log2: Presto vs Velox values differ

[peterenescu]

Bug description

log2 produces different values between Presto/Velox queries found during fuzzer execution.

Reproduction

(Presto followed by Velox):

presto:di> select log2(0.3213630791287869);
        _col0         
---------------------
 -1.6377239051944648 
(1 row)
presto:di> select log2(c0) from (values (0.3213630791287869) ) t(c0);
        _col0         
---------------------
 -1.6377239051944645 
(1 row)

log2 value differences affect the following fuzzer call: SELECT trail(to_base64(to_ieee754_64(log2(c0))), INTEGER '525854645')

[kagamiori]

The mismatch of log2() is actually OK because that is basically the imprecision of floating-point numbers (and we don't plan to match that 100%). But this is a good finding. It means that when generating a subexpression for an argument of floating-point type, we need some special handling (e.g., round the floating-point result) because the imprecision can propagate and be magnified.