Re: [tree-ssa] New test for dominator optimizations bug

[law at redhat dot com]

In message <1061667453.22500.12.camel@frodo.toronto.redhat.com>, Diego Novillo 
writes:
 >
 >This test case exposes a bug in the dominator optimizer.  We are
 >exposing symbols without SSA version numbers.
Arguably we shouldn't have propagated "c" since it's not an SSA
variable.  I'd probably tend to prefer that solution to trying to
copy the virtual operands.  To get the optimization we want, we
should be copying "c" into a temporary before we use it.  Both Jason
and I have bits lying around to do that.

[ I've got a one liner which will do that, but it regresses the gcov
  tests and I haven't been able to sort through how all the gcov stuff
  works well enough to understand why the test is failing. ]

In my experience the few times I've found myself looking to copy
virtual operands from one statement to another the better solution
has been to ensure the key operand is copied into an SSA temporary
variable and to use that SSA temporary rather than the original
variable.

 >A fix is in progress.  I'm going to factor out some code out of
 >optimize_stmt.  It's a bit difficult to follow as it is now.
Feel free, that's always been the plan -- I had planned on that 
once I had a better idea of what stuff we were missing so that
any restructuring would make it easier to add the missing bits.


Jeff