Fix ICE with widen_mult and -ftrapv (PR tree-optimization/83523)

Message ID 20171221175451.GB2353@tucnak
State New
Headers show
Series
  • Fix ICE with widen_mult and -ftrapv (PR tree-optimization/83523)
Related show

Commit Message

Jakub Jelinek Dec. 21, 2017, 5:54 p.m.
Hi!

Converting widening mult, or widen_{plus,minus}_expr, or integral mult to
fma is IMNSHO undesirable with -ftrapv for signed types - the expansion of
those doesn't detect overflows and by giving up we don't need to worry about
formerly trapping MULT_EXPR becoming something that isn't trapping and
performing EH cleanups etc.

Bootstrapped/regtested on x86_64-linux and i686-linux, ok for trunk?

2017-12-21  Jakub Jelinek  <jakub@redhat.com>

	PR tree-optimization/83523
	* tree-ssa-math-opts.c (is_widening_mult_p): Return false if
	for INTEGER_TYPE TYPE_OVERFLOW_TRAPS.
	(convert_mult_to_fma): Likewise.

	* g++.dg/tree-ssa/pr83523.C: New test.


	Jakub

Comments

Richard Biener Dec. 21, 2017, 6:57 p.m. | #1
On December 21, 2017 6:54:51 PM GMT+01:00, Jakub Jelinek <jakub@redhat.com> wrote:
>Hi!

>

>Converting widening mult, or widen_{plus,minus}_expr, or integral mult

>to

>fma is IMNSHO undesirable with -ftrapv for signed types - the expansion

>of

>those doesn't detect overflows and by giving up we don't need to worry

>about

>formerly trapping MULT_EXPR becoming something that isn't trapping and

>performing EH cleanups etc.

>

>Bootstrapped/regtested on x86_64-linux and i686-linux, ok for trunk?


OK. 

Richard. 

>2017-12-21  Jakub Jelinek  <jakub@redhat.com>

>

>	PR tree-optimization/83523

>	* tree-ssa-math-opts.c (is_widening_mult_p): Return false if

>	for INTEGER_TYPE TYPE_OVERFLOW_TRAPS.

>	(convert_mult_to_fma): Likewise.

>

>	* g++.dg/tree-ssa/pr83523.C: New test.

>

>--- gcc/tree-ssa-math-opts.c.jj	2017-12-21 09:43:19.000000000 +0100

>+++ gcc/tree-ssa-math-opts.c	2017-12-21 12:22:10.576488270 +0100

>@@ -2196,8 +2196,12 @@ is_widening_mult_p (gimple *stmt,

> {

>   tree type = TREE_TYPE (gimple_assign_lhs (stmt));

> 

>-  if (TREE_CODE (type) != INTEGER_TYPE

>-      && TREE_CODE (type) != FIXED_POINT_TYPE)

>+  if (TREE_CODE (type) == INTEGER_TYPE)

>+    {

>+      if (TYPE_OVERFLOW_TRAPS (type))

>+	return false;

>+    }

>+  else if (TREE_CODE (type) != FIXED_POINT_TYPE)

>     return false;

> 

>if (!is_widening_mult_rhs_p (type, gimple_assign_rhs1 (stmt),

>type1_out,

>@@ -2656,7 +2660,7 @@ convert_mult_to_fma (gimple *mul_stmt, t

> 

>   /* We don't want to do bitfield reduction ops.  */

>   if (INTEGRAL_TYPE_P (type)

>-      && !type_has_mode_precision_p (type))

>+      && (!type_has_mode_precision_p (type) || TYPE_OVERFLOW_TRAPS

>(type)))

>     return false;

> 

>/* If the target doesn't support it, don't generate it.  We assume that

>--- gcc/testsuite/g++.dg/tree-ssa/pr83523.C.jj	2017-12-21

>12:20:56.555403891 +0100

>+++ gcc/testsuite/g++.dg/tree-ssa/pr83523.C	2017-12-21

>12:20:02.000000000 +0100

>@@ -0,0 +1,25 @@

>+// PR tree-optimization/83523

>+// { dg-do compile }

>+// { dg-options "-O2 -fexceptions -fnon-call-exceptions -ftrapv" }

>+

>+#ifdef __SIZEOF_INT128__

>+typedef __int128 T;

