Mainline Regression in gcc.target/i386/pr32268.c

Mainline Regression in gcc.target/i386/pr32268.c

[Kaveh R. GHAZI]

Hi Uros,

This broke on x86_64 in the last day or so, I thought maybe you might know
what happened.  (You touched config/i386 FP stuff recently.)  I'm seeing a
regression in gcc.target/i386/pr32268.c as seen here:
http://gcc.gnu.org/ml/gcc-testresults/2008-02/msg01108.html

Right now the execute failures only appear with -fpic/-fPIC, however I
believe this is because inlining is masking the problem and allowing
things to get folded.  With -fpic/-fPIC the inlining is disabled and the
abort occurs.

You can trigger the bug in a regular pass by applying the following
noinline patch to the testcase.  Thoughts?

		Thanks,
		--Kaveh


diff -rup orig/egcc-SVN20080216/gcc/testsuite/gcc.target/i386/pr32268.c egcc-SVN20080216/gcc/testsuite/gcc.target/i386/pr32268.c
--- orig/egcc-SVN20080216/gcc/testsuite/gcc.target/i386/pr32268.c	2007-12-31 19:11:58.000000000 +0100
+++ egcc-SVN20080216/gcc/testsuite/gcc.target/i386/pr32268.c	2008-02-16 18:34:04.000000000 +0100
@@ -4,12 +4,14 @@

 extern void abort(void);

-int test_lt(__float128 x, __float128 y)
+int __attribute__ ((__noinline__))
+test_lt(__float128 x, __float128 y)
 {
   return x < y;
 }

-int test_gt (__float128 x, __float128 y)
+int __attribute__ ((__noinline__))
+test_gt (__float128 x, __float128 y)
 {
   return x > y;
 }

Re: Mainline Regression in gcc.target/i386/pr32268.c

[Richard Guenther]

On Feb 16, 2008 6:50 PM, Kaveh R. GHAZI <ghazi@caip.rutgers.edu> wrote:
> Hi Uros,
>
> This broke on x86_64 in the last day or so, I thought maybe you might know
> what happened.  (You touched config/i386 FP stuff recently.)  I'm seeing a
> regression in gcc.target/i386/pr32268.c as seen here:
> http://gcc.gnu.org/ml/gcc-testresults/2008-02/msg01108.html
>
> Right now the execute failures only appear with -fpic/-fPIC, however I
> believe this is because inlining is masking the problem and allowing
> things to get folded.  With -fpic/-fPIC the inlining is disabled and the
> abort occurs.
>
> You can trigger the bug in a regular pass by applying the following
> noinline patch to the testcase.  Thoughts?

How does it break?  It works for me with noinline and with -fPIC.

Richard.

>                 Thanks,
>                 --Kaveh
>
>
> diff -rup orig/egcc-SVN20080216/gcc/testsuite/gcc.target/i386/pr32268.c egcc-SVN20080216/gcc/testsuite/gcc.target/i386/pr32268.c
> --- orig/egcc-SVN20080216/gcc/testsuite/gcc.target/i386/pr32268.c       2007-12-31 19:11:58.000000000 +0100
> +++ egcc-SVN20080216/gcc/testsuite/gcc.target/i386/pr32268.c    2008-02-16 18:34:04.000000000 +0100
> @@ -4,12 +4,14 @@
>
>  extern void abort(void);
>
> -int test_lt(__float128 x, __float128 y)
> +int __attribute__ ((__noinline__))
> +test_lt(__float128 x, __float128 y)
>  {
>    return x < y;
>  }
>
> -int test_gt (__float128 x, __float128 y)
> +int __attribute__ ((__noinline__))
> +test_gt (__float128 x, __float128 y)
>  {
>    return x > y;
>  }
>

Re: Mainline Regression in gcc.target/i386/pr32268.c

[Uros Bizjak]

Richard Guenther wrote:

>> This broke on x86_64 in the last day or so, I thought maybe you might know
>> what happened.  (You touched config/i386 FP stuff recently.)  I'm seeing a
>> regression in gcc.target/i386/pr32268.c as seen here:
>> http://gcc.gnu.org/ml/gcc-testresults/2008-02/msg01108.html
>>
>> Right now the execute failures only appear with -fpic/-fPIC, however I
>> believe this is because inlining is masking the problem and allowing
>> things to get folded.  With -fpic/-fPIC the inlining is disabled and the
>> abort occurs.
>>
>> You can trigger the bug in a regular pass by applying the following
>> noinline patch to the testcase.  Thoughts?
>>     
>
> How does it break?  It works for me with noinline and with -fPIC.
>   
You need to patch the testcase. The change was somehow trivial 
(effectively changed return mode of TFmode soft-fp compare from "long" 
into __libgcc_cmp_return__ mode, which should be the same). I'll look 
into it tomorrow.

