********************************************************************************
********************************************************************************
********************************************************************************

To: mock@thistle.cs.washington.edu
cc: spin-comp-archive@cs
Subject: Discordant Variables and Loop Unrolling
Date: Wed, 01 Oct 1997 13:34:26 PDT
From: Brian Grant <grant@thistle.cs.washington.edu>

Two edges are mutually exclusive if the conjunction of their reachability
conditions is not satisfiable.  A merge is static if all of its incoming
edges are pairwise mutually exclusive.

So, with our new BTA, loop-head merges are never found to be "static", based
on reachability conditions. 

This has implications for loop unrolling in the current design.
	-- Loop-head merges are always unit boundaries.  There will be a
	   cache lookup at the loop-head merge unless the laziness policy is
	   Eager and the cache policy is CacheAllUnchecked.

	-- Derived static temps will always be killed at loop-head merges.

In the future, I might eliminate the unit boundary at the loop head in the
case where no cache lookup is needed.  The cache lookup could also be
eliminated if the loop is to be unrolled eagerly and the discordant variables
change monotonically (there must be an analysis in the literature that
determines this).  Loop laziness can be eliminated if all exit conditions
are static.

We might also want to address the second problem.

In general, how do we know we don't have to kill discordant derived static
temps at a discordant merge?  We must be able to figure that out from control-
dependence info and def-use chains.  We essentially need to know which derived
temps vary with the discordant ones.

--Brian