On Thursday, 12 May 2016 at 14:29:01 UTC, Steven Schveighoffer
wrote:
There was a question on the forums a while back about
equivalent C++ code that didn't work in D. The answer turned
out to be, you had to shoehorn everything into doubles in order
to get the same answer.
I can certainly see that being the case, especially when dealing
with SSE-based code. floats and doubles in XMM registers don't
get calculated at 80-bit precision internally, their storage size
dictates their calculation precision. Which has led MSVC to
promoting floats to doubles for CRT functions when it thinks it
can get away with it (and one instance where the compiler forgot
to convert back to float afterwards and thus the lower 32 bits of
a double were being treated as a float...)
It's fun comparing assembly too. There's one particular function
we have here that collapsed to something like 20-30 lines of
SSE-based code (half after I hand optimised it with branchless
SSE intrinsics and without a call to fmod). The same function in
D resulted in a significantly larger amount of x87 code.
I don't miss x87 at all. But this is getting OT.
