amd64: partially depessimize cpu_fetch_syscall_args and cpu_set_syscall_retval

[mjg]

Summary

cpu_fetch_syscall_args becoms jmp-free for the common case (modulo sv_mask which is going away soon). Bumps getuid rate on Broadwell from 105mln to 107.5mln.

This pessimizes 6 or more arg syscalls, which are a complete minority.

cpu_set_syscall_retval change removes 3 branches in favor of 1 and prevents reloads of ->td_frame.

Note the entire mechanism should be reworked - as it is there is a func call to amd64_syscall, then 3 indirect calls (fetch, syscall itself, set). The initial call can be indirect to a routine which does the fetch/set work and includes inlined amd64_syscall work.

[kib]

sys/amd64/amd64/trap.c

973 ↗

(On Diff #49082)

Does it need to be non-static ?

985 ↗

(On Diff #49082)

May be pass frame and sa as the arguments. You already calculated them in non-fallback version.

1005 ↗

(On Diff #49082)

Change to nitems() ?

1049 ↗

(On Diff #49082)

Can you replace all three if()s with single or-ed condition and change goto to direct return(fallback()); statement ?

1053 ↗

(On Diff #49082)

Change to nitems() ?

1060 ↗

(On Diff #49082)

Again, can you replace this with return () ?

1063 ↗

(On Diff #49082)

Why 6 and not regcnt ? Or, do you need reg and regcnt in this function at all ?

1069 ↗

(On Diff #49082)

Then, remove the label.

[mjg]

• also depessimize cpu_set_syscall_retval

[kib]

Commit _retval() change now, I agree with it.

sys/amd64/amd64/trap.c

985 ↗

(On Diff #49089)

There regcnt is used as a constant.

I suggest you to add a define to machine/proc.h e.g. NARGREGS 6, and use it consistently instead of the literal '6' or regcnt, in both functions.

1038 ↗

(On Diff #49089)

This and previous lines, () are excessive.

1052 ↗

(On Diff #49089)

argp is not needed, inline its initializator there.

[mjg]

I did not commit set retval due to the freeze.