build: write a gcc-style depfile when using clang-cl on windows

[piscisaureus]

Because clang-cl doesn't support msvc-style /showIncludes header dep detection, at least not in such way that it works properly with sccache.

[ry]

I can't say I really understand what's happening here. So I'm mostly concerned from a long term maintenance perspective: Should this patch be undone when mozilla/sccache#246 is fixed? Or is this something that could be sent upstream to Chrome?

I think you should describe the symptoms that this is fixing in the commit message - i.e. that libdeno doesn't get rebuilt properly on windows

[ry]

Does this fix #1424 ?

[piscisaureus]

I can't say I really understand what's happening here. So I'm mostly concerned from a long term maintenance perspective: Should this patch be undone when mozilla/sccache#246 is fixed? Or is this something that could be sent upstream to Chrome?

It could probably go upstream; there's no disadvantages to it.
The underlying issue can't be fixed in sccache; it should be (and has been) fixed in in clang (https://reviews.llvm.org/D46394) but that hasn't made it into a stable release yet.

[piscisaureus]

Does this fix #1424 ?

It fixes the rebuild issues I was experiencing locally, so I expect it solves #1424 also.
But who knows... the only way to find out is to try.

[ry]

So maybe add a comment that lets the future maintainer know when they can safely undo that patch. I guess it's when https://reviews.llvm.org/D46394 gets into stable?

[piscisaureus]

I can't say I really understand what's happening here.

MSVC (cl.exe) can't write a depfile, but it has the /showIncludes flag that prints included headers to stdout, and ninja parses that output.

sccache adds the /E flag that writes preprocessor output to stdout. It uses that compute the hash.

Now what happens if you pass both flags?
cl -> writes preprocessor output to stdout and headers to stderr
clang-cl -> writes preprocessor output to stdout, and does not write headers at all.

So dependency tracking was broken.

With some effort clang-cl can however write a gcc-style depfile, so that's what I'm making it do here.

[piscisaureus]

So maybe add a comment that lets the future maintainer know when they can safely undo that patch. I guess it's when https://reviews.llvm.org/D46394 gets into stable?

Did you see this denoland/chromium_build@75a7f82...1e3840b#diff-215e0f644aef180fe353893e8703f3f2R186 ?

[ry]

LGTM - thanks for the explanation.

Consider mentioning in the commit message what this fixes from a Deno user’s perspective.