lib/msun amd64: use AVX FMA instructions for fma(3) and fmaf(3) when available

[fuz]

Love it!

[fuz]

There could be some performance issues related to the SSE/AVX transition state machine, but checking the Intel Software Optimization Manual, it seems like executing 128 bit AVX instructions seems unproblematic in any steady-state scenario.

[kfv]

Brilliant work. Thanks!

[kib]

Put added symbols into private namespace.
It does not matter for shared library, but helps static linking.

[kib]

Switch to use libc/amd64/string ARCHLEVEL infra to control the selection.

[fuz]

Oh yeah, using the archlevel framework sounds like a good idea! I didn't even think about that, as it's specific to libc. Other libm functions may profit in the future.

[kib]

In D59462#1365021, @fuz wrote:

Oh yeah, using the archlevel framework sounds like a good idea! I didn't even think about that, as it's specific to libc. Other libm functions may profit in the future.

Do you have any notes about the updated patch?

BTW, I was not able to come up with a C expression that can be used to initialize the function table, so I resorted to filling it in assembler same as is done for strings.

[kib]

Add kargl@ patch from PR 298260

[fuz]

In D59462#1365088, @kib wrote:

In D59462#1365021, @fuz wrote:

Oh yeah, using the archlevel framework sounds like a good idea! I didn't even think about that, as it's specific to libc. Other libm functions may profit in the future.

Do you have any notes about the updated patch?

BTW, I was not able to come up with a C expression that can be used to initialize the function table, so I resorted to filling it in assembler same as is done for strings.

Oh the patch is fine, I just forgot to hit “approve.” I don't think the table can be initialised in C. I designed it this way to reduce the amounts of relocations and the table size, and I don't think C wants to do this stuff. We could do an inline assembly macro to build the table in C if you like.