First, a ping for <http://gcc.gnu.org/ml/gcc-patches/2011-06/msg00136.html>; updated regclass documentation. Ok?
Second, I updated the CRIS port to fit the proposed documentation update (adding a class as the patch you sent, but more complete), with regtest results clean for revisions before the revision where build started failing. But, after that revision, I get a SEGV; details added to <http://gcc.gnu.org/bugzilla/show_bug.cgi?id=49154>. At first I thought I messed up the regclass description but it appears to be just an obvious miss, fixed below. If it isn't, I think setup_regno_cost_classes_by_mode should have a comment as to who else should set up regno_cost_classes[regno] and that "who else" needs to be fixed to do the setup. Odd that this bug didn't trig before or for any of the targets on which you tested. Maybe that makes cris-elf qualify for the set of targets you test IRA changes on? By the way, can we do a s/_aclass/_class/ in applicable files, for example ira-costs.c? Someone appears to have done a botched word-replace of class to aclass; besides the identifier "class" it changed occurrences within identifiers too (at least those after a _) so we have e.g. cost_classes_aclass_cache and setup_regno_cost_classes_by_aclass vs. cost_classes_mode_cache and setup_regno_cost_classes_by_mode. Tested on cris-elf; together with the mentioned update it restores build with results consistent with those before the breakage. Ok to commit? gcc: PR rtl-optimization/49154 * ira-costs.c (setup_regno_cost_classes_by_mode): If there already is a matching slot in the hashtable, assign it to classes_ptr. diff --git gcc/ira-costs.c gcc/ira-costs.c index f517386..a22bb15 100644 --- gcc/ira-costs.c +++ gcc/ira-costs.c @@ -299,6 +299,8 @@ setup_regno_cost_classes_by_mode (int regno, enum machine_mode mode) classes_ptr = setup_cost_classes (&classes); *slot = classes_ptr; } + else + classes_ptr = *slot; cost_classes_mode_cache[mode] = (cost_classes_t) *slot; } regno_cost_classes[regno] = classes_ptr; brgds, H-P
