Re: [tree-ssa] bootstrap failure

[law at redhat dot com]

In message <1069870197.3158.6.camel@frodo.toronto.redhat.com>, Diego Novillo wr
ites:
 >On Wed, 2003-11-26 at 10:59, law@redhat.com wrote:
 >
 >>  >That seems unnecessary to me.  This is happening in expr_phi_insertion
 >>  >right after we compute the blocks that will contain EPHIs.  I'm going to
 >>  >try stopping it from requesting a PHI node of length 1.  If that fails,
 >>  >the quickest fix for me will be to add a check in make_phi_nodes and
 >>  >disable PRE.
 >> Just force it to allocate a PHI with a capacity of 2, which is safe, even
 >> if we only end up using one argument.
 >> 
 >There's definitely more than one bad interaction going on between PRE
 >and the new PHI manager.  After fixing the allocation problem, I'm still
 >getting various SEGVs on alpha, ia64 and amd64.  They all happen
 >somewhere in the GC, as it tries to mark roots.  It doesn't happen on
 >any of the 32bit hosts, though.
 >
 >Jeff, I am going to leave the branch closed until you can figure out
 >what may be going on.  This patch seems to let me go beyond the original
 >failures on these hosts.  It may probably be fastest to do try and fix
 >the problem on an alpha box as this is where we die the soonest (we ICE
 >building libgcc with the stage1 compiler).
 >
 >If you think it'll take you longer to find the problem, I'll go ahead
 >and revert the PHI manager patch.
Well, I did manage to trigger the failure on my alpha, but not on the one in
Mountain View (odd).  Unfortunately, the debugger on my alpha is so old that
it doesn't grok the debug symbols emitted by the current compiler.  So I 
can't debug it yet.  I'm in the process of upgrading things now.

Given the fact that tomorrow is Thanksgiving in the US and that I 
probably won't be online, if you want to revert the patch.  Go ahead.


jeff