Re: GCC 4.9 considered evil

[Carlos Rafael Giani <dv@pseudoterminal.org>]

[-- Attachment #1: Type: text/plain, Size: 5277 bytes --]

On 08/14/2014 10:22 AM, Peter A. Bigot wrote:
> On 08/14/2014 03:12 AM, Carlos Rafael Giani wrote:
>> On 08/14/2014 02:46 AM, Khem Raj wrote:
>>>
>>>
>>> On Wednesday, August 13, 2014, Otavio Salvador 
>>> <otavio@ossystems.com.br <mailto:otavio@ossystems.com.br>> wrote:
>>>
>>>     On Wed, Aug 13, 2014 at 6:24 PM, Gary Thomas <gary@mlbassoc.com
>>>     <javascript:;>> wrote:
>>>     > I've found that the latest GCC doesn't work very well, at
>>>     > least not on ARM (and obviously other architectures as well [1])
>>>     > When I build Google Chromium browser for my i.MX boards using
>>>     > GCC-4.9.x, no pages can be rendered - massive bloodshed and
>>>     > failures are shown on the console.  If I use the older GCC 4.8.2,
>>>     > everything else the same, all is well.
>>>     >
>>>     > Here's my configuration:
>>>     > BB_VERSION        = "1.23.1"
>>>     > BUILD_SYS         = "x86_64-linux"
>>>     > NATIVELSBSTRING   = "Ubuntu-13.10"
>>>     > TARGET_SYS        = "arm-amltd-linux-gnueabi"
>>>     > MACHINE           = "teton-p0382"
>>>     > DISTRO            = "amltd"
>>>     > DISTRO_VERSION    = "1.6+snapshot-20140812"
>>>     > TUNE_FEATURES     = "arm armv7a vfp neon callconvention-hard
>>>     cortexa9"
>>>     > TARGET_FPU        = "vfp-neon"
>>>     > meta              =
>>>     "master:86afd7eb7c679eb065706137f28f44248f3fbc5a"
>>>     > meta-amltd        =
>>>     "master:899529b4184dc9e3e291e7fcdbbf157233db056d"
>>>     > meta-teton-imx6-p0382 =
>>>     "master:9188e2f8fafd203c95dcd7a3c79cb38a1568e9b5"
>>>     > meta-fsl-arm      =
>>>     "master:b0cf5c78c5f98c5977ae556d331e7495648f154c"
>>>     > meta-fsl-arm-extra =
>>>     "master:12e560967b7136222c325d11633295fe3a0c701c"
>>>     > meta-browser      =
>>>     "master:da93c8e386133a15eff1414d9307c8f2c7a44787"
>>>     >
>>>     > Should this be filed as a bug?  I don't have much data other
>>>     than it
>>>     > simply breaks (and chrome is not the easiest thing to debug!).
>>>      Other
>>>     > applications seem OK, but I am loathe to trust it...
>>>     >
>>>     > I'm going to hold onto GCC-4.8.x for my $DISTRO at least (and
>>>     I hope
>>>     > it doesn't vanish like 4.7.x did too quickly).
>>>     >
>>>     > [1]
>>>     >
>>>     http://slashdot.org/story/14/07/27/1838219/linus-torvalds-gcc-490-seems-to-be-terminally-broken
>>>     >
>>>     > Note: I've also tried this on qemux86 (a totally different
>>>     > architecture) and chrome bombs just as badly!
>>>
>>>     I confirm that GCC 4.9 does NOT work for Chromium 35. At this
>>>     moment I
>>>     am not aware of any fix for it.
>>>
>>>
>>> Again Its a broad brush statement. Something concrete is needed if 
>>> some.action is to taken
>>>
>>>
>>>     IIRC the Chromium 37 works though.
>>>
>>>
>>> Well then its less chance.that someone will fix 35
>>>
>>>     --
>>>     Otavio Salvador                             O.S. Systems
>>>     http://www.ossystems.com.br http://code.ossystems.com.br
>>>     Mobile: +55 (53) 9981-7854            Mobile: +1 (347) 903-9750
>>>     --
>>>     _______________________________________________
>>>     Openembedded-core mailing list
>>>     Openembedded-core@lists.openembedded.org <javascript:;>
>>>     http://lists.openembedded.org/mailman/listinfo/openembedded-core
>>>
>>>
>>>
>>
>>
>> The problem is that narrowing down is very difficult, since the 
>> problems manifest themselves as seemingly random stack corruptions. I 
>> have tried to dig into it, but got nowhere. Pointers suddenly became 
>> null for no apparent reason, or were corrupted, free() calls failed, 
>> values on the stack suddenly changed without being modified by the 
>> code etc.
>>
>> I wouldn't rule out that Linus Torvald's find isn't the cause here. 
>> Look at this part of his message:
>>
>> "The x86-64 ABI specifies a 128-byte red-zone under the stack 
>> pointer, and this is ok by that limit. It looks like it's illegal 
>> (136 > 128), but the fact is, we've had four "pushq"s to update %rsp 
>> since loading the frame pointer, so it's just *barely* legal with the 
>> red-zoning."
>>
>> Perhaps gcc is pushing further outside of the red zone bounds, thus 
>> causing the problems. No idea how to check for that at the moment though.
>
> Simplest would be to apply the upstream fix to Yocto's gcc and see if 
> that helps.  You'd want commit 
> 556537c4ad0df4cbebb74197bb2bdea75cf5dd35 from 
> git://gcc.gnu.org/git/gcc.git.  (The patch went into gcc-4_9-branch 
> one day after 4.9.1 was released.)
>
> I'd add it to the current set but ATT Khem and I both have pending 
> patches that touch the same gcc files, so it'd just increase the 
> conflicts.
>
> Peter

Are you sure this is the right patch? Commit message:

[PATCH] 2014-07-17  Richard Biener  <rguenther@suse.de>

         PR rtl-optimization/61801
         * sched-deps.c (sched_analyze_2): For ASM_OPERANDS and
         ASM_INPUT don't set reg_pending_barrier if it appears in a
         debug-insn.

It does not apply cleanly on gcc in either master or master-next . 
Patching ChangeLog fails. I will try to use it with the ChangeLog patch 
part omitted.

[-- Attachment #2: Type: text/html, Size: 9296 bytes --]