public inbox for gcc-patches@gcc.gnu.org
 help / color / mirror / Atom feed
* [PATCH] tree-optimization/105198 - wrong code with predictive commoning
@ 2022-04-08 12:08 Richard Biener
  0 siblings, 0 replies; only message in thread
From: Richard Biener @ 2022-04-08 12:08 UTC (permalink / raw)
  To: gcc-patches

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.

Bootstrapped on x86_64-unknown-linux-gnu, testing in progress.

Richard.

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

	PR tree-optimization/105198
	* tree-predcom.c (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.
---
 gcc/testsuite/gcc.dg/torture/pr105198.c | 26 +++++++++++++++++
 gcc/tree-predcom.cc                     | 38 +++++++++++++++++++++----
 2 files changed, 59 insertions(+), 5 deletions(-)
 create mode 100644 gcc/testsuite/gcc.dg/torture/pr105198.c

diff --git a/gcc/testsuite/gcc.dg/torture/pr105198.c b/gcc/testsuite/gcc.dg/torture/pr105198.c
new file mode 100644
index 00000000000..91f92afc163
--- /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 bb3a1cb68fc..fb457250bbd 100644
--- 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;
 }
 
-- 
2.34.1

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

only message in thread, other threads:[~2022-04-08 12:08 UTC | newest]

Thread overview: (only message) (download: mbox.gz / follow: Atom feed)
-- links below jump to the message on this page --
2022-04-08 12:08 [PATCH] tree-optimization/105198 - wrong code with predictive commoning 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).