tree-ssa: a weird (mis)optimization

[Andrew Haley]

Richard Henderson writes:
 > On Wed, Dec 01, 2004 at 10:13:45AM +0000, Andrew Haley wrote:
 > > Let me put it this way: a trapping instruction has an abnormal edge in
 > > the CFG, so why would it get deleted?
 > 
 > Well, for one, I didn't think things worked this way.  It doesn't
 > make a whole lot of sense to me.  We work hard to *get rid* of 
 > instructions that trap, and to remove the associated abnormal edges.

I see.

 > I still think that the Java front end should be doing something in
 > order to indicate that these loads should not be removed, beyond
 > just marking them as trapping -- or rather, not marking them as not
 > trapping.  

Well, okay.  What I need is something that gets treated as

      (p ? *p : throw)
  
... I don't want to make every load volatile, because it will cause
unnecessary loads.  Once we've determined that a read from p won't
trap we don't want to read it again.  So how could a front end express
such a thing?

 > Otherwise -fnon-call-exceptions will suddenly generate
 > *significantly* worse code for everyone else.
 
I guess that depends on how many unnecessary reads from memory one can
eliminate with production code.  Yeah, I know...

Andrew.