[poky] BeagleBoard using GCC 4.6.0

[Darren Hart]

On 06/09/2011 04:32 AM, Gary Thomas wrote:
> On 06/09/2011 05:20 AM, Gary Thomas wrote:
>> On 06/09/2011 12:48 AM, Khem Raj wrote:
>>> On 06/08/2011 07:24 PM, Gary Thomas wrote:
>>>> On 06/08/2011 06:42 PM, Khem Raj wrote:
>>>>> On 06/08/2011 11:20 AM, Saul Wold wrote:
>>>>>> On 06/08/2011 10:43 AM, Gary Thomas wrote:
>>>>>>> Now that Poky is using GCC 4.6.0, has anyone actually checked the
>>>>>>> operation on the BeagleBoard? I suspect that you'll find that the
>>>>>>> EHCI USB does not work. I can't really check here as I don't trust
>>>>>>> the EHCI on my BeagleBoard rev C3 (not xM).
>>>>>>>
>>>>>>> That said, I've tried this new compiler (previously reported) on
>>>>>>> my own OMAP/3530 board which uses the same 2.6.37 kernel (just not
>>>>>>> the linux-yocto version). When I build my kernel with GCC 4.5.2,
>>>>>>> it works perfectly, but fails with GCC 4.6.0 (I trust the hardware).
>>>>>>>
>>>>>>> I've isolated it down to at least the function 'ehci_hub_control'
>>>>>>> (but I suspect the problem is more fundamental). Comparing the code
>>>>>>> generated by the two compilers with the same source tree, this function
>>>>>>> is dramatically different. I can see why it's failing, I just don't
>>>>>>> know why the compiler is doing what it's doing.
>>>>>>>
>>>>>>
>>>>>> Gary,
>>>>>>
>>>>>> I am not sure if Khem Raj is on this list, so I am forwarding it to him,
>>>>>> he might have some insight.
>>>>>>
>>>>>> Sau!
>>>>>>
>>>>>>
>>>>>>> The lines at drivers/usb/host/ehci-hub.c:841
>>>>>>> case GetPortStatus:
>>>>>>> if (!wIndex || wIndex > ports)
>>>>>>> goto error;
>>>>>>> wIndex--;
>>>>>>> status = 0;
>>>>>>> temp = ehci_readl(ehci, status_reg);
>>>>>>>
>>>>>>> are being compiled very differently.
>>>>>>>
>>>>>>> With GCC 4.5.2:
>>>>>>> 0xc0229810 <ehci_hub_control+672>: cmp r3, #0 ; 0x0
>>>>>>> 0xc0229814 <ehci_hub_control+676>: beq 0xc0229df8
>>>>>>> <ehci_hub_control+2184>
>>>>>>> 0xc0229818 <ehci_hub_control+680>: cmp r3, r0
>>>>>>> 0xc022981c <ehci_hub_control+684>: bgt 0xc0229df8
>>>>>>> <ehci_hub_control+2184>
>>>>>>> 0xc0229820 <ehci_hub_control+688>: sub r8, r3, #1 ; 0x1
>>>>>>> 0xc0229824 <ehci_hub_control+692>: ldr r5, [r7, #4]
>>>>>>> 0xc0229828 <ehci_hub_control+696>: uxth r8, r8
>>>>>>> 0xc022982c <ehci_hub_control+700>: dmb sy
>>>>>>>
>>>>>>> With GCC 4.6.0:
>>>>>>> 0xc0221630 <ehci_hub_control+808>: cmp r3, #0 ; 0x0
>>>>>>> 0xc0221634 <ehci_hub_control+812>: beq 0xc0221d7c
>>>>>>> <ehci_hub_control+2676>
>>>>>>> 0xc0221638 <ehci_hub_control+816>: cmp r3, r0
>>>>>>> 0xc022163c <ehci_hub_control+820>: bgt 0xc0221d7c
>>>>>>> <ehci_hub_control+2676>
>>>>>>> 0xc0221640 <ehci_hub_control+824>: sub r7, r3, #1 ; 0x1
>>>>>>> 0xc0221644 <ehci_hub_control+828>: add r3, r11, #16 ; 0x10
>>>>>>> 0xc0221648 <ehci_hub_control+832>: add r3, r8, r3, lsl #2
>>>>>>> 0xc022164c <ehci_hub_control+836>: uxth r7, r7
>>>>>>> 0xc0221650 <ehci_hub_control+840>: ldrb r5, [r3, #5]
>>>>>>> 0xc0221654 <ehci_hub_control+844>: ldrb r2, [r3, #4]
>>>>>>> 0xc0221658 <ehci_hub_control+848>: orr r5, r2, r5, lsl #8
>>>>>>> 0xc022165c <ehci_hub_control+852>: ldrb r2, [r3, #6]
>>>>>>> 0xc0221660 <ehci_hub_control+856>: ldrb r3, [r3, #7]
>>>>>>> 0xc0221664 <ehci_hub_control+860>: orr r5, r5, r2, lsl #16
