[LLVMdev] clang optimizer does not remove unused/uneeded variables(and accesses) from global scope

[Robinson, Paul]

Duncan Sands wrote:
>On 05/07/12 13:30, Dimitry Andric wrote:
>> On 2012-07-05 12:53, Dennis Luehring wrote:
>>> addition to my last post:
>>> http://groups.google.com/group/llvm-dev/browse_thread/thread/afd85c3f303ec2a9
>>>
>>> i know that using static on my globals will remove them - no other
>>> compilation unit can then extern the vars
>>>
>>> but my question is - in this closed szenario clang is "directly"
>>> producing the exe - isn't clang aware of that the vars are not in use?
>>
>> It can't know that, unless you use link-time optimization.  Which
>> optimization level did you use?
>
>it can know: if you looked at his command line you can see that it is producing
>a final executable from the file, not just compiling it to object code.  As far
>as I know clang relies on an LLVM aware linker to do the link-time optimization,
>though I'm not sure why: in this kind of case it could just do it itself.  On
>linux you can use the gold linker with the LLVM gold plugin, maybe that can be
>used on windows too, I wouldn't know.

It can't know.  It can make the assumption that all possible runtime
environments would be unaffected by this externally-available definition,
but it's an assumption not a known fact, even if you see that a single
translation unit is producing an executable image.

Just to postulate an example, suppose I have a runtime environment that
makes a weak reference to some symbol.  If the program defines that symbol,
then the behavior of my runtime is affected.  If clang/LLVM arrogantly assumes
that can't possibly happen, then the behavior of my program is affected 
when it should not be.

Link-time optimization can figure this out, but the compiler by itself
cannot and must not.
--paulr