Re: [patch] tree-ssa-sink.c: Simplify/improve code sinking.

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

On Wed, 2005-04-06 at 11:10 -0400, Kazu Hirata wrote:
> Hi,
> 
> Attached is a patch to simplify/improve code sinking.
> 
> Given a statement, the code sinking in tree-ssa-sink.c computes a
> location to sink a statement into by finding the nearest common
> dominator of all uses of the statement's result.  To do so, it goes
> through immediate uses of the statement's result.
> 
> It turns out that nearest_common_dominator_of_uses are going through
> things more than immediate uses.  Specifically, it was going through
> all PHI arguments of a PHI node where at least one use occurs.  Since
> uses of operands in PHI arguments are considered to occur in
> predecessors of edges going into a PHI node, we are computing the
> nearest common dominator of a larger set of basic blocks, which
> obviously makes the result of nearest_common_dominator_of_uses
> unnecessarily higher in the dominator tree.
> 
> The patch fixes this problem by visiting a PHI argument where a use
> actually occurs.  Fortunately, Andrew MacLeod introduced a new gadget
> PHI_ARG_INDEX_FROM_USE, so we don't have to use a "for" loop to find a
> PHI argument where a use occurs.
> 
> This patch is compile-time neutral, even when I use bootstrapped
> compilers.  After all, the "for" loop in question isn't a hot place.
> 
> Without patch, this pass sinks 2079 statements while compiling cc1-i
> files.  With this patch, it sinks 2179 statements.  A few statements
> are sunk deeper with this patch.
> 
> Tested on i686-pc-linux-gnu.  OK to apply?
> 
> Kazu Hirata
> 
> 2005-04-06  Kazu Hirata  <kazu@cs.umass.edu>
> 
> 	* tree-ssa-sink.c (nearest_common_dominator_of_uses): Look at
> 	a PHI argument where a use occurs instead of all PHI arguments.
This is fine.  THanks,

Jeff