2 Gold (installer from GOG). Tested GPUs were AMD RX 5500XT (NAVI14)
with AmdVLK as well as RADV, Nvidia Quadro M2000M with nvidia and Intel HD
Graphics 520 (Skylake) with ANV.

Regression symptoms include the missing background tree in the main menu, a
blackscreen in singleplayer and missing non-plane textures in multiplayer (no
movement control however).
> From the Protondb reports I figured, that this worked originally on at least

Intel and Nvidia, so I started looking for a regression since the reported wine
versions.

Bisected this to f99d307a3e1f9beb7fd9dc8892b5cfabbabf816b
# first bad commit: [f99d307a3e1f9beb7fd9dc8892b5cfabbabf816b] msvcrt: Use
parse_double for scanf floats.

[Bug 52393] Sacred 2 Gold: Textures largely missing since 5.0-rc3

[WineHQ Bugzilla]

https://bugs.winehq.org/show_bug.cgi?id=52393

[e-mail] changed:

What    |Removed                     |Added
----------------------------------------------------------------------------
  Attachment #71619|main menu without           |bisect log
        description|background textures         |

[Bug 52393] Sacred 2 Gold: Textures largely missing since 5.0-rc3

[WineHQ Bugzilla]

https://bugs.winehq.org/show_bug.cgi?id=52393

--- Comment #1 from [e-mail] ---
Created attachment 71620
  --> https://bugs.winehq.org/attachment.cgi?id=71620
Main menu since regression commit

[Bug 52393] Sacred 2 Gold: Textures largely missing since 5.0-rc3

[WineHQ Bugzilla]

https://bugs.winehq.org/show_bug.cgi?id=52393

--- Comment #2 from [e-mail] ---
Created attachment 71621
  --> https://bugs.winehq.org/attachment.cgi?id=71621
Multiplayer start since regression commit

[Bug 52393] Sacred 2 Gold: Textures largely missing since 5.0-rc3

[WineHQ Bugzilla]

https://bugs.winehq.org/show_bug.cgi?id=52393

--- Comment #3 from [e-mail] ---
Created attachment 71622
  --> https://bugs.winehq.org/attachment.cgi?id=71622
Main menu in last good commit

[Bug 52393] Sacred 2 Gold: Textures largely missing since 5.0-rc3

[WineHQ Bugzilla]

https://bugs.winehq.org/show_bug.cgi?id=52393

--- Comment #4 from [e-mail] ---
Created attachment 71623
  --> https://bugs.winehq.org/attachment.cgi?id=71623
Multiplayer start in last good commit

[Bug 52393] Sacred 2 Gold: Textures largely missing since 5.0-rc3

[WineHQ Bugzilla]

https://bugs.winehq.org/show_bug.cgi?id=52393

--- Comment #5 from [e-mail] ---
Unfortunately I could not get it to run without DXVK, but besides that, it’s
pure wine (well, with Lutris runtime).

[Bug 52393] Sacred 2 Gold: Textures largely missing since 5.0-rc3

[WineHQ Bugzilla]

https://bugs.winehq.org/show_bug.cgi?id=52393

[e-mail] changed:

What    |Removed                     |Added
----------------------------------------------------------------------------
                URL|                            |https://www.protondb.com/ap
                   |                            |p/225640

--- Comment #6 from [e-mail] ---
Many of the reports in ProtonDB are related to missing PhysX, but a few are
suspecting their drivers and describing the same symptoms. As said, they are
occurring on all drivers for the three main vendors however.

Some of those are:
https://www.protondb.com/app/225640#cZiEsAQGJf
https://www.protondb.com/app/225640#9TGvCuCRiB
https://www.protondb.com/app/225640#Yahn76ETHd

[Bug 52393] Sacred 2 Gold: Textures largely missing since 5.0-rc3

[WineHQ Bugzilla]

https://bugs.winehq.org/show_bug.cgi?id=52393

Henri Verbeet  changed:

What    |Removed                     |Added
----------------------------------------------------------------------------
           Keywords|                            |regression
          Component|-unknown                    |msvcrt
    Regression SHA1|                            |f99d307a3e1f9beb7fd9dc8892b
                   |                            |5cfabbabf816b

--- Comment #7 from Henri Verbeet  ---
(In reply to labre from comment #5)
> Unfortunately I could not get it to run without DXVK, but besides that, it’s
> pure wine (well, with Lutris runtime).

I think this game used to work with upstream Wine. (And the AppDB appears to
confirm that.) If it really doesn't, perhaps that's another regression.

I notice your bisection log starts at Wine 5.0, and you entered 6.23 for the
reported version. Did you also try this with the current wine-7.0-rc5 version?

[Bug 52393] Sacred 2 Gold: Textures largely missing since 5.0-rc3

[WineHQ Bugzilla]

https://bugs.winehq.org/show_bug.cgi?id=52393

Piotr Caban  changed:

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

--- Comment #8 from Piotr Caban  ---
While running the game without dxvk I had to install d3dx9_36 to avoid a crash.
I'm not able to reproduce the missing textures problem both with dxvk and with
native d3dx9_36.

