* [SH] Correct fabs and fneg insns in simulator
@ 2014-10-13 23:39 Oleg Endo
2014-11-28 15:51 ` Joel Brobecker
0 siblings, 1 reply; 3+ messages in thread
From: Oleg Endo @ 2014-10-13 23:39 UTC (permalink / raw)
To: gdb-patches; +Cc: Kaz Kojima
[-- Attachment #1: Type: text/plain, Size: 915 bytes --]
Hi,
It seems that the implementation of the SH fabs and fneg insns in the
simulator is not correct. They use the FP_UNARY macro which checks the
FPSCR.PR setting and raises an exception if PR = 1 (double precision)
and the register number is not even (i.e. a valid DF reg number).
For normal unary FP insns this is fine. However, fneg and fabs perform
the same (integer) operations regardless of the FPSCR.PR setting.
This issue initially popped up here
https://gcc.gnu.org/bugzilla/show_bug.cgi?id=63260
I've checked some of the failing tests mentioned in GCC PR 63260 above
with the patch applied and the failures go away.
Could somebody please apply it if it's OK?
I'm not subscribed to gdb-patches, please CC me when replying.
Cheers,
Oleg
sim/ChangeLog:
2014-10-14 Oleg Endo <olegendo@gcc.gnu.org>
* sh/gencode.c (fabs, fneg): Implement as integer operation
instead of using the FP_UNARY macro.
[-- Attachment #2: sh_sim_fneg_fabs.patch --]
[-- Type: text/x-patch, Size: 805 bytes --]
diff --git a/sim/sh/gencode.c b/sim/sh/gencode.c
index 738b718..bc65604 100644
--- a/sim/sh/gencode.c
+++ b/sim/sh/gencode.c
@@ -429,8 +429,14 @@ op tab[] =
/* sh2e */
{ "", "", "fabs <FREG_N>", "1111nnnn01011101",
- "FP_UNARY (n, fabs);",
- "/* FIXME: FR (n) &= 0x7fffffff; */",
+ " union",
+ " {",
+ " unsigned int i;",
+ " float f;",
+ " } u;",
+ " u.f = FR (n);",
+ " u.i &= 0x7fffffff;",
+ " SET_FR (n, u.f);",
},
/* sh2e */
@@ -662,7 +668,14 @@ op tab[] =
/* sh2e */
{ "", "", "fneg <FREG_N>", "1111nnnn01001101",
- "FP_UNARY (n, -);",
+ " union",
+ " {",
+ " unsigned int i;",
+ " float f;",
+ " } u;",
+ " u.f = FR (n);",
+ " u.i ^= 0x80000000;",
+ " SET_FR (n, u.f);",
},
/* sh4a */
^ permalink raw reply [flat|nested] 3+ messages in thread
* Re: [SH] Correct fabs and fneg insns in simulator
2014-10-13 23:39 [SH] Correct fabs and fneg insns in simulator Oleg Endo
@ 2014-11-28 15:51 ` Joel Brobecker
2014-11-30 16:35 ` Oleg Endo
0 siblings, 1 reply; 3+ messages in thread
From: Joel Brobecker @ 2014-11-28 15:51 UTC (permalink / raw)
To: Oleg Endo; +Cc: gdb-patches, Kaz Kojima
Hello Oleg,,
> It seems that the implementation of the SH fabs and fneg insns in the
> simulator is not correct. They use the FP_UNARY macro which checks the
> FPSCR.PR setting and raises an exception if PR = 1 (double precision)
> and the register number is not even (i.e. a valid DF reg number).
> For normal unary FP insns this is fine. However, fneg and fabs perform
> the same (integer) operations regardless of the FPSCR.PR setting.
>
> This issue initially popped up here
> https://gcc.gnu.org/bugzilla/show_bug.cgi?id=63260
>
> I've checked some of the failing tests mentioned in GCC PR 63260 above
> with the patch applied and the failures go away.
>
> Could somebody please apply it if it's OK?
>
> I'm not subscribed to gdb-patches, please CC me when replying.
>
> Cheers,
> Oleg
>
> sim/ChangeLog:
>
> 2014-10-14 Oleg Endo <olegendo@gcc.gnu.org>
>
> * sh/gencode.c (fabs, fneg): Implement as integer operation
> instead of using the FP_UNARY macro.
Thank you for the patch. We don't have an SH maintainer for GDB,
and in the absence of any feedback from other interested people,
we will just trust you on this. So I pushed this patch under the
"tiny patch" rule.
If you ever think of contributing other patches to GDB, we'll probably
have to start thinking about getting your copyright assignment to
cover GDB as well.
--
Joel
^ permalink raw reply [flat|nested] 3+ messages in thread
* Re: [SH] Correct fabs and fneg insns in simulator
2014-11-28 15:51 ` Joel Brobecker
@ 2014-11-30 16:35 ` Oleg Endo
0 siblings, 0 replies; 3+ messages in thread
From: Oleg Endo @ 2014-11-30 16:35 UTC (permalink / raw)
To: Joel Brobecker; +Cc: gdb-patches, Kaz Kojima
Hi,
On Fri, 2014-11-28 at 19:51 +0400, Joel Brobecker wrote:
> Hello Oleg,,
>
> > It seems that the implementation of the SH fabs and fneg insns in the
> > simulator is not correct. They use the FP_UNARY macro which checks the
> > FPSCR.PR setting and raises an exception if PR = 1 (double precision)
> > and the register number is not even (i.e. a valid DF reg number).
> > For normal unary FP insns this is fine. However, fneg and fabs perform
> > the same (integer) operations regardless of the FPSCR.PR setting.
> >
> > This issue initially popped up here
> > https://gcc.gnu.org/bugzilla/show_bug.cgi?id=63260
> >
> > I've checked some of the failing tests mentioned in GCC PR 63260 above
> > with the patch applied and the failures go away.
> >
> > Could somebody please apply it if it's OK?
> >
> > I'm not subscribed to gdb-patches, please CC me when replying.
> >
> > Cheers,
> > Oleg
> >
> > sim/ChangeLog:
> >
> > 2014-10-14 Oleg Endo <olegendo@gcc.gnu.org>
> >
> > * sh/gencode.c (fabs, fneg): Implement as integer operation
> > instead of using the FP_UNARY macro.
>
> Thank you for the patch. We don't have an SH maintainer for GDB,
> and in the absence of any feedback from other interested people,
> we will just trust you on this. So I pushed this patch under the
> "tiny patch" rule.
Thanks.
> If you ever think of contributing other patches to GDB, we'll probably
> have to start thinking about getting your copyright assignment to
> cover GDB as well.
Yes, that might be a good idea. Could you please send me the necessary
papers (or initiate the procedure)?
Cheers,
Oleg
^ permalink raw reply [flat|nested] 3+ messages in thread
end of thread, other threads:[~2014-11-30 16:35 UTC | newest]
Thread overview: 3+ messages (download: mbox.gz / follow: Atom feed)
-- links below jump to the message on this page --
2014-10-13 23:39 [SH] Correct fabs and fneg insns in simulator Oleg Endo
2014-11-28 15:51 ` Joel Brobecker
2014-11-30 16:35 ` Oleg Endo
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).