Re: gcc-4.4.x bug in ARM-EABI code generator/optimizer for Objective-C?

["Dr. H. Nikolaus Schaller" <hns@xxxxxxxxxxxxx>]

Solved (see below). Thanks for all the help.

Am 15.04.2012 um 19:14 schrieb Dr. H. Nikolaus Schaller:

> Anyone out there with a hint how to debug/solve this issue?
> 
> Tnx,
> Nikolaus
> 
> Am 09.04.2012 um 12:03 schrieb Dr. H. Nikolaus Schaller:
> 
>> Dear help,
>> I have built three cross-toolchains that run on Mac OS X and
>> cross-compile to i486, arm-eabi, mipsel. The goal is to make
>> them compatible to Debian Squeeze.
>> 
>> Therefore I have gcc-4.4.5, binutils-2.20.1, glibc-2.11.3, linux-headers-2.6.32
>> 
>> Everything works fine and I was even able to cross-compile
>> ARM-Linux-kernels (2.6.32 and 3.2).
>> 
>> Only one remaining issue is with Objective C.
>> 
>> As soon as I use the [super method] construct in a method, I
>> get this message from the assembler for each class using such
>> a super call in any method:
>> 
>> /var/folders/Q-/Q-r6LiEpG+S3E54UanorqU+++TM/-Tmp-//ccG6BCbh.s: Assembler messages:
>> /var/folders/Q-/Q-r6LiEpG+S3E54UanorqU+++TM/-Tmp-//ccG6BCbh.s:164: Error: symbol `_OBJC_CLASS_Test' is already defined
>> 
>> The really strange thing is that it is
>> * only on arm-linux-gnueabi target
>> * only for -O1 -O2 -O3 -Os but NOT for -O0
>> * by trying all -f options to control the optimizer I could not find a
>> single one that is responsible

I did not try -fsection-anchors.

>> 
>> The test code is very simple:
>> 
>> #define BUG 1
>> 
>> #include <objc/objc.h>
>> #include <objc/Object.h>
>> 
>> @interface Test : Object
>> @end
>> 
>> @implementation Test
>> 
>> - (id) init
>> {
>> #if BUG
>> 	[super init];
>> #endif
>> 	return nil;
>> }
>> 
>> @end
>> 
>> Compiling with -S option gives me the assembler sources. The i486 and the
>> mipsel code looks quite similar (except architecture specific codes and
>> variations), but the arm code is quite different.
>> 
>> It has indeed a second definition of _OBJC_CLASS_Test: and does define
>> additional symbols like .LANCHOR1 which is not seen in the unoptimized

ANCHOR was the important hint.

>> -O0 code.
>> 
>> I had also built gcc-4.4.7 and there is no difference.
>> 
>> So is this a known bug in the ARM-EABI code generator/optimizer?
>> Is there a workaround by giving some compiler options?

-fno-section-anchors

>> How can I further debug the issue to find a fix?

trial and error and afflatus.

It appears that my cross-toolchain enables -fsection-anchors by
default (no idea why),
while the 'apt-get install gobjc' compiler (gcc-4.4.5) on Debian/armel
rejects this option (for apparently good reasons) and tells:

gcc: GNU Objective C can't use -fsection-anchors


>> 
>> Thanks,
>> Nikolaus Schaller
>> 
>> arm for -O0:
>> 
>> <Test-arm-linux-gnueabi--O0.s>
>> 
>> generated assembler code (for -O1):
>> arm:
>> 
>> <Test-arm-linux-gnueabi--O1.s>
>> 
>> i486:
>> 
>> <Test-i486-linux-gnu--O1.s>
>> 
>> mipsel:
>> 
>> <Test-mipsel-linux-gnu--O1.s>
>> 
>