Re: [PATCH] Fix (?) load_kills_store

[Jim Wilson <wilson at tuliptree dot org>]

Zdenek Dvorak wrote:
> considering description of load_kills_store, it seems to me that
> arguments to true_dependence are passed in reversed order, which
> could cause problems when one of the mems is unchanging (actually
> switching the order seems to fix a misscompilation with my dce of
> mems patch).

This looks over simplistic to me.  Looking at the code in gcse.c, I see 
that both store_killed_before and store_killed_after end up here, so it 
isn't safe to assume that either mem comes first.  Which comes first 
depends on whether we started in store_killed_before or 
store_killed_after.  Also, despite the name, load_kills_store is used 
for both loads and stores, and clobbers too.

I think we need a more complicated fix, where we keep track of whether 
we are searching before or after the store.  Also, we probably need to 
keep track of whether we have a read or a write, since strictly speaking 
we should be calling different routines, for different cases.  It isn't 
clear if this is important though.  I think keeping track of the 
before/after info is the main problem.

Jim