[compiler-rt] [compiler-rt][builtins]Fix complex division for aarch64 (PR #106664)

[David Green via llvm-commits]

davemgreen wrote:

The "behaves differently for some pairs of (infinity, nan)" is an SNaN I believe. They do work differently than other QNaN values in fmaxnm (or fmaxf), but as far as I understand llvm's default floating point environment roughly assumes that floating point exceptions are not read and all nans are quite.  GCC assumes the same and has a -fsignaling-nans option to start "caring" about signalling nans.

Where do the snans come from that are causing problems with complex div? They shouldn't be generated from any operation, so are presumably coming from the input somehow.

https://github.com/llvm/llvm-project/pull/106664