>>>>>>> 0xc0221668 <ehci_hub_control+864>: orr r5, r5, r3, lsl #24
>>>>>>> 0xc022166c <ehci_hub_control+868>: dmb sy
>>>>>>>
>>>>>>> As you can see, the old compiler accesses the ehci status register
>>>>>>> in a single access, the new compiler dances around and makes multiple
>>>>>>> accesses, which in the end get very incorrect data (I think that this
>>>>>>> register is like many which clear bits on reads).
>>>>>>>
>>>>>>> Any ideas where I can go with this? I'm really trying to keep up with
>>>>>>> Poky/Yocto, but this move to GCC 4.6.0 has broken my ARM targets :-(
>>>>>>> I do have PowerPC targets as well - they seem fine (from limited
>>>>>>> testing)
>>>>>>> with the new compiler.
>>>>>>>
>>>>>>> Note: if you want to see the whole function disassembly, look at
>>>>>>> http://www.mlbassoc.com/poky/ehci_hub_control-disassembly-gcc-4.5.2
>>>>>>> http://www.mlbassoc.com/poky/ehci_hub_control-disassembly-gcc-4.6.0
>>>>>>>
>>>>>>> Thanks
>>>>>>>
>>>>>
>>>>> OK. thanks for sending the sample code.
>>>>> Can you add -fstrict-volatile-bitfields option to CFLAGS while
>>>>> compiling this file or even the kernel for test sake ?
>>>>> and see if the problem goes away ?
>>>>
>>>> Sorry, this doesn't seem to make any difference in the generated code.
>>>> Hardly surprising as the register in question is a simple u32, not a
>>>> bitfield.
>>>>
>>>
>>> I reduced the testcase a bit further and it seems the bug is similar to http://gcc.gnu.org/bugzilla/PR45819
>>>
>>> but a bit different that in this case -fstrict-volatile-bitfields does not help
>>>
>>> a fix for this particular problem is
>>>
>>> diff --git a/include/linux/usb/ehci_def.h b/include/linux/usb/ehci_def.h
>>> index 6563802..b8c1833 100644
>>> --- a/include/linux/usb/ehci_def.h
>>> +++ b/include/linux/usb/ehci_def.h
>>> @@ -194,7 +194,7 @@ struct ehci_dbg_port {
>>> u32 data47;
>>> u32 address;
>>> #define DBGP_EPADDR(dev, ep) (((dev)<<8)|(ep))
>>> -} __attribute__ ((packed));
>>> +} __attribute__ ((packed,aligned(__alignof__(int))));
>>>
>>> #ifdef CONFIG_EARLY_PRINTK_DBGP
>>> #include <linux/init.h>
>>>
>>>
>>> since this particular structure all elements are integers but there might be more packed structures which have say char or short ints
>>> and are not necessarily ordered in the descending order of alignment
>>>
>>> Try it out and see if it helps
>>
>> Yes, this does fix the problem :-) The EHCI USB port on the OMAP
>> is working again.
> 
> Note: it also works properly if the __attribute__ ((packed)) is simply removed.
> Looking at the layout of this structure, I don't see what that attribute is
> trying to accomplish - every item within the structure is u32[], so IMO there
> is nothing to pack.  Perhaps this observation can help the GCC folks figure
> out why things are getting confused.

I tried each of these and neither worked. I confirmed the kernel was
indeed different after each build and that gcc 4.6.0 was being used (by
adding CC --version to the kernel do_compile and checking it in the log).

Gary, are you certain that you were building with 4.6.0 ? I've found it
to be a bit tricky to switch back and forth between compilers.

-- 
Darren Hart
Intel Open Source Technology Center
Yocto Project - Linux Kernel