Investigate impact of removing shell32 dependency of libstd on Windows

[alexcrichton]

An excellent blog post recently showed that removing the gdi32.dll dependency from LLVM speed up its test suite by a huge amount. For us libstd doesn't depend on gdi32.dll directly but it does depend on shell32.dll which depends on gdi32.dll (according to the blog post). Specificaly it looks like CommandLineToArgvW, the same function as LLVM, is used by us!

It'd be great to investigate removing the shell32 dependency from libstd (may be a few other functions here and there). We should run some tests on Windows to see if this improves rustc cycle times!

[retep998]

Based on my limited investigation, it's likely that replacing CommandLineToArgvW with a pure Rust equivalent should be good enough to get rid of the gdi32 dependency in things that only use libstd. So really someone just needs to write that pure Rust implementation which should be easy to do.

[Zoxc]

It would be nice to get rid of this in the compiler too. We spawn a lot of compiler instances during bootstrapping.

[retep998]

So now that the change has been implemented, have we actually investigated how much of an impact the change made?

[Eh2406]

Perf, an lolbench are both linux only. So nether help with this question. So don't waste your time like I did.

[Zoxc]

Here are just number for run-pass on 8 cores / 16 threads on Windows 10 v1809:

Without the shell32 dependency:
333.012 325.587 320.650

With shell32:
310.852 319.165 323.645

[Mark-Simulacrum]

For what it's worth, though there's no good way to differentiate PR-to-PR due to very high variance, it does look like there was a recent regression in PR build times:



[Zoxc]

I see that the CPU usage averages about 40% when running run-pass, so maybe we'd see some real improvements if we got rid of shell32.dll in the compiler too. Something is certainly holding things back.

[notriddle]

@Zoxc

micha@DESKTOP-IIQA1VP MINGW64 ~/IdeaProjects/rust (master)
$ ldd build/x86_64-pc-windows-msvc/stage1/bin/rustc.exe | grep -i ntdll
        ntdll.dll => /c/WINDOWS/SYSTEM32/ntdll.dll (0x7ffbede90000)

micha@DESKTOP-IIQA1VP MINGW64 ~/IdeaProjects/rust (master)
$ ldd build/x86_64-pc-windows-msvc/stage1/bin/rustc.exe | grep -i gdi

micha@DESKTOP-IIQA1VP MINGW64 ~/IdeaProjects/rust (master)
$

[Zoxc]

Yeah, I saw that the compiler doesn't load gdi32.dll. So there's probably some other reason why compiletest doesn't load all the cores.