[Bug 52393] Sacred 2 Gold: Textures largely missing since 5.0-rc3

[WineHQ Bugzilla]

https://bugs.winehq.org/show_bug.cgi?id=52393

--- Comment #9 from Piotr Caban  ---
Please attach a log with "msvcrt" debug channel enabled. While collecting the
log, please quit the game as soon as menu with missing textures is displayed.

[Bug 52393] Sacred 2 Gold: Textures largely missing since 5.0-rc3

[WineHQ Bugzilla]

https://bugs.winehq.org/show_bug.cgi?id=52393

--- Comment #10 from [e-mail] ---
(In reply to Henri Verbeet from comment #7)
> (In reply to labre from comment #5)
> > Unfortunately I could not get it to run without DXVK, but besides that, it’s
> > pure wine (well, with Lutris runtime).
>
> I think this game used to work with upstream Wine. (And the AppDB appears to
> confirm that.) If it really doesn't, perhaps that's another regression.

Nope, it starts with d3dx9_36. Thanks for the hint @Piotr.
> I notice your bisection log starts at Wine 5.0, and you entered 6.23 for the
> reported version. Did you also try this with the current wine-7.0-rc5
> version?

Not yet, but I’ll try it on the weekend.

[Bug 52393] Sacred 2 Gold: Textures largely missing since 5.0-rc3

[WineHQ Bugzilla]

https://bugs.winehq.org/show_bug.cgi?id=52393

--- Comment #11 from [e-mail] ---
Created attachment 71631
  --> https://bugs.winehq.org/attachment.cgi?id=71631
Msvcrt debug log; compressed to stay in limit boundaries

Got it to run with d3dx9_36. However the bug is still reproducible for me.
Attaching the msvcrt debug log compressed for size reasons.

Sysinfo:
Gentoo amd64/hardened.
GCC:8.5.0 -pie -hardened ssp
These old Wine versions would not compile otherwise.
On:
Thinkpad P50 (Nvidia)
Thinkpad T460 (intel)
Custom desktop with Bulldozer era APU+RX5500XT (AMD)

[Bug 52393] Sacred 2 Gold: Textures largely missing since 5.0-rc3

[WineHQ Bugzilla]

https://bugs.winehq.org/show_bug.cgi?id=52393

--- Comment #12 from Piotr Caban  ---
It looks like scanf is working correctly in all calls from the log (I've
checked that native msvcr80 returns the same values as wine in all cases from
the log).

Could you please confirm that regression test was done correctly (e.g. by
running the game with git head set to f99d307a3e1f9beb7fd9dc8892b5cfabbabf816b
and f99d307a3e1f9beb7fd9dc8892b5cfabbabf816b^)?

[Bug 52393] Sacred 2 Gold: Textures largely missing since 5.0-rc3

[WineHQ Bugzilla]

https://bugs.winehq.org/show_bug.cgi?id=52393

[e-mail] changed:

What    |Removed                     |Added
----------------------------------------------------------------------------
            Version|6.23                        |5.0-rc3

[Bug 52393] Sacred 2 Gold: Textures largely missing since 5.0-rc3

[WineHQ Bugzilla]

https://bugs.winehq.org/show_bug.cgi?id=52393

--- Comment #13 from [e-mail] ---
Yes, I can confirm that.

wine --version gives me
5.0-rc3-22-g66c9c358ae5c # good
5.0-rc3-23-gf99d307a3e1f # bad

I also recreated the prefix from scratch with wine-vanilla-6.23, because
4.21/5.0 could not install physx.

I created a win32 prefix for 2.65.1 and a win64 prefix for 2.65.2 + community
patch and it is reproducible in both.
Used winetricks vcrun2005 physx d3dx9_36 win7 to set them up.

I also tried gcc-10, but unfortunately there were a few toolchain related
regressions in 5.0. The first version, which the Gentoo toolchain team
supported again, is 5.4. So I’m still compiling with gcc-8.5.0.
If my toolchain setup would be horribly wrong however, I probably could not
reproduce it with the lutris builds of 4.21 and 5.0.

I could provide the compiled binary packages (generic x64/32), if that helps
and I can post external links to, say, box.com, here.

I’ve created a diff of the msvcrt log from both versions, which I will attach
now.

[Bug 52393] Sacred 2 Gold: Textures largely missing since 5.0-rc3

[WineHQ Bugzilla]

https://bugs.winehq.org/show_bug.cgi?id=52393

[e-mail] changed:

What    |Removed                     |Added
----------------------------------------------------------------------------
  Attachment #71631|0                           |1
        is obsolete|                            |

--- Comment #14 from [e-mail] ---
Created attachment 71651
  --> https://bugs.winehq.org/attachment.cgi?id=71651
msvcrt log of the bad commit

[Bug 52393] Sacred 2 Gold: Textures largely missing since 5.0-rc3

[WineHQ Bugzilla]

https://bugs.winehq.org/show_bug.cgi?id=52393

--- Comment #15 from [e-mail] ---
Created attachment 71652
  --> https://bugs.winehq.org/attachment.cgi?id=71652
msvcrt log of the good commit

[Bug 52393] Sacred 2 Gold: Textures largely missing since 5.0-rc3

[WineHQ Bugzilla]

https://bugs.winehq.org/show_bug.cgi?id=52393

--- Comment #16 from [e-mail] ---
Created attachment 71653
  --> https://bugs.winehq.org/attachment.cgi?id=71653
diff of both logs

[Bug 52393] Sacred 2 Gold: Textures largely missing since 5.0-rc3

[WineHQ Bugzilla]

https://bugs.winehq.org/show_bug.cgi?id=52393

--- Comment #17 from [e-mail] ---
Created attachment 71654
  --> https://bugs.winehq.org/attachment.cgi?id=71654
warn+fixme channels for both versions

Just checked the warn+fixme log and I get a lot of GLSL warning stuff,
mentioning vs_out[n] possibly being used before initialization. Do you also get
these? I’m getting these in both versions however.

