[PATCH] Add default parameter to MCSymbol accessors to control IsUsed
Vedant Kumar via llvm-commits
llvm-commits at lists.llvm.org
Tue Aug 25 10:41:03 PDT 2015
Ahh, ok. Here is the NFC part (no problems with check-all on x86+arm).
I'll play around with the second part some more.
thanks!
vedant
-------------- next part --------------
A non-text attachment was scrubbed...
Name: MCSymbol-patch1-NFC.patch
Type: application/octet-stream
Size: 2856 bytes
Desc: not available
URL: <http://lists.llvm.org/pipermail/llvm-commits/attachments/20150825/28cf1ac2/attachment.obj>
-------------- next part --------------
> On Aug 25, 2015, at 6:36 AM, Rafael EspĂndola <rafael.espindola at gmail.com> wrote:
>
> With your patch we produce
>
> 000000000000 00010000000a R_X86_64_32 0000000000000000 zed + 0
> 000000000004 00010000000a R_X86_64_32 0000000000000000 zed + 0
>
> gas produces
>
> 000000000000 00040000000a R_X86_64_32 0000000000000000 bar + 0
> 000000000004 00050000000a R_X86_64_32 0000000000000000 zed + 0
>
> So we go from producing an error to producing an output that is
> different from gas, which is bad thing.
>
> The part about not setting IsUsed unconditionally is still nice, so if
> you don't mind finishing up a NFC patch that just refines what is a
> use that would be great.
>
> Cheers,
> Rafael
>
>
>
> On 25 August 2015 at 09:26, Rafael EspĂndola <rafael.espindola at gmail.com> wrote:
>> Taking a look.
>>
>> On 21 August 2015 at 17:34, Vedant Kumar <vsk at apple.com> wrote:
>>> Ping
>>>
>>> We were discussing "invalid reassignment to non-absolute variable" and whether we still need it around.
>>>
>>> thanks
>>> vedant
>>>
>>>> On Aug 13, 2015, at 9:31 AM, Vedant Kumar <vsk at apple.com> wrote:
>>>>
>>>> Hi Rafael,
>>>>
>>>> Sorry for the delay on this.
>>>>
>>>>> I think that dates from the early days of MC. My understanding is that
>>>>> it was added just as a safety so that we have to reason about code
>>>>> like
>>>>>
>>>>> foo = bar
>>>>> .long foo
>>>>> foo = zed
>>>>> .long foo
>>>>>
>>>>> Should that produce two relocations with bar, two with foo or one each?
>>>>
>>>> Here is what GAS does:
>>>>
>>>> SYMBOL TABLE:
>>>> 00000000 l d .text 00000000 .text
>>>> 00000000 l d .data 00000000 .data
>>>> 00000000 l d .bss 00000000 .bss
>>>> 00000000 *UND* 00000000 bar
>>>> 00000000 *UND* 00000000 zed
>>>>
>>>> RELOCATION RECORDS FOR [.text]:
>>>> OFFSET TYPE VALUE
>>>> 00000000 R_386_32 bar
>>>> 00000004 R_386_32 zed
>>>>
>>>> Here is a different test that initially triggered this issue:
>>>>
>>>>> foo = bar
>>>>> .long foo
>>>>> foo = bar
>>>>> .long foo
>>>>
>>>>
>>>> And here is GAS's output:
>>>>
>>>> RELOCATION RECORDS FOR [.text]:
>>>> OFFSET TYPE VALUE
>>>> 00000000 R_386_32 bar
>>>> 00000004 R_386_32 bar
>>>>
>>>> With the patch I suggested, both of these examples just work with llvm-mc. Here's what the relocation table looks like on Darwin for comparison:
>>>>
>>>> Relocation information (__TEXT,__text) 2 entries
>>>> address pcrel length extern type scattered symbolnum/value
>>>> 00000004 0 2 1 0 0 1
>>>> 00000000 0 2 1 0 0 1
>>>>
>>>> This looks sane to me. I've updated the patch to get rid of the "invalid reassignment to non-absolute variable" error and fixed the test cases. Could you review it for any issues?
>>>>
>>>> thanks
>>>> vedant
>>>>
>>>> <MCSymbol-patch1-NFC.patch><MCSymbol-patch2.patch>
>>>
More information about the llvm-commits
mailing list