I have a project to rewrite llvm output as OpenCL, https://github.com/hughperkins/cuda-on-cl It currently contains lots of conditional branches, whch become /labels / gotos in the output OpenCL.
What options are available to ‘reloop’ this to contain strictly ifs/fors/switches/breaks/whiles ?
There is the Böhm-Jacopini theorem, but that will probably be
inefficient, but necessary as a last resort for cases that can't
The cbe (https://github.com/Ace17/llvm-cbe) seems to actually put some
loop constructs in the output, even though teh actual control flow is
done via goto, but the loop constructs there might help as a starting point.
You might try:
which claims to remove goto's without code replication.
the method avoids problems of code replication