From mboxrd@z Thu Jan 1 00:00:00 1970 Message-ID: <38CF94BE.E62C4353@cray.com> Date: Wed, 15 Mar 2000 07:48:46 -0600 From: Mike Wolf MIME-Version: 1.0 To: embedded Subject: Re: compiling glibc for 4xx core powerpc References: <38CEEF3B.2665CD22@m2tech.co.nz> Content-Type: text/plain; charset=us-ascii Sender: owner-linuxppc-embedded@lists.linuxppc.org List-Id: Ben, I've been hitting this problem also. I've been working is a cross compiling enviroment. I found an posting from Marcus Sundberg that outlined the problems and how to fix them. I have included that post at the bottom of this. I got a patch from Wolfgang Denk. He also has a site that talks about what needs to be done. http://www.denx.de/solutions-en.html The patch is here: ----------------------------------------------------------------------------- --- dl-machine.c.ORIG Fri Mar 5 00:26:43 1999 +++ dl-machine.c Sat Nov 6 14:33:47 1999 @@ -250,7 +250,11 @@ PowerPC processors have line sizes of exactly 32 bytes. */ size_modified = lazy ? rel_offset_words : PLT_INITIAL_ENTRY_WORDS; +#ifdef PPC_CACHELINESIZE_32 for (i = 0; i < size_modified; i+= 8) +#else + for (i = 0; i < size_modified; i+= 4) +#endif PPC_DCBST (plt + i); PPC_DCBST (plt + size_modified - 1); PPC_SYNC; ----------------------------------------------------------------------------- --- Dist.ORIG Mon Sep 14 18:37:37 1998 +++ Dist Sat Nov 6 14:35:45 1999 @@ -1,7 +1,3 @@ dl-machine.c dl-start.S -fenv_const.c -fenv_libc.h ppc-mcount.S -fe_nomask.c -t_sqrt.c ----------------------------------------------------------------------------- --- Makefile.ORIG Tue Sep 1 16:30:52 1998 +++ Makefile Sat Nov 6 14:39:22 1999 @@ -1,7 +1,3 @@ -ifeq ($(subdir),math) -libm-support += fenv_const fe_nomask t_sqrt -endif - ifeq ($(subdir),gmon) sysdep_routines += ppc-mcount endif I applied that patch from the glibc/sysdeps/powerpc directory. Then I ran the following script. I tried to run the one Marcus had in his email, but it is bash specific and I am running under IRIX. #!/bin/ksh echo "mv bits/fenv.h bits/mathdef.h bits/mathinline.h fpu/bits" mv bits/fenv.h bits/mathdef.h bits/mathinline.h fpu/bits echo "mv Versions e_sqrt.c e_sqrtf.c fclrexcpt.c fe_nomask.c fegetenv.c fegetround.c feholdexcpt.c fenv_const.c fenv_libc.h fesetenv.c fesetround.c feupdateenv.c fgetexcptflg.c fpu_control.h fraiseexcpt.c fsetexcptflg.c ftestexcept.c s_copysign.S s_copysignf.S s_fabs.S s_fabsf.S s_fmax.S s_fmaxf.S s_fmin.S s_fminf.S s_isnan.c s_isnanf.S s_lrint.c s_lrintf.S s_rint.c s_rintf.c t_sqrt.c w_sqrt.c w_sqrtf.c fpu" mv Versions e_sqrt.c e_sqrtf.c fclrexcpt.c fe_nomask.c fegetenv.c fegetround.c feholdexcpt.c fenv_const.c fenv_libc.h fesetenv.c fesetround.c feupdateenv.c fgetexcptflg.c fpu_control.h fraiseexcpt.c fsetexcptflg.c ftestexcept.c s_copysign.S s_copysignf.S s_fabs.S s_fabsf.S s_fmax.S s_fmaxf.S s_fmin.S s_fminf.S s_isnan.c s_isnanf.S s_lrint.c s_lrintf.S s_rint.c s_rintf.c t_sqrt.c w_sqrt.c w_sqrtf.c fpu echo "cat fpu/fenv_libc.h fpu/fe_nomask.c fpu/t_sqrt.c >> fpu/Dist" cat fpu/fenv_libc.h fpu/fe_nomask.c fpu/t_sqrt.c >> fpu/Dist echo "ifeq ($(subdir),math)\nlibm-support += fenv_const fe_nomask t_sqrt\nendif\n" > fpu/Makefile echo "Created fpu/Makefile" Then I configured and compiled and it compiled up fine. I havent gotten a chance to test it yet. Following is the post from Marcus: Hi, first note the "Reply To" - we have a new list, so why not use it. Scott Wood wrote: > > Graham Stoney wrote: > > > > Scott Wood writes: > > > The MPC8xx have no floating point, so it should be compiled with -msoft-float, > > > (I think)... ./configure --without-fp will do it. > > > > Is this implied when gcc has been configured --with-cpu=860? Configuring gcc with --with-cpu=860 and --nfp will make the _compiler_ default to -msoft-float. It will however _not_ make the pre-processor define -D_SOFT_FLOAT by default, which will cause all variable arguments functions taking floating point arguments (like the *print[fs] family) to be mis-compiled. The fix is to add this code to your gcc's 'specs' file under the section '*cpp_sysv': %{!mhard-float: -D_SOFT_FLOAT} Likewise, passing -mcpu=860 to gcc will imply -msoft-float, but not -D_SOFT_FLOAT. So either fix your specs file