On Tue, Feb 6, 2018 at 4:33 PM, Dan Williams [off-list ref] wrote:

Should we go with array_element_nospec() in the meantime? So we're not
depending on jump labels? With the constraint fix and killing that
superfluous AND the assembly is now:

     e26:       48 81 fd 4d 01 00 00    cmp    $0x14d,%rbp
     e2d:       48 19 d2                sbb    %rdx,%rdx
                        NR_syscalls);
        if (likely(call))
     e30:       48 21 d0                and    %rdx,%rax
     e33:       74 1e                   je     e53 <do_syscall_64+0x73>
                regs->ax = (*call)(regs->di, regs->si, regs->dx,
     e35:       48 8b 4b 38             mov    0x38(%rbx),%rcx
     e39:       48 8b 53 60             mov    0x60(%rbx),%rdx
     e3d:       48 8b 73 68             mov    0x68(%rbx),%rsi
     e41:       48 8b 7b 70             mov    0x70(%rbx),%rdi
     e45:       4c 8b 4b 40             mov    0x40(%rbx),%r9
     e49:       4c 8b 43 48             mov    0x48(%rbx),%r8
     e4d:       ff 10                   callq  *(%rax)

That looks fairly optimal, except for the fact that the callq is
through a register.

Of course, that register-indirect calling convention is forced on us
by retpoline anyway (which you don't have enabled, likely because of a
lack of compiler). But without retpoline that callq could be

         callq  sys_call_table(,%rax,8)

if the masking is done on the index (and if the conditional jump had
been done on the cmp rather than the later 'and').

Instead, you have a

        leaq    sys_call_table(,%rbp,8),%rax

hiding somewhere earlier that doesn't show in your asm snippet.

Oh well. We'll have an extra instruction however we do this. I guess
that's just something we'll have to live with. No more bikeshedding..

            Linus