m4: add gcc support

[becker33]

For gcc
m1 --> armv8.4-a (available from 8.0)
m2 --> armv8.6-a (available from 10.1)
m3 == m2
m4 --> armv8.6-a+sme2

[alalazo]

Can you confirm you tried a build with Spack, before I merge this? I don't have any access to m4

[becker33]

I don't have access either, going to ask @haampie to test. This came up from a conversation he and I were having on slack.

[alalazo]

It seems gcc 15 added support up to -mcpu=apple-m3, see https://gcc.gnu.org/gcc-15/changes.html

[alalazo]

A quick search gave me this. Wondering if adding +crc+crypto to our baseline gives improved performance.

Apple Chip	Architecture	First GCC to support -mcpu	Generic Fallback Flag (For older GCC)	First GCC for Fallback
M1	Armv8.5-A	GCC 15.1 (apple-m1)	-march=armv8.5-a+crc+crypto	GCC 9.1
M2	Armv8.6-A	GCC 15.1 (apple-m2)	-march=armv8.6-a+crc+crypto	GCC 10.1
M3	Armv8.6-A	GCC 15.1 (apple-m3)	-march=armv8.6-a+crc+crypto	GCC 10.1
M4	Armv9.2-A	TBD (apple-m4*)	-march=armv9.2-a+nosve	GCC 14.1

[alalazo]

See the table for another suggestion. For what is worth, it seems that using the armv8.6 flag as above is needed for GCC older than v14

[becker33]

I think based on the table in the gcc manual pages that these two descriptions are identical, except that the one based on 8.6-a makes explicit that sme2 is included even though sve is not.

See

arch value | Architecture | Includes by default
-- | -- | --
‘armv8-a’   | Armv8-A   | ‘+fp’, ‘+simd’
‘armv8.1-a’ | Armv8.1-A | ‘armv8-a’, ‘+crc’, ‘+lse’, ‘+rdma’
‘armv8.2-a’ | Armv8.2-A | ‘armv8.1-a’
‘armv8.3-a’ | Armv8.3-A | ‘armv8.2-a’, ‘+pauth’
‘armv8.4-a’ | Armv8.4-A | ‘armv8.3-a’, ‘+flagm’, ‘+fp16fml’, ‘+dotprod’
‘armv8.5-a’ | Armv8.5-A | ‘armv8.4-a’, ‘+sb’, ‘+ssbs’, ‘+predres’
‘armv8.6-a’ | Armv8.6-A | ‘armv8.5-a’, ‘+bf16’, ‘+i8mm’
‘armv8.7-a’ | Armv8.7-A | ‘armv8.6-a’, ‘+ls64’
‘armv8.8-a’ | Armv8.8-a | ‘armv8.7-a’, ‘+mops’
‘armv8.9-a’ | Armv8.9-a | ‘armv8.8-a’
‘armv9-a’.  | Armv9-A   | ‘armv8.5-a’, ‘+sve’, ‘+sve2’
‘armv9.1-a’ | Armv9.1-A | ‘armv9-a’, ‘+bf16’, ‘+i8mm’
‘armv9.2-a’ | Armv9.2-A | ‘armv9.1-a’, ‘+ls64’
‘armv9.3-a’ | Armv9.3-A | ‘armv9.2-a’, ‘+mops’
‘armv9.4-a’ | Armv9.4-A | ‘armv9.3-a’
‘armv8-r’   | Armv8-R   | ‘armv8-r’

[becker33]

Actually looking more closely, 9.2-a+nosve would be the same as 8.7-a, not 8.6-a. But the 8.7-a and 9.2-a tables include +wfxt+xs which we don't list as flags on the m4 architecture in archspec, so I'm hesitant to make that change.

[becker33]

Ok after a bunch more reading:

• There is no way to use gcc to target sme instructions while only allowing sve in streaming mode (the only mode that apple-m4 supports).
• Apple only promises that their hardware supports the armv8.6-a ISA
• Based on info from sysctl hw.optional it looks like the m4 does support the flags that distinguish armv8.7-a from armv8.6-a.
• Apple mostly supports the armv9.2-a ISA, with the exception of SVE non-streaming instructions
• You can mostly disable SVE non-streaming instructions with -fno-tree-vectorize -fno-tree-slp-vectorize, but that's not a guarantee.

With all that said, I think the best option for us would be to use -march=armv9.2-a+sme2 -fno-tree-vectorize -fno-tree-slp-vectorize. This will do no autovectorization but would allow sme intrinsics to work.

If we want to be 100% safe, we would need to stay with -march=armv8.6-a+wfxt.

I'll update the PR with my recommendation.

[alalazo]

@becker33 JSON LGTM. Feel free to merge if you don't think we need to test this further. I plan to release a new version of archspec soon-ish.