VLA test breaks assertion of decl being used without being marked used

[Quuxplusone]

Bugzilla Link	PR31042
Status	CONFIRMED
Importance	P normal
Reported by	Paulo Matos (pmatos@linki.tools)
Reported on	2016-11-17 04:01:23 -0800
Last modified on	2017-01-30 17:00:37 -0800
Version	trunk
Hardware	PC Linux
CC	hans@chromium.org, llvm-bugs@lists.llvm.org
Fixed by commit(s)
Attachments	vla-15-9b115d-e3f212.c (3462 bytes, text/plain) vla-15-9b115d-e3f212.sh (1695 bytes, application/x-shellscript) pr31042.zip (1080 bytes, application/zip)
Blocks
Blocked by
See also

Created attachment 17597
reproducer source

The attached test causes an ice on 4.0.0 nightly. This is extracted from GCC
VLA tests.

[Quuxplusone]

Attached vla-15-9b115d-e3f212.c (3462 bytes, text/plain): reproducer source

[Quuxplusone]

Attached vla-15-9b115d-e3f212.sh (1695 bytes, application/x-shellscript): reproducer script

[Quuxplusone]

Attached pr31042.zip (1080 bytes, application/zip): Minimal reproducer

[Quuxplusone]

After a short discussion on this in the mailing list:
http://lists.llvm.org/pipermail/cfe-dev/2016-November/051579.html

It seems that the problem is that even though we try to special case typeof,
in order for type of to find if it's evaluated it checks the previous context
(in Sema::TransformToPotentiallyEvaluated). If this previous context is
unevaluated then it does not transform the typeof context. When the typeof is
inside a sizeof context, which is unevaluated, typeof is never transformed and
therefore you can't do a sizeof ( typeof ( <vla> )).

I wonder if it's possible for us to mark sizeof as potentially evaluated
instead? What are the downsides of this?

[Quuxplusone]

Created review request https://reviews.llvm.org/D26843

[Quuxplusone]

Reproduces on trunk and 4.0, but also on 3.9 so not strictly a regression.