llvm dev - Jul 2012 - [LLVMdev] Problem with getting a result of an instruction.

Hello,

I'm a newbie in LLVM. Now I'm trying to implement a pass that does some 
simple form of dynamic dataflow analysis.

In my dataflow analysis I want to know if a specific variable is 
"dependent" on another one. "Dependent" means the following:
if we have
three variables in a program - a, b and c and in some places of this 
program we have the following assignments: b = a and then c = b. Then 
variable b is dependent on a and variable c is dependent on both a and 
b. Basically "dependent" means that a result in some variable is 
calculated basing on some other one.

While implementing this pass I have faced the following problem - let's 
say we have the following piece of code:

store i32 0, i32* %a, align 4
%0 = load i32* %a, align 4
store i32 %0, i32* %b, align 4

This piece of code basically means an assignment b = a. My problem is 
that I need to detect what kind of virtual register we use in load 
instruction (in that example we use virtual register 0) and my pass 
cannot do it. I'm using the following piece of code in my pass:

   for (inst_iterator I = inst_begin(F), E = inst_end(F); I != E; ++I){
           if((*I).getOpcode() == Instruction::Load){
             errs() << cast<LoadInst>(*I).getValueName() <<
"\n";
           }
       }

The method getValueName() gives nothing as well as getName() or 
getValueID(). Because of this I'm not able to detect that b is dependent 
on a.

However if I have something like this:

store i32 0, i32* %a, align 4
%tmp = load i32* %a, align 4
store i32 %tmp, i32* %b, align 4

I that case my pass is able to tell that we are using "tmp" to perform
this assignment.

How can I resolve this issue and get a name or an ID of result?

Thanks in advance.

Best,
Yaroslav.
-------------- next part --------------
An HTML attachment was scrubbed...
URL:
<http://lists.llvm.org/pipermail/llvm-dev/attachments/20120709/4ae020d5/attachment.html>

Yaroslav Markov wrote:
> Hello,
>
> I'm a newbie in LLVM. Now I'm trying to implement a pass that does
some
> simple form of dynamic dataflow analysis.
>
> In my dataflow analysis I want to know if a specific variable is
> "dependent" on another one. "Dependent" means the
following: if we have
> three variables in a program - a, b and c and in some places of this
> program we have the following assignments: b = a and then c = b. Then
> variable b is dependent on a and variable c is dependent on both a and
> b. Basically "dependent" means that a result in some variable is
> calculated basing on some other one.
>
> While implementing this pass I have faced the following problem - let's
> say we have the following piece of code:
>
> store i32 0, i32* %a, align 4
> %0 = load i32* %a, align 4
> store i32 %0, i32* %b, align 4
>
> This piece of code basically means an assignment b = a. My problem is
> that I need to detect what kind of virtual register we use in load
> instruction (in that example we use virtual register 0) and my pass
> cannot do it. I'm using the following piece of code in my pass:
>
> for (inst_iterator I = inst_begin(F), E = inst_end(F); I != E; ++I){
> if((*I).getOpcode() == Instruction::Load){
> errs() << cast<LoadInst>(*I).getValueName() <<
"\n";
> }
> }
>
> The method getValueName() gives nothing as well as getName() or
> getValueID(). Because of this I'm not able to detect that b is
dependent
> on a.
As for getName(), see this reply:

   http://lists.cs.uiuc.edu/pipermail/llvmdev/2012-July/051558.html

but you don't need getName. You never need getName, except for debugging.
> However if I have something like this:
>
> store i32 0, i32* %a, align 4
> %tmp = load i32* %a, align 4
> store i32 %tmp, i32* %b, align 4
>
> I that case my pass is able to tell that we are using "tmp" to
perform
> this assignment.
>
> How can I resolve this issue and get a name or an ID of result?
An Instruction* is its result, the equals sign is a lie. The Value* 
serves as the ID you're looking for.

It's a common mistake to interpret the '%x = add i32 %a, %b' format
as
having two llvm values and an assignment occurring between them. In 
reality there is one llvm::Value (an Instruction) and it is the add. The 
instruction's name is 'x'. You can uniquely refer to the result of
the
add by the Instruction* given. Recall that LLVM IR is in static single 
assignment form, so there is no '%tmp' to be reassigned; it *is* the 
load instruction (well, it's the name of it).

As for the actual analysis you're trying to do, see the 
MemoryDependenceAnalysis pass, which in turn uses AliasAnalysis. You 
have many things to consider, such as the case where we call an external 
function and want to decide whether it could have modified the data 
inside pointer %a.

Nick