High-level shading languages (GLSL, HLSL, MetalSL, etc...) are one of the few areas in games and graphics programming that, despite their name, are still very low level and have evaded tighter, more meaningful integration with applications. There have been past attempts to provide richer semantic information, but those are largely couched in the realm of being wrappers over an existing language: e.g. FX and Cg. The biggest issue with this approach is that the semantic information isn't a first-class citizen of the language. These wrappers are just fancy variations on the C preprocessor, which by definition makes them lossy. What we really want is a language that has application-level semantics built into it. Let's start by examining why, especially through the lens of game development.
Many game developers make games for multiple platforms. Some of them stay in the safe zone of Windows, PS4, and Xbox One. This is really beneficial from a shader standpoint because HLSL (or a close variant thereof) can be used on all three of these platforms. In a lot of cases, the PC version won't have additional features that necessitate different shaders. That being said, there's still the possibility that different hardware optimizations will be chosen for each, making platform-specific code necessary. There are also developers that release games across a wider range of platforms, including iOS, Android, Linux, MacOS, etc. This brings in other shader languages including desktop GLSL, GLSL ES, and Apple's Metal Shading Language. With this, the problem of supporting multiple platforms becomes much more pressing. Whereas with the first situation you'd be able to just use #defines for blocks of platform-specific functionality, you now need to either have some sort of conversion pipeline, or manually author multiple, platform-specific shaders.
Fortunately, for developers that just need something to work right now, there's a rich ecosystem of solutions for shader conversion and cross-compilation:
There's a couple problems with the above solutions. The first is that cross-platform support isn't really there. Things like pixel-local storage and framebuffer fetch are encoded through specific usage patterns instead of with function calls. This is honestly a terrible approach because it requires knowledge of something with a magical outcome: i.e. a given, very specific configuration will translate into the desired usage in the output, but it must match exactly. It also means that the language semantics need to be further overloaded to work with future platform-specific extensions. The second problem is that HLSL is largely viewed as the "right" language to use as the input language. While I definitely agree that it has better syntax and semantics than GLSL, it's not without its own problems, especially with regards to type safety and safety in general. Additionally, the existence of wrappers like Cg and FX should make it obvious that HLSL doesn't have any real application-level facilities. What I mean by application-level is features for users that aid in developing applications: e.g. support for variations of a shader with/without a given feature (henceforth called "variants"), material-level features (stuff like constant buffer header creation/serialization), and render-pass level features (techniques). These are all possible in some form in FX and Cg, but they're not tightly integrated and instead applied as a layer above the shader code itself.
The last aspect of all of this, largely motivated by prior experience with a codebase of very complicated shaders with many variants controlled by boolean feature definitions, is a desire to discourage use of the preprocessor as much as possible. It's a rather natural leap to want to control enabling/disabling features of a shader through the use of preprocessor definitions, especially since the preprocessor is the most powerful metaprogramming tool that current shading languages have available. The problem with this is that it's a lossy transformation. Let's say you want to know what samplers and uniform buffers are used by a given configuration of a shader. Using the preprocessor for variants means that you'll need to build and then reflect the information from every single variant, the union of which is the information that you wanted to know. If your shader has thousands of variants you'll need to compile all of them just to get this information, meaning that your shader build is potentially now a bottleneck for your material build.
The alternative here is having variants be a first-class citizen of the language, such that an un-varianted version of a shader already represents the union of all possible shaders that can be generated from it. In practice, this means having some type called Variant that the user uses like any other type (with some semantic restrictions).
An additional aspect of having extra semantic information in the shader is that we can even strip unnecessary vertex inputs based on their usage in the pipeline. For example, if texture mapping is an optional feature, we can strip not only the code to do the texture fetches from the pixel shader, but also the code to read/copy/generate the texture coordinates themselves from the vertex shader. This same idea extends naturally into deeper pipelines which include geometry or tessellation shaders. In current systems, this can be achieved through having defines in vertex input structures to eliminate elements. In this system, it would be automatic and would require no additional work besides the user varianting the code in the pixel shader.
In light of all of this, I propose writing a new high-level language, designed around some core tenets:
If we look to the design of Apple's Metal Shading Language (from a software engineering standpoint), it fits the first four points pretty well. The reason for this is that it's a modified subset of C++ parsed with a special version of Clang. I propose that this is currently the best direction to go in when developing this new language: a modified subset of C++, parsed by an existing, well-tested parser (Clang). I imagine the pipeline as looking like [Clang] -> [Abstract Representation] -> [Backend Transpilers].
Since we'd have an intermediate representation, there are all sorts of interesting things that could be done:
Hopefully I've made clear the case for having a shader language beyond what the various platform langauges offer, and moving to an authoring model that treats the platform-level shading languages as something akin to assembly: a low-level language to compile to rather than author directly. There are definitely benefits beyond what I've outlined that this approach offers and having control over the language (versus dealing with a platform/API vendor) opens the door to implementing high-level constructs that the platform vendors wouldn't consider approaching.