[PATCH 2/3] Add AMDGPU generic targets

[Simon Marchi]

On 11/7/25 8:37 AM, Jan Beulich wrote:
> On 31.10.2025 18:54, Luis Machado wrote:
>> --- a/bfd/archures.c
>> +++ b/bfd/archures.c
>> @@ -560,19 +560,24 @@ DESCRIPTION
>>  .#define bfd_mach_loongarch64	2
>>  .  bfd_arch_amdgcn,     {* AMDGCN *}
>>  .#define bfd_mach_amdgcn_unknown 0x000
>> +.#define bfd_mach_amdgcn_gfx9_generic    0x051
>>  .#define bfd_mach_amdgcn_gfx900          0x02c
>>  .#define bfd_mach_amdgcn_gfx904          0x02e
>>  .#define bfd_mach_amdgcn_gfx906          0x02f
>>  .#define bfd_mach_amdgcn_gfx908          0x030
>>  .#define bfd_mach_amdgcn_gfx90a          0x03f
>> +.#define bfd_mach_amdgcn_gfx9_4_generic  0x05f
>>  .#define bfd_mach_amdgcn_gfx942          0x04c
>>  .#define bfd_mach_amdgcn_gfx950          0x04f
>> +.#define bfd_mach_amdgcn_gfx10_1_generic 0x052
>>  .#define bfd_mach_amdgcn_gfx1010         0x033
>>  .#define bfd_mach_amdgcn_gfx1011         0x034
>>  .#define bfd_mach_amdgcn_gfx1012         0x035
>> +.#define bfd_mach_amdgcn_gfx10_3_generic 0x053
>>  .#define bfd_mach_amdgcn_gfx1030         0x036
>>  .#define bfd_mach_amdgcn_gfx1031         0x037
>>  .#define bfd_mach_amdgcn_gfx1032         0x038
>> +.#define bfd_mach_amdgcn_gfx11_generic   0x054
>>  .#define bfd_mach_amdgcn_gfx1100         0x041
>>  .#define bfd_mach_amdgcn_gfx1101         0x046
>>  .#define bfd_mach_amdgcn_gfx1102         0x047
>> @@ -580,6 +585,7 @@ DESCRIPTION
>>  .#define bfd_mach_amdgcn_gfx1151         0x04a
>>  .#define bfd_mach_amdgcn_gfx1152         0x055
>>  .#define bfd_mach_amdgcn_gfx1153         0x058
>> +.#define bfd_mach_amdgcn_gfx12_generic   0x059
>>  .#define bfd_mach_amdgcn_gfx1200         0x048
>>  .#define bfd_mach_amdgcn_gfx1201         0x04e
> 

Hi Jan,

I will pick up this patch from Luis.

> Here I'm getting curious: Where are these numbers coming from? I'm asking in
> particular in order to understand whether there's a risk of future collisions.

The canonical source for these numbers is:

https://llvm.org/docs/AMDGPUUsage.html#amdgpu-ef-amdgpu-mach-table

> I further wonder by what criteria _generic entries are added (or not).

Generic processor versioning is described here:

https://llvm.org/docs/AMDGPUUsage.html#generic-processor-versioning
https://llvm.org/docs/AMDGPUUsage.html#amdgpu-generic-processor-table

We just mirror what is there.  Whether a generic processor variant
exists is an engineering decision at AMD.  As far as I understand, a
binary compiled for gfx12_generic, for example, will run on any gfx120*
card.

I will add those links as comments in the code (if it is allowed to have
external links), I think it will be useful for others who wonder where
these numbers come from in the future.

> 
>> --- a/include/elf/amdgpu.h
>> +++ b/include/elf/amdgpu.h
>> @@ -79,6 +79,12 @@
>>  #define EF_AMDGPU_MACH_AMDGCN_GFX942  0x04c
>>  #define EF_AMDGPU_MACH_AMDGCN_GFX1201 0x04e
>>  #define EF_AMDGPU_MACH_AMDGCN_GFX950  0x04f
>> +#define EF_AMDGPU_MACH_AMDGCN_GFX9_GENERIC 0x051
>> +#define EF_AMDGPU_MACH_AMDGCN_GFX10_1_GENERIC 0x052
>> +#define EF_AMDGPU_MACH_AMDGCN_GFX10_3_GENERIC 0x053
>> +#define EF_AMDGPU_MACH_AMDGCN_GFX11_GENERIC 0x054
>> +#define EF_AMDGPU_MACH_AMDGCN_GFX12_GENERIC 0x059
>> +#define EF_AMDGPU_MACH_AMDGCN_GFX9_4_GENERIC 0x05f
> 
> Why is it that these constants aren't used in bfd/archures.c?

Good question, if we want to switch to that I'll make a patch on top of
all this.

Simon