Two same "struct" are generated with different names when an arithmetic math is applied to the generic-argument about slang

[swoods-nv]

Jay, is there an estimate for this one?

[jkwak-work]

This is probably an easy one but I don't have much confidence.
I put 3 days for now.

[csyonghe]

The issue is usually because we failed to identify the two specializations are the same thing.

In IR, we usually deduplicate specialize insts. When we specialize the Reduce function, we want to make sure we didn't create duplicated specialized structs. To trace this down, you can set a breakpoint after specialization pass, and inspect the IR for duplicate struct types. Find the duplicate struct type with the bigger UID, and try to track down why it gets created. There is a lookup process before creating the specialized struct type to see if a specialization is already created. Check why that lookup failed to determine the case.

[csyonghe]

Set a breakpoint on the call to specializeModule in slang-emit.cpp:424 and start from there.
Then specializeGeneric will be called on everything that needs specialization, this include the generic functions and the generic struct type. Try to set a break point in this function when we are trying to specialize a struct type, and see why the lookup failed in slang-ir-specialize.cpp:238.

[swoods-nv]

Jay, is this issue still blocking you?

[jkwak-work]

Yes it blocks my UE task.
I will start working on this tomorrow.

[jkwak-work]

I am having some trouble reproducing the issue.
When I target "HLSL", it is easy to reproduce.
But when I use gfx with slang-test, both dxc and fxc don't print the error and everything works fine.
I am little puzzled.
The bug might be in the code where HLSL gets emitted.

[jkwak-work]

It appears that the issue stems from the fact that the duplication detecting logic uses hash values.
And the hash values are calculated from the memory pointer values rather than the values the pointers point to.

HashCode IRSimpleSpecializationKey::getHashCode() const
{
    auto valCount = vals.getCount();
    HashCode hash = Slang::getHashCode(valCount);
    for( Index ii = 0; ii < valCount; ++ii )
    {
        hash = combineHash(hash, Slang::getHashCode(vals[ii])); // <== This generates a hash value from memory pointer values; not the value kIROp_IntLit represents.
    }
    return hash;
}

When this happens, the following is where the issue happens.

    IRInst* specializeGeneric(
        IRGeneric* genericVal,
        IRSpecialize* specializeInst)
    {
        // First, we want to see if an existing specialization
        // has already been made. To do that we will construct a key
        // for lookup in the generic specialization context.
        //
        // Our key will consist of the identity of the generic
        // being specialized, and each of the argument values
        // being pased to it. In our hypothetical example of
        // `specialize(g, a, b, c)` the key will then be
        // the array `[g, a, b, c]`.
        //
        Key key;
        key.vals.add(specializeInst->getBase());
        UInt argCount = specializeInst->getArgCount();
        for (UInt ii = 0; ii < argCount; ++ii)
        {
            key.vals.add(specializeInst->getArg(ii));
        }

        {
            // We use our generated key to look for an
            // existing specialization that has been registered.
            // If one is found, our work is done.
            //
            IRInst* specializedVal = nullptr;
            if (genericSpecializations.tryGetValue(key, specializedVal)) // <== This didn't properly work
                return specializedVal;
        }

[jkwak-work]

I found another problem regarding this issue.
When I wrap the generic argument, which is a compile-time constant, with "int", it becomes not-compile-time constant.

tests/bugs/gh-3935.slang(14): error 39999: expression does not evaluate to a compile-time constant
MyStruct<int(ArraySize / 2)> Reduce(MyStruct<ArraySize> o)

Note that the error message above shows "int(ArraySize /2)".
It compiles fine when I just did "MyStruct<ArraySize / 2> Reduce..."

[jkwak-work]

Fix is merged.
Closing the issue.