or always explicitly pass -msoft-float to gcc. > I don't think so... I put -msoft-float into the sysdeps/powerpc/Makefile and I see > it twice below, so I would imagine that if you use both flags then -msoft-float > will be used. > > I am using YellowDog Linux now and instead of an 'internal compiler error', > I get: > > make[1]: Entering directory `/usr/src/redhat/SOURCES/glibc-2.1.1/math' > gcc ../sysdeps/powerpc/s_isnan.c -c -Wall -Winline -Wstrict-prototypes -Wwrite-strings -mcpu=860 -msoft-float -fpic > -msoft-float -Wno-uninitialized -Wno-write-strings -I../include -I. -I.. -I../libio -I../sysdeps/powerpc/elf > -I../sysdeps/unix/sysv/linux/powerpc -I../sysdeps/unix/sysv/linux -I../sysdeps/gnu -I../sysdeps/unix/common -I../sysdeps/unix/mman > -I../sysdeps/unix/inet -I../sysdeps/unix/sysv -I../sysdeps/unix -I../sysdeps/posix -I../sysdeps/powerpc -I../sysdeps/wordsize-32 > -I../sysdeps/ieee754 -I../sysdeps/libm-ieee754 -I../sysdeps/generic/elf -I../sysdeps/generic -include > ../include/libc-symbols.h -DPIC -D__NO_MATH_INLINES -D__LIBC_INTERNAL_MATH_INLINES -DNO_LONG_DOUBLE -D_Mlong_double_=double -o > s_isnan.os > ../sysdeps/powerpc/s_isnan.c: In function `__isnan': > ../sysdeps/powerpc/s_isnan.c:32: fixed or forbidden register 32 (0) was spilled for class FLOAT_REGS. > This may be due to a compiler bug or to impossible asm > statements or clauses. > make[1]: *** [s_isnan.os] Error 1 Ok, here's how to build glibc for embedded PPC: First you must remove the assumption that cachelines are 32 bytes: Apply this diff, and simply move sysdeps/powerpc/memset.S out of the way for now: diff -ur orig/glibc-2.1.1/sysdeps/powerpc/dl-machine.c glibc-2.1.1/sysdeps/powerpc/dl-machine.c --- orig/glibc-2.1.1/sysdeps/powerpc/dl-machine.c Fri Mar 5 23:41:23 1999 +++ glibc-2.1.1/sysdeps/powerpc/dl-machine.c Mon May 17 20:59:06 1999 @@ -250,7 +250,11 @@ PowerPC processors have line sizes of exactly 32 bytes. */ size_modified = lazy ? rel_offset_words : PLT_INITIAL_ENTRY_WORDS; +#ifdef PPC_CACHELINESIZE_32 for (i = 0; i < size_modified; i+= 8) +#else + for (i = 0; i < size_modified; i+= 4) +#endif PPC_DCBST (plt + i); PPC_DCBST (plt + size_modified - 1); PPC_SYNC; (The ifdef is bogus - PPC_CACHELINESIZE_32 is actually never defined, and this should really be handled runtime anyway. The problem is that reading the PVR is a supervisor level operation for some stupid reason, and afaik there is no other way to find out what the line size is. My vote is to have a special sysctl entry for the cacheline size, for fast and easy access (one syscall compared to the open()/read()/close() triplet for normal /proc entries, and you also don't have to have the /proc fs mounted), and then cache the result in a static variable.) Secondly you will want to remove the floating point assembler. This is done by piping the following diff into the attached script (with sysdeps/powerpc as your cwd): diff -ur orig/glibc-2.1.1/sysdeps/powerpc/Dist glibc-2.1.1/sysdeps/powerpc/Dist --- orig/glibc-2.1.1/sysdeps/powerpc/Dist Tue Sep 15 00:57:08 1998 +++ glibc-2.1.1/sysdeps/powerpc/Dist Tue May 4 17:00:33 1999 @@ -1,7 +1,3 @@ dl-machine.c dl-start.S -fenv_const.c -fenv_libc.h ppc-mcount.S -fe_nomask.c -t_sqrt.c diff -ur orig/glibc-2.1.1/sysdeps/powerpc/Makefile glibc-2.1.1/sysdeps/powerpc/Makefile --- orig/glibc-2.1.1/sysdeps/powerpc/Makefile Wed Nov 18 00:48:00 1998 +++ glibc-2.1.1/sysdeps/powerpc/Makefile Tue May 4 17:01:07 1999 @@ -1,7 +1,3 @@ -ifeq ($(subdir),math) -libm-support += fenv_const fe_nomask t_sqrt -endif - ifeq ($(subdir),gmon) sysdep_routines += ppc-mcount endif Now you can simply configure glibc with --without-fp and build it. It is possible that some of the more advanced stuff in the math library will not work - I get warnings about fe_* functions being stubs, but as I have no idea of what a "floating point environment" is or how to implement it when using soft floats it's nothing I can do about it. Most programs work fine here. //Marcus -- -------------------------------+------------------------------------ Marcus Sundberg | http://www.stacken.kth.se/~mackan/ Royal Institute of Technology | Phone: +46 707 295404 Stockholm, Sweden | E-Mail: mackan@stacken.kth.se 
#!/bin/sh

