caiyinyu
|
a53451559d
|
LoongArch: Add glibc.cpu.hwcap support.
Key Points:
1. On lasx & lsx platforms, We must use _dl_runtime_{profile, resolve}_{lsx, lasx}
to save vector registers.
2. Via "tunables", users can choose str/mem_{lasx,lsx,unaligned} functions with
`export GLIBC_TUNABLES=glibc.cpu.hwcaps=LASX,...`.
Note: glibc.cpu.hwcaps doesn't affect _dl_runtime_{profile, resolve}_{lsx, lasx}
selection.
Usage Notes:
1. Only valid inputs: LASX, LSX, UAL. Case-sensitive, comma-separated, no spaces.
2. Example: `export GLIBC_TUNABLES=glibc.cpu.hwcaps=LASX,UAL` turns on LASX & UAL.
Unmentioned features turn off. With default ifunc: lasx > lsx > unaligned >
aligned > generic, effect is: lasx > unaligned > aligned > generic; lsx off.
3. Incorrect GLIBC_TUNABLES settings will show error messages.
For example: On lsx platforms, you cannot enable lasx features. If you do
that, you will get error messages.
4. Valid input examples:
- GLIBC_TUNABLES=glibc.cpu.hwcaps=LASX: lasx > aligned > generic.
- GLIBC_TUNABLES=glibc.cpu.hwcaps=LSX,UAL: lsx > unaligned > aligned > generic.
- GLIBC_TUNABLES=glibc.cpu.hwcaps=LASX,UAL,LASX,UAL,LSX,LASX,UAL: Repetitions
allowed but not recommended. Results in: lasx > lsx > unaligned > aligned >
generic.
|
2023-09-19 09:11:49 +08:00 |
|
