Re: [RFC PATCH] gdb/py: Add pretty printer for __mfp8 for aarch64

[Ezra Sitorus <ezra.sitorus@arm.com>]

> I don't really know anything about this type.  Could you say what it
> really is?
> 
> If it's just a float of some kind, then I don't see why it couldn't be
> emitted that way (i.e., base type with DW_ATE_Float and the appropriate
> size), and then if gdb needs special support, that could be added using
> the usual approach (floatformat or whatever).

There are 2 FP8 formats: e4m3 and e5m2. However, instead of introducing 2
new types, we have a single FP8 type called __mfp8. To specify which format
is being used for instructions, there is a new Floating Point Mode Register
(FPMR), which specifies which format is used for source/destination register.
Support for this has just been added to GDB.

Whenever we have an __mfp8/mfloat8_t variable, it's not clear which format
it is using so things like arithmetic or casting aren't allowed. In GCC,
we have also used a uint8_t as the base type which is why it's emitting
DW_ATE_unsigned.

My previous approach was to add new e4m3 and e5m2 floatformats but this was
rejected because:
* there might be slight differences between different architectures
* it's still not a 'real' type so addition etc shouldn't really be allowed.
* all I'm trying to do really is just to present the __mfp8 (uint8_t) in a
  specific way so it's better as a pretty printer.

I think that's a brief summary of Pedro's email here:
https://inbox.sourceware.org/gdb-patches/5ec9ce76-d614-405f-96e8-f490e035aefb@palves.net/

Ezra