How does slang reduce compilation time?

[zznewclear13]

It is said that slang reduces compilation time by reusing slang IR during module importing. But slang IR will eventually be translated to HLSL/SPRIV/GLSL and get compiled again. Does that mean , slang generates really optimized HLSL/SPRIV/GLSL code that can be compiled super fast?

[jkwak-work]

I am not sure what you meant by "super fast".

The compilation steps for HLSL, as an example, is following,

1. Slang reads the input shaders and generates IR.
2. Slang emits HLSL shader from the generated IR.
3. Slang calls DXC compiler to compile the emitted HLSL; and you will get DXIL.

As a comparison, if your shaders are just HLSL, you can skip the step 1 and 2 entirely.
In other words, compiling HLSL shader with Slang will always be slower than compiling with DXC directly.

Regarding the compilation speed up with the IR modules, if you have more than one shader that shares the same module, you can generate the slang-module once and reuse it for the other shader compilations.

[zznewclear13]

In other words, compiling HLSL shader with Slang will always be slower than compiling with DXC directly.

Yeah that is what I thought.

But I watched Getting Started With Slang: Modulations by The Khronos Group(https://youtu.be/ABFc0WRHbH8?t=560), the speaker said Slang with cached modules compiles or hot reloads faster than hlsl with dxc. I wonder how could that happen.

[juliusikkala]

That presentation refers to SPIR-V, which the Slang compiler can target directly without going through HLSL or GLSL. The compilation paths compared there are Slang => SPIR-V and DXC => SPIR-V, where Slang with modules can reuse a lot of the compilation work while DXC with #include cannot. Currently, you probably wouldn't see similar benefits if you targeted DXIL instead.

[zznewclear13]

Thanks, that makes sense.