Add support for GLSL interface blocks

[cheneym2]

This adds support for GLSL style syntax for shader inputs, e.g.:

in VertexData
{
vec2 texcoord_0;
vec2 texcoord_1;
} vd;

[csyonghe]

In slang-ir-translate-global-var.cpp:97, instead of creating a new field type layout for the global var, there should already be a layout created on input (stored in a IRLayoutDecoration), we should just use that as fieldTypeLayout instead of building it.

[cheneym2]

In slang-ir-translate-global-var.cpp:97, instead of creating a new field type layout for the global var, there should already be a layout created on input (stored in a IRLayoutDecoration), we should just use that as fieldTypeLayout instead of building it.

Okay, thanks, that seems to be working, though I don't know what to do about the call to "addResourceUsage" on the fieldTypeLayout builder which would go unused in the case that I reuse a layout from IRLayoutDecoration.

[csyonghe]

varLayoutBuilder takes a TypeLayout to initialize, so you just delete the existing fieldTypeLayout completely, and pass the type layout you get from input's layout decoration to initailize varLayoutBuilder.

[cheneym2]

varLayoutBuilder takes a TypeLayout to initialize, so you just delete the existing fieldTypeLayout completely, and pass the type layout you get from input's layout decoration to initailize varLayoutBuilder.

That's what I think I'm doing, and it mostly works. I'll just upload a commit to explain.

[cheneym2]

The doubt I was having was about this call to addResourceUsage. I think it doesn't make sense if I'm pulling fieldTypeLayout from IRLayoutDecoration.

[cheneym2]

glsl-intrinsic/fragment-processing/fragment-processing.slang is failing due to the missing shadowed globals replacement optimization. I believe the output of the test is still correct, but doesn't match expect output for that reason.

[cheneym2]

With the change to use tryReplaceUsesofStageInput for the case of nested input, compilation of the test case outputs this:

#version 450
layout(row_major) uniform;
layout(row_major) buffer;

#line 31 0
vec4 _S1;


#line 31
layout(location = 0)
out vec4 entryPointParam_fragmentMain_out1_0;


#line 1974 1
layout(location = 0)
in vec2 vd_texcoord_0_0;


#line 1974
layout(location = 1)
in vec2 vd_texcoord_1_0;


#line 1974
layout(location = 2)
in vec4 vd_inner_texcoord_2_0;


#line 19 0
struct innerData_0
{
    vec4 texcoord_2_0;
};


#line 33
void main()
{

#line 33
    innerData_0 _S2 = { vd_inner_texcoord_2_0 };

#line 33
    _S3.texcoord_0_0 = vd_texcoord_0_0;

#line 33
    _S3.texcoord_1_0 = vd_texcoord_1_0;

#line 33
    _S3.inner_0 = _S2;

    vec4 _S4 = vec4(vd_texcoord_0_0, vd_texcoord_1_0.x, vd_inner_texcoord_2_0.y);

#line 35
    _S1 = _S4;

#line 35
    entryPointParam_fragmentMain_out1_0 = _S4;

#line 35
    return;
}

[csyonghe]

I guess the left over global variable _S3 can be removed if we add an optimization pass that removes non-output global variables that has only writes but no reads.

[csyonghe]

we can also follow uses of any FieldAddr on the globalVar and remove the stores to those field addrs to get rid of the store to _S3 in the test case.

[cheneym2]

/format

[slangbot]

🌈 Formatted, please merge the changes from this PR