Inline globals

[molikto]

• warn on inline(never) that are not respected
• inline global variables so that we can inline less functions caused by illegal arg/return type

fix https://github.com/EmbarkStudios/rust-gpu/issues/847

tests

My shader which compile to a single file of 2500 lines (dis glsl) now respect the inline(never) and result in 1400 lines of glsl

TODO:

1. some compiletests produce different result??

2. I was hoping by force inlining less functions, some shader can compile faster, but it is not the case yet, (need to look into if it actually result in less inlines?) af3cf2b74e8ef4f4bd98a3cd2cda64b9379d20d7 compile time: before: 222s -> after: 287s

[eddyb]

some compiletests produce different result??

For the SPIR-V disasm checks, that makes sense - you're removing some inlining requirements.

---

I've tried looking at this PR, and while I believe the general idea makes sense (i.e. as a form of "parameter weakening", all call sites agreeing on a quasi-constant value for some parameter does allow removing the parameter from the signature), I can't easily reason about the specifics of the code in this PR at this time.

---

At the very least, I feel like it should go into the param_weakening module: https://github.com/EmbarkStudios/rust-gpu/blob/5ac500d5b96145d72da6fe2a1059bd4e2a34b5e8/crates/rustc_codegen_spirv/src/linker/param_weakening.rs#L1-L4

And use the CallGraph abstraction to correctly visit the module in the right order for the right operations. Ideally we would have a framework for computing general function parameter weakenings from either callee uses (like the existing remove_unused_params) or call sites, and correctly propagating them through any number of nested calls.

---

I'm leaving this comment now as-is because I'm worried I might not be able to meaningfully review this any time soon, or at least not without doing some of the generalization work, to be able to convince myself this transform won't have subtle failure modes.

[AdrianEddy]

Any update on this? I have a shader which compiles to 6000 lines of GLSL or WGSL code, while hand-crafted one was only 300 lines, and it would really help to not inline the functions, as my only parameter to the inlined functions is the single uniform struct reference