[Bug 54660] New: Call of Juarez shows 'Assertion failed: deref->offset_regset == HLSL_REGSET_NUMERIC'

[WineHQ Bugzilla]

https://bugs.winehq.org/show_bug.cgi?id=54660

Bug ID: 54660
           Summary: Call of Juarez shows 'Assertion failed:
                    deref->offset_regset == HLSL_REGSET_NUMERIC'
           Product: vkd3d
           Version: 1.6
          Hardware: x86-64
                OS: Linux
            Status: NEW
          Severity: normal
          Priority: P2
         Component: hlsl
          Assignee: [e-mail]
          Reporter: [e-mail]
      Distribution: ---

Assertion failed: deref->offset_regset == HLSL_REGSET_NUMERIC, file 
../vkd3d/libs/vkd3d-shader/hlsl_codegen.c, line 3143

1.6-199-gda243570

[Bug 54660] Call of Juarez shows 'Assertion failed: deref->offset_regset == HLSL_REGSET_NUMERIC'

[WineHQ Bugzilla]

https://bugs.winehq.org/show_bug.cgi?id=54660

--- Comment #1 from Nikolay Sivov  ---
Please attach +d3dcompiler,+vkd3d log, I'd like to see what's happening in the
shader to trigger this.

[Bug 54660] Call of Juarez shows 'Assertion failed: deref->offset_regset == HLSL_REGSET_NUMERIC'

[WineHQ Bugzilla]

https://bugs.winehq.org/show_bug.cgi?id=54660

--- Comment #2 from Andrey Gusev  ---
Created attachment 74192
  --> https://bugs.winehq.org/attachment.cgi?id=74192
+d3dcompiler,+vkd3d

[Bug 54660] Call of Juarez shows 'Assertion failed: deref->offset_regset == HLSL_REGSET_NUMERIC'

[WineHQ Bugzilla]

https://bugs.winehq.org/show_bug.cgi?id=54660

Francisco Casas  changed:

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

--- Comment #3 from Francisco Casas  ---
In the log I see three shaders that fail to compile.

These relate to bugs #54647, #54648, and this one, respectively.

These are the first lines in the output assembly of the last one:

out half4 output : COLOR0
      2:     float2 | <input-TEXCOORD0>
      3:            | = (input.xy @2)
  ...

The problem is that, in this shader, the variable "input" is a struct, so it
shouldn't be stored to directly (without indexation).

The assertion fails when the call to write_sm4_store() for instruction 3, calls
write_sm4_store(), which calls sm4_register_from_deref().

Currently, deref->offset_regset is 0 for derefs to structs. Which makes sense
since:

/* Instructions that directly refer to structs or arrays (instead of
single-register components)
    * are removed later by dce. So it is not a problem to just cleanup their
derefs. */

[Bug 54660] Call of Juarez shows 'Assertion failed: deref->offset_regset == HLSL_REGSET_NUMERIC'

[WineHQ Bugzilla]

https://bugs.winehq.org/show_bug.cgi?id=54660

--- Comment #4 from Francisco Casas  ---
The problem is that we need support for input semantics in arrays, so that in a
shader like this:

struct pixelInput
  {
      float2 uvTexcoords[8] : TEXCOORD0;
  };

void main(in pixelInput input)
  {
    ...
  }

uvTexcoords[0] is mapped to TEXCOORD0, uvTexcoords[1] to TEXCOORD1,
uvTexcoords[2] to TEXCOORD2, and so on...

[Bug 54660] Call of Juarez shows 'Assertion failed: deref->offset_regset == HLSL_REGSET_NUMERIC'

[WineHQ Bugzilla]

https://bugs.winehq.org/show_bug.cgi?id=54660

--- Comment #5 from Francisco Casas  ---
Properly supporting arrays (while at it, also multi-dimensional ones) in
prepend_input_copy() and append_output_copy() solves that problem and
derivative ones.

I am cleaning a patch series for that.

However, after fixing it, these fixmes arise when trying to compile this third
shader:

E5017: Aborting due to not yet implemented feature: Dereference with
non-constant offset of type HLSL_IR_EXPR.
E5017: Aborting due to not yet implemented feature: SM4 combined sample
expression.
E5017: Aborting due to not yet implemented feature: SM4 cast to half.

[Bug 54660] Call of Juarez shows 'Assertion failed: deref->offset_regset == HLSL_REGSET_NUMERIC'

[WineHQ Bugzilla]

https://bugs.winehq.org/show_bug.cgi?id=54660

--- Comment #6 from Nikolay Sivov  ---
(In reply to Francisco Casas from comment #5)
> Properly supporting arrays (while at it, also multi-dimensional ones) in
> prepend_input_copy() and append_output_copy() solves that problem and
> derivative ones.
>
> I am cleaning a patch series for that.
>
> However, after fixing it, these fixmes arise when trying to compile this
> third shader:
>
> E5017: Aborting due to not yet implemented feature: Dereference with
> non-constant offset of type HLSL_IR_EXPR.

That's bug 54832.
> E5017: Aborting due to not yet implemented feature: SM4 combined sample
> expression.

That's addressed with https://gitlab.winehq.org/wine/vkd3d/-/merge_requests/209
once merged.
> E5017: Aborting due to not yet implemented feature: SM4 cast to half.

This one should be fixed with
https://source.winehq.org/git/vkd3d.git/commit/24c1eb562f1493f60e6e93514b2ef....

So right now, isnan() is definitely still missing (bug 54648).