public inbox for gcc-patches@gcc.gnu.org
 help / color / mirror / Atom feed
* [PATCH] Fix wrong-code issue in VN
@ 2023-02-17 11:30 Richard Biener
  0 siblings, 0 replies; only message in thread
From: Richard Biener @ 2023-02-17 11:30 UTC (permalink / raw)
  To: gcc-patches

The following fixes the wrong removed dead store discovered on the
PR108657 testcase when the reported DSE issue is not fixed.
The issue is we were using ssa_undefined_value_p on virtual operands
which returns a result based on the real definition of the definition
statement.  That doesn't make sense so this patch guards the calls
properly and makes sure nobody else does the same mistake.

Bootstrapped and tested on x86_64-unknown-linux-gnu, pushed.

	* tree-ssa.cc (ssa_undefined_value_p): Assert we are not
	called on virtual operands.
	* tree-ssa-sccvn.cc (vn_phi_lookup): Guard
	ssa_undefined_value_p calls.
	(vn_phi_insert): Likewise.
	(set_ssa_val_to): Likewise.
	(visit_phi): Avoid extra work with equivalences for
	virtual operand PHIs.
---
 gcc/tree-ssa-sccvn.cc | 20 +++++++++++++++-----
 gcc/tree-ssa.cc       |  2 ++
 2 files changed, 17 insertions(+), 5 deletions(-)

diff --git a/gcc/tree-ssa-sccvn.cc b/gcc/tree-ssa-sccvn.cc
index 8ee77fd2b78..d5b081a309f 100644
--- a/gcc/tree-ssa-sccvn.cc
+++ b/gcc/tree-ssa-sccvn.cc
@@ -4824,7 +4824,8 @@ vn_phi_lookup (gimple *phi, bool backedges_varying_p)
       if (TREE_CODE (def) == SSA_NAME
 	  && (!backedges_varying_p || !(e->flags & EDGE_DFS_BACK)))
 	{
-	  if (ssa_undefined_value_p (def, false))
+	  if (!virtual_operand_p (def)
+	      && ssa_undefined_value_p (def, false))
 	    def = VN_TOP;
 	  else
 	    def = SSA_VAL (def);
@@ -4877,7 +4878,8 @@ vn_phi_insert (gimple *phi, tree result, bool backedges_varying_p)
       if (TREE_CODE (def) == SSA_NAME
 	  && (!backedges_varying_p || !(e->flags & EDGE_DFS_BACK)))
 	{
-	  if (ssa_undefined_value_p (def, false))
+	  if (!virtual_operand_p (def)
+	      && ssa_undefined_value_p (def, false))
 	    def = VN_TOP;
 	  else
 	    def = SSA_VAL (def);
@@ -5059,6 +5061,7 @@ set_ssa_val_to (tree from, tree to)
 	}
       curr_invariant = is_gimple_min_invariant (currval);
       curr_undefined = (TREE_CODE (currval) == SSA_NAME
+			&& !virtual_operand_p (currval)
 			&& ssa_undefined_value_p (currval, false));
       if (currval != VN_TOP
 	  && !curr_invariant
@@ -5081,6 +5084,7 @@ set_ssa_val_to (tree from, tree to)
       else if (currval != VN_TOP
 	       && !curr_undefined
 	       && TREE_CODE (to) == SSA_NAME
+	       && !virtual_operand_p (to)
 	       && ssa_undefined_value_p (to, false))
 	{
 	  if (dump_file && (dump_flags & TDF_DETAILS))
@@ -5116,6 +5120,7 @@ set_and_exit:
          PR82320 for a testcase were we'd otherwise not terminate iteration.  */
       && !(curr_undefined
 	   && TREE_CODE (to) == SSA_NAME
+	   && !virtual_operand_p (to)
 	   && ssa_undefined_value_p (to, false))
       /* ???  For addresses involving volatile objects or types operand_equal_p
          does not reliably detect ADDR_EXPRs as equal.  We know we are only
@@ -5880,7 +5885,14 @@ visit_phi (gimple *phi, bool *inserted, bool backedges_varying_p)
 	    sameval = def;
 	    sameval_e = e;
 	  }
-	else if (!expressions_equal_p (def, sameval))
+	else if (expressions_equal_p (def, sameval))
+	  sameval_e = NULL;
+	else if (virtual_operand_p (def))
+	  {
+	    sameval = NULL_TREE;
+	    break;
+	  }
+	else
 	  {
 	    /* We know we're arriving only with invariant addresses here,
 	       try harder comparing them.  We can do some caching here
@@ -5957,8 +5969,6 @@ visit_phi (gimple *phi, bool *inserted, bool backedges_varying_p)
 	    sameval = NULL_TREE;
 	    break;
 	  }
-	else
-	  sameval_e = NULL;
       }
 
   /* If the value we want to use is flowing over the backedge and we
diff --git a/gcc/tree-ssa.cc b/gcc/tree-ssa.cc
index 5126020d40f..a5cad2d344e 100644
--- a/gcc/tree-ssa.cc
+++ b/gcc/tree-ssa.cc
@@ -1320,6 +1320,8 @@ ssa_undefined_value_p (tree t, bool partial)
 {
   gimple *def_stmt;
 
+  gcc_checking_assert (!virtual_operand_p (t));
+
   if (ssa_defined_default_def_p (t))
     return false;
 
-- 
2.35.3

^ permalink raw reply	[flat|nested] only message in thread

only message in thread, other threads:[~2023-02-17 11:31 UTC | newest]

Thread overview: (only message) (download: mbox.gz / follow: Atom feed)
-- links below jump to the message on this page --
2023-02-17 11:30 [PATCH] Fix wrong-code issue in VN Richard Biener

This is a public inbox, see mirroring instructions
for how to clone and mirror all data and code used for this inbox;
as well as URLs for read-only IMAP folder(s) and NNTP newsgroup(s).