I’m currently testing with the Nvidia Quadro M2000M from the P50.
Mesa is 21.2.6 with nvidia-drivers 470.94.
The msvcrt logs were also captured with this system.

[Bug 52393] Sacred 2 Gold: Textures largely missing since 5.0-rc3

[WineHQ Bugzilla]

https://bugs.winehq.org/show_bug.cgi?id=52393

--- Comment #18 from [e-mail] ---
Forgot to mention that I also tested with yesterday’s git HEAD, although I lost
the specific commit. Unfortunately it’s also reproducible with it.

[Bug 52393] Sacred 2 Gold: Textures largely missing since 5.0-rc3

[WineHQ Bugzilla]

https://bugs.winehq.org/show_bug.cgi?id=52393

[e-mail] changed:

What    |Removed                     |Added
----------------------------------------------------------------------------
       Distribution|---                         |Gentoo

[Bug 52393] Sacred 2 Gold: Textures largely missing since 5.0-rc3

[WineHQ Bugzilla]

https://bugs.winehq.org/show_bug.cgi?id=52393

--- Comment #19 from Piotr Caban  ---
Created attachment 71665
  --> https://bugs.winehq.org/attachment.cgi?id=71665
part of regression commit that modifies parse_double function

I still can't see anything wrong in results returned by scanf. The patch also
makes some changes to parse_double helper, maybe the regression is caused by
it.

While reading C:\users\manu\AppData\Local\Ascaron Entertainment\Sacred
2\options.txt file volume_group01 value is different between good and bad log
(68 in good log and 69 in bad log (on my machine it's always 68)). I have no
clue if this file changes between runs.

Could you please try the attached patch applied on top of 66c9c358ae5c? I
wonder if this change already breaks the game.

[Bug 52393] Sacred 2 Gold: Textures largely missing since 5.0-rc3

[WineHQ Bugzilla]

https://bugs.winehq.org/show_bug.cgi?id=52393

--- Comment #20 from [e-mail] ---
(In reply to Piotr Caban from comment #19)
> Could you please try the attached patch applied on top of 66c9c358ae5c? I
> wonder if this change already breaks the game.

Sure, but I forgot my external system drive, so I get to this as early as this
weekend.

[Bug 52393] Sacred 2 Gold: Textures largely missing since 5.0-rc3

[WineHQ Bugzilla]

https://bugs.winehq.org/show_bug.cgi?id=52393

--- Comment #21 from [e-mail] ---
> Could you please try the attached patch applied on top of 66c9c358ae5c? I
> wonder if this change already breaks the game.

I applied the patch and it does not break the game. Has to be something else.

[Bug 52393] Sacred 2 Gold: Textures largely missing since 5.0-rc3

[WineHQ Bugzilla]

https://bugs.winehq.org/show_bug.cgi?id=52393

--- Comment #22 from [e-mail] ---
Created attachment 71712
  --> https://bugs.winehq.org/attachment.cgi?id=71712
Fixing patch against regression commit

Hm, I strongly suspect, that make_double/fpnum_double returns infinity due to
some unexpected or invalid double format, which might be given by the game,
leading to the textures being misplaced, well, by infinity.

The major difference made by the commit is, that the _control87 call is
removed, which allows most likely invalid double formats or maths operations.
Moving this call into the function and reverting it before any returns fixes
this bug.

It’s still odd, that you can’t reproduce it. At first I thought the decimal
separator might be the difference, because my locale (de_DE.utf8) uses comma,
however setting LC_ALL=en_US.utf8 in environment or registry did not make a
difference, even when installing the English version of the game. What version
of GCC/C library are you using and what is your locale?

I’m currently compiling wine-vanilla-7.0 with the ported version of the patch.
Let’s see, whether it works.

[Bug 52393] Sacred 2 Gold: Textures largely missing since 5.0-rc3

[WineHQ Bugzilla]

https://bugs.winehq.org/show_bug.cgi?id=52393

--- Comment #23 from [e-mail] ---
Created attachment 71715
  --> https://bugs.winehq.org/attachment.cgi?id=71715
Port attempt to 7.0 non-working

Nope, not working. Tried to guess the functions, which could be based on
make_double, but even my bruteforce approach is not working. Some traces should
be inserted to find the root problem, especially to find infinity returns,
which might be wrong.