Thanks,
Uros.

Re: Mainline Regression in gcc.target/i386/pr32268.c

[Richard Guenther]

On Feb 16, 2008 7:52 PM, Uros Bizjak <ubizjak@gmail.com> wrote:
> Richard Guenther wrote:
>
> >> This broke on x86_64 in the last day or so, I thought maybe you might know
> >> what happened.  (You touched config/i386 FP stuff recently.)  I'm seeing a
> >> regression in gcc.target/i386/pr32268.c as seen here:
> >> http://gcc.gnu.org/ml/gcc-testresults/2008-02/msg01108.html
> >>
> >> Right now the execute failures only appear with -fpic/-fPIC, however I
> >> believe this is because inlining is masking the problem and allowing
> >> things to get folded.  With -fpic/-fPIC the inlining is disabled and the
> >> abort occurs.
> >>
> >> You can trigger the bug in a regular pass by applying the following
> >> noinline patch to the testcase.  Thoughts?
> >>
> >
> > How does it break?  It works for me with noinline and with -fPIC.
> >
> You need to patch the testcase. The change was somehow trivial
> (effectively changed return mode of TFmode soft-fp compare from "long"
> into __libgcc_cmp_return__ mode, which should be the same). I'll look
> into it tomorrow.

I'm confused - I tested on updated trunk exactly as Kaveh reported - so you
say I need additional patches to trigger the failure?

Richard.

Re: Mainline Regression in gcc.target/i386/pr32268.c

[Andrew Pinski]

On Feb 16, 2008 10:57 AM, Richard Guenther <richard.guenther@gmail.com> wrote:
> > You need to patch the testcase. The change was somehow trivial
> > (effectively changed return mode of TFmode soft-fp compare from "long"
> > into __libgcc_cmp_return__ mode, which should be the same). I'll look
> > into it tomorrow.
>
> I'm confused - I tested on updated trunk exactly as Kaveh reported - so you
> say I need additional patches to trigger the failure?

No, Uros is saying that the fix will a change long to __libgcc_cmp_return__ .

-- Pinski

Re: Mainline Regression in gcc.target/i386/pr32268.c

[Kaveh R. Ghazi]

From: "Richard Guenther" <richard.guenther@gmail.com>

>> You can trigger the bug in a regular pass by applying the following
>> noinline patch to the testcase.  Thoughts?
>
> How does it break?  It works for me with noinline and with -fPIC.
> Richard.

I'm running with mainline revision 132367 on x86_64-unknown-linux-gnu (i.e. 
lp64, not ilp32).  When I run the failing testcase under gdb I get:

(gdb) run
Starting program: /tmp/kg/build/gcc/testsuite/gcc/pr32268.exe

Program received signal SIGABRT, Aborted.
0x00002b2ae231d07b in raise () from /lib/libc.so.6
(gdb) up
#1  0x00002b2ae231e84e in abort () from /lib/libc.so.6
(gdb) up
#2  0x0000000000400575 in main () at 
/tmp/kg/egcc-SVN20080216/gcc/testsuite/gcc.target/i386/pr32268.c:32
32          abort();
(gdb) list
27        if (r != ((double) a < (double) b))
28          abort();
29
30        r = test_gt (a, b);
31        if (r != ((double) a > (double) b))
32          abort();
33
34        return 0;
35      }
(gdb) p a
$1 = 0
(gdb) p b
$2 = 0


"b" should be 1.0.

        --Kaveh
--
Kaveh R. Ghazi

Re: Mainline Regression in gcc.target/i386/pr32268.c

[Richard Guenther]

