[PATCH 2/5] d3dx9: Always use type->Rows as D3DXPC_MATRIX_ROWS parameters size. (resend)

[Matteo Bruni]

2014-03-05 17:37 GMT+01:00 Rico Schüller <kgbricola at web.de>:
> On 03.03.2014 17:48, Matteo Bruni wrote:
>>
>> diff --git a/dlls/d3dx9_36/tests/shader.c b/dlls/d3dx9_36/tests/shader.c
>> index 74ada6e..80c0f64 100644
>> --- a/dlls/d3dx9_36/tests/shader.c
>> +++ b/dlls/d3dx9_36/tests/shader.c
>> @@ -120,7 +120,7 @@ static const DWORD ctab_matrices[] = {
>>       0x000000ac,
>>       0x00000058, 0x00070002, 0x00000001, 0x00000064, 0x00000000,
>> /* Constant 1 desc (fmatrix3x1) */
>>       0x00000074, 0x00000002, 0x00000004, 0x00000080, 0x00000000,
>> /* Constant 2 desc (fmatrix4x4) */
>> -    0x00000090, 0x00040002, 0x00000003, 0x0000009c, 0x00000000,
>> /* Constant 3 desc (imatrix2x3) */
>> +    0x00000090, 0x00040002, 0x00000002, 0x0000009c, 0x00000000,
>> /* Constant 3 desc (imatrix2x3) */
>
> Why this? Was the shader blob wrong (invalid)? Or did you just compile with
> vs without optimization?
>

AFAICT that shader bytecode was hand-generated and specifically that
constant was wrongly given 3 registers instead of 2.

We still differ from native in that, in cases like this where a
constant is assigned more registers than it would naturally need, we
only allocate the minimum required, while native sometimes just uses
the register size stored in the shader (which is the reason this test
passed on native). Since that kind of overallocation doesn't seem to
happen in real shaders generated by D3DCompile() / fxc.exe I don't
think that is an issue we should care of.