[Bug tree-optimization/52734] [4.7/4.8 Regression] Incorrect optimization of uClibc sbrk()
http://gcc.gnu.org/bugzilla/show_bug.cgi?id=52734 --- Comment #13 from Sedat Dilek 2012-04-10 22:57:15 UTC --- Just some remarks: Freetz is an OSS project supporting router (platforms) of a popular German company. The target-system is mostly MIPS(EL). The majority of our developers build on AMD64 hosts (32bit-toolchain). The full patchset for gcc-4.7.0 can be found in [1]. Jörg provided his adapted/backported patch in [2]. Thank you for testing! [0] http://freetz.org/ [1] http://freetz.org/browser/trunk/toolchain/make/target/gcc/4.7.0/ [2] http://freetz.org/attachment/ticket/1752/r8889-gcc-4.7-fix.patch
[Bug tree-optimization/52734] [4.7/4.8 Regression] Incorrect optimization of uClibc sbrk()
http://gcc.gnu.org/bugzilla/show_bug.cgi?id=52734 --- Comment #17 from Sedat Dilek 2012-04-13 19:38:24 UTC --- Thanks, Tom!
[Bug tree-optimization/49618] New: When building uClibc with GCC 4.6.1 old_atexit is miscompiled
http://gcc.gnu.org/bugzilla/show_bug.cgi?id=49618
Summary: When building uClibc with GCC 4.6.1 old_atexit is
miscompiled
Product: gcc
Version: unknown
Status: UNCONFIRMED
Severity: normal
Priority: P3
Component: tree-optimization
AssignedTo: unassig...@gcc.gnu.org
ReportedBy: sedat.di...@gmail.com
Created attachment 24661
--> http://gcc.gnu.org/bugzilla/attachment.cgi?id=24661
Preprocessed file "old_atexit.i"
Hi,
this issue kept me UP for several weeks. It first occured when generating a
mipsel target-toolchain based on gcc-4.6.0 with uClibc-0.9.32 for a router
project called freetz. Finally, I could track the problem with the assistance
from Edwin Török. A big thank you, Edwin.
### Problem description:
When building uClibc with GCC 4.6.1 old_atexit is miscompiled, which causes
this testprogram to crash when calling old_atexit:
#include
void foo() {}
int main() { return atexit(foo);}
This is a regression from GCC 4.5.3 which compiled old_atexit() fine.
Attached is the preprocessed file old_atexit.i.
Commandline to create old_atexit.os:
mipsel-linux-uclibc-gcc -S old_atexit.i -o old_atexit.os -funsigned-char
-fno-builtin -fno-asm -msoft-float -std=gnu99 -march=4kc -mtune=4kc -mabi=32
-fno-stack-protector -Os -funit-at-a-time
-fmerge-all-constants -fstrict-aliasing -fno-tree-loop-optimize
-fno-tree-dominator-opts -fno-strength-reduce -mno-split-addresses -fPIC
As seen below with GCC 4.6.1 &__dso_handle is assumed to be non-NULL
and the branch (beqz) eliminated, but it is in fact NULL at runtime
which causes the crash.