[Bug 52393] Sacred 2 Gold: Textures largely missing since 5.0-rc3

[WineHQ Bugzilla]

https://bugs.winehq.org/show_bug.cgi?id=52393

--- Comment #24 from Piotr Caban  ---
(In reply to labre from comment #22)
> Created attachment 71712 [details]
> Fixing patch against regression commit

There are few problems with attached patch:
fpcontrol = _control87(0, 0);
_control87(MSVCRT__EM_DENORMAL|MSVCRT__EM_INVALID|MSVCRT__EM_ZERODIVIDE
    |MSVCRT__EM_OVERFLOW|MSVCRT__EM_UNDERFLOW|MSVCRT__EM_INEXACT, 0xffffffff);
if (!m) return sign * 0.0;
You're not restoring fp control word in m==0 case.

if (exp < -(1<<EXP_BITS))
{
    if (err) *err = MSVCRT_ERANGE;
    result = sign * 0.0;
    _control87(fpcontrol, 0xffffffff);
}
You have removed return sign * 0.0 in this if statement.

The patch is removing the _control87 calls because it's no longer needed. The
conversion is done using integers. All floating operations were removed except
of sign*0.0 and sign*INFINITY (and none of this calls depends on fp control
word value). Maybe it will be cleaner to change it to something like (sign == 1
? 0.0 : -0.0) and (sign == 1 ? INFINITY : -INFINITY) but it shouldn't matter.

[Bug 52393] Sacred 2 Gold: Textures largely missing since 5.0-rc3

[WineHQ Bugzilla]

https://bugs.winehq.org/show_bug.cgi?id=52393

[e-mail] changed:

What    |Removed                     |Added
----------------------------------------------------------------------------
  Attachment #71712|0                           |1
        is obsolete|                            |

--- Comment #25 from [e-mail] ---
Created attachment 71730
  --> https://bugs.winehq.org/attachment.cgi?id=71730
Fixing patch against regression commit

[Bug 52393] Sacred 2 Gold: Textures largely missing since 5.0-rc3

[WineHQ Bugzilla]

https://bugs.winehq.org/show_bug.cgi?id=52393

--- Comment #26 from [e-mail] ---
(In reply to Piotr Caban from comment #24)
> (In reply to labre from comment #22)
> > Created attachment 71712 [details]
> > Fixing patch against regression commit
>
> There are few problems with attached patch:

Addressed the problems and it still works.
> The patch is removing the _control87 calls because it's no longer needed.
> The conversion is done using integers. All floating operations were removed
> except of sign*0.0 and sign*INFINITY (and none of this calls depends on fp
> control word value). Maybe it will be cleaner to change it to something like
> (sign == 1 ? 0.0 : -0.0) and (sign == 1 ? INFINITY : -INFINITY) but it
> shouldn't matter.

I agree, that it shouldn’t matter, but at least in this commit it apparently
does. In wine-7.0 however it does indeed make no difference. Of course it has
been a long time since the regression commit. I probably could also bisect the
commit, which introduced that.

[Bug 52393] Sacred 2 Gold: Textures largely missing since 5.0-rc3

[WineHQ Bugzilla]

https://bugs.winehq.org/show_bug.cgi?id=52393

--- Comment #27 from Piotr Caban  ---
Created attachment 71736
  --> https://bugs.winehq.org/attachment.cgi?id=71736
float scanf tests with 24-bit precision and status word check

Could you please attach output of msvcr90_test.exe compiled with attached patch
(applied on top of current wine)? The patch adds some tests for +-0, +-INFINITY
and random float while using 24-bit precision (the game uses 24-bit precision
mode).

Here's output on my machine:
msvcr90.c:2370: 0 (0)
msvcr90.c:2374: 80000000 (0)
msvcr90.c:2378: 7f800000 (5)
msvcr90.c:2382: 7f800000 (0)
msvcr90.c:2386: ff800000 (5)
msvcr90.c:2390: ff800000 (0)
msvcr90.c:2394: 43a313b6 (1)
0020:msvcr90: 82 tests executed (0 marked as todo, 0 failures), 0 skipped.

[Bug 52393] Sacred 2 Gold: Textures largely missing since 5.0-rc3

[WineHQ Bugzilla]

https://bugs.winehq.org/show_bug.cgi?id=52393

--- Comment #28 from Piotr Caban  ---
Created attachment 71737
  --> https://bugs.winehq.org/attachment.cgi?id=71737
another approach on setting control word befory every operation on floats

(In reply to labre from comment #25)
> Created attachment 71730 [details]
> Fixing patch against regression commit

You can also try the attached patch (to be applied on top of
f99d307a3e1f9beb7fd9dc8892b5cfabbabf816b). Is it still fixing the regression?
Is yes, it should be possible to find what part of the patch fixes it by
commenting some control87 calls.

Taking in account what's in the logs I guess that:
result = *((double*)&bits);
assignment is causing problems.

Could you please attach msvcrt.dll (or msvcrt.dll.so in case of no-mingw build)
from your build?

