x86/split lock detection: #AC: fasmg/78798 took a split_lock trap at address: 0x4031ea    

#!/bin/bash
sysctl split_lock_mitigate=0 ; fasmg "$@" ; sysctl split_lock_mitigate=1    

If a memory operand is referenced, the processor’s locking protocol is automatically implemented for the duration of the exchange operation, regardless of the presence or absence of the LOCK prefix or of the value of the IOPL.

; xchg eax,[mem]
push ebx
mov ebx,[mem] ; if mem=esp then mov ebx,[mem+4]
mov [mem],eax ; if mem=esp then mov [mem+4],eax
mov eax,ebx
pop ebx    

You could try rewriting the xchg's with equivalent code.

Code:

; xchg eax,[mem]
push ebx
mov ebx,[mem] ; if mem=esp then mov ebx,[mem+4]
mov [mem],eax ; if mem=esp then mov [mem+4],eax
mov eax,ebx
pop ebx    

If you're going to push/pop anyway, why not just push and pop directly with memory operand?

Furs wrote:

If you're going to push/pop anyway, why not just push and pop directly with memory operand?

- Three xor's requires four memory ops, all unaligned.

- With push reg/push mem/pop reg/pop mem that is six mem ops, two of those unaligned

- With push/mov/mov/pop that is four mem ops, two of those unaligned.

The the last one has fewer mem ops, and minimises the unaligned accesses.

It needs to be tested, but often minimising mem ops, and especially minimising non-aligned mem ops, can be a great efficiency win, and maybe also a runtime win.

; xchg eax, [mem]
push dword [mem]
mov [mem], eax
pop eax    

push dword [mem]    

I mean just:

Code:

; xchg eax, [mem]
push dword [mem]
mov [mem], eax
pop eax    

I don't see how it's more memory ops than yours?

Code:

push dword [mem]    

The same command with the same limitations as xchg

x86/split lock detection: #AC: fasmg/78798 took a split_lock trap at address: 0x4031ea

0x4031df: 0x67 0xf6 0x43 0x01 0x01  test   [ebx+0x1],byte 0x1
0x4031e4: 0x74 0x12                 je     0x12
0x4031e6: 0x67 0x8d 0x46 0xfc       lea    eax,[esi-0x4]
0x4031ea: 0x67 0x87 0x02            xchg   [edx],eax
0x4031ed: 0x67 0x8d 0x56 0xfc       lea    edx,[esi-0x4]    

Please let me know if there are more traps, I may have missed something.

$ dmesg | grep 'Linux version'
[    0.000000] Linux version 6.6.52-calculate (root@localhost) (gcc (Gentoo 13.3.1_p20240614 p17) 13.3.1 20240614, GNU ld (Gentoo 2.42 p3) 2.42.0) #1 SMP PREEMPT_DYNAMIC Tue Sep 24 21:17:50 +04 2024
$ cat /proc/cpuinfo | grep 'model name' | head --lines=1 -
model name      : 11th Gen Intel(R) Core(TM) i5-11600KF @ 3.90GHz
$ fasmg fasmg.asm ./fasmg
flat assembler  version g.kd3c
5 passes, 0.4 seconds, 70638 bytes.
$ ./fasmg fasmg.asm ./fasmg.x64
flat assembler  version g.kkod
5 passes, 0.4 seconds, 70638 bytes.
$ dmesg | grep 'split lock detection'
$    

x86/split lock detection: #AC: fasmg/212583 took a split_lock trap at address: 0x408323