[llvm-dev] Bug or incorrect use of inline asm?

[Eric Christopher via llvm-dev]

On Sat, Aug 5, 2017 at 5:28 AM Johan Engelen <jbc.engelen at gmail.com> wrote:

> On Fri, Aug 4, 2017 at 11:55 AM, Johan Engelen <jbc.engelen at gmail.com>
> wrote:
>
>> On Thu, Aug 3, 2017 at 6:19 PM, Tim Northover <t.p.northover at gmail.com>
>> wrote:
>>
>>> 2017-08-03 8:58 GMT-07:00 Johan Engelen via llvm-dev <
>>> llvm-dev at lists.llvm.org>:
>>> > The error is gone after removing (or reducing) the alignment of `%a`.
>>> This
>>> > makes me believe that our inline asm syntax is correct to add an
>>> offset to a
>>> > pointer: " 4+$0 ".
>>>
>>> The AT&T syntax for a displaced address doesn't have the '+'; it's
>>> just "4(%rsp)" so you should change the IR to "4$0".
>>>
>>> LLVM appears to be optimizing the addressing mode, and the optimizer
>>> is being more lenient on this detail so you're getting lucky when the
>>> +4 can be combined with an existing %rsp offset.
>>
>>
>> Things break down when part of the offset is a linktime constant:
>>
>> ...
>>
>
>> So depending on the pointer argument, I need to write a + or without +,
>> complicating automatic codegen.
>>
>
> The problem is worse. The following code is accepted depending on PIC or
> non-PIC compilation:
>
> ```
> source_filename = "asanasm.d"
> target datalayout = "e-m:o-i64:64-f80:128-n8:16:32:64-S128"
> target triple = "x86_64-linux-gnu"
>
> @_D7asanasm9constantsG2k = global [2 x i32] [i32 66051, i32 66051]
>
> define void @_D7asanasm8offconstFZv() {
>   call void asm sideeffect "movdqa 4$0, %xmm7", "*m,~{xmm7}"([2 x i32]*
> @_D7asanasm9constantsG2k)
>   ret void
> }
> ```
>
> I get no errors with PIC:
> > llc test.ll -o test.s -O0 -relocation-model=pic
> The asm output:
>   movq    _D7asanasm9constantsG2k at GOTPCREL(%rip), %rax
>   #InlineASM
>   movdqa  4(%rax), %xmm7
>
>
> But without PIC:
> > llc test.ll -o test.s -O0 -relocation-model=static
> <inline asm>:1:10: error: unexpected token in argument list
>         movdqa 4_D7asanasm9constantsG2k(%rip), %xmm7
>                 ^
>
> So it appears that applying an offset to a pointer variable will break
> using either "4+$0" or "4$0", and you have to be lucky to guess correctly
> what's going to happen with the pointer argument passed into the inline asm
> call. (my problem would be solved when "4+$0" would always be accepted)
>
> I now feel this is a bug in LLVM, do you agree?
> Unfortunately, I do not know how to work around the current problem.
>

Oh I haven't disagreed with you at all. Sorry I wasn't clear :)

Mostly the x86 asm parser needs work to support this. I -think- it's just a
matter of typing, but I haven't really tried to fix it.
You might be able to use a register constraint to put the address into a
register and go from there. It's not ideal, but might work.

-eric



>
> Thanks,
>   Johan
>
>
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.llvm.org/pipermail/llvm-dev/attachments/20170807/e59a0eef/attachment.html>