On Feb 16, 2008 8:01 PM, Kaveh R. Ghazi <ghazi@caip.rutgers.edu> wrote:
> From: "Richard Guenther" <richard.guenther@gmail.com>
>
> >> You can trigger the bug in a regular pass by applying the following
> >> noinline patch to the testcase.  Thoughts?
> >
> > How does it break?  It works for me with noinline and with -fPIC.
> > Richard.
>
> I'm running with mainline revision 132367 on x86_64-unknown-linux-gnu (i.e.
> lp64, not ilp32).  When I run the failing testcase under gdb I get:
>
> (gdb) run
> Starting program: /tmp/kg/build/gcc/testsuite/gcc/pr32268.exe
>
> Program received signal SIGABRT, Aborted.
> 0x00002b2ae231d07b in raise () from /lib/libc.so.6
> (gdb) up
> #1  0x00002b2ae231e84e in abort () from /lib/libc.so.6
> (gdb) up
> #2  0x0000000000400575 in main () at
> /tmp/kg/egcc-SVN20080216/gcc/testsuite/gcc.target/i386/pr32268.c:32
> 32          abort();
> (gdb) list
> 27        if (r != ((double) a < (double) b))
> 28          abort();
> 29
> 30        r = test_gt (a, b);
> 31        if (r != ((double) a > (double) b))
> 32          abort();
> 33
> 34        return 0;
> 35      }
> (gdb) p a
> $1 = 0
> (gdb) p b
> $2 = 0
>
>
> "b" should be 1.0.

Ok.  So this works for me with r132367, with the original testcase and flags,
with the original testcase and -fPIC and with the modified testcase
and both cases.

Richard.

Re: Mainline Regression in gcc.target/i386/pr32268.c

[Kaveh R. Ghazi]

From: "Richard Guenther" <richard.guenther@gmail.com>

>>
>> "b" should be 1.0.
>
> Ok.  So this works for me with r132367, with the original testcase and 
> flags,
> with the original testcase and -fPIC and with the modified testcase
> and both cases.
> Richard.

I can't explain why it passes for you.  (x86_64 lp64 mode, right?)  I just 
redid the tests and I can trigger the failure with either the noinline patch 
or -fpic.  I can also trigger it in the unpatched case by adding -O0 instead 
of -fpic which avoids the inline as well.

Can anyone else reproduce it?

        Thanks,
        --Kaveh
--
Kaveh R. Ghazi

Re: Mainline Regression in gcc.target/i386/pr32268.c

[Richard Guenther]

On Feb 16, 2008 8:19 PM, Kaveh R. Ghazi <ghazi@caip.rutgers.edu> wrote:
> From: "Richard Guenther" <richard.guenther@gmail.com>
>
> >>
> >> "b" should be 1.0.
> >
> > Ok.  So this works for me with r132367, with the original testcase and
> > flags,
> > with the original testcase and -fPIC and with the modified testcase
> > and both cases.
> > Richard.
>
> I can't explain why it passes for you.  (x86_64 lp64 mode, right?)  I just
> redid the tests and I can trigger the failure with either the noinline patch
> or -fpic.  I can also trigger it in the unpatched case by adding -O0 instead
> of -fpic which avoids the inline as well.
>
> Can anyone else reproduce it?

Ok, if I rebuild libgcc I can reproduce the fail.  Care to open a PR for this?

Thanks,
Richard.

[PATCH] Re: Mainline Regression in gcc.target/i386/pr32268.c

[Uros Bizjak]

Richard Guenther wrote:

>> I can't explain why it passes for you.  (x86_64 lp64 mode, right?)  I just
>> redid the tests and I can trigger the failure with either the noinline patch
>> or -fpic.  I can also trigger it in the unpatched case by adding -O0 instead
>> of -fpic which avoids the inline as well.
>>
>> Can anyone else reproduce it?
>>     
>
> Ok, if I rebuild libgcc I can reproduce the fail.  Care to open a PR for this?
>   

Please don't. The fix is to change soft-float/soft-fp.h a bit. Jakub, is 
this change acceptable for glibc?

Index: soft-fp.h
===================================================================
--- soft-fp.h   (revision 132367)
+++ soft-fp.h   (working copy)
@@ -189,7 +189,7 @@
 typedef USItype UHWtype;
 #endif
 
-#ifndef CMPtype
+#ifdef _LIBC
 #define CMPtype                int
 #endif
 
Uros.

Re: [PATCH] Re: Mainline Regression in gcc.target/i386/pr32268.c

[Kaveh R. Ghazi]

From: "Uros Bizjak" <ubizjak@gmail.com>

>> Ok, if I rebuild libgcc I can reproduce the fail.  Care to open a PR for 
>> this?
>>
>
> Please don't. The fix is to change soft-float/soft-fp.h a bit. Jakub, is 
> this change acceptable for glibc?
>
> Index: soft-fp.h
> ===================================================================
> --- soft-fp.h   (revision 132367)
> +++ soft-fp.h   (working copy)
> @@ -189,7 +189,7 @@
> typedef USItype UHWtype;
> #endif
>
> -#ifndef CMPtype
> +#ifdef _LIBC
> #define CMPtype                int
> #endif
>
> Uros.

