Only compile assembly codes as armv8.2-a

[asdfugil]

Compile C code as base armv8-a, and compile assembly codes as armv8.2-a to allow m1n1 to be properly compiled for older SoCs.

[asdfugil]

The build failure appears to be caused by https://src.fedoraproject.org is down/not working properly.

[asdfugil]

updated to specify armv8.2-a directly whenever needed

[asdfugil]

the build failures appears to be because of the build environment not supplying some runtime builtins that are required when LSE atomics are unavailable. however the m1n1.spec file is outside of this repo so nothing can be done here.

[jannau]

I'm not sure if I'm happy with this change. I would prefer if the HV continues to use LSE atomics if available.
Could you live with enabling this only for old SoCs via a config variable?

[jannau]

please break the asm statement into 2 lines like

        __asm__ volatile(".arch armv8.2-a\n"                                                       \
                         "mrs\t%0, " #reg : "=r"(val));                                            \

[jannau]

I don't think this is enough for armv8-a under all circumstances. I think you have to explicitly set -march=armv8-a and add -mno-outline-atomics to make sure we end up with base armv8-a

[asdfugil]

I'm not sure if I'm happy with this change. I would prefer if the HV continues to use LSE atomics if available. Could you live with enabling this only for old SoCs via a config variable?

Actually the HV is already not using LSE atomics in some cases when it could have:
HV calls into spin_lock(), and the asm block there was changed to not use LSE atomics

How about compiling HV C files with -march=armv8.2-a, and then define a hv_spin_lock() that restores the old spin_lock() behavior (and use it only inside HV C files)?

[jannau]

spinlock (and hv) still use built-in atomics and with -march=armv8.2-a those translate to LSE atomics

[asdfugil]

spinlock (and hv) still use built-in atomics and with -march=armv8.2-a those translate to LSE atomics

yeah but the asm block used to use LSE atomics as well but this was changed so m1n1 could work on armv8.0 869d2ae

[asdfugil]

updates:

split the mrs/msr marco inline asm statement into two lines
use -mno-outline-atomics and -march=armv8-a in general
HV codes are compiled with -march=armv8.2-a still, and there's hv versions of spin_lock() and spin_unlock() (spin_lock() explicitly uses casa in its inline asm block which is the behavior before a7-a11 support is added)