[Bug 52393] Sacred 2 Gold: Textures largely missing since 5.0-rc3

[WineHQ Bugzilla]

https://bugs.winehq.org/show_bug.cgi?id=52393

--- Comment #29 from [e-mail] ---
(In reply to Piotr Caban from comment #27)
> Created attachment 71736 [details]
> float scanf tests with 24-bit precision and status word check
>
> Could you please attach output of msvcr90_test.exe compiled with attached
> patch (applied on top of current wine)? The patch adds some tests for +-0,
> +-INFINITY and random float while using 24-bit precision (the game uses
> 24-bit precision mode).
>
> Here's output on my machine:
> msvcr90.c:2370: 0 (0)
> msvcr90.c:2374: 80000000 (0)
> msvcr90.c:2378: 7f800000 (5)
> msvcr90.c:2382: 7f800000 (0)
> msvcr90.c:2386: ff800000 (5)
> msvcr90.c:2390: ff800000 (0)
> msvcr90.c:2394: 43a313b6 (1)
> 0020:msvcr90: 82 tests executed (0 marked as todo, 0 failures), 0 skipped.

Okay, here is the output from patched wine-7.0 compiled with gcc-11; 64-bit
prefix.
msvcr90.c:2370: 0 (0)
msvcr90.c:2374: 80000000 (0)
msvcr90.c:2378: 7f800000 (5)
msvcr90.c:2382: 7f800000 (0)
msvcr90.c:2386: ff800000 (5)
msvcr90.c:2390: ff800000 (0)
msvcr90.c:2394: 43a313b6 (1)
00f4:msvcr90: 82 tests executed (0 marked as todo, 0 failures), 0 skipped.
Seems to be the same, except for the last line prefix.

[Bug 52393] Sacred 2 Gold: Textures largely missing since 5.0-rc3

[WineHQ Bugzilla]

https://bugs.winehq.org/show_bug.cgi?id=52393

--- Comment #30 from [e-mail] ---
Ah, okay. You used a 32 bit prefix, due to the game.
msvcr90.c:2370: 0 (0)
msvcr90.c:2374: 80000000 (0)
msvcr90.c:2378: 7f800000 (5)
msvcr90.c:2382: 7f800000 (0)
msvcr90.c:2386: ff800000 (5)
msvcr90.c:2390: ff800000 (0)
msvcr90.c:2394: 43a313b6 (1)
0020:msvcr90: 82 tests executed (0 marked as todo, 0 failures), 0 skipped.
32-bit on my machine.

[Bug 52393] Sacred 2 Gold: Textures largely missing since 5.0-rc3

[WineHQ Bugzilla]

https://bugs.winehq.org/show_bug.cgi?id=52393

--- Comment #31 from [e-mail] ---
Created attachment 71739
  --> https://bugs.winehq.org/attachment.cgi?id=71739
Fixing patch minimized

So, we’re requiring this in sign * 0.0, if the mantissa is zero. Maybe sign
becomes NaN. The trace wouldn’t catch this case.

[Bug 52393] Sacred 2 Gold: Textures largely missing since 5.0-rc3

[WineHQ Bugzilla]

https://bugs.winehq.org/show_bug.cgi?id=52393

[e-mail] changed:

What    |Removed                     |Added
----------------------------------------------------------------------------
  Attachment #71730|0                           |1
        is obsolete|                            |

--- Comment #32 from [e-mail] ---
Comment on attachment 71730
  --> https://bugs.winehq.org/attachment.cgi?id=71730
Fixing patch against regression commit

Obsoleted by minimized version

[Bug 52393] Sacred 2 Gold: Textures largely missing since 5.0-rc3

[WineHQ Bugzilla]

https://bugs.winehq.org/show_bug.cgi?id=52393

--- Comment #33 from [e-mail] ---
(In reply to labre from comment #31)
> Created attachment 71739 [details]
> Fixing patch minimized
>
> So, we’re requiring this in sign * 0.0, if the mantissa is zero. Maybe sign
> becomes NaN. The trace wouldn’t catch this case.

On the other hand, a TRACE("sign is %d", sign); before it shows, that it is
always 1 or -1, as expected. Hm, I dunno… Have to sleep.

[Bug 52393] Sacred 2 Gold: Textures largely missing since 5.0-rc3

[WineHQ Bugzilla]

https://bugs.winehq.org/show_bug.cgi?id=52393

--- Comment #34 from [e-mail] ---
Tried a few things without success:

– Returning 0.0 instead of sign * 0.0 does not fix it.
– Result is the same between patched and unpatched.
> TRACE("result is now %f", result)
> return result;

gives zero differences after sort | uniq for both outputs.

– Had the assumption, that the code flow might be changed by a floating
exception, e.g. not returning 0.0, but unpatched always emits traces before and
after the calculation. So I’m not sure, what the _control87 function does, even
after looking up the Microsoft docs. Is there a way to catch occurring floating
exceptions with gdb? I have no experience with it, but I could find some
tutorials.

[Bug 52393] Sacred 2 Gold: Textures largely missing since 5.0-rc3

[WineHQ Bugzilla]

https://bugs.winehq.org/show_bug.cgi?id=52393

--- Comment #35 from Piotr Caban  ---
(In reply to labre from comment #34)
> – Returning 0.0 instead of sign * 0.0 does not fix it.

The game still works for me when 0.0 is returned (even so it's incorrect).
> – Had the assumption, that the code flow might be changed by a floating
> exception, e.g. not returning 0.0, but unpatched always emits traces before
> and after the calculation. So I’m not sure, what the _control87 function
> does, even after looking up the Microsoft docs. Is there a way to catch
> occurring floating exceptions with gdb? I have no experience with it, but I
> could find some tutorials.

