Re: replace_ssa_names abort that started last week

[Jeffrey A Law <law at redhat dot com>]

On Wed, 2004-05-26 at 20:51, Bradley Lucier wrote:
> Jeff (and others);
> 
> PR 15643 began last week some time, which correlates when you were 
> making changes to the tree optimization code. There are short testcases 
> in C and Java that fail on multiple platforms; perhaps this could be 
> investigated while the problem is still fresh.
The underlying problem was that kill_redundant_phi_nodes was doing
copy propagation even though the result and/or an argument to a
PHI was marked with SSA_NAME_OCCURS_IN_ABNORMAL_PHI.  When the code
actually tried to do the replacements it aborted.

This patch fixes the problem by raising the destination of a PHI node
to the top of the lattice if the destination or any argument is
marked with SSA_NAME_OCCURS_IN_ABNORMAL_PHI.  Raising the value to the
top of the lattice prevents the problematical copy propagation.

Bootstrapped and regression tested on i686-pc-linux-gnu.




> 
> Brad

Attachment: ZZZ
Description: Text document