llvm dev - Nov 2013 - [LLVMdev] [Propose] Add address-taken bit to GlobalVariable for disambiguation purpose

Nadav:

    I don't think this is right approach for engineering.
The time-complexity of re-analyzing addr_taken for each single alias 
query depends on
    1. how many global variable
    2. how many occurrence of these global variables.
    3. how many queries the compiler have.

   3) depends on compiler.  You never know what we will have in the 
following few years.
1 and 2 depends on the program.  You never know what kind of program you 
will run into.
How can we use what we have today the extrapolate the future ignoring 
the highly
unpredictable complexity.

    It's interesting that recently, many EE magazine (circuit cellar, 
Elector, EE times) are
discussing buggy SW kill people.  I remember some posts complaining that 
some buggy program
have amazingly large # of global variables. I can find one post in 
Chinese website:

    http://forum.xitek.com/thread-1226816-5-1-1.html

    The 1st post says,  "a program has 11000 global variables"!

    As to "Can you provide this data"?  My answer is no, and I will
not
to implement the analysis
which perform on-the-fly analysis unless I'm convinced that saving 
addr_taken bit to llvm::GlobalVariable
is fundamentally flawed.

Shuxin





On 10/30/13 9:20 AM, Nadav Rotem wrote:
>
> On Oct 29, 2013, at 4:39 PM, Shuxin Yang <shuxin.llvm at gmail.com 
> <mailto:shuxin.llvm at gmail.com>> wrote:
>
>>   That said, I don't think it is still inexpensive if we re-analyze
the
>>    addr-taken again and again *ON THE FLY".  It is really
difficult to
>>    predict compile-time impact. You never know how many global
variables
>>    in a program, and you never know how extensive they are used.
>>    Poorly-written program tends to have lots of global-vars.
>
> Actually, it is not that difficult to predict the effect on compile 
> time.  All you need to do is implement the analysis “on the fly” and 
> measure the compile time.  Can you provide this data ?  Compile time 
> measurements can help us make a decision.
-------------- next part --------------
An HTML attachment was scrubbed...
URL:
<http://lists.llvm.org/pipermail/llvm-dev/attachments/20131030/292a7d20/attachment.html>

On Oct 30, 2013, at 10:37 AM, Shuxin Yang <shuxin.llvm at gmail.com>
wrote:
> Nadav:
> 
>    I don't think this is right approach for engineering.  
> The time-complexity of re-analyzing addr_taken for each single alias query
depends on
>    1. how many global variable
>    2. how many occurrence of these global variables.
>    3. how many queries the compiler have.
> 
>   3) depends on compiler.  You never know what we will have in the
following few years.
> 1 and 2 depends on the program.  You never know what kind of program you
will run into.
> How can we use what we have today the extrapolate the future ignoring the
highly
> unpredictable complexity.
This logic doesn't make sense to me.  You can implement it both ways and get
empirical results on *programs we have today* and *in our compiler*.  This is
not a theoretical exercise.

In practice, walking the use list of a global variable is very fast.  As
you've noticed, we already use this approach (in an admittedly ad-hoc and
decentralized way) throughout the compiler.
>    It's interesting that recently, many EE magazine (circuit cellar,
Elector, EE times) are
> discussing buggy SW kill people.  I remember some posts complaining that
some buggy program
> have amazingly large # of global variables. I can find one post in Chinese
website:
> 
>    http://forum.xitek.com/thread-1226816-5-1-1.html
> 
>    The 1st post says,  "a program has 11000 global variables"!
This is just FUD and completely unrelated to the discussion.
>    As to "Can you provide this data"?  My answer is no, and I
will not to implement the analysis
> which perform on-the-fly analysis unless I'm convinced that saving
addr_taken bit to llvm::GlobalVariable
> is fundamentally flawed. 
You don't have to be convinced.  The burden of proof is on you - not on us
to convince you.

Here's the deal: there are tons of "potentially useful" things
that could be encoded in the IR.  Each thing added to IR has a complexity
increase on the entire compiler.  Passes that work on global variables will have
to reason about this bit, and transformations that could invalidate it (e.g.
global merging) will have to have code added to update/preserve it.

