diff mbox series

rtl-ssa: Fix split_clobber_group tree insertion [PR116044]

Message ID mptttgfypcv.fsf@arm.com
State New
Headers show
Series rtl-ssa: Fix split_clobber_group tree insertion [PR116044] | expand

Commit Message

Richard Sandiford July 24, 2024, 11:57 a.m. UTC
PR116044 is a regression in the testsuite on AMD GCN caused (again)
by the split_clobber_group code.  The first patch in this area
(g:71b31690a7c52413496e91bcc5ee4c68af2f366f) fixed a bug caused
by carrying the old group over as one the split ones.  That patch
instead:

- created two new groups
- inserted them in the splay tree as neighbours of the old group
- removed the old group, and
- invalidated the old group (to force lazy recomputation when
  a clobber's parent group is queried)

However, this left add_def trying to insert the new definition
relative to a stale splay tree root.  The second patch
(g:34f33ea801563e2eabb348e8d3e9344a91abfd48) attempted to fix
that by inserting it relative to the new root.  But that's not
always correct either.  We specifically want to insert it after
the first of the two new groups, whether that group is the root
or not.

This patch does that, and tries to refactor the code to make
it a bit less brittle.

Bootstrapped & regression-tested on aarch64-linux-gnu and
x86_64-linux-gnu.  OK to install?

Sorry for all the trouble that this code has caused :-(

Richard

gcc/
	PR rtl-optimization/116044
	* rtl-ssa/accesses.h (function_info::split_clobber_group): Return
	an array of two clobber_groups.
	* rtl-ssa/accesses.cc (function_info::split_clobber_group): Return
	the new clobber groups.  Don't modify the splay tree here.
	(function_info::add_def): Update call accordingly.  Generalize
	the splay tree insertion code so that the new definition can be
	inserted as a child of any existing node, not just the root.
	Fix the insertion used after calling split_clobber_group.
---
 gcc/rtl-ssa/accesses.cc | 66 +++++++++++++++++++++++------------------
 gcc/rtl-ssa/functions.h |  3 +-
 2 files changed, 39 insertions(+), 30 deletions(-)

Comments

Jeff Law July 24, 2024, 8:21 p.m. UTC | #1
On 7/24/24 5:57 AM, Richard Sandiford wrote:
> PR116044 is a regression in the testsuite on AMD GCN caused (again)
> by the split_clobber_group code.  The first patch in this area
> (g:71b31690a7c52413496e91bcc5ee4c68af2f366f) fixed a bug caused
> by carrying the old group over as one the split ones.  That patch
> instead:
> 
> - created two new groups
> - inserted them in the splay tree as neighbours of the old group
> - removed the old group, and
> - invalidated the old group (to force lazy recomputation when
>    a clobber's parent group is queried)
> 
> However, this left add_def trying to insert the new definition
> relative to a stale splay tree root.  The second patch
> (g:34f33ea801563e2eabb348e8d3e9344a91abfd48) attempted to fix
> that by inserting it relative to the new root.  But that's not
> always correct either.  We specifically want to insert it after
> the first of the two new groups, whether that group is the root
> or not.
> 
> This patch does that, and tries to refactor the code to make
> it a bit less brittle.
> 
> Bootstrapped & regression-tested on aarch64-linux-gnu and
> x86_64-linux-gnu.  OK to install?
> 
> Sorry for all the trouble that this code has caused :-(
We're both on the hot seat for different reasons.  Better now than late 
fall though.


> 
> Richard
> 
> gcc/
> 	PR rtl-optimization/116044
> 	* rtl-ssa/accesses.h (function_info::split_clobber_group): Return
> 	an array of two clobber_groups.
> 	* rtl-ssa/accesses.cc (function_info::split_clobber_group): Return
> 	the new clobber groups.  Don't modify the splay tree here.
> 	(function_info::add_def): Update call accordingly.  Generalize
> 	the splay tree insertion code so that the new definition can be
> 	inserted as a child of any existing node, not just the root.
> 	Fix the insertion used after calling split_clobber_group.
You know this code far better than anyone.  OK.

jeff
diff mbox series

Patch

diff --git a/gcc/rtl-ssa/accesses.cc b/gcc/rtl-ssa/accesses.cc
index 0bba8391b00..5450ea118d1 100644
--- a/gcc/rtl-ssa/accesses.cc
+++ b/gcc/rtl-ssa/accesses.cc
@@ -792,12 +792,12 @@  function_info::merge_clobber_groups (clobber_info *clobber1,
 }
 
 // GROUP spans INSN, and INSN now sets the resource that GROUP clobbers.
-// Split GROUP around INSN, to form two new groups, and return the clobber
-// that comes immediately before INSN.
+// Split GROUP around INSN, to form two new groups.  The first of the
+// returned groups comes before INSN and the second comes after INSN.
 //
-// The resource that GROUP clobbers is known to have an associated
-// splay tree.  The caller must remove GROUP from the tree on return.
-clobber_info *
+// The caller is responsible for updating the def_splay_tree and chaining
+// the defs together.
+std::array<clobber_group *, 2>
 function_info::split_clobber_group (clobber_group *group, insn_info *insn)
 {
   // Search for either the previous or next clobber in the group.
@@ -835,14 +835,10 @@  function_info::split_clobber_group (clobber_group *group, insn_info *insn)
   auto *group1 = allocate<clobber_group> (first_clobber, prev, tree1.root ());
   auto *group2 = allocate<clobber_group> (next, last_clobber, tree2.root ());
 
