k-FP | Benchmarks for the k-FP WF attack | Functional Programming library

Caveat: I have never used a PDP-11, so the following is based purely on the documentation linked in the question.

The PDP-11 was a 16-bit machine using 16-bit words. 32-bit single-precision floating-point data was stored in a "mixed-endian" format: the more significant word was stored at the lower address, but within each word, the less significant byte was stored at the lower address. In other words, in order of increasing addresses, the four bytes are stored in order 2, 3, 0, 1.

The PDP-11 floating-point format is similar to the IEEE-754 binary32 format, using sign-magnitude representation with eight exponent bits and a significand (mantissa) whose most significant bit is assumed to be 1 and therefore not stored. The exponent bias is 128 instead of 127 for IEEE-754 binary32, and the significand is normalized to [0.5, 1) instead of [1, 2) for IEEE-754 binary32. Also, subnormals, infinities, and NaNs are not supported.

This means that conversion can never overflow, but it can underflow (with an accompanying potential reduction in accuracy) to an IEEE-754 binary subnormal. The following program assumes that we are running on a machine with IEEE-754 floating-point, and are presented with PDP-11 floating-point data as a sequence of bytes in memory order.