http://bugs.winehq.org/show_bug.cgi?id=58622
Bug ID: 58622 Summary: A shader with large loops is very slow to compile Product: vkd3d Version: 1.17 Hardware: x86-64 OS: Linux Status: NEW Keywords: download, source, testcase Severity: normal Priority: P2 Component: hlsl Assignee: wine-bugs@winehq.org Reporter: z.figura12@gmail.com Distribution: ---
Created attachment 79172 --> http://bugs.winehq.org/attachment.cgi?id=79172 shader
Attached shader is from a real application, meant to be compiled with ps_5_1.
Disabling loop unrolling helps. That doesn't mean that we need to target loop unrolling specifically, though. There may be other things we can improve. Here's what's been observed so far:
* copy-prop is doing a *lot* of work; it's an important part of loop unrolling. Copy-prop is doing a lot of work partly because we're generating an awful lot of synthetic variables for various reasons, not all of which are necessary:
- <index-val> variables can be avoided in 99% of cases. I've slapped together a heuristic that checks if we're indexing a LOAD and if there's nothing potentially problematic between the LOAD and INDEX, and it gets rid of all the <index-val> cases.
- <deref> variables exist for the same reason IIRC and can probably be avoided with the same heuristic.
- Temp copies for the uniforms [and there's some very big uniform arrays in this shader, which are also largely responsible for the many index-vals.] The uniforms aren't written, but we still generate the copies in case they are. DCE before anything else might help, but even better would be avoiding generating them in the first place.
* We can do this trivially because the uniforms are implicitly const, although this is something we'll need to fix with new API. However it may generally be possible to do this without having to rely on that knowledge. I suspect there are a decent number of large shaders in the wild that use backcompat flags or old compilers but whose uniforms are still constant. Well, maybe. Maybe we should just implement implicit const first.
- That first fromTangentToWorld assignment, and I think several other assignments, are resulting in a cast for some reason, which is generating a bunch of <cast> synthetics when splitting. That shouldn't... cast? Why is that casting?
* Native seems to struggle with unrolling this shader quickly as well.
* Native is known to limit loop unrolling by execution time at least under some circumstances. If we have to do that here, so be it.