rustc_codegen_spirv taking a long time processing my (large) shader

[hatoo]

This issue is from Discord conversation https://discordapp.com/channels/750717012564770887/750717499737243679/943879544006381638

Expected Behaviour

I want to compile my shader faster.

Example & Steps To Reproduce

1. clone my repo https://github.com/hatoo/rene
2. cargo build --release -Z timings
3. See compile duration

I tested on af3cf2b74e8ef4f4bd98a3cd2cda64b9379d20d7.

Completed rene v0.1.0 build script (run) in 263.2s

In console in VSCode and rust-analyzer plugin. It takes more time.

Completed rene v0.1.0 build script (run) in 407.3s

System Info

• Rust: rustc 1.60.0-nightly (1bd4fdc94 2022-01-12)
• OS: Windows 11
• CPU: 3950X
• GPU: RTX2080ti
• SPIR-V: SPIRV-Tools v2021.4-dev v2021.3-86-g21e3f681

[hatoo]

After I added codes to my shader, it takes very longer time to compile.
In this commit, it takes about 1000s.

Completed rene v0.1.0 build script (run) in 1009.6s

And I added one commit to this in this branch.
I've waited for few hours and the build didn't end.

[hatoo]

I enabled -Zself-profile and build it and I found rustc is spending almost all time to link (link_block_ordering_pass_and_mem2reg).

Procedure

1. Enable -Zself-profile on spirv-builder. It can be done by like hatoo@e34bda5
2. Change rene's spirv-builder dependency to use it.
3. cargo install --git https://github.com/rust-lang/measureme crox flamegraph summarize
4. cargo build --release on rene
5. crox .\rene_shader-xxxx.mm_profdata and chrome_profiler.json will be produced.
6. Open Chrome Developper tool and click the Performance tab. And upload chrome_profiler.json by clicking ⬆️ button.

[eddyb]

I've been working on a mem2reg replacement recently, relying on SPIR-T qptrs (based on EmbarkStudios/spirt#29 / EmbarkStudios/spirt#41 - no PR yet for the mem2reg replacement though, but I really hope I can get it into Rust-GPU 0.9, behind the qptr opt-in).

$ rg mem2reg 2023-07-08-spirt-disagg-baseline 
512:time:   0.011; rss:  187MB ->  187MB (   +0MB)      link_block_ordering_pass_and_mem2reg-before-inlining
515:time: 126.704; rss:  189MB ->  158MB (  -31MB)      link_block_ordering_pass_and_mem2reg-after-inlining
$ rg qptr::partition_and_propagate 2023-07-08-spirt-disagg-qptr-pnp 
538:time:   0.570; rss:  220MB ->  221MB (   +1MB)      qptr::partition_and_propagate

so that's a 222x speed-up (for the same version of rene-shader I used to demo the initial impact from SPIR-T, over a year ago)

(RUSTGPU_CODEGEN_ARGS="--spirt-passes=qptr --no-infer-storage-classes --no-legacy-mem2reg" is the opt-in, at least on my local branch, I'm hoping we can at least simplify it to RUSTGPU_CODEGEN_ARGS=--qptr for 0.9)

EDIT: to be perfectly clear: those flags are not useful without the necessary combination of SPIR-T + Rust-GPU changes and I don't even have an up-to-date Rust-GPU branch anywhere comparable to that one.