Re: [PATCH dwarves v2 0/3] pahole: Fix arm64 argument register allocation

["Alexei Starovoitov" <alexei.starovoitov@gmail.com>]

On Thu Sep 24, 2026 at 4:15 PM UTC, Alan Maguire wrote:
> On 17/09/2026 17:40, Yonghong Song wrote:
>> 
>> The arm64 ABI starts an argument twice the register size on an even-numbered
>> argument register, leaving a hole when the next free register is an odd one:
>> 
>>   u64 f(u64 a, __int128 v, u64 b);
>> 
>> passes a in x0, v in x2:x3 -- skipping x1 -- and b in x4.  pahole did not
>> model the hole, so b looked like it was in an unexpected register and the
>> whole function was dropped from BTF.  Patch 1 skips the register the ABI
>> leaves behind, patch 3 adds a test for it.  Only arm64 has this rule; the
>> other architectures are unaffected.
>> 
>> Patch 2 is a small follow-up.  Floating point and vector arguments are
>> allocated from a register bank of their own -- v0-v7 on arm64, xmm0-xmm7 on
>> x86-64 -- so a wide one never occupies a general purpose register pair and
>> must not trigger the even-register rule either.  It does not change the BTF
>> emitted today, because such an argument still advances the general purpose
>> register index by its own size, so it comes with no test of its own; it
>> keeps the alignment rule scoped to the bank it applies to.
>>
>
> one small issue here; we record alignment for structs from DW_AT_align
> in associated tags; we should ensure that we utilize that alignment
> where available. However it's a bit broken in that we don't appear
> to record alignment for typedefs, just structs.
>
> I think we can tackle that with a followup though; unless anyone objects I'll
> land these patches later today.

Pls go ahead.
Also pls land LOC* things too, so you can post the rest of LOC* kernel bits.
So that everything lands and ready for the merge window in a couple weeks.

People will be traveling soon, so the sooner the better.