tree-optimization/105198 - wrong code with predictive commoning

When predictive commoning looks for a looparound PHI it tries
to match the entry value definition (a load) up with the appropriate
member of the chain.  But it fails to consider stmts clobbering
the very same memory location inbetween the load and loop entry.

In theory we could be more clever on must aliases that would be
also picked up from a load (so not exactly stmt_kills_ref_p) and
use the stored value from that if it is an exact match.  But we
currently have no way to propagate this information inside predcom.

2022-04-08  Richard Biener  <rguenther@suse.de>

	PR tree-optimization/105198
	* tree-predcom.cc (find_looparound_phi): Check whether
	the found memory location of the entry value is clobbered
	inbetween the value we want to use and loop entry.

	* gcc.dg/torture/pr105198.c: New testcase.

diff --git a/gcc/testsuite/gcc.dg/torture/pr105198.c b/gcc/testsuite/gcc.dg/torture/pr105198.c
new file mode 100644 (file)
index 0000000..91f92af
--- /dev/null
+++ b/gcc/testsuite/gcc.dg/torture/pr105198.c
@@ -0,0 +1,26 @@
+/* { dg-do run } */
+/* { dg-additional-options "-fno-tree-pre -fpredictive-commoning" } */
+
+static  __attribute__ ((noipa)) void
+next_set(int *x, int n, int k)
+{
+  int j = k - 1;
+  int tmp = x[j]++;
+  while (j > 0)
+    {
+      if (x[j] < n - (k - 1 -j))
+        break;
+      j--;
+      x[j]++;
+      tmp = x[j];
+    }
+  if (tmp != 2 || j != 1 || x[0] != 0 || x[1] != 2 || x[2] != 5)
+    __builtin_abort ();
+}
+
+int main()
+{
+  int x[3] = {0, 1, 4};
+  next_set(x, 5, 3);
+  return 0;
+}

diff --git a/gcc/tree-predcom.cc b/gcc/tree-predcom.cc
index bb3a1cb..fb45725 100644 (file)
--- a/gcc/tree-predcom.cc
+++ b/gcc/tree-predcom.cc
@@ -1377,7 +1377,6 @@ gphi *
 pcom_worker::find_looparound_phi (dref ref, dref root)
 {
   tree name, init, init_ref;
-  gphi *phi = NULL;
   gimple *init_stmt;
   edge latch = loop_latch_edge (m_loop);
   struct data_reference init_dr;
@@ -1395,14 +1394,19 @@ pcom_worker::find_looparound_phi (dref ref, dref root)
   if (!name)
     return NULL;
 
+  tree entry_vuse = NULL_TREE;
+  gphi *phi = NULL;
   for (psi = gsi_start_phis (m_loop->header); !gsi_end_p (psi); gsi_next (&psi))
     {
-      phi = psi.phi ();
-      if (PHI_ARG_DEF_FROM_EDGE (phi, latch) == name)
+      gphi *p = psi.phi ();
+      if (PHI_ARG_DEF_FROM_EDGE (p, latch) == name)
+       phi = p;
+      else if (virtual_operand_p (gimple_phi_result (p)))
+       entry_vuse = PHI_ARG_DEF_FROM_EDGE (p, loop_preheader_edge (m_loop));
+      if (phi && entry_vuse)
        break;
     }
-
-  if (gsi_end_p (psi))
+  if (!phi || !entry_vuse)
     return NULL;
 
   init = PHI_ARG_DEF_FROM_EDGE (phi, loop_preheader_edge (m_loop));
@@ -1430,6 +1434,30 @@ pcom_worker::find_looparound_phi (dref ref, dref root)
   if (!valid_initializer_p (&init_dr, ref->distance + 1, root->ref))
     return NULL;
 
+  /* Make sure nothing clobbers the location we re-use the initial value
+     from.  */
+  if (entry_vuse != gimple_vuse (init_stmt))
+    {
+      ao_ref ref;
+      ao_ref_init (&ref, init_ref);
+      unsigned limit = param_sccvn_max_alias_queries_per_access;
+      tree vdef = entry_vuse;
+      do
+       {
+         gimple *def = SSA_NAME_DEF_STMT (vdef);
+         if (limit-- == 0 || gimple_code (def) == GIMPLE_PHI)
+           return NULL;
+         if (stmt_may_clobber_ref_p_1 (def, &ref))
+           /* When the stmt is an assign to init_ref we could in theory
+              use its RHS for the initial value of the looparound PHI
+              we replace in prepare_initializers_chain, but we have
+              no convenient place to store this info at the moment.  */
+           return NULL;
+         vdef = gimple_vuse (def);
+       }
+      while (vdef != gimple_vuse (init_stmt));
+    }
+
   return phi;
 }