Michael C said:
Are you saying excel doesn't use the FP processor? How then did the
previous intel floating point bug appear in excel?
Excel uses the FPU for INTERMEDIATE calculations, but the final value
of any cell formula gets passed through some, er, MSFT value-added
processing before it becomes the cell's value. This is how
nonparenthesized expressions are implicitly rounded to 15 decimal
digits. Also, and quite unfortunately, some built-in function calls
perform additional and usually unnecessary postprocessing of FPU
results, e.g., MOD(2^30-1,3), which returns #NUM! because MOD can't
handle arguments that produce quotients comfortably within 32-bit
integer or 64-bit double precision bounds even though this is well
within the FPU's capabilities.
I don't think this is the case for 850*77.1. I suspect some bright
spark on the Excel development team decided to rewrite their machine
representation-to-human representation internal number formatting
function, and they screwed it up somehow.
What really concerns me is what Dana DeLouis pointed out: with A1
containing =850*77.1, =A1+1 returns 100001 while =A1-1 returns 65534.
That's VERY, VERY BAD because it implies the value in A1 isn't just
fed directly to the FPU stack. It's subjected to, er, value-added
preprocessing which seems to differ depending on what the precise
operation is. What I'm most worried about, given Dana's other
examples, is whether someone thought they were doing Excel users a
favor by treating increment (+1), decrement (-1), idempotent
operations (+0, *1, ^1), binary left shift (*2) and binary right shift
(/2) as special cases in the formula parser. MSFT *DID* rewrite parts
of the formula parser in XL2007, which is how they lifted the 7 nested
function call limit etc. I'd bet they rewrote too much of it.
