A way to discover the label of a block (in ValueRef)

br instructions have an argument telling what label(s) they are branching to, and phi instructions have arguments telling what labels they might have arrived from, but there is no way to match up these labels with the blocks they belong to.

I am trying to do control-flow analysis, including determining how many different paths are possible through any given block, and which blocks are located inside loops. Code sample:

int sample(int arg) {
   for(int i=0; i<10; i++) {
      arg = (2*arg + 3*i) % 71;
   }
   return arg;
}

Compile this with -O1 and generate LL output, then use llvm.parse_assembly(). In the i32 @sample(i32) function there will be at least three blocks and a phi statement that “comes from” two of the blocks. Task: write a Python program that uses llvmlite to answer the question: “Which two blocks are the ones that might branch to that phi statement?”

Disclaimer: I don’t know how to do this within the “real” LLVM API, and I am a new to Python3 programming. I don’t see the answer in the (rather limited) documentation of the ValueRef class.

I am using Python 3.6.8, llvmlite 0.33.0, and clang 9.0.1 on a CentOS system. I built the clang from source but I don’t know how to build python3 or llvmlite.

[esc]

@mrob27 thanks for submitting this! I have labelled it as a question for now.

[lwerdna]

I also want this feature.

Until then, you can parse the first line of the string representation of a block:

def get_block_label(block):
    lines = str(block).split('\n')
    try:
        i = 1 if lines[0] == '' else 0
        if m := re.match(r'^(\d+):', lines[i]):
            return m.group(1)
    except:
        return None