FPUFILES="
Versions
bits/fenv.h
bits/mathdef.h
bits/mathinline.h
e_sqrt.c
e_sqrtf.c
fclrexcpt.c
fe_nomask.c
fegetenv.c
fegetround.c
feholdexcpt.c
fenv_const.c
fenv_libc.h
fesetenv.c
fesetround.c
feupdateenv.c
fgetexcptflg.c
fpu_control.h
fraiseexcpt.c
fsetexcptflg.c
ftestexcept.c
s_copysign.S
s_copysignf.S
s_fabs.S
s_fabsf.S
s_fmax.S
s_fmaxf.S
s_fmin.S
s_fminf.S
s_isnan.c
s_isnanf.S
s_lrint.c
s_lrintf.S
s_rint.c
s_rintf.c
t_sqrt.c
w_sqrt.c
w_sqrtf.c
"

FPUDISTFILES="fenv_const.c
fenv_libc.h
fe_nomask.c
t_sqrt.c
"

MAKEFILE='-ifeq ($(subdir),math)
-libm-support += fenv_const fe_nomask t_sqrt
-endif'

patch -p1

cd sysdeps/powerpc #|| exit 1
mkdir -p fpu/bits || exit 2

for a in $FPUFILES; do
  mv "$a" "fpu/$a" && echo "Moved $a -> fpu/$a"
done

fail=
for a in $FPUDISTFILES; do
  echo "$a" >> fpu/Dist || fail=1
done
test "$fail" || echo "Created fpu/Dist"

Hi,

first note the "Reply To" - we have a new list, so why not use it.

