[fpc-devel] Maybe a bug, maybe not

[J. Gareth Moreton via fpc-devel]

Hi everyone,

So I've found a small inconsistency on x86_64 with inline assembly,
specificially the PINSRW instruction.  Under AT&T notation, you can
write something like this and it will compile fine with the internal
assembler:

pinsrw  $0x0, %dx, %xmm0

But if you use -a so the external assembler is used, it throws an error,
because the second parameter can either be a 16-bit memory location, or
a 32-bit register... not 16-bit.  In other words, even though the upper
16 bits aren't used on the register for this instruction, you must
specify the full 32-bit register:

pinsrw  $0x0, %edx, %xmm0

This is confirmed on page 4-296 of Vol. 2B of the Intel® 64 and IA-32
Architectures Software Developer’s Manual, showing the second parameter
is of the form "r32/m16".

Long story short, "pinsrw $0x0, %dx, %xmm0" is technically illegal even
if it makes more logical sense.  I put this down to PINSRW being a
carry-over from the MMX instruction set and possibly not designed as
well as it could be.  Should the internal assembler / syntax checker be
changed to trap this?

Gareth aka. Kit