[Migrated] Optimize `spirv-builder` clean build wall time

Issue automatically imported from old repo: https://github.com/EmbarkStudios/rust-gpu/issues/858

Originally creatd by repi on 2022-04-10T11:01:16Z

We should look into improve our clean build times of when building spriv-builder & rustc_codegen_spirv as that is the main scenario in CI and also useful for users if our crates rebuild fast.

Right now it looks like we have a couple of quite unfortunate serial dependencies that significantly increases the wall time. Here is the current results with cargo build --release -p spirv-builder -Z timings on my AMD 5950x (32 vCPU):

Potential optimizations

• Could spirv-builder be the one depending on spirv-tools instead of rustc_codegen_spirv?
  • I.e. move the responsibility of running spirv-opt/spirv-val and final linking to the builder.
  • Does that sound like it could be doable @eddyb ?
• Remove/replace serde, quite crazy it takes 20 seconds to compile
  • Think we only have two use cases, our ModuleResult output and then the internal decorations that is using more complex serde implementation (CustomDecoration).
  • have no idea how serde+serde_json is taking such a crazy long time to build here though, can't see any other crates enabling tons of types that serde derive macros are used on or similar
  • ModuleResult would be trivial to replace serde with nanoserde, could work for CustomDecoration also?
  • #860
• Remove/replace sanitize-filename dependency that compiles and uses regex crate for super trivial things
• Optimize builds of rspirv itself somehow?
  • Unfortunate that it is so slow to build by itself but could be hard to fix and tons of types in it. No open issues on it in the rspirv repo as far as I could see
• Determine why rustc_codegen_spirv is not split in frontend/codegen sections in the profile report.
  • This would enable spriv-builder itself to start building earlier and before the codegen of rustc_codegen_spirv is done.
  • Update: This is not possible as it it needs to be built as a dylib

If we do get rid of serde and manage to make sure rustc_codegen_spirv doesn't have to wait for the spirv-tools-sys build, we could get get a 15+ second wall time improvement here as rustc_codegen_spirv could start as soon as frontend section of rspirv has been built:
-

Commencez par exécuter cargo build --release -p spirv-builder -Z timings et examinez le temps des dépendances lors d’une compilation propre pour spirv-builder et rustc_codegen_spirv. Analysez les relations de dépendance impliquant spirv-tools, serde, serde_json, sanitize-filename, regex et rspirv ; le travail est terminé lorsqu’une réduction mesurée du temps d’exécution d’une compilation propre est obtenue grâce à une optimisation clairement choisie.