Glibc Picks Up Some More FMA Performance Optimizations

[vasc]

That's accurate. Though, IIRC, both Intel and AMD began working on it around the same time, but didn't communicate with one another. Intel began working with FMA4 switching to FMA3 later on while AMD began working on FMA3 and then switched to FMA4. All that was before any of the two had announced the instruction to the public. When AMD announced FMA4 Intel has already switched to working on FMA3 and released FMA3 some years later (around 2 years after AMD had released FMA4). Since then, with AMD's worse market share and developers picking up FMA3 over FMA4, AMD ended up having to implement FMA3 as well. As of Zen, FMA4 has been deprecated. It seems the instruction still works if you submit it to the CPU but the result is wrong.

I remember doing some research on the subject a while ago and that's what I remember from it.

Actually that's not true. AMD AFAIK originally planned to include FMA in the original Sledgehammer as part of the TFP (technical floating point) extension. This was supposed to be a three operand register based FP ISA to replace X87. But later they decided to simply use Intel's SSE2 as it would require less effort with tools and porting software to it. Later AMD designed the FMA4 ISA extension to SSE2 before Intel designed FMA3 and the situation you describe occured.