Write seperate pass for function dot file generation. Don't overwrite dot files.

[zjuchenyuan]

Different binaries may have the functions with same name, the most common one is the main function.

When generating dot files, only function name is kept in naming the dot file for this function. Therefore, for preprocessing building process, only the last compiled binary's main function's dot file will be retained.

For example, following the instruction from README, to build xmllint, you give make all, and I get this cfg.main.dot:

digraph "CFG for 'main' function" {
        label="CFG for 'main' function";

        Node0x38c0b10 [shape=record,label="{gjobread.c:288:}"];
        Node0x38c0b10 -> Node0x39ce120;
        Node0x39ce120 [shape=record,label="{gjobread.c:296:}"];
        Node0x39ce120 -> Node0x39ce400;
        Node0x39ce120 -> Node0x39ce020;
        Node0x39ce400 [shape=record,label="{gjobread.c:297:}"];
        Node0x39ce400 -> Node0x39cd710;
        Node0x39ce400 -> Node0x39cd7b0;
        Node0x39cd710 [shape=record,label="{gjobread.c:299:}"];
        Node0x39cd710 -> Node0x39cd760;
        Node0x39cd7b0 [shape=record,label="{gjobread.c:301:}"];
        Node0x39cd7b0 -> Node0x39cd760;
        Node0x39cd760 [shape=record,label="{gjobread.c:303:}"];
        Node0x39cd760 -> Node0x39ce370;
        Node0x39ce370 [shape=record,label="{gjobread.c:296:}"];
        Node0x39ce370 -> Node0x39ce120;
        Node0x39ce020 [shape=record,label="{gjobread.c:306:}"];
}

Since example/gjobread is the last built binary, its main function dot is kept, not the target binary xmllint.

I think this is a problem leading to wrong CFG distance analaysis, and I suggest:

• use make xmllint to avoid building other binaries
• change the generated dot file names (maybe keeping same name files with .1 .2?)

[mboehme]

Thanks ChenYuan! I'll add make xmlint to our Readme.md.

In the longer run, we need to update our scripts so we don't override dot-files.

[mboehme]

Just reopening, so we don't loose track.

[zjuchenyuan]

Do not override dot files is easier said than done. When compiling a single compile unit, I think the compiler do not know which binary this unit will be built into. So, the compiler cannot just name the dot file as format cfg.<binary>_<function name>.dot

Even if we keep same named files using like .1, .2, which file should distance calculation use?

I think the easist way to solve this issue is to require user to just compile target binary, like make xmllint, just add the target binary as a make argument should be enough.

[Liblor]

One way to address this, is to extract the function dot file generation part of the pass into a separate pass. Then one could use opt with this new pass to create the dot files in the genDistance.sh/gen_distance_fast.py, the files could then be put into a dedicated subfolder.

Something like

opt -gen-func-dot -name xmllint xmllint.0.0.preopt.bc

[mboehme]

@Liblor, this is a great suggestion!

@zjuchenyuan You are welcome to submit a pull request if this works for you.