If you install a fix, please consider also installing my noinline patch to 
the testcase so that problems are caught in the default scenario 
without -fpic/-fPIC.

        Thanks,
        --Kaveh
--
Kaveh R. Ghazi

Re: [PATCH] Re: Mainline Regression in gcc.target/i386/pr32268.c

[Jakub Jelinek]

On Sat, Feb 16, 2008 at 10:46:15PM +0100, Uros Bizjak wrote:
> Please don't. The fix is to change soft-float/soft-fp.h a bit. Jakub, is 
> this change acceptable for glibc?

No.  You should IMHO instead
#define CMPtype int __attribute__ ((mode (__libgcc_cmp_return__)))
in config/i386/sfp-machine.h.
There is nothing libc specific on having a default CMPtype definition.

> -#ifndef CMPtype
> +#ifdef _LIBC
> #define CMPtype                int
> #endif

	Jakub

Re: Re: [PATCH] Re: Mainline Regression in gcc.target/i386/pr32268.c

[David Edelsohn]

>>>>> Jakub Jelinek writes:

> No.  You should IMHO instead
> #define CMPtype int __attribute__ ((mode (__libgcc_cmp_return__)))
> in config/i386/sfp-machine.h.
> There is nothing libc specific on having a default CMPtype definition.

	Does this imply that the same change should be made to
sfp-machine.h for every target that supports soft-fp?  If so, all targets
should be fixed at the same time.

Thanks, David

Re: [PATCH] Re: Mainline Regression in gcc.target/i386/pr32268.c

[Uros Bizjak]

[-- Attachment #1: Type: text/plain, Size: 1191 bytes --]

Jakub Jelinek wrote:

>> Please don't. The fix is to change soft-float/soft-fp.h a bit. Jakub, is 
>> this change acceptable for glibc?
>>     
>
> No.  You should IMHO instead
> #define CMPtype int __attribute__ ((mode (__libgcc_cmp_return__)))
> in config/i386/sfp-machine.h.
> There is nothing libc specific on having a default CMPtype definition.
>   

Hm, this won't fly. But following will:

typedef int __gcc_CMPtype __attribute__ ((mode (__libgcc_cmp_return__)));
#define CMPtype __gcc_CMPtype

Attached patch adds this construct to i386/sfp-machine.h and 
rs6000/sfp-machine.h. It also changes test as Kaveh suggested.

2008-02-17  Uros Bizjak  <ubizjak@gmail.com>

        * config/i386/sfp-machine.h (__gcc_CPMtype): Rename from CMPtype.
        (CMPtype): Define as __gcc_CMPtype.
        * config/rs6000/sfp-machine.h (__gcc_CMPtype): New typedef.
        (CMPtype): Define as __gcc_CMPtype.

testsuite/ChangeLog:

2008-02-17  Uros Bizjak  <ubizjak@gmail.com>

        * gcc.target/i386/pr32268.c (test_lt): Add noinline attribute.
        (test_gt): Ditto.

The patch was tested on x86_64-pc-linux-gnu.

OK for mainline (patch needs rs6000 maintainer approval)?

Thanks,
Uros.

[-- Attachment #2: p.diff.txt --]
[-- Type: text/plain, Size: 1837 bytes --]

Index: config/i386/sfp-machine.h
===================================================================
--- config/i386/sfp-machine.h	(revision 132376)
+++ config/i386/sfp-machine.h	(working copy)
@@ -10,9 +10,9 @@
 
 /* The type of the result of a floating point comparison.  This must
    match `__libgcc_cmp_return__' in GCC for the target.  */
+typedef int __gcc_CMPtype __attribute__ ((mode (__libgcc_cmp_return__)));
+#define CMPtype __gcc_CMPtype
 
-typedef int CMPtype __attribute__ ((mode (__libgcc_cmp_return__)));
-
 #define _FP_MUL_MEAT_Q(R,X,Y)                           \
   _FP_MUL_MEAT_2_wide(_FP_WFRACBITS_Q,R,X,Y,umul_ppmm)
 
Index: config/rs6000/sfp-machine.h
===================================================================
--- config/rs6000/sfp-machine.h	(revision 132376)
+++ config/rs6000/sfp-machine.h	(working copy)
@@ -3,6 +3,11 @@
 #define _FP_WS_TYPE		signed long
 #define _FP_I_TYPE		long
 
