History of Berkeley SoftFloat, to Release 3e

John R. Hauser
2018 January 20

Changed the default numeric code for optional rounding mode odd (round to odd, also known as jamming) from 5 to 6.
Modified the behavior of rounding mode odd when rounding to an integer value (either conversion to an integer format or a ‘roundToInt’ function). Previously, for those cases only, rounding mode odd acted the same as rounding to minimum magnitude. Now all operations are rounded consistently.
Fixed some errors in the specialization code modeling Intel x86 floating-point, specifically the integers returned on invalid operations and the propagation of NaN payloads in a few rare cases.
Added specialization code modeling ARM floating-point, conforming to VFPv2 or later.
Added an example target for ARM processors.
Fixed a minor bug whereby function f16_to_ui64 might return a different integer than expected in the case that the floating-point operand is negative.
Added example target-specific optimization for GCC, employing GCC instrinsics and support for 128-bit integer arithmetic.
Made other minor improvements.

Fixed bugs in the square root functions for 64-bit double-precision, 80-bit double-extended-precision, and 128-bit quadruple-precision. For 64-bit double-precision (f64_sqrt), the result could sometimes be off by 1 unit in the last place (1 ulp) from what it should be. For the larger formats, the square root could be wrong in a large portion of the less-significant bits. (A bug in f128_sqrt was first reported by Alexei Sibidanov.)

Added optional rounding mode odd (round to odd, also known as jamming).
Corrected the documentation concerning non-canonical representations in 80-bit double-extended-precision.

Implemented the common 16-bit “half-precision” floating-point format (float16_t).
Made the integer values returned on invalid conversions to integer formats be determined by the port-specific specialization instead of being the same for all ports.
Added preprocessor macro THREAD_LOCAL to allow the floating-point state (modes and exception flags) to be made per-thread.
Modified the provided Makefiles to allow some options to be overridden from the make command.
Made other minor improvements.

Complete rewrite, funded by the University of California, Berkeley, and consequently having a different use license than earlier releases. Major changes included renaming most types and functions, upgrading some algorithms, restructuring the source files, and making SoftFloat into a true library.
Added functions to convert between floating-point and unsigned integers, both 32-bit and 64-bit (uint32_t and uint64_t).
Added functions for fused multiply-add, for all supported floating-point formats except 80-bit double-extended-precision.
Added support for a fifth rounding mode, near_maxMag (round to nearest, with ties to maximum magnitude, away from zero).
Dropped the timesoftfloat program (now part of the Berkeley TestFloat package).

Fixed mistakes affecting some 64-bit processors.
Further improved the documentation and the wording for the legal restrictions on using SoftFloat releases through 2c (not applicable to Release 3 or later).

Made minor updates to the documentation, including improved wording for the legal restrictions on using SoftFloat.

Added functions to convert between 64-bit integers (int64) and all supported floating-point formats.
Fixed a bug in all 64-bit-version square root functions except float32_sqrt that caused the result sometimes to be off by 1 unit in the last place (1 ulp) from what it should be. (Bug discovered by Paul Donahue.)
Improved the Makefiles.

Created the 64-bit (bits64) version, adding the floatx80 and float128 formats.
Changed the source directory structure, splitting the sources into a bits32 and a bits64 version. Renamed environment.h to milieu.h to avoid confusion with environment variables.
Fixed a small error that caused float64_round_to_int often to round the wrong way in nearest/even mode when the operand was between 2²⁰ and 2²¹ and halfway between two integers.

Corrected a mistake that caused borderline underflow cases not to raise the underflow flag when they should have. (Problem reported by Doug Priest.)
Added the float_detect_tininess variable to control whether tininess is detected before or after rounding.

Original release, based on work done for the International Computer Science Institute (ICSI) in Berkeley, California.