This fixes a bug in code adding edges to the dependence graph. Values for this_dir can be -1 (backward edge), 0 (no edge), 1 (forward edge), and 2 (both backward and forward edges). There can be multiple dependencies checked, creating multiple edges that have to be merged together. this_dir contains the current edge. dir contains the previous edges. The code fails to handle the case where this_dir is 2, in which case we can return immediately. This is a minor optimization to improve compile time. The code handles the case where dir is non-zero and this_dir is zero by returning 2, which is incorrect. dir should be unmodified in this case. We can return 2 only if both dir and this_dir are non-zero and unequal, i.e. one is -1 and the other is 1. This problem creates extra unnecessary edges, which can prevent loops from being distributed. The patch fixes both problems.
This passed an x86_64 gcc bootstrap with -ftree-loop-distribution added to BOOT_CFLAGS and a testsuite regression check. Curiously, I see that I get different results for the C/C++ ubsan tests every time I run them, but this has nothing to do with my patch, as it happens with or without my patch. I haven't tried debugging this yet. Might be related to my recent upgrade to Ubuntu 16.04 LTS. Otherwise, there are no regressions. On SPEC CPU2006, on aarch64, I see 5879 loops distributed without the patch, and 5906 loops distributed with the patch. So 27 extra loops are distributed which is about 0.5% more loop distributions. There is no measurable performance gain from the bug fix on the CPU2006 run time though I plan to spend some more time looking at this code to see if I can find other improvements. OK? Jim
2016-11-09 Jim Wilson <jim.wil...@linaro.org> * tree-loop-distribution.c (pg_add_dependence_edges): Return 2 if this_dir is 2. Check for this_dir non-zero before dir != this_dir check. Index: gcc/tree-loop-distribution.c =================================================================== --- gcc/tree-loop-distribution.c (revision 242025) +++ gcc/tree-loop-distribution.c (working copy) @@ -1408,9 +1408,11 @@ pg_add_dependence_edges (struct graph *rdg, vec<lo else this_dir = 0; free_dependence_relation (ddr); - if (dir == 0) + if (this_dir == 2) + return 2; + else if (dir == 0) dir = this_dir; - else if (dir != this_dir) + else if (this_dir != 0 && dir != this_dir) return 2; /* Shuffle "back" dr1. */ dr1 = saved_dr1;
