If the intrinsic _itof is passed an argument that is truncated to a size smaller than 32 bits, the compiler may incorrectly discard the truncation. This truncation can occur with a cast to short or a bit-mask to the lower 16 bits.

For example:

/*
   build with
   cl6x -mv6600 -O3 -s file.cpp
*/
#include <stdint.h>
float bug(uint32_t a, uint32_t b)
{      
    const uint32_t lo16 = b & 0x0000FFFFU;
    return _itof(lo16);
}

You'll get

bug:

           RETNOP          B3,4 
           MV      .L1X    B4,A4 

The & 0x0000ffff is ignored. So, if the input value has bits set in the upper 16-bits, they are incorrectly part of the result returned from the function.