[llvm] r193251 - Use address-taken to disambiguate global variable and indirect memops.

[Shuxin Yang]

> "does" not "dose". A dose is a dosage of a drug. "does" is the verb to 
> perform something.

Sorry this is one of my common typo.  Eric spotted this typo as well, 
unfortunately I forgot to fix.

>
> Also, while I am sure this is the right approach, I made some comments 
> on this patch:
>
>
> http://lists.cs.uiuc.edu/pipermail/llvm-commits/Week-of-Mon-20131021/192186.html 
>
>
> it looks like you committed without replying or acting on any of my 
> review comments? Huh?
>
I somehow miss this mail.  There is no reason to ignore any response. I 
have never ever
ignore feedback before. Actually I'm expecting your feedback as you are 
original author of the globalopt.

 >Could I ask for the keyword to be worded "addr not taken" instead of
 >"not addr taken"?

Ok.

 >Does addr_not_taken imply unnamed_addr? I'm pretty sure it does. If so,
 >please make sure that this works (see how readonly/readnone are handled
 >-- readnone implies readonly).
I will double check next week, and go back to you.

 >I'm not convinced that you actually need to cache this result. You say
 >that it isn't efficient to determine whether the address is taken on the
 >fly.
 >How bad is it if you just loop over the direct users of the global
 >variable, and bail if you see anything that isn't a load or store? Yes,
 >it would have false negatives on load/store of constantexpr GEPs. How
 >big a problem is that really? If that really doesn't work, is it okay if
 >you do a depth-first scan over uses, and bail on anything which isn't
 >load/store/constant-gep/constant-bitcast?

The compile time is really depends on how many global variables a module 
have,
and how many reference it has. C-thinking programs tend to have lots 
global-variables,
making compile-time impact hard to predict.

I also try to analyze address-taken and cache the result during AA phase.
However, in this case, we also need to provide a interface to let AA client
to invalidate the the cached result. (e.g. a imaginary out-lining pass makes
a addr-not-taken variable address-taken).

Thank you again for your feedback!

Shuxin