Bugzilla Bug 59793

iozone 3.218 and 3.221 get an internal compiler error in in try_split at
emit-rtl.c:3326 when compiling with gcc-3.4.1; compiling with gcc-3.3.4 works
perfectly.

Reproducible: Always
Steps to Reproduce:
1. gcc-config i686-pc-linux-gnu-3.4.1
2. source /etc/profile
3. emerge iozone

Actual Results:  
gcc -march=athlon-xp -O2 -pipe -frename-registers -fomit-frame-pointer   -c -o
libasync.o libasync.c
iozone.c: In function `do_compute':
iozone.c:16294: internal compiler error: in try_split, at emit-rtl.c:3326
Please submit a full bug report,
with preprocessed source if appropriate.
See <URL:http://bugs.gentoo.org/> for instructions.
Preprocessed source stored into /var/tmp/portage/iozone-3.218/temp/ccXvVxrH.out
file, please attach this to your bugreport.
make: *** [iozone_linux.o] Error 1

!!! ERROR: app-benchmarks/iozone-3.218 failed.
!!! Function src_compile, Line 63, Exitcode 2
!!! Compile failed

Expected Results:  
successful emerge

gcc-3.4.1-r2 (non-hardened); glibc-2.3.3.20040420; linux26-headers-2.6.7-r3;
CFLAGS="-march=athlon-xp -O2 -pipe -frename-registers -fomit-frame-pointer"

Created an attachment (id=37053) [edit]
preprocessed source file, iozone-3.218

Preprocessed source stored into /var/tmp/portage/iozone-3.218/temp/ccXvVxrH.out
file, please attach this to your bugreport.

Created an attachment (id=37054) [edit]
preprocessed source file, iozone-3.221

Preprocessed source stored into /var/tmp/portage/iozone-3.221/temp/ccVKnU1o.out
file, please attach this to your bugreport.

both emerge fine for me with gcc 3.4.1 20040803 (gentoo Linux 3.4.1-r2,
ssp-3.4.2, pie-8.7.6.5)

Portage 2.0.51_pre17 (default-x86-2004.2, gcc-3.4.1, glibc-2.3.3.20040420-r1,
2.6.7-gentoo-r11 i686 Celeron (Coppermine))
=================================================================
System uname: 2.6.7-gentoo-r11 i686 Celeron (Coppermine)
Gentoo Base System version 1.4.16
distcc 2.13 i686-pc-linux-gnu (protocols 1 and 2) (default port 3632) [enabled]
ccache version 2.3 [enabled]
Autoconf: sys-devel/autoconf-2.59-r4
Automake: sys-devel/automake-1.8.3
Binutils: sys-devel/binutils-2.14.90.0.8-r1
Headers:  sys-kernel/linux-headers-2.4.21-r1,sys-kernel/linux-headers-2.4.19-r1
Libtools: sys-devel/libtool-1.4.3-r4
ACCEPT_KEYWORDS="x86"
AUTOCLEAN="yes"
CFLAGS="-march=pentium3 -O2 -pipe"

Works for me too.

OK, the bug only seems to trigger when I have -march=athlon-xp with gcc-3.4.1;
if I set CFLAGS to "-march=i686 -O2 -pipe -frename-registers
-fomit-frame-pointer", it compiles fine.
Perhaps some GCC people could look at this? Why is -march=athlon-xp unsafe with
gcc-3.4 ?

Still no problems:

env ACCEPT_KEYWORDS=~x86 CFLAGS="-march=athlon-xp -O2 -pipe -frename-registers -fomit-frame-pointer" emerge iozone
Calculating dependencies ...done!
>>> emerge (1 of 1) app-benchmarks/iozone-3.221 to /
>>> md5 src_uri ;-) iozone3_221.tar
>>> Unpacking source...
>>> Unpacking iozone3_221.tar to /var/tmp/portage/iozone-3.221/work
>>> Source unpacked.

Building iozone for Linux

gcc -march=athlon-xp -O2 -pipe -frename-registers -fomit-frame-pointer -Dunix -c -DHAVE_ANSIC_C -DASYNC_IO -DHAVE_PREAD \
        -DSHARED_MEM -Dlinux -D_LARGEFILE64_SOURCE iozone.c \
        -DNAME='"linux"' -o iozone_linux.o
gcc -march=athlon-xp -O2 -pipe -frename-registers -fomit-frame-pointer   -c -o libbif.o libbif.c
gcc -march=athlon-xp -O2 -pipe -frename-registers -fomit-frame-pointer   -c -o libasync.o libasync.c
gcc -march=athlon-xp -O2 -pipe -frename-registers -fomit-frame-pointer -Dunix -c -DHAVE_ANSIC_C -DASYNC_IO -D_LARGEFILE64_SOURCE \
        -DSHARED_MEM -Dlinux libbif.c -o libbif.o
gcc -march=athlon-xp -O2 -pipe -frename-registers -fomit-frame-pointer -Dunix -c -Dlinux -DHAVE_ANSIC_C -DASYNC_IO \
        -D_LARGEFILE64_SOURCE libasync.c  -o libasync.o
gcc -march=athlon-xp -O2 -pipe -frename-registers -fomit-frame-pointer -Dunix -DHAVE_ANSIC_C -DSHARED_MEM -DASYNC_IO \
        -D_LARGEFILE64_SOURCE -Dlinux \
        iozone_linux.o libasync.o libbif.o -lpthread \
        -lrt -o iozone
>>> Test phase [not enabled]: app-benchmarks/iozone-3.221

>>> Install iozone-3.221 into /var/tmp/portage/iozone-3.221/image/ category app-benchmarks