+/* The type of the result of a floating point comparison.  This must
+   match `__libgcc_cmp_return__' in GCC for the target.  */
+typedef int __gcc_CMPtype __attribute__ ((mode (__libgcc_cmp_return__)));
+#define CMPtype __gcc_CMPtype
+
 #define _FP_MUL_MEAT_S(R,X,Y)				\
   _FP_MUL_MEAT_1_wide(_FP_WFRACBITS_S,R,X,Y,umul_ppmm)
 #define _FP_MUL_MEAT_D(R,X,Y)				\
Index: testsuite/gcc.target/i386/pr32268.c
===================================================================
--- testsuite/gcc.target/i386/pr32268.c	(revision 132376)
+++ testsuite/gcc.target/i386/pr32268.c	(working copy)
@@ -4,12 +4,14 @@
 
 extern void abort(void);
 
-int test_lt(__float128 x, __float128 y)
+int __attribute__ ((__noinline__))
+test_lt(__float128 x, __float128 y)
 {
   return x < y;
 }
 
-int test_gt (__float128 x, __float128 y)
+int __attribute__ ((__noinline__))
+test_gt (__float128 x, __float128 y)
 {
   return x > y;
 }

Re: [PATCH] Re: Mainline Regression in gcc.target/i386/pr32268.c

[Jakub Jelinek]

On Sun, Feb 17, 2008 at 08:27:03AM +0100, Uros Bizjak wrote:
> Jakub Jelinek wrote:
> 
> >>Please don't. The fix is to change soft-float/soft-fp.h a bit. Jakub, is 
> >>this change acceptable for glibc?
> >>    
> >
> >No.  You should IMHO instead
> >#define CMPtype int __attribute__ ((mode (__libgcc_cmp_return__)))
> >in config/i386/sfp-machine.h.
> >There is nothing libc specific on having a default CMPtype definition.
> >  
> 
> Hm, this won't fly. 

Ok, you're right.  Should have tried that in fn return value :(

> But following will:
> 
> typedef int __gcc_CMPtype __attribute__ ((mode (__libgcc_cmp_return__)));
> #define CMPtype __gcc_CMPtype

So would:
typedef int CMPtype __attribute__ ((mode (__libgcc_cmp_return__)));
#define CMPtype CMPtype

> Attached patch adds this construct to i386/sfp-machine.h and 
> rs6000/sfp-machine.h. It also changes test as Kaveh suggested.

Do we really need to change rs6000 before 4.3 release?  AFAIK soft-fp
is used there just for 32-bit stuff, and __libgcc_cmp_return__ will be there
always int.  All I'm saying is that in rs6000 case this isn't a regression
bugfix of any kind (and I doubt
http://gcc.gnu.org/viewcvs?root=gcc&view=rev&rev=132350
has been a regression bugfix either).  I'd say that's just a cleanup
we can do once 4.4 reopens (Monday).

	Jakub

Re: [PATCH] Re: Mainline Regression in gcc.target/i386/pr32268.c

[Uros Bizjak]

[-- Attachment #1: Type: text/plain, Size: 991 bytes --]

Jakub Jelinek wrote:

> Do we really need to change rs6000 before 4.3 release?  AFAIK soft-fp
> is used there just for 32-bit stuff, and __libgcc_cmp_return__ will be there
> always int.  All I'm saying is that in rs6000 case this isn't a regression
> bugfix of any kind (and I doubt
> http://gcc.gnu.org/viewcvs?root=gcc&view=rev&rev=132350
> has been a regression bugfix either).  I'd say that's just a cleanup
> we can do once 4.4 reopens (Monday).
>   

I have reverted mentioned change for now (and fixed pr32268.c):

2008-02-17  Uros Bizjak  <ubizjak@gmail.com>

        Revert:

        2008-02-15  Uros Bizjak  <ubizjak@gmail.com>
        * config/i386/sfp-machine.h (CMPtype): Define as typedef using
        libgcc_cmp_return mode.

testsuite/ChnageLog:

2008-02-17  Uros Bizjak  <ubizjak@gmail.com>

        * gcc.target/i386/pr32268.c (test_lt): Add noinline attribute.
        (test_gt): Ditto.


Patch was tested on x86_64-pc-linux-gnu, committed to mainline.

Thanks,
Uros.



