haiku/docs/develop/kernel/arch/long double.md
Augustin Cavalier 18a8edbf0e docs/develop: More reorganization.
* There is now a 'busses' folder, and the extant USB/SDHCI/Bluetooth/etc.
   docs now live in it, instead of various other places.
 * kernel/ports is now kernel/arch, like it is in src/system.
   SPARC documentation is now in there, too.
 * VM files (these are rather outdated) are now in kernel/vm.
 * SCSI ASC info removed, this is easily available online and
   it doesn't seem to be very relevant.
2019-03-30 18:00:46 -04:00

49 lines
1.5 KiB
Markdown

Notes on long double support
============================
The "long double" type is different on each architecture. Depending on the
available hardware and ABI conventions, performance compromises, etc, there
may be many implementations of it. Here is a summary for our convenience.
128-bit IEEE
------------
Platforms: Sparc, ARM64, RISC-V
This is the standard long double type from IEEE754. It has 1 sign bit,
15 exponent bit, and 112 fractional part bits. It is the natural extension
of the 64bit double.
Sparc specifies this type in their ABI but no implementation actually has
the instructions, they instead trigger a trap which would software emulate
them. However, gcc short circuits this by default and calls C library
support functions directly.
64-bit IEEE
-----------
Platforms: ARM
This is the same representation as plain "double". ARM uses this for simplicity.
80-bit
------
Platform: x86, x86\_64, m68k
This intermediate format is used by x86 CPUs internally. It may end up being
faster than plain double there. It consists of a 64bit fractional part, 15
exponent bits, and 1 sign bit. This is convenient because the fractional part
is a relatively easy to handle 64bit number.
m68k uses a similar format, but padded to 96 bits (the extra 16 bits are unused).
double double
-------------
Platforms: PowerPC?
This is also a 128bit type, but the representation is just two 64bit doubles.
The value is the sum of the two halves. This format allows faster emulation
than a "true" 128bit long double, and the precision is almost as good.