[gentoo-user] Re: Flash and cpu_flags_x86_sse2

[James]

Neil Bothwick <neil <at> digimed.co.uk> writes:



> CPU specific USE flags have moved, there was a news item about it
> eselect news read
 
yea, ok.

so I ran the script and got:

# cpuinfo2cpuflags-x86
CPU_FLAGS_X86="aes avx fma3 fma4 mmx mmxext popcnt sse sse2 sse3 sse4_1
sse4_2 sse4a ssse3 xop"

But, looking at : /usr/portage/profiles/desc/cpu_flags_x86.desc
I surmise:

CPU_FLAGS_X86="mmx sse sse2 sse3 ssse3 sse4 sse4_1 sse4_2 sse4a ssse3 \
mmx mmxext xop 3dnowprefetch  3dnow 3dnowext aes avx  avx2 fma3 fma4 \
padlock popcnt "


Some discussion would be keen, as I'd just assume if any ebuild called
a specific instruction set feature, then it should be listed globally
for that processor. In fact, I'll bet there are actually few packages
that need any of those (amd64)CPU flags turned off; so why not just
list all amd64 CPU flags by default and have the user use a -feature,
flag setting, per package? 

After all, if a software dev goes to the trouble to
research how to make their software run (faster), via specific instruction
sets of a given processor, it's a pretty good bet that those ought to be 
invoked. Ok of coarse I have no idea if GCC(++) is doing anything
"squirrely" in the compilers lately.... (related to CPU flag settings);
mostly because that is some voluminous research and reading.

What would be cool, is if the devs took the entire list of cpu flags
from proc/cpuinfo and piped it thru a simple validation script (equery
hasuse) to generate the maximum flags to set by default for a given cpu. I
see no reason this should be limited to either x86 or amd64. Arm64 is out
in the wild and it will not be too much longer before this list has
numerous questions about arm64 processor flags settings, for the myriad
of arm64 processors headed to the linux distros.


James

[1] http://thread.gmane.org/gmane.linux.gentoo.user/165425/focus=165441