[PATCH] tree-optimization PR/101014 - Limit new value calculations to first order effects.

[Andrew MacLeod <amacleod@redhat.com>]

[-- Attachment #1: Type: text/plain, Size: 1760 bytes --]

As mentioned in the Text from the PR:

"When a range is being calculated for an ssa-name, the propagation 
process often goes along back edges. These back edges sometime require 
other ssa-names which have not be processed yet. These are flagged as 
"poor values" and when propagation is done, we visit the list of poor 
values, calculate them, and see if that may result if a better range for 
the original ssa-name.

The problem is that calculating these poor values may also spawn another 
set of requests since the block at the far end of the back edge has not 
been processed yet... its highly likely that some additional unprocessed 
ssa-names are used in the calculation of that name, but typically they 
do not affect the current range in a significant way.

Thus we mostly we care about the first order effect only.  It turns out 
to be very rare that a 2nd order effect on a back edge affects anything 
that we don't catch later.

This patch turns off poor-value tagging when looking up the first order 
values, thus avoiding the 2nd order and beyond cascading effects.

I haven't found a test case we miss yet because of this change, yet it 
probably resolves a number of the outstanding compilation problems in a 
significant way.

I think this will probably apply to gcc 11 in some form as well, so I'll 
look at an equivalent patch for there."


This patch simplifies the enable_new_value routines.. replacing the 
enable/disable with an enable with flag routine, which returns the 
previous value.This lets us change the mode and then set it back to what 
it was before.  Seems better in general.

Then disables new values for 2nd+ order effects. GCC11 patch forthcoming.

Bootstraps on x86_64-pc-linux-gnu, no regressions.  pushed.

Andrew


[-- Attachment #2: 101014.patch --]
[-- Type: text/x-patch, Size: 3760 bytes --]

commit 1d65d96fc3a0bc7139e619738793bf3c0ec3d0fa
Author: Andrew MacLeod <amacleod@redhat.com>
Date:   Mon Jun 14 15:33:59 2021 -0400

    Limit new value calculations to first order effects.
    
    When utilzing poor values during propagation, we mostly care about values that
    were undefined/processed directly used in calcualting the SSA_NAME being
    processed.  2nd level derivations of such poor values rarely affect the
    inital calculation.  Leave them to when they are directly encountered.
    
            * gimple-range-cache.cc (ranger_cache::ranger_cache): Adjust.
            (ranger_cache::enable_new_values): Set to specified value and
            return the old value.
            (ranger_cache::disable_new_values): Delete.
            (ranger_cache::fill_block_cache): Disable non 1st order derived
            poor values.
            * gimple-range-cache.h (ranger_cache): Adjust prototypes.
            * gimple-range.cc (gimple_ranger::range_of_expr): Adjust.

diff --git a/gcc/gimple-range-cache.cc b/gcc/gimple-range-cache.cc
index 249515fbaf5..d9a57c294df 100644
--- a/gcc/gimple-range-cache.cc
+++ b/gcc/gimple-range-cache.cc
@@ -727,7 +727,7 @@ ranger_cache::ranger_cache (gimple_ranger &q) : query (q)
       if (bb)
 	m_gori.exports (bb);
     }
-  enable_new_values ();
+  enable_new_values (true);
 }
 
 ranger_cache::~ranger_cache ()
@@ -748,21 +748,15 @@ ranger_cache::dump (FILE *f)
   fprintf (f, "\n");
 }
 
-// Allow the cache to flag and query new values when propagation is forced
-// to use an unknown value.
+// Allow or disallow the cache to flag and query new values when propagation
+// is forced to use an unknown value.  The previous state is returned.
 
-void
-ranger_cache::enable_new_values ()
-{
-  m_new_value_p = true;
-}
-
-// Disable new value querying.
-
-void
-ranger_cache::disable_new_values ()
+bool
+ranger_cache::enable_new_values (bool state)
 {
-  m_new_value_p = false;
+  bool ret = m_new_value_p;
+  m_new_value_p = state;
+  return ret;
 }
 
 // Dump the caches for basic block BB to file F.
@@ -1343,7 +1337,12 @@ ranger_cache::fill_block_cache (tree name, basic_block bb, basic_block def_bb)
 
 	  // Calculate a range at the exit from the block so the caches feeding
 	  // this block will be filled, and we'll get a "better" value.
+	  // Disallow additonal "poor values" during this phase to avoid
+	  // iterations that are unlikely to be profitable for this name.
+	  // See PR 101014.
+	  bool state = enable_new_values (false);
 	  query.range_on_exit (tmp, calc_bb, rec.calc);
+	  enable_new_values (state);
 	  
 	  // Then ask for NAME to be re-evaluated on outgoing edges and 
 	  // use any new values.
diff --git a/gcc/gimple-range-cache.h b/gcc/gimple-range-cache.h
index ce4449a08db..1a2aaceb3ed 100644
--- a/gcc/gimple-range-cache.h
+++ b/gcc/gimple-range-cache.h
@@ -101,8 +101,7 @@ public:
   bool get_non_stale_global_range (irange &r, tree name);
   void set_global_range (tree name, const irange &r);
 
-  void enable_new_values ();
-  void disable_new_values ();
+  bool enable_new_values (bool state);
   non_null_ref m_non_null;
   gori_compute m_gori;
 
diff --git a/gcc/gimple-range.cc b/gcc/gimple-range.cc
index b534b8e0a2c..481b89b2b80 100644
--- a/gcc/gimple-range.cc
+++ b/gcc/gimple-range.cc
@@ -1173,9 +1173,9 @@ gimple_ranger::range_of_expr (irange &r, tree expr, gimple *stmt)
   // trigger new value calculations.  PR 100781.
   if (is_gimple_debug (stmt))
     {
-      m_cache.disable_new_values ();
+      bool state = m_cache.enable_new_values (false);
       m_cache.range_of_expr (r, expr, stmt);
-      m_cache.enable_new_values ();
+      m_cache.enable_new_values (state);
       return true;
     }
   basic_block bb = gimple_bb (stmt);