Portage 2.0.51_pre17 (default-x86-2004.2, gcc-3.4.1, glibc-2.3.3.20040420-r0, 2.6.7-gentoo-r11 i686 AMD Athlon(tm) XP 2500+)
=================================================================
System uname: 2.6.7-gentoo-r11 i686 AMD Athlon(tm) XP 2500+
Gentoo Base System version 1.4.16
distcc 2.13 i686-pc-linux-gnu (protocols 1 and 2) (default port 3632) [disabled]
Autoconf: sys-devel/autoconf-2.59-r4
Automake: sys-devel/automake-1.8.3
Binutils: sys-devel/binutils-2.14.90.0.8-r1
Headers:  sys-kernel/linux26-headers-2.6.7-r3
Libtools: sys-devel/libtool-1.4.3-r4
ACCEPT_KEYWORDS="x86 ~x86"
AUTOCLEAN="yes"
CFLAGS="-march=athlon-xp"
CHOST="i686-pc-linux-gnu"
COMPILER="gcc3"
CONFIG_PROTECT="/etc /usr/X11R6/lib/X11/xkb /usr/kde/2/share/config /usr/kde/3.2/share/config /usr/kde/3/share/config /usr/lib/mozilla/defaults/pref /usr/share/config /var/qmail/control"
CONFIG_PROTECT_MASK="/etc/gconf /etc/terminfo /etc/env.d"
CXXFLAGS="-O2 -march=athlon-xp -fomit-frame-pointer -pipe"


bash-2.05b# cat /proc/cpuinfo
processor       : 0
vendor_id       : AuthenticAMD
cpu family      : 6
model           : 10
model name      : AMD Athlon(tm) XP 2500+
stepping        : 0
cpu MHz         : 1837.612
cache size      : 512 KB
fdiv_bug        : no
hlt_bug         : no
f00f_bug        : no
coma_bug        : no
fpu             : yes
fpu_exception   : yes
cpuid level     : 1
wp              : yes
flags           : fpu vme de pse tsc msr pae mce cx8 apic sep mtrr pge mca cmov pat pse36 mmx fxsr sse syscall mmxext 3dnowext 3dnow
bogomips        : 3629.05

Is this a gcc-3.4.1-r2 problem or gcc-3.4.1

Still cannot repoduce this on a Athlon XP. Can you try remerging gcc-3.4.1-r1 just to be sure.

1. gcc-3.4.1-r1 is not available
2. same errors with hardened gcc-3.4.1 CFLAGS="-march=athlon-xp -O2 -pipe -frename-registers -fomit-frame-pointer"
3. removing or adding -frename-registers and -fomit-frame-pointer has no effects with gcc-3.4.1
4. changing to "-march=i686 -O2" has no effects in 3.4.1 (i.e. doesn't compile)
5. "-march=athlon-xp -O1" does not compile
6. "-march=athlon-xp -O0" *does* compile fine with gcc-3.4.1
7. In general, setting optimization to -O0 seems to be the only way I can avoid this with gcc-3.4.1

Anyway, I have the same CPU (athlon xp 2500), same kernel and headers, and presumably the same compiler (gcc-3.4.1 hardened), and presumably the same glibc (2.3.3.20040420 hardened). The only thing I can think of is that my glibc was compiles with hardened gcc-3.3.3, which shouldn't matter because the C ABI didn't change.
I have no idea what could be causing the errror.

OK, I've isolated the problem: compiling the following with gcc-3.4.1
-mtune=i686 -O (or -mtune=athlon-xp etc, but not -mtune=i586) on my system
produces an internal compiler error:

---snip iozone-test.c---
float
foo(float x)
{
 if(x == (float)0)
  return(0);
}
---snip---
xrostov@sunnyvale tmp $ /usr/libexec/gcc/i686-pc-linux-gnu/3.4.1/cc1 foo.c
-mtune=i686 -O -o -
        .file   "foo.c"  foo

foo.c: In function `foo':
foo.c:6: internal compiler error: in try_split, at emit-rtl.c:3326
Please submit a full bug report,
with preprocessed source if appropriate.
See <URL:http://bugs.gentoo.org/> for instructions.

Dragonheart, do you have any problems compiling foo.c with any compiler
switches?

A couple of comments. First, the function is not in good C style - there is no
default return value. Adding return(0) at the end of the function removes the
compiler errror (note that do_compute() in iozone does not have a return at the
end of the function). Second, the gcc manual says that "-O" means "-fdefer-pop
-fmerge-constants -fthread-jumps -floop-optimize -fif-conversion
-fif-conversion2 -fdelayed-branch -fguess-branch-probability -fcprop-registers
-fomit-frame-pointer". I tried to compile with those switches, and there were
no errors. So it seems that -O combined with -mtune=i686 turns on something
hidden in my version of gcc-3.4.1

Sorry, typo, that should have been
---snip foo.c---

very good analysis. Hardened peoples - what are your thoughts?
I'm compiling gcc-3.4.1-r1(hardened) now.

compiles fine here with gcc-3.4.1-r2, binutils-2.14.90.0.8 20040114,
glibc-2.3.4.20040619-r1

iozone-3.221-r1.ebuild commited with a patch to put return values on functions.

Note also that 0 is an int and 0.0 is a float in your code. Maybe an inplicit conversion problem in your version of gcc for some reason.

Reopen if there are problems.

Passing onto a few arch testers too. FYI includes src_test routine.

arm added to 3.221-r1

I just tried to build app-benchmark/iozone-3.226 on ppc64, but I get this:

make: *** No rule to make target `linux-ppc64'. Stop

So we need a ppc64 rule. Anybody volunteer for writing one? 0:-)

Markus

this has been easier than I thought. I've added app-bechmark/iozone-3.226-r1
which includes the ppc64 patch and have added ~ppc64 to KEYWORDS.

Markus

updating version in title. Good one Markus. Includes src_test routine for
testing.