public inbox for libc-hacker@sourceware.org
 help / color / mirror / Atom feed
* [PATCH] Fix ia64 feraiseexcept
@ 2007-05-08 21:30 Jakub Jelinek
  2007-05-09  6:47 ` [PATCH] Fix ia64 feraiseexcept (take 2) Jakub Jelinek
  0 siblings, 1 reply; 2+ messages in thread
From: Jakub Jelinek @ 2007-05-08 21:30 UTC (permalink / raw)
  To: Ulrich Drepper, H. J. Lu; +Cc: Glibc hackers

Hi!

While feraiseexcept (FE_OVERFLOW); or feraiseexcept (FE_UNDERFLOW) are
allowed to raise additionally inexact exception, feraiseexcept (FE_INEXACT)
shouldn't raise any exception but inexact.

DBL_MAX + 1.0 overflows and is inexact though, while (DBL_MAX / 2.0) + 1.0
is only inexact, but doesn't overflow.

Fixes test-fenv.

2007-05-08  Jakub Jelinek  <jakub@redhat.com>

	* sysdeps/ia64/fpu/fraiseexcpt.c (feraiseexcept): Don't raise overflow
	exception in addition to inexact when asked to raise only FE_INEXACT.

--- libc/sysdeps/ia64/fpu/fraiseexcpt.c.jj	2003-11-29 08:28:46.000000000 +0100
+++ libc/sysdeps/ia64/fpu/fraiseexcpt.c	2007-05-08 23:22:55.000000000 +0200
@@ -1,5 +1,6 @@
 /* Raise given exceptions.
-   Copyright (C) 1997, 1998, 2000, 2001, 2002 Free Software Foundation, Inc.
+   Copyright (C) 1997, 1998, 2000, 2001, 2002, 2007
+   Free Software Foundation, Inc.
    This file is part of the GNU C Library.
    Contributed by Jes Sorensen <Jes.Sorensen@cern.ch>, 2000.
 
@@ -69,7 +70,7 @@ feraiseexcept (int excepts)
   /* Last: inexact.  */
   if (FE_INEXACT & excepts)
     {
-      dummy = DBL_MAX;
+      dummy = DBL_MAX / 2.0;
       __asm__ __volatile__ ("fadd.d.s0 %0=%1,f1" : "=f" (dummy) : "0" (dummy));
     }
 

	Jakub

^ permalink raw reply	[flat|nested] 2+ messages in thread

* [PATCH] Fix ia64 feraiseexcept (take 2)
  2007-05-08 21:30 [PATCH] Fix ia64 feraiseexcept Jakub Jelinek
@ 2007-05-09  6:47 ` Jakub Jelinek
  0 siblings, 0 replies; 2+ messages in thread
From: Jakub Jelinek @ 2007-05-09  6:47 UTC (permalink / raw)
  To: Ulrich Drepper; +Cc: Glibc hackers

On Tue, May 08, 2007 at 11:38:15PM +0200, Jakub Jelinek wrote:
> While feraiseexcept (FE_OVERFLOW); or feraiseexcept (FE_UNDERFLOW) are
> allowed to raise additionally inexact exception, feraiseexcept (FE_INEXACT)
> shouldn't raise any exception but inexact.
> 
> DBL_MAX + 1.0 overflows and is inexact though, while (DBL_MAX / 2.0) + 1.0
> is only inexact, but doesn't overflow.

Small correction, feraiseexcept without the patch misbehaves with fesetround
(FE_UPWARD) only, but IMHO it should be fixed even for that case.
Another alternative is to subtract f1 from DBL_MAX rather than add it,
which has the advantage that only 2 double constants are in .sdata rather
than 3 as with previous patch.

2007-05-09  Jakub Jelinek  <jakub@redhat.com>

	* sysdeps/ia64/fpu/fraiseexcpt.c (feraiseexcept): Don't raise overflow
	exception in addition to inexact when asked to raise only FE_INEXACT.

--- libc/sysdeps/ia64/fpu/fraiseexcpt.c.jj	2003-11-29 08:28:46.000000000 +0100
+++ libc/sysdeps/ia64/fpu/fraiseexcpt.c	2007-05-09 08:38:43.000000000 +0200
@@ -1,5 +1,6 @@
 /* Raise given exceptions.
-   Copyright (C) 1997, 1998, 2000, 2001, 2002 Free Software Foundation, Inc.
+   Copyright (C) 1997, 1998, 2000, 2001, 2002, 2007
+   Free Software Foundation, Inc.
    This file is part of the GNU C Library.
    Contributed by Jes Sorensen <Jes.Sorensen@cern.ch>, 2000.
 
@@ -70,7 +71,7 @@ feraiseexcept (int excepts)
   if (FE_INEXACT & excepts)
     {
       dummy = DBL_MAX;
-      __asm__ __volatile__ ("fadd.d.s0 %0=%1,f1" : "=f" (dummy) : "0" (dummy));
+      __asm__ __volatile__ ("fsub.d.s0 %0=%1,f1" : "=f" (dummy) : "0" (dummy));
     }
 
   /* Success.  */

	Jakub

^ permalink raw reply	[flat|nested] 2+ messages in thread

end of thread, other threads:[~2007-05-09  6:47 UTC | newest]

Thread overview: 2+ messages (download: mbox.gz / follow: Atom feed)
-- links below jump to the message on this page --
2007-05-08 21:30 [PATCH] Fix ia64 feraiseexcept Jakub Jelinek
2007-05-09  6:47 ` [PATCH] Fix ia64 feraiseexcept (take 2) Jakub Jelinek

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).