There are no floating point exceptions. First of all, 0.0 * (any integer)
should not cause any exceptions. Also the game is already masking all floating
point exceptions. The control87 call affects 2 things:
 - it may affect how compiler optimizes code
 - it changes precision mode

You can try checking following things:
 - maybe adding _control87(0, 0) call is enough to make the bug gone
 - instead of changing whole control word you can try setting only the
precision (_control87(_PC_64, _MCW_PC)), it's the only setting changed by the
game
 - please upload msvcr80.dll (or msvcr80.dll.so in case of non-mingw build)
from wine f99d307a3e1f9beb7fd9dc8892b5cfabbabf816b
 - change "return sign * 0.0" to "return sign==-1 ? -0.0 : 0.0"

Are you using any non-standard compilation options or CPU specific
optimizations?

[Bug 52393] Sacred 2 Gold: Textures largely missing since 5.0-rc3

[WineHQ Bugzilla]

https://bugs.winehq.org/show_bug.cgi?id=52393

--- Comment #36 from [e-mail] ---
Created attachment 71750
  --> https://bugs.winehq.org/attachment.cgi?id=71750
msvcr80.dll.so from my unpatched regression commit build

Sorry for forgetting that request.

[Bug 52393] Sacred 2 Gold: Textures largely missing since 5.0-rc3

[WineHQ Bugzilla]

https://bugs.winehq.org/show_bug.cgi?id=52393

--- Comment #37 from [e-mail] ---
> Are you using any non-standard compilation options or CPU specific
> optimizations?

x86_64-pc-linux-gnu-gcc -m32 -c -o isfinite.o
/var/tmp/portage/app-emulation/wine-vanilla-9999/work/wine-9999/libs/port/isfinite.c
\
  -I.
-I/var/tmp/portage/app-emulation/wine-vanilla-9999/work/wine-9999/libs/port
-I../../include \
  -I/var/tmp/portage/app-emulation/wine-vanilla-9999/work/wine-9999/include
-D__WINESRC__ \
  -D_REENTRANT -fno-PIC -Wall -pipe -fcf-protection=none -fno-stack-protector \
  -fno-tree-loop-distribute-patterns -fno-strict-aliasing
-Wdeclaration-after-statement -Wempty-body \
  -Wignored-qualifiers -Wno-packed-not-aligned -Wshift-overflow=2
-Wstrict-prototypes -Wtype-limits \
  -Wunused-but-set-parameter -Wvla -Wwrite-strings -Wpointer-arith -Wlogical-op
\
  -fno-omit-frame-pointer -march=x86-64 -mtune=generic -O2 -pipe
-U_FORTIFY_SOURCE -D_FORTIFY_SOURCE=0

This is what the portage ebuild generates. No custom-cflags or optimizations
beyond -O2 on my behalf.

# /etc/portage/make.conf/custom
CFLAGS="-march=x86-64 -mtune=generic -O2 -pipe"
CXXFLAGS="${CFLAGS}"

[Bug 52393] Sacred 2 Gold: Textures largely missing since 5.0-rc3

[WineHQ Bugzilla]

https://bugs.winehq.org/show_bug.cgi?id=52393

--- Comment #38 from [e-mail] ---
> You can try checking following things:
> - maybe adding _control87(0, 0) call is enough to make the bug gone

Nope, it wasn’t.
> - instead of changing whole control word you can try setting only the
>
> precision (_control87(_PC_64, _MCW_PC)), it's the only setting changed by
> the game

Nope, that did not help.
> - change "return sign * 0.0" to "return sign==-1 ? -0.0 : 0.0"

Did not help.

[Bug 52393] Sacred 2 Gold: Textures largely missing since 5.0-rc3

[WineHQ Bugzilla]

https://bugs.winehq.org/show_bug.cgi?id=52393

--- Comment #39 from [e-mail] ---
Created attachment 71751
  --> https://bugs.winehq.org/attachment.cgi?id=71751
Patches used for the above testing.

[Bug 52393] Sacred 2 Gold: Textures largely missing since 5.0-rc3

[WineHQ Bugzilla]

https://bugs.winehq.org/show_bug.cgi?id=52393

--- Comment #40 from [e-mail] ---
Using -O0 instead of the default -O2 with custom-cflags USE flag does not fix
it either. Here are the flags of that compilation.

x86_64-pc-linux-gnu-gcc -m32 -c -o dllfunc.o
/var/tmp/portage/app-emulation/wine-vanilla-9999/work/wine-9999/dll
s/strmbase/dllfunc.c \
  -I.
-I/var/tmp/portage/app-emulation/wine-vanilla-9999/work/wine-9999/dlls/strmbase
\
  -I../../include
-I/var/tmp/portage/app-emulation/wine-vanilla-9999/work/wine-9999/include \

-I/var/tmp/portage/app-emulation/wine-vanilla-9999/work/wine-9999/include/msvcrt
-D__WINESRC__ \
  -D_REENTRANT -fno-PIC -fno-builtin -fshort-wchar -Wall -pipe
-fcf-protection=none \
  -fno-stack-protector -fno-tree-loop-distribute-patterns -fno-strict-aliasing
\
  -Wdeclaration-after-statement -Wempty-body -Wignored-qualifiers
-Wno-packed-not-aligned \
  -Wshift-overflow=2 -Wstrict-prototypes -Wtype-limits
-Wunused-but-set-parameter -Wvla \
  -Wwrite-strings -Wpointer-arith -Wlogical-op -fno-omit-frame-pointer
-march=x86-64 -pipe -O0

[Bug 52393] Sacred 2 Gold: Textures largely missing since 5.0-rc3

[WineHQ Bugzilla]

https://bugs.winehq.org/show_bug.cgi?id=52393

--- Comment #41 from Piotr Caban  ---
Created attachment 71757
  --> https://bugs.winehq.org/attachment.cgi?id=71757
patches to test

I don't see anything wrong in your msvcr80 build.

