x86 regression: optimizer bug in 1.1 and dev

[Toon Moene]

>> Excess precision on x86 issue I suspect.
>>
>> -ffloat-store makes it go away.
>>

>  I completely forgot about this little gem.
>
>  Is this something that we should always be adding to our
>  compile flags on x86?

No, because:

1. It generates slower code.

2. It isn't correct in more convoluted examples anyway.

The bottom line, which I have repeated on comp.arch and  
comp.compilers, is:  Unless you don't use floating point registers  
*at all*, reasonable floating point behaviour on the ix86  
architecture is impossible.

[ There's a processor specific flag you can set prior to the main  
program executing, but it'll only restrict floating point arithmetic  
to double precision for the mantissa - the range of the exponent is  
still the 80-bit variant one ]

Cheers,
Toon.