Apparently it happens quite easily that different NaNs are produced in
the interpreter than in the execution of the compiled code. This
non-determinism caused problems for the fuzzer which compares the
equality of the results of the interpreter and the compiled code.

I decided therefore to refactor the detection of non-determinism in the
interpreter. Instead of tracking whether potentially non-deterministic
NaNs were produced, I track now whether potentially non-deterministic
NaNs could have been observed. The only way the NaN non-determinism can
be observed is by observing the non-deterministic bit pattern of the
NaN. AFAICT the only way to observe the bit pattern is with a
I(32|64)_REINTERPRET_F(32|64) instruction or with a F(32|64)_STORE
followed by a load. Therefore I flag an execution as potentially
non-deterministic when either a NaN is reinterpreted to an int, or when
a NaN is stored to memory.

R=titzer@chromium.org, eholk@chromium.org
BUG=682180

Review-Url: https://codereview.chromium.org/2671803002
Cr-Commit-Position: refs/heads/master@{#42917}