[committed] PR 61095: tsan fallout from wide-int merge

[committed] PR 61095: tsan fallout from wide-int merge

[Richard Sandiford]

This PR was due to code in which -(int) foo was suposed to be sign-extended,
but was being ORed with an unsigned int and so ended up being zero-extended.
Fixed by using the proper-width type.

Tested on x86_64-linux-gnu and applied as obvious.  Sorry for the breakage.

Thanks,
Richard


gcc/
	PR tree-optimization/61095
	* tree-ssanames.c (get_nonzero_bits): Fix type extension in wi::shwi.

Index: gcc/tree-ssanames.c
===================================================================
--- gcc/tree-ssanames.c	2014-05-07 16:50:15.136064484 +0100
+++ gcc/tree-ssanames.c	2014-05-07 16:50:15.422063737 +0100
@@ -271,7 +271,8 @@ get_nonzero_bits (const_tree name)
     {
       struct ptr_info_def *pi = SSA_NAME_PTR_INFO (name);
       if (pi && pi->align)
-	return wi::shwi (-(int) pi->align | pi->misalign, precision);
+	return wi::shwi (-(HOST_WIDE_INT) pi->align
+			 | (HOST_WIDE_INT) pi->misalign, precision);
       return wi::shwi (-1, precision);
     }
 

Re: [committed] PR 61095: tsan fallout from wide-int merge

[Richard Sandiford]

Richard Sandiford <rdsandiford@googlemail.com> writes:
> This PR was due to code in which -(int) foo was suposed to be sign-extended,
> but was being ORed with an unsigned int and so ended up being zero-extended.
> Fixed by using the proper-width type.

As Kostya rightly said in the PR, this should have had a testcase too.

Tested on x86_64-linux-gnu.  It failed before the patch on x86_64,
passes after it, and is skipped for -m32.  OK to install?

Thanks,
Richard


gcc/testsuite/
	PR tree-optimization/61095
	* gcc.dg/torture/pr61095.c: New test.

Index: gcc/testsuite/gcc.dg/torture/pr61095.c
===================================================================
--- /dev/null	2014-05-03 11:58:38.033951363 +0100
+++ gcc/testsuite/gcc.dg/torture/pr61095.c	2014-05-08 08:46:01.203827892 +0100
@@ -0,0 +1,23 @@
+/* { dg-do run } */
+/* { dg-require-effective-target lp64 } */
+
+extern void __attribute__ ((noreturn)) abort (void);
+
+int __attribute__ ((noinline, noclone))
+foo (unsigned long addr) {
+    unsigned long *p = (unsigned long*)((addr & 0xffff83fffffffff8UL) * 4);
+    unsigned long xxx = (unsigned long)(p + 1);
+    return xxx >= 0x3c000000000UL;
+}
+
+int
+main (void)
+{
+  if (foo (0))
+    abort ();
+  if (foo (0x7c0000000000UL))
+    abort ();
+  if (!foo (0xfc0000000000UL))
+    abort ();
+  return 0;
+}

Re: [committed] PR 61095: tsan fallout from wide-int merge

[Richard Biener]

On Thu, May 8, 2014 at 9:48 AM, Richard Sandiford
<rdsandiford@googlemail.com> wrote:
> Richard Sandiford <rdsandiford@googlemail.com> writes:
>> This PR was due to code in which -(int) foo was suposed to be sign-extended,
>> but was being ORed with an unsigned int and so ended up being zero-extended.
>> Fixed by using the proper-width type.
>
> As Kostya rightly said in the PR, this should have had a testcase too.
>
> Tested on x86_64-linux-gnu.  It failed before the patch on x86_64,
> passes after it, and is skipped for -m32.  OK to install?

Ok.

Thanks,
Richard.

> Thanks,
> Richard
>
>
> gcc/testsuite/
>         PR tree-optimization/61095
>         * gcc.dg/torture/pr61095.c: New test.
>
> Index: gcc/testsuite/gcc.dg/torture/pr61095.c
> ===================================================================
> --- /dev/null   2014-05-03 11:58:38.033951363 +0100
> +++ gcc/testsuite/gcc.dg/torture/pr61095.c      2014-05-08 08:46:01.203827892 +0100
> @@ -0,0 +1,23 @@
> +/* { dg-do run } */
> +/* { dg-require-effective-target lp64 } */
> +
> +extern void __attribute__ ((noreturn)) abort (void);
> +
> +int __attribute__ ((noinline, noclone))
> +foo (unsigned long addr) {
> +    unsigned long *p = (unsigned long*)((addr & 0xffff83fffffffff8UL) * 4);
> +    unsigned long xxx = (unsigned long)(p + 1);
> +    return xxx >= 0x3c000000000UL;
> +}
> +
> +int
> +main (void)
> +{
> +  if (foo (0))
> +    abort ();
> +  if (foo (0x7c0000000000UL))
> +    abort ();
> +  if (!foo (0xfc0000000000UL))
> +    abort ();
> +  return 0;
> +}