When Bruno designed the AVX spec, it looks like he created names like:
VADDPSrm
VADDPSYrm
to distinguish 128-bit and 256-bit operation. It's sort of become a
personal hobby of mine to curse the binutils maintainers for
"vcvtpd2psy" because it's so easy to miss the trailing suffix on a
quick
scan.
I find that same trouble reading dumps with the current AVX names.
Before regalloc it is harder to determine the vector length of an
instruction at a glance.
To overcome that, I started using a naming convention of:
VxADDSSrm
VxADDPSrm
VyADDPSrm
What do people think? I realize it's "only" naming but when
scannig
debug dumps I've found it really helpful to quicky determine how
something was vectorized. From a performance debugging standpoint,
vector length is often an indication of problems or missed
opportunities.
I'm willing to do a gradual mechanical change if this seems reasonable
to people.
-Dave
