On Monday, 4 January 2021 at 05:58:09 UTC, Walter Bright wrote:
On 1/3/2021 8:37 PM, 9il wrote:
I didn't believe it when I got a similar answer about IEEE
floating-point numbers: D doesn't pertinent to be IEEE 754
compatible language and the extended precision bug is declared
to be a language feature.
The "extended precision bug" is how all x87 code works, C to
C++ to Java. The reason is simple - to remove the problem
requires all intermediate results to be written to memory and
read back in, which is a terrible performance problem. Early
Java implementations did this write/read, and were forced to
change it.
The advent of the XMM registers resolved this issue, and all
the x86 D compilers now use XMM for 32 and 64 bit floating
point math, when compiled for a CPU that has XMM registers.
Extended precision only happens when the `real` 80 bit type is
used, and that is IEEE conformant.
But you still have to deal with things like ARM, so maybe the
better option is to figure out what the differences are between
various hardware and define "floating point conformance levels"
that library can test for, including what SIMD instructions are
available. For instance, the accuracy of functions like
log/exp/sin/cos/arcsin/… can vary between implementations. It
would be useful for libraries to know.