Patchwork Fix stringop profile feedback optimization (PR bootstrap/47187)

login
register
mail settings
Submitter Jakub Jelinek
Date Jan. 6, 2011, 12:10 p.m.
Message ID <20110106121003.GS16156@tyan-ft48-01.lab.bos.redhat.com>
Download mbox | patch
Permalink /patch/77705/
State New
Headers show

Comments

Jakub Jelinek - Jan. 6, 2011, 12:10 p.m.
Hi!

If a memcpy/mempcpy/memset has lhs and profile feedback estimates
some likely size, the GIMPLE is invalid, because both calls
use the same SSA_NAME on the lhs.
Fixed thusly, bootstrapped/regtested on i686-linux (x86_64-linux regtest
pending), ok for trunk?

2011-01-06  Jakub Jelinek  <jakub@redhat.com>

	PR bootstrap/47187
	* value-prof.c (gimple_stringop_fixed_value): Handle
	lhs of the call properly.

	* gcc.dg/tree-prof/pr47187.c: New test.


	Jakub
Jeff Law - Jan. 7, 2011, 5:26 p.m.
-----BEGIN PGP SIGNED MESSAGE-----
Hash: SHA1

On 01/06/11 05:10, Jakub Jelinek wrote:
> Hi!
> 
> If a memcpy/mempcpy/memset has lhs and profile feedback estimates
> some likely size, the GIMPLE is invalid, because both calls
> use the same SSA_NAME on the lhs.
> Fixed thusly, bootstrapped/regtested on i686-linux (x86_64-linux regtest
> pending), ok for trunk?
> 
> 2011-01-06  Jakub Jelinek  <jakub@redhat.com>
> 
> 	PR bootstrap/47187
> 	* value-prof.c (gimple_stringop_fixed_value): Handle
> 	lhs of the call properly.
> 
> 	* gcc.dg/tree-prof/pr47187.c: New test.
OK.
Jeff
-----BEGIN PGP SIGNATURE-----
Version: GnuPG v1.4.11 (GNU/Linux)
Comment: Using GnuPG with Fedora - http://enigmail.mozdev.org/

iQEcBAEBAgAGBQJNJ0zCAAoJEBRtltQi2kC7zI4H/RAPBjEr2eV2IPQfxZ+/CXyN
2ji+zwlxolM8O0P6RbXyJAMZSLc4yS9otum1+hMglN7TY0WkCMzkRgQAyeSqf+wa
K9iM/TWc471CI2p+QEudBgzwczHNF/vj8pam+bGuq8yArUUcafPF1+GzKdhJmoQv
KQK8HdjofouwRruzRJmPy6g/2XrWvavIFlt8B4BoIUYBJ5s9nBW+BHLvDOj6qStK
XGOk/KKJMkrZak4Rm3vCYQfPZNbBzRi0NX35vDaCtkEHmv9TyvmwpMCFB3g8ScEi
iBzG8pmtkWyKVZSNlRYwGvIS0fmXWsI9rMDh+Z0KlgbdNzjQk+48jg03KvK827Y=
=49gv
-----END PGP SIGNATURE-----

Patch

--- gcc/value-prof.c.jj	2010-12-22 09:54:58.000000000 +0100
+++ gcc/value-prof.c	2011-01-06 11:12:39.000000000 +0100
@@ -1401,6 +1401,21 @@  gimple_stringop_fixed_value (gimple vcal
   e_vj->probability = REG_BR_PROB_BASE;
   e_vj->count = all - count;
 
+  /* Insert PHI node for the call result if necessary.  */
+  if (gimple_call_lhs (vcall_stmt)
+      && TREE_CODE (gimple_call_lhs (vcall_stmt)) == SSA_NAME)
+    {
+      tree result = gimple_call_lhs (vcall_stmt);
+      gimple phi = create_phi_node (result, join_bb);
+      SSA_NAME_DEF_STMT (result) = phi;
+      gimple_call_set_lhs (vcall_stmt,
+			   make_ssa_name (SSA_NAME_VAR (result), vcall_stmt));
+      add_phi_arg (phi, gimple_call_lhs (vcall_stmt), e_vj, UNKNOWN_LOCATION);
+      gimple_call_set_lhs (icall_stmt,
+			   make_ssa_name (SSA_NAME_VAR (result), icall_stmt));
+      add_phi_arg (phi, gimple_call_lhs (icall_stmt), e_ij, UNKNOWN_LOCATION);
+    }
+
   /* Because these are all string op builtins, they're all nothrow.  */
   gcc_assert (!stmt_could_throw_p (vcall_stmt));
   gcc_assert (!stmt_could_throw_p (icall_stmt));
--- gcc/testsuite/gcc.dg/tree-prof/pr47187.c.jj	2011-01-06 11:23:10.000000000 +0100
+++ gcc/testsuite/gcc.dg/tree-prof/pr47187.c	2011-01-06 11:20:14.000000000 +0100
@@ -0,0 +1,23 @@ 
+/* PR bootstrap/47187 */
+/* { dg-options "-O2" } */
+
+char buf[64];
+char buf2[64];
+
+void *
+foo (char *p, long size)
+{
+  return __builtin_memcpy (buf, p, size);
+}
+
+int
+main (void)
+{
+  long i;
+  for (i = 0; i < 65536; i++)
+    if (foo ("abcdefghijkl", 12) != buf)
+      __builtin_abort ();
+  if (foo (buf2, 64) != buf)
+    __builtin_abort ();
+  return 0;
+}