Metal: Texture write fix

[Dynamitos]

In Metal, there is no overload for different component counts of textures, only 4-component writes are allowed. For textures that do not have four components the unused components get discarded, as far as I can gather from the specs.

[csyonghe]

Can we move this if statement to the loop above, instead of doing a fresh loop?

[Dynamitos]

for now i implemented a split between ImageStore and ImageStoreArray, because i cant split the array component from the coordinate in hlsl.meta.slang, but also i dont know when to split it off in the legalize because i dont know if the target texture is an array texture or not. this way if the swizzle does work in the future the entire ImageStoreArray can be deleted and the array coordinate passed the aux coord in ImageStore

[csyonghe]

You should be able to check the type of the image operand to get its shape and arrayness.

auto type = imageStoreInst->getOperand(0)->getDataType();
auto imageType = as<IRTextureType>(type);
imageType->xxx //get info about the texture type.

[Dynamitos]

Now a much better solution, not needing the ImageStoreArray anymore, but IMO it would be a bit better to swizzle the location in the hlsl.meta.slang instead if doing it in the legalize step

[csyonghe]

I think you should be able to do this completely in stdlib without implementing any logic for kIROp_ImageStore in the compiler backend.

[require(metal, texture_sm_4_1)]
static void __metalImageStore(This val, vector<uint, Shape.dimensions> location, T value, uint arrayIndex)
{
     __intrinsic_asm "...";
}

//...
case metal:
    __metalImageStore(this, __vectorReshape<Shape.dimensions>(location), newValue, location[Shape.dimensions+isArray-1]);
     return;

[csyonghe]

You should be able to use vectorReshape<> on the value to write into in stdlib as well? That should remove the need for this leglaization step, right?

[csyonghe]

@Dynamitos Do you plan to clean up the image store inst with vectorShape to change the incoming value into float4?
If not we can also just merge the PR as is.