[PATCH, x86_64]: Use soft-fp for TFmode (128bit) FP.

[Joseph S. Myers]

On Sat, 14 Apr 2007, Uros Bizjak wrote:

> There is one condition, I'm not sure about (I have CCed rth, as a primary
> author of op-common.h). In FP_EXTEND and FP_TRUNC macros, the top condition
> always aborts due to last ORed part of the condition. I'm not sure, if the

This part asserts that denormals truncate to 0 and extend to normals.

> then the conversions work OK for all operands (including inf and nan).

Did you test denormals?

My belief (having written those macro definitions in their present forms) 
is that the code needs extending to handle denormals before you can remove 
the condition.  It doesn't need extending in full generality - you could 
for example only extend for the case where the two types have the same 
bias, which covers the quad/extended pair (so denormals extend to 
denormals and truncate to denormals or 0) and adjust the assertion to 
allow that additional case only.

-- 
Joseph S. Myers
joseph@codesourcery.com