decimal floating point

["Ivan Godard" <igodard@xxxxxxxxxxx>]

After looking at the arguments for supporting decimal FP at
http://www2.hursley.ibm.com/decimal, I find them relatively persuasive and
the value encoding downright clever. On the other hand, I sure don't want to
put an implementation in the way of our first tapeout.

So I'm inclined to include decimal support in the base architecture, and
leave the actual implementation to emulation in our first product. We do
know that the first product will have to support binary FP regardless due to
customer demand. However, this raises some questions for the Committee if
you don't mind.

Firstly, in a hardware environment that supports both formats, how does the
application source (that is, how do you anticipate the application source
will) specify which format to use in any given host language?

Secondly, what is the interpretation of an operation that is applied to
mixed formats? In particular, assuming there are interformat conversion
operations in hardware, what are the promotion rules?

Lastly, do you envisage a mixed format machine environment being implemented
as multiple machine opcodes (e.g. BFADD-DFADD, etc.), or as single opcodes
(FADD etc.) with a machine mode bit specifying the anticipated data format.
The former would support imported data in foreign formats and mixed format
computation; the latter would only required one copy of the code of any
libraries but would play havoc with constants in the code and compiler.

Or have you looked at mixed formats already and considered them hopeless?

Your help and advice appreciated.

Ivan Godard