Scott Wood wrote:
>
> Graham Stoney wrote:
> >
> > Scott Wood writes:
> > > The MPC8xx have no floating point, so it should be
compiled with -msoft-float,
> > > (I think)...  ./configure --without-fp  will do it.
> >
> > Is this implied when gcc has been configured --with-cpu=860?

Configuring gcc with --with-cpu=860 and --nfp will make the
_compiler_ default to -msoft-float. It will however _not_ make the
pre-processor define -D_SOFT_FLOAT by default, which will cause
all variable arguments functions taking floating point arguments
(like the *print[fs] family) to be mis-compiled.

The fix is to add this code to your gcc's 'specs' file under the
section '*cpp_sysv':
%{!mhard-float: -D_SOFT_FLOAT}

Likewise, passing -mcpu=860 to gcc will imply -msoft-float, but
not -D_SOFT_FLOAT. So either fix your specs file or always
explicitly pass -msoft-float to gcc.

> I don't think so...  I put -msoft-float into the
sysdeps/powerpc/Makefile and I see
> it twice below, so I would imagine that if you use both flags then
-msoft-float
> will be used.
>
> I am using YellowDog Linux now and instead of an 'internal compiler
error',
> I get:
>
> make[1]: Entering directory
`/usr/src/redhat/SOURCES/glibc-2.1.1/math'
> gcc ../sysdeps/powerpc/s_isnan.c -c -Wall -Winline
-Wstrict-prototypes -Wwrite-strings -mcpu=860 -msoft-float   -fpic
> -msoft-float   -Wno-uninitialized -Wno-write-strings  -I../include
-I.  -I.. -I../libio  -I../sysdeps/powerpc/elf
> -I../sysdeps/unix/sysv/linux/powerpc -I../sysdeps/unix/sysv/linux
-I../sysdeps/gnu -I../sysdeps/unix/common -I../sysdeps/unix/mman
> -I../sysdeps/unix/inet -I../sysdeps/unix/sysv -I../sysdeps/unix
-I../sysdeps/posix -I../sysdeps/powerpc -I../sysdeps/wordsize-32
> -I../sysdeps/ieee754 -I../sysdeps/libm-ieee754
-I../sysdeps/generic/elf -I../sysdeps/generic    -include
> ../include/libc-symbols.h  -DPIC   -D__NO_MATH_INLINES
-D__LIBC_INTERNAL_MATH_INLINES -DNO_LONG_DOUBLE -D_Mlong_double_=double
-o
> s_isnan.os
> ../sysdeps/powerpc/s_isnan.c: In function `__isnan':
> ../sysdeps/powerpc/s_isnan.c:32: fixed or forbidden register 32 (0)
was spilled for class FLOAT_REGS.
> This may be due to a compiler bug or to impossible asm
> statements or clauses.
> make[1]: *** [s_isnan.os] Error 1

Ok, here's how to build glibc for embedded PPC:

First you must remove the assumption that cachelines are 32 bytes:
Apply this diff, and simply move sysdeps/powerpc/memset.S out of the
way for now:

diff -ur orig/glibc-2.1.1/sysdeps/powerpc/dl-machine.c
glibc-2.1.1/sysdeps/powerpc/dl-machine.c
--- orig/glibc-2.1.1/sysdeps/powerpc/dl-machine.c       Fri Mar  5
23:41:23 1999
+++ glibc-2.1.1/sysdeps/powerpc/dl-machine.c    Mon May 17 20:59:06 1999
@@ -250,7 +250,11 @@
         PowerPC processors have line sizes of exactly 32 bytes.  */

       size_modified = lazy ? rel_offset_words :
PLT_INITIAL_ENTRY_WORDS;
+#ifdef PPC_CACHELINESIZE_32
       for (i = 0; i < size_modified; i+= 8)
+#else
+      for (i = 0; i < size_modified; i+= 4)
+#endif
        PPC_DCBST (plt + i);
       PPC_DCBST (plt + size_modified - 1);
       PPC_SYNC;

(The ifdef is bogus - PPC_CACHELINESIZE_32 is actually never defined,
and this should really be handled runtime anyway. The problem is that
reading the PVR is a supervisor level operation for some stupid reason,
and afaik there is no other way to find out what the line size is.
My vote is to have a special sysctl entry for the cacheline size,
for fast and easy access (one syscall compared to the
open()/read()/close()
triplet for normal /proc entries, and you also don't have to have the
/proc fs mounted), and then cache the result in a static variable.)

