------- Comment #11 from law at redhat dot com 2006-11-07 21:03 ------- Subject: Re: bootstrap comparision fails with "-ftree-vectorize -maltivec" on ppc
On Tue, 2006-11-07 at 08:31 +0000, irar at il dot ibm dot com wrote: > > ------- Comment #9 from irar at il dot ibm dot com 2006-11-07 08:31 ------- > In r110758 (and r110705) bootstrap with BOOT_CFLAGS="-O2 -g -ftree-vectorize > -maltivec" fails with another error: [ ... ] I've been unable to make any progress with this. It's not even clear to me right now that the .i file you sent is valid. Every compiler I've tried to use to process that .i file complains that the size of the test_real_width array is negative. Clearly I'm doing something different than you are. I can speculate that one of a few things might be happening. 1. We might have some kind of memory corruption or similar bug causing us to zero the SSA_NAME_DEF_STMT/TREE_CHAIN field of the affected SSA_NAME. Possible, but unlikely. 2. We have an SSA_NAME in the IL which was deleted. 3. Something else totally unexpected. #2 has happened in the past and I'd consider it the most likely culprit. What is interesting is that a checking failure was not triggered -- our SSA verifiers should be checking for case #2 and issuing an error. So let's start with checking a couple things. 1. Put a breakpoint in tree_ssa_cd_dce when compiling the offending function from recog.c. When that breakpoint triggers issue: verify_ssa (true) I can't see any way for that to fail, but better safe than sorry. 2. When the gcc_assert triggers, issue: debug_tree (op) debug_generic_stmt (op) 3. Send me the phi-only copyprop dumps. If #1 passes and #2 indicates the name in question is on the free list, then either DCE is mucking things up or we have out-of-date operands in the operand cache relative to the IL (which I can detect with the dumps). Jeff -- http://gcc.gnu.org/bugzilla/show_bug.cgi?id=28752