Re: [Xen-devel] [PATCH 02/34] xen: clang: Disable built-in assembler

[Julien Grall <julien.grall@xxxxxxxxxx>]

On 04/01/2014 04:28 PM, Ian Campbell wrote:
> On Tue, 2014-04-01 at 15:50 +0100, Julien Grall wrote:
>> On 04/01/2014 02:11 PM, Ian Campbell wrote:
>>> On Sat, 2014-03-29 at 22:55 +0000, Julien Grall wrote:
>>>> Hi Tim,
>>>>
>>>> On 27/03/14 18:01, Tim Deegan wrote:
>>>>>
>>>>> The patch below works for me (at least as far as building
>>>>> asm-offsets.h on x86) by wrapping everything in a string.  I did try
>>>>> just prefixing with '#' but clang 3.5 also strips the comments out.
>>>>> That seems unhelpful, since I know some people put comments in their
>>>>> inline assembler too. :(
>>>>
>>>> I'm able to build correctly x86 with your patch, and this patch (e.g #2) 
>>>> reverted.
>>>>
>>>> But for ARM ... it breaks in another place :(
>>>>
>>>> vfp.c:8:25: error: invalid operand for instruction
>>>>      v->arch.vfp.fpexc = READ_CP32(FPEXC);
>>>> <inline asm>:1:6: note: instantiated into assembly here
>>>>          mrc p10, 7, r1, c8, c0, 0;
>>>>              ^
>>>>
>>>> Coprocessor p10 (and p11) are used for Neon instruction are clang 
>>>> doesn't allow to use it directly. 
>>>> (http://lists.cs.uiuc.edu/pipermail/llvm-commits/Week-of-Mon-20131104/194170.html).
>>>>
>>>> Depends on which part of the ARM ARM you are reading, p10 and p11 should 
>>>> not be used directly with mrc/mcr instruction ... but gas accept it.
>>>>
>>>> I guess the best solution is to use directly the VFP instructions but it 
>>>> would mean to re-enable VFP at compile time in Xen (see 
>>>> http://www.gossamer-threads.com/lists/xen/devel/284653?do=post_view_threaded).
>>>
>>> Do the VFP instructions have different encodings? I thought this was an
>>> assembler mnemonic difference only.
>>
>> Strictly speaking it works only because vldm* as the same encoding as
>> ldc*. But the processor will decode the instruction as vldm* (and
>> execute as it is).
> 
> Are those the instructions in question though? The thing you quoted was
> about reading FPEXC which is an mrc cp instruction vs vmrs I think. They
> do indeed have identical encodings, so I'm not entirely sure what the
> big deal is about which assembler mnemonic gets used is, the process has
> no idea what we wrote in the source.

Oh sorry, I took one in random. Every VFP instructions has the same
issue. I will raise the bug to Clang. And see if why they choose this
solution.

-- 
Julien Grall

_______________________________________________
Xen-devel mailing list
Xen-devel@xxxxxxxxxxxxx
http://lists.xen.org/xen-devel