I'm attaching 2 patches generated on top of
f99d307a3e1f9beb7fd9dc8892b5cfabbabf816b for testing:
 - first patch adds traces with x87/sse status and control words. Please attach
console output while running the game with the patch (the game should fail to
display menu textures with this patch).
 - second patch sets and resets control word in make_double function, it should
have the same performance impact as the patch that fixes the issue, maybe it's
a race

[Bug 52393] Sacred 2 Gold: Textures largely missing since 5.0-rc3

[WineHQ Bugzilla]

https://bugs.winehq.org/show_bug.cgi?id=52393

--- Comment #42 from [e-mail] ---
Created attachment 71761
  --> https://bugs.winehq.org/attachment.cgi?id=71761
Debug output patched with test1.

Assumed msvcrt debug channel.

[Bug 52393] Sacred 2 Gold: Textures largely missing since 5.0-rc3

[WineHQ Bugzilla]

https://bugs.winehq.org/show_bug.cgi?id=52393

--- Comment #43 from [e-mail] ---
> I don't see anything wrong in your msvcr80 build.

Glad to hear that.
> I'm attaching 2 patches generated on top of
> f99d307a3e1f9beb7fd9dc8892b5cfabbabf816b for testing:
> - first patch adds traces with x87/sse status and control words. Please
>
> attach console output while running the game with the patch (the game should
> fail to display menu textures with this patch).

Hope this helps.
> - second patch sets and resets control word in make_double function, it
>
> should have the same performance impact as the patch that fixes the issue,
> maybe it's a race

Yes, that works.

[Bug 52393] Sacred 2 Gold: Textures largely missing since 5.0-rc3

[WineHQ Bugzilla]

https://bugs.winehq.org/show_bug.cgi?id=52393

--- Comment #44 from Piotr Caban  ---
It turns out that the problem is caused by MXCSR register value. In non working
case, it's value is 0x108001f. On my machine it's always 0x8001f. The code for
saving/restoring x87 was incorrect and was also changing MXCSR. It means that
the regression commit fixed a bug that was allowing the application to run on
your machine.

I've tried setting MXCSR to 0x108001f as on your machine. It breaks menu
background as on your screenshot.

According to your logs MXCSR is not set to this value by msvcr* functions
(probably the game uses ldmxcsr assembly instruction to set it). Anyway, in
order to fix this bug, it will be needed to find what and why is setting MXCSR
register.

[Bug 52393] Sacred 2 Gold: Textures largely missing since 5.0-rc3

[WineHQ Bugzilla]

https://bugs.winehq.org/show_bug.cgi?id=52393

--- Comment #45 from Piotr Caban  ---
The 0x108001f value is not really mcxsr value, this is what _control87_2
returns (so it has _DN_FLUSH set).

[Bug 52393] Sacred 2 Gold: Textures largely missing since 5.0-rc3

[WineHQ Bugzilla]

https://bugs.winehq.org/show_bug.cgi?id=52393

--- Comment #46 from [e-mail] ---
(In reply to Piotr Caban from comment #44)
> It turns out that the problem is caused by MXCSR register value. In non
> working case, it's value is 0x108001f. On my machine it's always 0x8001f.
> The code for saving/restoring x87 was incorrect and was also changing MXCSR.
> It means that the regression commit fixed a bug that was allowing the
> application to run on your machine.

I’m not sure, what register means in this context. Are these variables declared
with the 'register' keyword, which might end up in a processor register?
> I've tried setting MXCSR to 0x108001f as on your machine. It breaks menu
> background as on your screenshot.

Glad to hear, that you could reproduce it. :)
> According to your logs MXCSR is not set to this value by msvcr* functions
> (probably the game uses ldmxcsr assembly instruction to set it). Anyway, in
> order to fix this bug, it will be needed to find what and why is setting
> MXCSR register.

So, what are possible candidates for this? Are they limited to other wine
components or could this be also driver/firmware/hardware related? Can I do
anything to help with that?

[Bug 52393] Sacred 2 Gold: Textures largely missing since 5.0-rc3

[WineHQ Bugzilla]

https://bugs.winehq.org/show_bug.cgi?id=52393

--- Comment #47 from [e-mail] ---
Oh, and might be helpful to have a minimal C program, which sets the keyword to
the working value, both for mitigating the bug temporarily without performance
impact and to see, whether and at which point something resets it.