[-- Attachment #2: p.diff.txt --]
[-- Type: text/plain, Size: 1148 bytes --]

Index: testsuite/gcc.target/i386/pr32268.c
===================================================================
--- testsuite/gcc.target/i386/pr32268.c	(revision 132376)
+++ testsuite/gcc.target/i386/pr32268.c	(working copy)
@@ -4,12 +4,14 @@
 
 extern void abort(void);
 
-int test_lt(__float128 x, __float128 y)
+int __attribute__ ((__noinline__))
+test_lt(__float128 x, __float128 y)
 {
   return x < y;
 }
 
-int test_gt (__float128 x, __float128 y)
+int __attribute__ ((__noinline__))
+test_gt (__float128 x, __float128 y)
 {
   return x > y;
 }
Index: config/i386/sfp-machine.h
===================================================================
--- config/i386/sfp-machine.h	(revision 132376)
+++ config/i386/sfp-machine.h	(working copy)
@@ -10,8 +10,7 @@ typedef unsigned int UTItype __attribute
 
 /* The type of the result of a floating point comparison.  This must
    match `__libgcc_cmp_return__' in GCC for the target.  */
-
-typedef int CMPtype __attribute__ ((mode (__libgcc_cmp_return__)));
+#define CMPtype			long
 
 #define _FP_MUL_MEAT_Q(R,X,Y)                           \
   _FP_MUL_MEAT_2_wide(_FP_WFRACBITS_Q,R,X,Y,umul_ppmm)

Re: [PATCH] Re: Mainline Regression in gcc.target/i386/pr32268.c

[David Edelsohn]

        * config/rs6000/sfp-machine.h (__gcc_CMPtype): New typedef.
        (CMPtype): Define as __gcc_CMPtype.

Yes, please go ahead and apply the equivalent of whatever fix you and
Jakub decide to apply to i386.

Thanks, David

Re: [PATCH] Re: Mainline Regression in gcc.target/i386/pr32268.c

[Uros Bizjak]

On Feb 19, 2008 2:29 AM, David Edelsohn <dje@watson.ibm.com> wrote:
>         * config/rs6000/sfp-machine.h (__gcc_CMPtype): New typedef.
>         (CMPtype): Define as __gcc_CMPtype.
>
> Yes, please go ahead and apply the equivalent of whatever fix you and
> Jakub decide to apply to i386.

> > typedef int __gcc_CMPtype __attribute__ ((mode (__libgcc_cmp_return__)));
> > #define CMPtype __gcc_CMPtype

> So would:
> typedef int CMPtype __attribute__ ((mode (__libgcc_cmp_return__)));
> #define CMPtype CMPtype

IMO, in the second version, the "#define X X" is a bit unusual. I
think that the first version describes the intention and avoids
unwanted cleanups in the future.

2008-02-19  Uros Bizjak  <ubizjak@gmail.com>

	* config/i386/sfp-machine.h (__gcc_CMPtype): New typedef.
	(CMPtype): Define as __gcc_CMPtype.
	* config/rs6000/sfp-machine.h (__gcc_CMPtype): New typedef.
	(CMPtype): Define as __gcc_CMPtype.

The patch was bootstrapped and regression tested on
x86_64-pc-linux-gnu, committed to 4.4.

Uros.

Index: config/i386/sfp-machine.h
===================================================================
--- config/i386/sfp-machine.h   (revision 132412)
+++ config/i386/sfp-machine.h   (working copy)
@@ -10,7 +10,8 @@

 /* The type of the result of a floating point comparison.  This must
    match `__libgcc_cmp_return__' in GCC for the target.  */
-#define CMPtype                        long
+typedef int __gcc_CMPtype __attribute__ ((mode (__libgcc_cmp_return__)));
+#define CMPtype __gcc_CMPtype

 #define _FP_MUL_MEAT_Q(R,X,Y)                           \
   _FP_MUL_MEAT_2_wide(_FP_WFRACBITS_Q,R,X,Y,umul_ppmm)
Index: config/rs6000/sfp-machine.h
===================================================================
--- config/rs6000/sfp-machine.h (revision 132412)
+++ config/rs6000/sfp-machine.h (working copy)
@@ -3,6 +3,11 @@
 #define _FP_WS_TYPE            signed long
 #define _FP_I_TYPE             long

+/* The type of the result of a floating point comparison.  This must
+   match `__libgcc_cmp_return__' in GCC for the target.  */
+typedef int __gcc_CMPtype __attribute__ ((mode (__libgcc_cmp_return__)));
+#define CMPtype __gcc_CMPtype
+
 #define _FP_MUL_MEAT_S(R,X,Y)                          \
   _FP_MUL_MEAT_1_wide(_FP_WFRACBITS_S,R,X,Y,umul_ppmm)
 #define _FP_MUL_MEAT_D(R,X,Y)                          \