-  // Insert GROUP2 into the splay tree as an immediate successor of GROUP1.
-  def_splay_tree::insert_child (group, 1, group2);
-  def_splay_tree::insert_child (group, 1, group1);
-
   // Invalidate the old group.
   group->set_last_clobber (nullptr);
 
-  return prev;
+  return { group1, group2 };
 }
 
 // Add DEF to the end of the function's list of definitions of
@@ -899,7 +895,7 @@  function_info::add_def (def_info *def)
   insn_info *insn = def->insn ();
 
   int comparison;
-  def_node *root = nullptr;
+  def_node *neighbor = nullptr;
   def_info *prev = nullptr;
   def_info *next = nullptr;
   if (*insn > *last->insn ())
@@ -909,8 +905,8 @@  function_info::add_def (def_info *def)
       if (def_splay_tree tree = last->splay_root ())
 	{
 	  tree.splay_max_node ();
-	  root = tree.root ();
-	  last->set_splay_root (root);
+	  last->set_splay_root (tree.root ());
+	  neighbor = tree.root ();
 	}
       prev = last;
     }
@@ -921,8 +917,8 @@  function_info::add_def (def_info *def)
       if (def_splay_tree tree = last->splay_root ())
 	{
 	  tree.splay_min_node ();
-	  root = tree.root ();
-	  last->set_splay_root (root);
+	  last->set_splay_root (tree.root ());
+	  neighbor = tree.root ();
 	}
       next = first;
     }
@@ -931,8 +927,8 @@  function_info::add_def (def_info *def)
       // Search the splay tree for an insertion point.
       def_splay_tree tree = need_def_splay_tree (last);
       comparison = lookup_def (tree, insn);
-      root = tree.root ();
-      last->set_splay_root (root);
+      last->set_splay_root (tree.root ());
+      neighbor = tree.root ();
 
       // Deal with cases in which we found an overlapping live range.
       if (comparison == 0)
@@ -943,22 +939,34 @@  function_info::add_def (def_info *def)
 	      add_clobber (clobber, group);
 	      return;
 	    }
-	  prev = split_clobber_group (group, insn);
-	  next = prev->next_def ();
+	  auto new_groups = split_clobber_group (group, insn);
+
+	  // Insert the two new groups immediately after GROUP.
+	  def_splay_tree::insert_child (group, 1, new_groups[1]);
+	  def_splay_tree::insert_child (group, 1, new_groups[0]);
+
+	  // Remove GROUP.
 	  tree.remove_root ();
-	  root = tree.root ();
-	  last->set_splay_root (root);
+	  last->set_splay_root (tree.root ());
+
+	  prev = new_groups[0]->last_clobber ();
+	  next = new_groups[1]->first_clobber ();
+
+	  // DEF comes after the first group.  (new_groups[1] and -1 would
+	  // also work.)
+	  neighbor = new_groups[0];
+	  comparison = 1;
 	}
-      // COMPARISON is < 0 if DEF comes before ROOT or > 0 if DEF comes
-      // after ROOT.
+      // COMPARISON is < 0 if DEF comes before NEIGHBOR or > 0 if DEF comes
+      // after NEIGHBOR.
       else if (comparison < 0)
 	{
-	  next = first_def (root);
+	  next = first_def (neighbor);
 	  prev = next->prev_def ();
 	}
       else
 	{
-	  prev = last_def (root);
+	  prev = last_def (neighbor);
 	  next = prev->next_def ();
 	}
     }
@@ -973,12 +981,12 @@  function_info::add_def (def_info *def)
     append_clobber_to_group (clobber, need_clobber_group (prev_clobber));
   else if (clobber && next_clobber)
     prepend_clobber_to_group (clobber, need_clobber_group (next_clobber));
-  else if (root)
+  else if (neighbor)
     {
-      // If DEF comes before ROOT, insert DEF to ROOT's left,
-      // otherwise insert DEF to ROOT's right.
+      // If DEF comes before NEIGHBOR, insert DEF to NEIGHBOR's left,
+      // otherwise insert DEF to NEIGHBOR's right.
       def_node *node = need_def_node (def);
-      def_splay_tree::insert_child (root, comparison >= 0, node);
+      def_splay_tree::insert_child (neighbor, comparison >= 0, node);
     }
   if (prev)
     insert_def_after (def, prev);
diff --git a/gcc/rtl-ssa/functions.h b/gcc/rtl-ssa/functions.h
index 8be04f1aa96..567701c7995 100644
--- a/gcc/rtl-ssa/functions.h
+++ b/gcc/rtl-ssa/functions.h
@@ -257,7 +257,8 @@  private:
   void append_clobber_to_group (clobber_info *, clobber_group *);
   void merge_clobber_groups (clobber_info *, clobber_info *,
 			     def_info *);
-  clobber_info *split_clobber_group (clobber_group *, insn_info *);
+  std::array<clobber_group *, 2> split_clobber_group (clobber_group *,
+						      insn_info *);
 
   void append_def (def_info *);
   void add_def (def_info *);