Control flow graph

Project Infrastructure LLVM Dev List Archives
1

Hi Chris,

Hi. FYI, it's usually better to email the llvmdev list with generic llvm questions.

Is it possible to get control flow graph of the application with the llvm
infrastructure in terms of basic blocks?

Yes, given a BasicBlock*, you can iterate over the pred/succ blocks in the CFG like this:

#include "llvm/Support/CFG.h"

   for (pred_iterator PI = pred_begin(BB), E = pred_end(BB); PI != E; ++PI) {
     BasicBlock *OnePredecessor = *PI;
     ...

For successors, s/pred/succ/

Is there any way of viewing the
Control flow graph of the application.

Yes, use:
   analyze -print-cfg X.(bc|ll)
or:
   analyze -print-cfg-only X.(bc|ll)

These will emit "dot" files that can be turned into a variety of graphics formats with the graphviz toolset that you can get on the net.

If you're running on a system with graphviz installed and with 'gv' (ghostview), you can even call functions like this from your code or from the debugger:

   LLVMFunction->viewCFG()

also ->viewCFGOnly();

Also, when I compile a C file to llvm assembly code i.e using llvmgcc -S
option then the output file has lot of "declare" and "typedef" and then
the "main" starts. Sorry for my naive questions but can you tell me what
are those.

I wouldn't suggest looking at this code. This is the raw output of the C front-end and has not been cleaned up at all. I would suggest using 'llvm-gcc -c', and using llvm-dis on the '.o' file.

Any document which I can go through so that I can understand
the llvm assembly file better.

http://llvm.cs.uiuc.edu/docs/LangRef.html

Actually I am trying to generate VLIW words looking at the assembly file of llvm. Is it possible to do that??

Sure, you can either write your own code generator (not recommended) or port the target independent code generator to your target (described in the docs). We currently have no specific support for VLIW targets, but it should not be difficult to add (and can probably be shared with the IA64 target).

-Chris

2

Hi,

I have generated llvm assembly file using following commands:
llvm-ld main.o fir.o
llvm-dis a.out.bc

I get a.out.ll as the llvm assembly file. When i go through this assembly
file I come across the following code
fir.entry: ; preds = %loopexit.1.i
  call void %print( sbyte* getelementptr ([2 x sbyte]* %.str_3, int
0, int 0), float* getelementptr ([97 x float]* %y, int 0, int 0), int 97 )
  call void %exit( int 0 )
  unreachable

Can anybody please tell me what is meant by "unreachable"? The output is
coming fine.

Regards,
manvi

3

Manvi Agarwal wrote:

Hi,

I have generated llvm assembly file using following commands:
llvm-ld main.o fir.o
llvm-dis a.out.bc

I get a.out.ll as the llvm assembly file. When i go through this assembly
file I come across the following code
fir.entry: ; preds = %loopexit.1.i
  call void %print( sbyte* getelementptr ([2 x sbyte]* %.str_3, int
0, int 0), float* getelementptr ([97 x float]* %y, int 0, int 0), int 97 )
  call void %exit( int 0 )
  unreachable

Can anybody please tell me what is meant by "unreachable"? The output is
coming fine.

The unreachable instruction tells the LLVM compiler that the statement cannot be reached. It allows LLVM to optimize the program more aggressively.

In this case, the unreachable instruction was probably inserted by the LLVM GCC frontend because it knows that the call to exit() will never return.

-- John T.

4

Hi,

Is it possible to generate a data dependence graph of llvm assembly file?
How can I get the data flow information in X.ll file? Also, in .ll file I
see everyline ends with something like this:
<type> [#uses=3]
What is the significance of this and is it used anyway by the compiler for
data dependence analysis.

I would be grateful for any help.

Regards,
Manvi

5

Is it possible to generate a data dependence graph of llvm assembly
file?

I am not aware of a general framework for data dependence analysis in
LLVM. It's been talked about, but hasn't (yet) been implemented.

How can I get the data flow information in X.ll file?

LLVM assembly is in SSA, data flow (use-def and def-use chains) is
explicit. What exactly are you looking for?

Also, in .ll file I see everyline ends with something like this:
<type> [#uses=3] What is the significance of this and is it used
anyway by the compiler for data dependence analysis.

Since it follows the semicolon ';' it's an LLVM comment and ignored by
llvm-as and gccas. However, it's not any extra information -- LLVM is
fully-typed so the <type> is the type of the statement on the line it
appears, and since def-use chains are explicit in LLVM (see above), the
# uses is also explicit in the code, so it's not any extra information
that isn't present already, either.

This paper is a good introduction to LLVM:
http://llvm.cs.uiuc.edu/pubs/2004-01-30-CGO-LLVM.html

6

Is it possible to generate a data dependence graph of llvm assembly file?
How can I get the data flow information in X.ll file? Also, in .ll file I
see everyline ends with something like this:
<type> [#uses=3]
What is the significance of this and is it used anyway by the compiler for
data dependence analysis.

There is no general framework at the moment. I wrote some basic data dependence analysis stuff for my research with software pipelining, but unfortunately its somewhat Sparc specific at the moment. You can take a look at:

lib/Target/SparcV9/ModuloScheduling/MSchedGraph.h
lib/Target/SparcV9/ModuloScheduling/MSchedGraph.cpp
lib/Target/SparcV9/ModuloScheduling/DependenceAnalyzer.cpp
lib/Target/SparcV9/ModuloScheduling/DependenceAnalyzer.h

Eventually it should be target independent, but some additional pieces to the puzzle need to fall in place first.

-Tanya