Reported by @darominski
The lowest-level interrupt handling code from Xilinx uses lazy save/restore of floating-point registers. There's a dedicated buffer to store the FP context, but it's only good enough to save one context. If there are nested interrupts both using FP, the buffer will overflow and corrupt subsequent memory, leading to a crash upon IRQ handler exit.
We need to provide a mechanism for the application to tweak the maximum interrupt nesting depth (at compile time). It might also be nice to detect the overflow and abort in a predictable way.