hypot(): simple sqrt(u*u + v*v) this is not the same as the original mlib provides

[denizzzka]

Module std.math.algebraic provides well known function hypot(). Internally it returns just:

return ratio * sqrt(u*u + v*v);

But as I know the purpose of hypot() function's within the original implementation in mlib was significant accelerating executon of the hypotenuse length by reusing intermediate values ​​in calculations appeared during calculation of power and then root. Also CPUs (x86?) provided instruction for this.

Has anything changed since then? (Maybe modern compilers automatically replace by some intrinsic code like sqrt(u*u + v*v)?)

[ibuclaw]

What's the issue here?

sqrt is and has always been an intrinsic.

https://github.com/dlang/dmd/blob/master/druntime/src/core/math.d#L89-L91

Depending on target architecture, u*u + v*v can be pattern matched into a fmadd instruction.

[denizzzka]

What's the issue here?

I still thought that hypot() function can be calculated faster than the naive calculation of sqrt(u*u + v*v).
This was true only for pre-FPU era :-)

Depending on target architecture, uu + vv can be pattern matched into a fmadd instruction.

It may be a false memory, but seems there were architectures that provided hardware implementation of hypot().

Also, seems, using SIMD can increase it performance. So I think we should rely on libm and don't need to maintain our own implementation of this function. Just as we do not maintain our own cos(), etc...

But existing libm implementations seem to do the same thing as Phobos, so case closed

[denizzzka]

But existing libm implementations seem to do the same thing as Phobos

Nope:
https://github.com/bminor/glibc/blob/master/sysdeps/x86_64/fpu/svml_d_hypot2_core.S
https://github.com/bminor/glibc/blob/master/sysdeps/aarch64/fpu/hypot_advsimd.c

They really do use all sorts of optimizing things!

[ibuclaw]

Glibc has an advantage that they use ifuncs/target clones. It's a jumptable that runs cpuid during C init ctors and select the correct implementation. D(MD) supports neither ifunc dispatching nor target clones.

[denizzzka]

But we are already loading libm? It is part of glibc and provides trigonometry and hypot too

[ibuclaw]

Libm is not pure.

[denizzzka]

Libm is not pure.

Can you explain what you mean?