Previously using the new "lower 8 bits" registers (SIL, SPL, ...) caused SETcc
to write to other registers (for example, SETcc SIL would generate SETcc DH).
The new NOP emitter breaks when called with a negative count. As it
turns out, it did happen when deoptimizing 8 bit MOVs because they are
only 4 bytes long and need no BSWAP.
Previously he function was misbehaving because of a missing check for
whether an 8-bit operand was a register operand; it would therefore
emit unnecessary REX prefixes, incorrectly assert on 32-bit targets, and
could potentially emit wrong code in rare cases (like a memory to register
operation involving AH.)
Also, some cleanup while I was in the area to make the function easier to
read.
Our defines were never clear between what meant 64bit or x86_64
This makes a clear cut between bitness and architecture.
This commit also has the side effect of bringing up aarch64 compiling support.
Floating-point is complicated...
Some background: Denormals are floats that are too close to zero to be
stored in a normalized way (their exponent would need more bits). Since
they are stored unnormalized, they are hard to work with, even in
hardware. That's why both PowerPC and SSE can be configured to operate
in faster but non-standard-conpliant modes in which these numbers are
simply rounded ('flushed') to zero.
Internally, we do the same as the PowerPC CPU and store all floats in
double format. This means that for loading and storing singles we need a
conversion. The PowerPC CPU does this in hardware. We previously did
this using CVTSS2SD/CVTSD2SS. Unfortunately, these instructions are
considered arithmetic and therefore flush denormals to zero if non-IEEE
mode is active. This normally wouldn't be a problem since the next
arithmetic floating-point instruction would do the same anyway but as it
turns out some games actually use floating-point instructions for
copying arbitrary data.
My idea for fixing this problem was to use x87 instructions since the
x87 FPU never supported flush-to-zero and thus doesn't mangle denormals.
However, there is one more problem to deal with: SNaNs are automatically
converted to QNaNs (by setting the most-significant bit of the
fraction). I opted to fix this by manually resetting the QNaN bit of all
values with all-1s exponent.