>+typedef long long int U;

>+#else

>+typedef long long int T;

>+typedef int U;

>+#endif

>+

>+struct S { S (); ~S (); };

>+void bar ();

>+

>+T

>+foo (U x, U y)

>+{

>+  T z = x;

>+  S s;

>+  bar ();

>+  z *= y;

>+  bar ();

>+  return z;

>+}

>

>	Jakub
Jeff Law Dec. 21, 2017, 11:09 p.m. | #2
On 12/21/2017 10:54 AM, Jakub Jelinek wrote:
> Hi!

> 

> Converting widening mult, or widen_{plus,minus}_expr, or integral mult to

> fma is IMNSHO undesirable with -ftrapv for signed types - the expansion of

> those doesn't detect overflows and by giving up we don't need to worry about

> formerly trapping MULT_EXPR becoming something that isn't trapping and

> performing EH cleanups etc.

> 

> Bootstrapped/regtested on x86_64-linux and i686-linux, ok for trunk?

> 

> 2017-12-21  Jakub Jelinek  <jakub@redhat.com>

> 

> 	PR tree-optimization/83523

> 	* tree-ssa-math-opts.c (is_widening_mult_p): Return false if

> 	for INTEGER_TYPE TYPE_OVERFLOW_TRAPS.

> 	(convert_mult_to_fma): Likewise.

> 

> 	* g++.dg/tree-ssa/pr83523.C: New test.

OK.
jeff

Patch

--- gcc/tree-ssa-math-opts.c.jj	2017-12-21 09:43:19.000000000 +0100
+++ gcc/tree-ssa-math-opts.c	2017-12-21 12:22:10.576488270 +0100
@@ -2196,8 +2196,12 @@  is_widening_mult_p (gimple *stmt,
 {
   tree type = TREE_TYPE (gimple_assign_lhs (stmt));
 
-  if (TREE_CODE (type) != INTEGER_TYPE
-      && TREE_CODE (type) != FIXED_POINT_TYPE)
+  if (TREE_CODE (type) == INTEGER_TYPE)
+    {
+      if (TYPE_OVERFLOW_TRAPS (type))
+	return false;
+    }
+  else if (TREE_CODE (type) != FIXED_POINT_TYPE)
     return false;
 
   if (!is_widening_mult_rhs_p (type, gimple_assign_rhs1 (stmt), type1_out,
@@ -2656,7 +2660,7 @@  convert_mult_to_fma (gimple *mul_stmt, t
 
   /* We don't want to do bitfield reduction ops.  */
   if (INTEGRAL_TYPE_P (type)
-      && !type_has_mode_precision_p (type))
+      && (!type_has_mode_precision_p (type) || TYPE_OVERFLOW_TRAPS (type)))
     return false;
 
   /* If the target doesn't support it, don't generate it.  We assume that
--- gcc/testsuite/g++.dg/tree-ssa/pr83523.C.jj	2017-12-21 12:20:56.555403891 +0100
+++ gcc/testsuite/g++.dg/tree-ssa/pr83523.C	2017-12-21 12:20:02.000000000 +0100
@@ -0,0 +1,25 @@ 
+// PR tree-optimization/83523
+// { dg-do compile }
+// { dg-options "-O2 -fexceptions -fnon-call-exceptions -ftrapv" }
+
+#ifdef __SIZEOF_INT128__
+typedef __int128 T;
+typedef long long int U;
+#else
+typedef long long int T;
+typedef int U;
+#endif
+
+struct S { S (); ~S (); };
+void bar ();
+
+T
+foo (U x, U y)
+{
+  T z = x;
+  S s;
+  bar ();
+  z *= y;
+  bar ();
+  return z;
+}