after-rounding tinyness detection and directed rounding

[Michel Hack (1-914-784-7648) <hack@xxxxxxxxxxxxxx>]

It seems I was confused about the meaning of after-rounding tininess
detection -- I have no experience with it.  Well, I do use Linux on
x86, but x86 FP behaviour is strange in a number of ways, and I only
just found out that it does use the after-rounding rule.  (I also use
AMD 64 systems, which presumably use the same rule, but perhaps don't
exhibit the x87 properties, or do they?)

Mostly I use S/390, z/Architecture, or PowerPC, because on those
systems I can program in assembler and find out precisely how they
behave; I'm also quite familiar with their Principles of Operation.

So the band of confusion for directed rounding is from Nmin-Dmin to
Nmin, twice as wide as for round-to-nearest, but otherwise not a big
deal, and I withdraw my example.

I will now have to look again at what it takes to emulate one rule
(for reproducibility reasons) when the native rule is the other way...

Michel.
Sent: 2007-10-15 20:35:23 UTC