Variable auto-init: also auto-init alloca

[jfb]

Summary

alloca isn’t auto-init’d right now because it’s a different path in clang that all the other stuff we support (it’s a builtin, not an expression). Interestingly, alloca doesn’t have a type (as opposed to even VLA) so we can really only initialize it with memset.

rdar://problem/49794007

[jfb]

One downside to alloca is that we can's use __attribute__((uninitialized)) because it's a builtin. Maybe we need a separate uninitialized alloca? What do you all think?

[pcc]

I probably wouldn't do anything about suppressing init on alloca for now, but if we did do something I think I'd be in favour of the separate builtin for uninitialized alloca. I also considered the alternative of allowing __attribute__((uninitialized)) to appear on a function or block statement, but it seemed clearer for the uninitialized-ness to be as close to the alloca as possible.

[jfb]

In D60548#1462124, @jfb wrote:

One downside to alloca is that we can's use __attribute__((uninitialized)) because it's a builtin. Maybe we need a separate uninitialized alloca? What do you all think?

Actually I'm wondering if we should just implement:

#pragma clang attribute push (__attribute__((uninitialized)), apply_to = variable(unless(is_global)))

And lean on that for alloca, instead of having a new uninitialized alloca builtin. The pragma is useful for debugging regardless, so I think overall it's better.

[jfb]

In D60548#1463181, @jfb wrote:

In D60548#1462124, @jfb wrote:

One downside to alloca is that we can's use __attribute__((uninitialized)) because it's a builtin. Maybe we need a separate uninitialized alloca? What do you all think?

Actually I'm wondering if we should just implement:

#pragma clang attribute push (__attribute__((uninitialized)), apply_to = variable(unless(is_global)))

And lean on that for alloca, instead of having a new uninitialized alloca builtin. The pragma is useful for debugging regardless, so I think overall it's better.

Although the pragma syntax doesn't seem to support builtins, only variables, so it's still; weird for alloca... I'd really want the pragma to apply to everything in its scope including alloca.

[rjmccall]

lib/CodeGen/CGBuiltin.cpp

58

Can this value be taken from some common source with the existing code? I mean, even if it's just a constant in some header, that'd be a start.

[jfb]

• Move patternFor to a shared file as requested by @rjmccall

[jfb]

lib/CodeGen/CGBuiltin.cpp

58

Done.

[rjmccall]

lib/CodeGen/PatternInit.cpp

18

Please use a qualified declaration here instead of defining it in the namespace.

lib/CodeGen/PatternInit.h

23

Please choose names that mean something outside of the mental context you were in when you wrote the patch. :)

[jfb]

• Change name, qualify declaration.

[jfb]

lib/CodeGen/PatternInit.h

23

I was just copy / pasting!

[rjmccall]

LGTM.

lib/CodeGen/PatternInit.h

23

Your own code. :)

[jfb]

lib/CodeGen/PatternInit.h

23

Hey you can't pull that one on me! Pulling it on you is *my* hobby!
And like... it's communal code now! ;-)

[jfb]

I got an lgtm from @pcc on IRC, so I'll commit this.