llvm dev - Mar 2013 - [LLVMdev] Problems with 64-bit register operands of inline asm on ARM

Jim Grosbach <grosbach at apple.com> writes:
> On Mar 13, 2013, at 11:01 AM, Renato Golin <renato.golin at
linaro.org>
> wrote:
>
>> On 13 March 2013 17:57, Jim Grosbach <grosbach at apple.com>
wrote:
>>> It seems to me that LLVM doesn’t parse the inline asm body. It just
>>> checks the constraints, (ie. Input/output interface). During ASM
>>> writing, it then binding those constraints to placeholders like %0,
>>> %1.
>> This is correct.
>> 
>> Ok, so maybe checking all possible ways to require paired registers
>> is not such a bad idea after all.
>> 
>
> The constraints are the right way to do it. There shouldn't be any
> magic beyond that.
Since there is no special operand constraint for a register pair, there
is no way to tell at that level.

GCC has (implicitly) defined 64-bit register operands as residing in
even/odd pairs, thus leaving inline asm free to make all manner of
assumptions based on this.  The only way I see to guarantee
compatibility is to mimic the gcc behaviour here.  It may be slightly
suboptimal in a few cases, but it's the safe choice.

-- 
Måns Rullgård
mans at mansr.com

On Mar 13, 2013, at 11:21 AM, Måns Rullgård <mans at mansr.com> wrote:
> Jim Grosbach <grosbach at apple.com> writes:
> 
>> On Mar 13, 2013, at 11:01 AM, Renato Golin <renato.golin at
linaro.org>
>> wrote:
>> 
>>> On 13 March 2013 17:57, Jim Grosbach <grosbach at apple.com>
wrote:
>>>> It seems to me that LLVM doesn’t parse the inline asm body. It
just
>>>> checks the constraints, (ie. Input/output interface). During
ASM
>>>> writing, it then binding those constraints to placeholders like
%0,
>>>> %1.
>>> This is correct.
>>> 
>>> Ok, so maybe checking all possible ways to require paired registers
>>> is not such a bad idea after all.
>>> 
>> 
>> The constraints are the right way to do it. There shouldn't be any
>> magic beyond that.
> 
> Since there is no special operand constraint for a register pair, there
> is no way to tell at that level.
> 
> GCC has (implicitly) defined 64-bit register operands as residing in
> even/odd pairs, thus leaving inline asm free to make all manner of
> assumptions based on this.  The only way I see to guarantee
> compatibility is to mimic the gcc behaviour here.  It may be slightly
> suboptimal in a few cases, but it's the safe choice.
Sure, that's fine for ARM mode. No realistic other option there. So long as
Thumb2 code can get the more expressive syntax for the more relaxed regalloc
availability, it's all good. This basically falls into "using the
constraints to figure it out."

0Jim
-------------- next part --------------
An HTML attachment was scrubbed...
URL:
<http://lists.llvm.org/pipermail/llvm-dev/attachments/20130313/3e934c0a/attachment.html>

Jim Grosbach <grosbach at apple.com> writes:
> On Mar 13, 2013, at 11:21 AM, Måns Rullgård <mans at mansr.com>
wrote:
>
>> Jim Grosbach <grosbach at apple.com> writes:
>> 
>>> On Mar 13, 2013, at 11:01 AM, Renato Golin <renato.golin at
linaro.org>
>>> wrote:
>>> 
>>>> On 13 March 2013 17:57, Jim Grosbach <grosbach at
apple.com> wrote:
>>>>> It seems to me that LLVM doesn’t parse the inline asm body.
It just
>>>>> checks the constraints, (ie. Input/output interface).
During ASM
>>>>> writing, it then binding those constraints to placeholders
like %0,
>>>>> %1.
>>>> This is correct.
>>>> 
>>>> Ok, so maybe checking all possible ways to require paired
registers
>>>> is not such a bad idea after all.
>>>> 
>>> 
>>> The constraints are the right way to do it. There shouldn't be
any
>>> magic beyond that.
>> 
>> Since there is no special operand constraint for a register pair, there
>> is no way to tell at that level.
>> 
>> GCC has (implicitly) defined 64-bit register operands as residing in
>> even/odd pairs, thus leaving inline asm free to make all manner of
>> assumptions based on this.  The only way I see to guarantee
>> compatibility is to mimic the gcc behaviour here.  It may be slightly
>> suboptimal in a few cases, but it's the safe choice.
>
> Sure, that's fine for ARM mode. No realistic other option there. So
> long as Thumb2 code can get the more expressive syntax for the more
> relaxed regalloc availability, it's all good. This basically falls
> into "using the constraints to figure it out."
So let's at least make this pair allocation unconditional in ARM mode.
A lot of existing inline asm doesn't work in Thumb mode anyway, so if
that fails, it's less of an issue.

-- 
Måns Rullgård
mans at mansr.com