search for: bz2_bzbufftobuffcompress

.... At the moment run time is comparable to dominance tree calculation. If you want, have a look into some results I got with a pass extracting maximal non trivial regions that do not contain loops from libbz2: http://tobias.osaft.eu/llvm/region/bz2NoLoops/ Interesting ones: regions_without_loops.BZ2_bzBuffToBuffCompress.dot.png has a lot of exit edges. regions_without_loops.bzopen_or_bzdopen.dot.png the first region has two entry edges. One is the loop latch. (Keep in mind all regions have the same color, so if it seems there is an edge into a region, there are just two regions close by) Without a prepass that...

...>> If you want, have a look into some results I got with a pass extracting >> maximal non trivial regions that do not contain loops from libbz2: >> >> http://tobias.osaft.eu/llvm/region/bz2NoLoops/ >> >> Interesting ones: >> >> regions_without_loops.BZ2_bzBuffToBuffCompress.dot.png >> has a lot of exit edges. > > I think this example proves the strengths and weaknesses of both > approaches. Making that region into structured control flow would add a lot > of additional blocks. This will also happen after generating code > from the polyhedral mode...

...got with a pass extracting >>>> maximal non trivial regions that do not contain loops from libbz2: >>>> >>>> http://tobias.osaft.eu/llvm/region/bz2NoLoops/ >>>> >>>> Interesting ones: >>>> >>>> regions_without_loops.BZ2_bzBuffToBuffCompress.dot.png >>>> has a lot of exit edges. >>> >>> I think this example proves the strengths and weaknesses of both >>> approaches. Making that region into structured control flow would add a lot >>> of additional blocks. This will also happen after genera...