[optimize 1/3] Fix phi to min/max

[optimize 1/3] Fix phi to min/max

[Nathan Sidwell]

[-- Attachment #1: Type: text/plain, Size: 653 bytes --]

Richard,

This patch fixes the problem I described earlier about min/max generation 
propagating incorrect range information 
(https://gcc.gnu.org/ml/gcc/2015-08/msg00024.html).

I went with creating a new name, if the PHI being modified has more than 2 
edges.  I also modified the testcase that my min/max vrp patch tickled -- it 
wasn't calling an init function and hence copying zeroes rather than the small 
values it thought it was copying.  Also added a new test to convert zeroes, 
which was the failure mode I observed in the DFP lib.

This patch was tested in combination with the new min max optimization I'll 
post momentarily.

ok?

nathan

[-- Attachment #2: phi-min-max-fix.patch --]
[-- Type: text/x-patch, Size: 2452 bytes --]

2015-08-10  Nathan Sidwell  <nathan@acm.org>

	* tree-ssa-phiopt.c (minmax_replacement): Create new ssa name if
	we're not the only contributor to target phi.

	testsuite/
	* c-c++-common/dfp/operator-comma.c: Call init function.
	* c-c++-common/dfp/convert-dfp-2.c: New test.

Index: tree-ssa-phiopt.c
===================================================================
--- tree-ssa-phiopt.c	(revision 226749)
+++ tree-ssa-phiopt.c	(working copy)
@@ -1277,8 +1277,16 @@ minmax_replacement (basic_block cond_bb,
       gsi_move_before (&gsi_from, &gsi);
     }
 
+  /* Create an SSA var to hold the min/max result.  If we're the only
+     things setting the target PHI, then we  can clone the PHI
+     variable.  Otherwise we must create a new one.  */
+  result = PHI_RESULT (phi);
+  if (EDGE_COUNT (gimple_bb (phi)->preds) == 2)
+    result = duplicate_ssa_name (result, NULL);
+  else
+    result = make_ssa_name (TREE_TYPE (result));
+
   /* Emit the statement to compute min/max.  */
-  result = duplicate_ssa_name (PHI_RESULT (phi), NULL);
   new_stmt = gimple_build_assign (result, minmax, arg0, arg1);
   gsi = gsi_last_bb (cond_bb);
   gsi_insert_before (&gsi, new_stmt, GSI_NEW_STMT);
Index: testsuite/c-c++-common/dfp/convert-dfp-2.c
===================================================================
--- testsuite/c-c++-common/dfp/convert-dfp-2.c	(revision 0)
+++ testsuite/c-c++-common/dfp/convert-dfp-2.c	(working copy)
@@ -0,0 +1,45 @@
+/* { dg-options "-O0" } */
+
+/* Test decimal fp conversions of zero.  */
+
+#include "dfp-dbg.h"
+
+volatile _Decimal32 d32a, d32c;
+volatile _Decimal64 d64a, d64c;
+volatile _Decimal128 d128a, d128c;
+
+int
+main ()
+{
+  d32a = d32c;
+  if (d32a)
+    FAILURE
+  d32a = d64c;
+  if (d32a)
+    FAILURE
+  d32a = d128c;
+  if (d32a)
+    FAILURE
+
+  d64a = d32c;
+  if (d64a)
+    FAILURE
+  d64a = d64c;
+  if (d64a)
+    FAILURE
+  d64a = d128c;
+  if (d64a)
+    FAILURE
+  
+  d128a = d32c;
+  if (d128a)
+    FAILURE
+  d128a = d64c;
+  if (d128a)
+    FAILURE
+  d128a = d128c;
+  if (d128a)
+    FAILURE
+  
+  FINISH
+}
Index: testsuite/c-c++-common/dfp/operator-comma.c
===================================================================
--- testsuite/c-c++-common/dfp/operator-comma.c	(revision 226749)
+++ testsuite/c-c++-common/dfp/operator-comma.c	(working copy)
@@ -24,6 +24,8 @@ init ()
 int
 main ()
 {
+  init ();
+  
   d32a = (d32b, d32c);
   if (d32a != d32c)
     FAILURE

Re: [optimize 1/3] Fix phi to min/max

[Richard Biener]

On Mon, 10 Aug 2015, Nathan Sidwell wrote:

> Richard,
> 
> This patch fixes the problem I described earlier about min/max generation
> propagating incorrect range information
> (https://gcc.gnu.org/ml/gcc/2015-08/msg00024.html).
> 
> I went with creating a new name, if the PHI being modified has more than 2
> edges.  I also modified the testcase that my min/max vrp patch tickled -- it
> wasn't calling an init function and hence copying zeroes rather than the small
> values it thought it was copying.  Also added a new test to convert zeroes,
> which was the failure mode I observed in the DFP lib.
> 
> This patch was tested in combination with the new min max optimization I'll
> post momentarily.
> 
> ok?

Ok.

Thanks,
Richard.