I figured out why this test is failing, it's a problem with fractional array indices. Not sure why this didn't fail when I did my pre-submission tests. I have a fix almost done.
Created attachment 16811 [details] patch with change log and tests
Comment on attachment 16811 [details] patch with change log and tests r=me
Committed revision 26899.
I'm not sure the float/double variants are actually any benefit now that this has to do the blind-assignment and equality check. I bet the compiler was previously optimizing out the float to double back to int conversion into just a float to int conversion... but we'd have to look at the assembler to be sure.
(In reply to comment #4) > I'm not sure the float/double variants are actually any benefit now that this > has to do the blind-assignment and equality check. I bet the compiler was > previously optimizing out the float to double back to int conversion into just > a float to int conversion... but we'd have to look at the assembler to be sure. I'm almost sure it was not. I saw conversion to float in the assembler. But please investigate!