Secondly you will want to remove the floating point assembler.
This is done by piping the following diff into the attached script
(with sysdeps/powerpc as your cwd):

diff -ur orig/glibc-2.1.1/sysdeps/powerpc/Dist
glibc-2.1.1/sysdeps/powerpc/Dist
--- orig/glibc-2.1.1/sysdeps/powerpc/Dist       Tue Sep 15 00:57:08 1998
+++ glibc-2.1.1/sysdeps/powerpc/Dist    Tue May  4 17:00:33 1999
@@ -1,7 +1,3 @@
 dl-machine.c
 dl-start.S
-fenv_const.c
-fenv_libc.h
 ppc-mcount.S
-fe_nomask.c
-t_sqrt.c
diff -ur orig/glibc-2.1.1/sysdeps/powerpc/Makefile
glibc-2.1.1/sysdeps/powerpc/Makefile
--- orig/glibc-2.1.1/sysdeps/powerpc/Makefile   Wed Nov 18 00:48:00 1998
+++ glibc-2.1.1/sysdeps/powerpc/Makefile        Tue May  4 17:01:07 1999
@@ -1,7 +1,3 @@
-ifeq ($(subdir),math)
-libm-support += fenv_const fe_nomask t_sqrt
-endif
-
 ifeq ($(subdir),gmon)
 sysdep_routines += ppc-mcount
 endif

Now you can simply configure glibc with --without-fp and build it.
It is possible that some of the more advanced stuff in the math
library will not work - I get warnings about fe_* functions being
stubs, but as I have no idea of what a "floating point environment"
is or how to implement it when using soft floats it's nothing I can
do about it. Most programs work fine here.

//Marcus
--
-------------------------------+------------------------------------
        Marcus Sundberg        | http://www.stacken.kth.se/~mackan/
 Royal Institute of Technology |       Phone: +46 707 295404
       Stockholm, Sweden       |   E-Mail: mackan@stacken.kth.se
#!/bin/sh

FPUFILES="
Versions
bits/fenv.h
bits/mathdef.h
bits/mathinline.h
e_sqrt.c
e_sqrtf.c
fclrexcpt.c
fe_nomask.c
fegetenv.c
fegetround.c
feholdexcpt.c
fenv_const.c
fenv_libc.h
fesetenv.c
fesetround.c
feupdateenv.c
fgetexcptflg.c
fpu_control.h
fraiseexcpt.c
fsetexcptflg.c
ftestexcept.c
s_copysign.S
s_copysignf.S
s_fabs.S
s_fabsf.S
s_fmax.S
s_fmaxf.S
s_fmin.S
s_fminf.S
s_isnan.c
s_isnanf.S
s_lrint.c
s_lrintf.S
s_rint.c
s_rintf.c
t_sqrt.c
w_sqrt.c
w_sqrtf.c
"

FPUDISTFILES="fenv_const.c
fenv_libc.h
fe_nomask.c
t_sqrt.c
"

MAKEFILE='-ifeq ($(subdir),math)
-libm-support += fenv_const fe_nomask t_sqrt
-endif'

patch -p1

cd sysdeps/powerpc #|| exit 1
mkdir -p fpu/bits || exit 2

for a in $FPUFILES; do
  mv "$a" "fpu/$a" && echo "Moved $a -> fpu/$a"
done

fail=
for a in $FPUDISTFILES; do
  echo "$a" >> fpu/Dist || fail=1
done
test "$fail" || echo "Created fpu/Dist"

echo "$MAKEFILE" >fpu/Makefile &&echo "$MAKEFILE"
>fpu/Makefile && echo "Created fpu/Makefile"

--
Mike Wolf
Cray Research
655-F Lone Oak Drive, Eagan, MN 55121
mwolf@cray.com          (651)683-5313

** Sent via the linuxppc-embedded mail list. See http://lists.linuxppc.org/