Why is 1.0+0.25*spacing(1.0) not rounded to 1.0?

[kengo]

Why is 1.0+0.25*spacing(1.0) not rounded to 1.0?
This is observed only in quadruple precision.
I use Absoft Pro FORTRAN 11.1 on OSX10.7Lion.

[Mike Linacre]

Kengo, ouch! Your results indicate that something has been converted to double-precision floating point ....

1. How about displaying the expressions: 1.0_16, 0.25_16, spacing(1.0_16), 0.25_16*spacing(1.0_16) to verify that the intermediate computations are what we expect?

2. How about stepping through the Absoft debugger's assembler code? This may reveal where there are unexpected numerical conversions.

[kengo]

Mike,
Thank you for your replying.

1.0_16, 0.25_16, spacing(1.0_16) and 0.25_16*spacing(1.0_16) seem to be correct.
1.0_16+0.25_16*spacing(1.0_16), however, seems to be incorrect.

I have attached a screen shot.
I think the fourth of the Value field of e16 must be 0x00000000.
It is equal to the second of d16.

There may be something wrong in procedures for exponent bits.

[kengo]

I have attached the second screenshot showing the assembler code.

e16 in the previous message by me is likely calculated between MAIN__+330 and MAIN__+429 in this screenshot.

I haven't determined the cause of the problem.