With GCC 4.5.3 there is a beqz that tests for &__dso_handle == NULL:
000537d0 :
537d0: 3c1c0003lui gp,0x3
537d4: 279c8d10addiu gp,gp,-29424
537d8: 0399e021addugp,gp,t9
537dc: 8f828a2clw v0,-30164(gp)
537e0: 8f9989aclw t9,-30292(gp)
537e4: 8c46lw a2,0(v0)
^^ SIGSEGV here, with a NULL dereference
537e8: 2821movea1,zero
537ec: 0328jr t9
537f0: 0002300amovza2,zero,v0
old_atexit.os with GCC 4.6.1:
.file 1 "old_atexit.c"
.section .mdebug.abi32
.previous
.gnu_attribute 4, 3
.abicalls
.text
.align 2
.globl old_atexit
.setnomips16
.entold_atexit
.type old_atexit, @function
old_atexit:
.frame $sp,0,$31 # vars= 0, regs= 0/0, args= 0, gp= 0
.mask 0x,0
.fmask 0x,0
.setnoreorder
.cpload $25
.setnomacro
lw $2,%got(__dso_handle)($28)
lw $25,%call16(__cxa_atexit)($28)
lw $6,0($2)
move$5,$0
.reloc 1f,R_MIPS_JALR,__cxa_atexit
1: jr $25
movz$6,$0,$2
.setmacro
.setreorder
.endold_atexit
.size old_atexit, .-old_atexit
.weak atexit
atexit = old_atexit
.weak __dso_handle
.ident "GCC: (GNU) 4.6.1"
old_atexit.os with GCC 4.5.3:
.file 1 "old_atexit.c"
.section .mdebug.abi32
.previous
.gnu_attribute 4, 3
.abicalls
.text
.align 2
.globl old_atexit
.setnomips16
.entold_atexit
.type old_atexit, @function
old_atexit:
.frame $sp,0,$31 # vars= 0, regs= 0/0, args= 0, gp= 0
.mask 0x,0
.fmask 0x,0
.setnoreorder
.cpload $25
.setnomacro
lw $2,%got(__dso_handle)($28)
beq $2,$0,$L2
move$6,$0
lw $6,0($2)
$L2:
lw $25,%call16(__cxa_atexit)($28)
.reloc 1f,R_MIPS_JALR,__cxa_atexit
1: jr $25
move$5,$0
.setmacro
.setreorder
.endold_atexit
.size old_atexit, .-old_atexit
.weak atexit
atexit = old_atexit
.weak __dso_handle
.ident "GCC: (GNU) 4.5.3"
### GCC versions (with gcc -v output) for target and host:
$
/mnt/sdb3/freetz/freetz-trunk_gcc-4.6.1/toolchain/build/mipsel_gcc-4.6.1_uClibc-0.9.32/mipsel-linux-uclibc/bin/mipsel-linux-uclibc-gcc
-v
Using built-in specs.
COLLECT_GCC=/mnt/sdb3/freetz/freetz-trunk_gcc-4.6.1/toolchain/build/mipsel_gcc-4.6.1_uClibc-0.9.32/mipsel-linux-uclibc/bin/mipsel-linux-uclibc-gcc
COLLECT_LTO_WRAPPER=/mnt/sdb3/freetz/freetz-trunk_gcc-4.6.1/toolchain/build/mipsel_gcc-4.6.1_uClibc-0.9.32/mipsel-linux-uclibc/bin/../libexec/gcc/mipsel-linux-uclibc/4.6.1/lto-wrapper
Target: mipsel-linux-uclibc
Configured with:
/mnt/sdb3/freetz/freetz-trunk/source/toolchain-mipsel_gcc-4.6.1_uClibc-0.9.32/gcc-4.6.1/configure
--prefix=/mnt/sdb3/freetz/freetz-trunk/toolchain/build/mipsel_gcc-4.6.1_uClibc-0.9.32/mipsel-linux-uclibc
--with-sysroot=/mnt/sdb3/
[Bug tree-optimization/49618] When building uClibc with GCC 4.6.1 old_atexit is miscompiled
http://gcc.gnu.org/bugzilla/show_bug.cgi?id=49618
--- Comment #3 from dileks 2011-07-03 15:39:02
UTC ---
Short feedback: -O{0,1,2,} works (beq* is there), -Os doesn't (beq deleted).
