LLVMCallGraph

LLVMCallGraph copied to clipboard

#+TITLE: Call Graph Construction

The code is tested under =LLVM 3.9.0=

• Compile and Run To compile the tool itself, run the following commands in root directory: #+BEGIN_SRC shell cmake . make #+END_SRC

This will generate =hebicg/libHebiCallGraphPass.so=

** LLVM Bitcode file The tool performs on LLVM bitcode file(=.bc= file). To compile a =.c= or =.cpp= file into =.bc= file, simply run:

#+BEGIN_SRC shell clang -g -O0 -emit-llvm xx.c -c -o xx.bc #+END_SRC

To compile a whole benchmark program, we can make use of the Makefile shipped with the benchmark. To do that, use the following configure and make idioms:

#+BEGIN_SRC shell tar zxvf gawk-4.1.3.tar.gz cd gawk-4.1.3 ./configure CC=clang make CC=clang CFLAGS="-g -O0 -emit-llvm" #+END_SRC

Note that all the generated =.o= files are actually =.bc= files, because we give =clang= the =-emit-llvm= option. The reason the suffix is left with =.o= is that they are written in the Makefile, which is tedious to modify. And on Unix system, the suffix really doesn't matter.

Finally, to link all the bitcode files into one single bitcode file to enable call graph generation across file boundary, use the =llvm-link= tool:

#+BEGIN_SRC shell llvm-link *.bc -o single.bc #+END_SRC

** Run the tool To run the tool against the bitcode file, issue the following command

#+BEGIN_SRC shell opt -load ../hebicg/libHebiCallGraphPass.so -hebicg /path/to/benchmark.bc #+END_SRC

The call graph will be outputed as text format, as well as an =out.dot= file. Issue the command to convert to =out.png= for easy viewing.

#+BEGIN_SRC shell dot -Tpng out.dot -o out.png #+END_SRC

** Using scripts The script =benchmark/test.sh= is a simple wrapper for the above commands. The common workflow is:

#+BEGIN_SRC shell cd benchmark make # compile the c and cpp file in benchmark folder ./test.sh test1.bc make png # => out.png ./test.sh simple-c-func.bc make png # => out.png ./test.sh gawk-xxx/main.o make png # => out.png #+END_SRC

• Experiment Result | Benchmark | part | node count | edge count | Time(s) | |------------+--------+------------+------------+---------| | gawk-4.1.3 | main.o | 24 | 41 | 0.077 | | gawk-4.1.3 | all | 470 | 1065 | 2.734 | | make-4.1 | main.o | 17 | 25 | 0.112 | | make-4.1 | all | 254 | 768 | 0.966 |

For the =gawk/main.o= file, the generated png file is: [[./gawk-main.png]]

After llvm-link all the bitcode file to a single bitcode file, we get the complete call graph: [[./gawk-single.png]]

Similarly, for another real world benchmark, the =make-4.1=, call graph for =make-4.1/main.o=: [[./make-main.png]]

the complete call graph is: [[./make-single.png]]