We are very conservative about changing IR for good reason.  We don't add
caches to IR unless there is pretty much no other way to achieve the result.  In
a perfect world, we would have nothing redundant in the IR at all.

That said, I'm open to this attribute, because I think the semantics can be
nailed down tightly (though your "volatile" discussion doesn't
make any sense to me) it is widely useful, and I don't think the burden of
maintaining it will be that high.  However, before we do it, you need to
demonstrate that lazily computing it from use-def chains is *empirically worse*.

-Chris
-------------- next part --------------
An HTML attachment was scrubbed...
URL:
<http://lists.llvm.org/pipermail/llvm-dev/attachments/20131030/bdbd7314/attachment.html>

Hi, all:

    Per Chris and Nadav's request, I begin to write the code about 
analyzing address-taken
lazily.  I realize the alias query could be initiated from any context 
(*function* pass, loop pass etc),
however, the analysis for global-variable-address-taken is conducted in 
*module* scope.
Is there any potential problem over here? (For instance,  function foo() 
and bar() comprise module m,
however, at time optimizer is working on foo(), bar() is not physically 
in that module. In this case,
analyze global-variable on the fly doesn't make sense.)

Thanks in advance!
Shuxin

On 10/30/13 2:08 PM, Chris Lattner wrote:
> On Oct 30, 2013, at 10:37 AM, Shuxin Yang <shuxin.llvm at gmail.com 
> <mailto:shuxin.llvm at gmail.com>> wrote:
>> Nadav:
>>
>>    I don't think this is right approach for engineering.
>> The time-complexity of re-analyzing addr_taken for each single alias 
>> query depends on
>>    1. how many global variable
>>    2. how many occurrence of these global variables.
>>    3. how many queries the compiler have.
>>
>>   3) depends on compiler.  You never know what we will have in the 
>> following few years.
>> 1 and 2 depends on the program.  You never know what kind of program 
>> you will run into.
>> How can we use what we have today the extrapolate the future ignoring 
>> the highly
>> unpredictable complexity.
>
> This logic doesn't make sense to me.  You can implement it both ways 
> and get empirical results on *programs we have today* and *in our 
> compiler*.  This is not a theoretical exercise.
>
> In practice, walking the use list of a global variable is very fast. 
>  As you've noticed, we already use this approach (in an admittedly 
> ad-hoc and decentralized way) throughout the compiler.
>
>>    It's interesting that recently, many EE magazine (circuit
cellar,
>> Elector, EE times) are
>> discussing buggy SW kill people.  I remember some posts complaining 
>> that some buggy program
>> have amazingly large # of global variables. I can find one post in 
>> Chinese website:
>>
>> http://forum.xitek.com/thread-1226816-5-1-1.html
>>
>>    The 1st post says,  "a program has 11000 global
variables"!
>
> This is just FUD and completely unrelated to the discussion.
>
>>    As to "Can you provide this data"?  My answer is no, and I
will
>> not to implement the analysis
>> which perform on-the-fly analysis unless I'm convinced that saving 
>> addr_taken bit to llvm::GlobalVariable
>> is fundamentally flawed.
>
> You don't have to be convinced.  The burden of proof is on you - not 
> on us to convince you.
>
> Here's the deal: there are tons of "potentially useful"
things that
> could be encoded in the IR.  Each thing added to IR has a complexity 
> increase on the entire compiler.  Passes that work on global variables 
> will have to reason about this bit, and transformations that could 
> invalidate it (e.g. global merging) will have to have code added to 
> update/preserve it.
>
> We are very conservative about changing IR for good reason.  We don't 
> add caches to IR unless there is pretty much no other way to achieve 
> the result.  In a perfect world, we would have nothing redundant in 
> the IR at all.
>
> That said, I'm open to this attribute, because I think the semantics 
> can be nailed down tightly (though your "volatile" discussion
doesn't
> make any sense to me) it is widely useful, and I don't think the 
> burden of maintaining it will be that high.  However, before we do it, 
> you need to demonstrate that lazily computing it from use-def chains 
> is *empirically worse*.
>
> -Chris
-------------- next part --------------
An HTML attachment was scrubbed...
URL:
<http://lists.llvm.org/pipermail/llvm-dev/attachments/20131104/c4a1d7a1/attachment.html>