I removed some constant folding optimizations for float instruction in
https://codereview.chromium.org/2647353007 because they were incorrect
if the input was a signalling NaN. Removing these optimizations, however
had an unexpectedly big impact on asm.js performance. With this CL I
restore the optimizations again when the source origin is not wasm. In
JavaScript signalling NaNs are not observable and therefore the
optimizations are correct.

R=titzer@chromium.org
BUG=chromium:686654

Review-Url: https://codereview.chromium.org/2666903002
Cr-Commit-Position: refs/heads/master@{#42850}