901092 - IonMonkey: Fold ToFloat32(ToDouble(x)) into x, when x is already a Float

Status: RESOLVED DUPLICATE of bug 888109

(Core :: JavaScript Engine, defect)

Description

[Benjamin Bouvier [:bbouvier] (inactive)]

If x is a Float32, it seems like ToFloat32(ToDouble(x)) == x, as we never lose precision during the process.

As we need to be cautious with this kind of optimizations, I have been running a C++ program in background that tries to find a float value which doesn't verify this condition. If the program doesn't find anything in a reasonable amount of time and nobody has any objections against it, I'll implement it.

Comment 1

[Jeff Walden [:Waldo]]

The structure of single- and double-precision IEEE-754 floating point values is such that single-precision is strictly a subset of double-precision.  The only slight exception is that NaN values, where the exact bit pattern matters, won't roundtrip.  But it *is* the case that any NaN float will convert to a NaN double, and that will convert to a NaN float, even if the start and end values don't have the same bit patterns.

Comment 2

[Luke Wagner [:luke]]

Or (based on the proof of the float/double commutative property of addition [1]):
  x = x+0 = ToFloat32(ToDouble(x) + ToDouble(0)) = ToFloat32(ToDouble(x))

[1] http://dl.acm.org/citation.cfm?id=221334

Comment 3

[Jeff Walden [:Waldo]]

x = x-0, you mean.  -0 + 0 is +0.  :-)  Good ol' IEEE-754!

http://goodolcharliebrown.tumblr.com/post/420445025

Comment 4

[Benjamin Bouvier [:bbouvier] (inactive)]

As the mir patch wasn't reviewed in bug 888109, I just wrote this two lines patch there.