On 18/08/17 15:16, Edward Cree wrote: > On 18/08/17 04:21, Alexei Starovoitov wrote: >> It seems you're trying to sort-of do per-fake-basic block liveness >> analysis, but our state_list_marks are not correct if we go with >> canonical basic block definition, since we mark the jump insn and >> not insn after the branch and not every basic block boundary is >> properly detected. > I think the reason this works is that jump insns can't do writes. > [snip] > the sl->state will never have any write marks and it'll all just work. > But I should really test that! I tested this, and found that, no, sl->state can have write marks, and the algorithm will get the wrong answer in that case. So I've got a patch to make the first iteration ignore write marks, as part of a series which I will post shortly. When I do so, please re-do your tests with adding state_list_marks in strange and exciting places; it should work wherever you put them. Like you say, it "magically doesn't depend on proper basic block boundaries", and that's because really pruning is just a kind of checkpointing that just happens to be most effective when done just after a jump (pop_stack).
Can I have a SOB for your "grr" test program, so I can include it in the series? -Ed