[Bug 52393] Sacred 2 Gold: Textures largely missing since 5.0-rc3

[WineHQ Bugzilla]

https://bugs.winehq.org/show_bug.cgi?id=52393

--- Comment #48 from Piotr Caban  ---
Created attachment 71824
  --> https://bugs.winehq.org/attachment.cgi?id=71824
print mxcsr in relay traces

The mxcsr register is set on process startup so it's not something that can be
preset with another application. In order to fix the bug it will be needed to
find what and why sets it to incorrect value.

Probably the first step is to find why the game doesn't work in current wine
(it works for me). You can try to find what patch is breaking it again after
setting control word in scanf/make_double.

Another option is to run with attached patch and WINEDEBUG=relay. The patch
prints if _DN_FLUSH is set. The trace look like this:
002c:Call ntdll.NtQueryDefaultLocale(00000000,0031f41c) ret=7b02c3f2 mxcsr=0
After mxcsr is changed it will look like:
002c:Call ntdll.NtQueryDefaultLocale(00000000,0031f41c) ret=7b02c3f2 mxcsr=1
Hopefully it will help find what is setting the _DN_FLUSH flag.

[Bug 52393] Sacred 2 Gold: Textures largely missing since 5.0-rc3

[WineHQ Bugzilla]

https://bugs.winehq.org/show_bug.cgi?id=52393

--- Comment #49 from [e-mail] ---
(In reply to Piotr Caban from comment #48)
> The mxcsr register is set on process startup so it's not something that can
> be preset with another application. In order to fix the bug it will be
> needed to find what and why sets it to incorrect value.

Scraping my basic OS knowledge together, I figured this too in the meantime.
Due to this being an illegal OS operation, it would cause a segmentation fault
and because of that require an exploit to work; thus this is not an option.
> Probably the first step is to find why the game doesn't work in current wine
> (it works for me). You can try to find what patch is breaking it again after
> setting control word in scanf/make_double.

I’ll try bisecting with the _control86_2 patch after looking for openal related
causes.
> Another option is to run with attached patch and WINEDEBUG=relay. The patch
> prints if _DN_FLUSH is set. The trace look like this:
> 002c:Call ntdll.NtQueryDefaultLocale(00000000,0031f41c) ret=7b02c3f2 mxcsr=0
> After mxcsr is changed it will look like:
> 002c:Call ntdll.NtQueryDefaultLocale(00000000,0031f41c) ret=7b02c3f2 mxcsr=1
> Hopefully it will help find what is setting the _DN_FLUSH flag.

Could be openal.
0009:Call openal32.alcGetString(00000000,00001013) ret=027d614d mxcsr=0
0009:Ret  openal32.alcGetString() retval=f5cd7a58 ret=027d614d mxcsr=1
…
0009:Call openal32.alcOpenDevice(0032e590 "JACK Default") ret=027d6824 mxcsr=1
0009:Ret  openal32.alcOpenDevice() retval=f3305010 ret=027d6824 mxcsr=1
Could be related to it trying to open a Jack device. I have no pulseaudio
support. Also I’m using pipewire, so the Jack server is actually the Pipewire
Jack emulation, which could imply, that pipewire sets it. I’ll try recompiling
openal with pulseaudio support, which might cause other code paths in pipewire.

versions (revisions indicate ebuild fixes or build patches):
pipewire-0.3.44-r1
openal-1.21.1-r2

[Bug 52393] Sacred 2 Gold: Textures largely missing since 5.0-rc3

[WineHQ Bugzilla]

https://bugs.winehq.org/show_bug.cgi?id=52393

--- Comment #50 from [e-mail] ---
Created attachment 71825
  --> https://bugs.winehq.org/attachment.cgi?id=71825
relay log at mxcsr switch with ±100 lines context

Could not add the complete start, because it is still 14M with xz -9
compression. If you need more context lines, leave a note.

[Bug 52393] Sacred 2 Gold: Textures largely missing since 5.0-rc3

[WineHQ Bugzilla]

https://bugs.winehq.org/show_bug.cgi?id=52393

--- Comment #51 from [e-mail] ---
So this was obviously caused by the default portaudio backend of openal, which
tries to use a Jack Audio device, probably in combination with pipewire, but
since pipewire is an external process, it can’t set the flag. Possibly its
library sets it or openal itself.

The workaround is to set
#~/.config/alsoft.conf
[General]
drivers="alsa,-jack,-oss,-port,-null,-wave,"

[Bug 52393] Sacred 2 Gold: Textures largely missing since 5.0-rc3

[WineHQ Bugzilla]

https://bugs.winehq.org/show_bug.cgi?id=52393

--- Comment #52 from [e-mail] ---
Note: this also works in current wine-7.1 including wine-staging.

[Bug 52393] Sacred 2 Gold: Textures largely missing since 5.0-rc3

[WineHQ Bugzilla]

https://bugs.winehq.org/show_bug.cgi?id=52393

[e-mail] changed:

What    |Removed                     |Added
----------------------------------------------------------------------------
             Status|UNCONFIRMED                 |RESOLVED
         Resolution|---                         |NOTOURBUG

--- Comment #53 from [e-mail] ---
Note 2: Had a DXVK crash bug after the game loading screen (singleplayer or
multiplayer) with resolutions above 1080p, which has been fixed in 1.9.4.