[ -O0 ]
beq$2,$0,$L2
[ -O1 ]
beql$2,$0,$L3
[ -O2 ]
beql$2,$0,$L2
[ -Os ]
[ empty ]
[Bug tree-optimization/49618] When building uClibc with GCC 4.6.1 old_atexit is miscompiled
http://gcc.gnu.org/bugzilla/show_bug.cgi?id=49618 --- Comment #4 from Sedat Dilek 2011-07-03 17:46:07 UTC --- With the Patch from [1] I could boot a firmware-image on my router. [1] http://freetz.org/attachment/ticket/1310/gcc-4.6.x-disable-if-conversion.patch
[Bug tree-optimization/49618] When building uClibc with GCC 4.6.1 old_atexit is miscompiled
http://gcc.gnu.org/bugzilla/show_bug.cgi?id=49618 --- Comment #8 from Sedat Dilek 2011-07-03 20:47:37 UTC --- Confirmed... Image built with gcc-4.6.x-disable-opt-flags.patch works also fine.
[Bug tree-optimization/49618] When building uClibc with GCC 4.6.1 old_atexit is miscompiled
http://gcc.gnu.org/bugzilla/show_bug.cgi?id=49618 --- Comment #13 from Sedat Dilek 2011-07-04 14:29:27 UTC --- Confirmed... gcc46-pr49618.patch solves the problem here, workaround patches from freetz project are no more necessary.
[Bug tree-optimization/49618] When building uClibc with GCC 4.6.1 old_atexit is miscompiled
http://gcc.gnu.org/bugzilla/show_bug.cgi?id=49618 --- Comment #14 from Sedat Dilek 2011-07-04 14:38:05 UTC --- Thank you gcc folks and all involved people to fix this bug.
[Bug tree-optimization/49618] When building uClibc with GCC 4.6.1 old_atexit is miscompiled
http://gcc.gnu.org/bugzilla/show_bug.cgi?id=49618 --- Comment #15 from Sedat Dilek 2011-07-04 16:47:25 UTC --- Created attachment 24677 --> http://gcc.gnu.org/bugzilla/attachment.cgi?id=24677 Simple testcase-script for gcc-PR49618
[Bug tree-optimization/49618] When building uClibc with GCC 4.6.1 old_atexit is miscompiled
http://gcc.gnu.org/bugzilla/show_bug.cgi?id=49618 --- Comment #17 from Sedat Dilek 2011-07-04 16:48:44 UTC --- Created attachment 24679 --> http://gcc.gnu.org/bugzilla/attachment.cgi?id=24679 old_atexit.os_OptLevel-O1
[Bug tree-optimization/49618] When building uClibc with GCC 4.6.1 old_atexit is miscompiled
http://gcc.gnu.org/bugzilla/show_bug.cgi?id=49618 --- Comment #20 from Sedat Dilek 2011-07-04 16:50:10 UTC --- Created attachment 24682 --> http://gcc.gnu.org/bugzilla/attachment.cgi?id=24682 old_atexit.os_OptLevel-Os
[Bug tree-optimization/49618] When building uClibc with GCC 4.6.1 old_atexit is miscompiled
http://gcc.gnu.org/bugzilla/show_bug.cgi?id=49618 --- Comment #18 from Sedat Dilek 2011-07-04 16:49:13 UTC --- Created attachment 24680 --> http://gcc.gnu.org/bugzilla/attachment.cgi?id=24680 old_atexit.os_OptLevel-O2
[Bug tree-optimization/49618] When building uClibc with GCC 4.6.1 old_atexit is miscompiled
http://gcc.gnu.org/bugzilla/show_bug.cgi?id=49618 --- Comment #19 from Sedat Dilek 2011-07-04 16:49:44 UTC --- Created attachment 24681 --> http://gcc.gnu.org/bugzilla/attachment.cgi?id=24681 old_atexit.os_OptLevel-O3
[Bug tree-optimization/49618] When building uClibc with GCC 4.6.1 old_atexit is miscompiled
http://gcc.gnu.org/bugzilla/show_bug.cgi?id=49618 --- Comment #16 from Sedat Dilek 2011-07-04 16:48:15 UTC --- Created attachment 24678 --> http://gcc.gnu.org/bugzilla/attachment.cgi?id=24678 old_atexit.os_OptLevel-O0
