Use fibers to guarantee stack size on Windows

[alexreinking]

We used to use #pragma comment(linker, "/STACK:8388608") to enforce an 8MB stack on Windows. Unfortunately, clang-cl and lld-link don't support this pragma. I have opened a PR with upstream LLVM to correct this^[1], but it is clear we need a more robust solution than a transitive linker flag.

This PR adds a helper, run_with_large_stack to execute a std::function<void()> in a context where it will have at least 8 MB of stack space. Currently, it merely calls the function on Linux, but on Windows it creates a new fiber with the appropriate stack size and runs it there, blocking until the fiber returns.

It is applied in two places: around lower() and around the call to compile_func in CodeGen LLVM. This was sufficient to build locally on Windows, but there might be more places it's required. We'll let the buildbots sort that out.

^[1] https://reviews.llvm.org/D99680

[alexreinking]

Looks like some of the workers were not started in a virtualenv or otherwise numpy is not installed.

[alexreinking]

I just got paranoid about what happens if the function throws an exception. Is it valid to catch the exception, move it to the parent fiber and then re-throw it?

[abadams]

Looks like Raymond Chen wrote a series of blog posts on our use case. Part 4 seems to cover exceptions. https://devblogs.microsoft.com/oldnewthing/20200605-00/?p=103840

[alexreinking]

Looks like Raymond Chen wrote a series of blog posts on our use case. Part 4 seems to cover exceptions. https://devblogs.microsoft.com/oldnewthing/20200605-00/?p=103840

Sorry I didn't see this before implementing. I'll take a look, but does this look wrong to you?

[alexreinking]

Oh, good, looks like I did the right thing. Raymond's implementation supports return values, which is nice, but maybe overkill for what we're doing.

[steven-johnson]

LGTM. Would sure be nice if there was a way to do something similar on Linux/OSX/etc so that we could avoid requiring the linker flags there,too.

[alexreinking]

LGTM. Would sure be nice if there was a way to do something similar on Linux/OSX/etc so that we could avoid requiring the linker flags there,too.

We don't require any stack-size linker flags on Linux, though, do we?

[abadams]

LGTM. Would sure be nice if there was a way to do something similar on Linux/OSX/etc so that we could avoid requiring the linker flags there,too.

We don't require any stack-size linker flags on Linux, though, do we?

That's correct. 8MB is the default already.

We do have a custom stack size for the fft app on os x only, but that's to reduce it substantially to act as a canary if we add a new pass that uses too much stack.

[alexreinking]

@steven-johnson looks like we've got some zombies?

[steven-johnson]

@steven-johnson looks like we've got some zombies?

It's zombies all the way down.

I've commented on an existing bug on buildbot re: the zombies that won't die.
I've open a ticket at Google re: the network that is dropping intermittently.

Unfortunately, the only way to kill the zombies is to stop and restart the master, which I'll do now.

[alexreinking]

Thanks! Looks like all the checks are passing now... good to merge?

[steven-johnson]

LGTM