Add vmla_* / vmlal_*

[makotokato]

Add Vector multiply accumulate intrinsic.

[rust-highfive]

r? @Amanieu

(rust_highfive has picked a reviewer for you, use r? to override)

[Amanieu]

VMLA is not a fused multiply-add. It involves an intermediate rounding step just like if you use two separate operations. Only VFMA is a fused multiply-add.

[Amanieu]

Ping. The vmla_* functions needs to use simd_add and simd_mul. The vfma_* functions need to use simd_fma.

[Amanieu]

You can use the mul_add method on f64.

[makotokato]

Sorry for delay an thank you for review. I will update this PR.

[makotokato]

unrelated CI is failure due to OOM...

[Amanieu]

CI should be fixed now.

However you still haven't fixed the definitions of the functions:

• vfma_f{32,64} should always do a fused-multiply add. (vfma on ARM, fmadd/fmla on AArch64`)
• vmla_f{32,64} should always do a separate multiple and add. (vmla on ARM, fmul + fadd on AArch64).

The vfma_f{32,64} functions are also available on 32-bit ARM if the vfpv4 feature is enabled.

Have a look at this: https://godbolt.org/z/34Wo5n

[bors]

☔ The latest upstream changes (presumably 594ff85) made this pull request unmergeable. Please resolve the merge conflicts.