Confused about fsystemcalls

Hi David,
We're tryign to use the fast system call stuff here, but our
measurements show way more overhead than the standard `break' mechanism. I suspect we're doing something badly wrong.

We've written a little stub to call the fast system call entry point directly (to save trying to get into libc).

Here it is (I know there are more stop bits than necessary).

#include <asm/asmmacro.h>
	GLOBAL_ENTRY(mygetpid)
        .prologue
	.save ar.pfs, r11
	mov r11 = ar.pfs
	.body
	movl r2 = 0xa000000000020000;; // start of gate page
	mov b7 = r2;;
	mov r15 = 1041;;
	br.call.sptk.many b6 = b7;;
	.restore sp
	mov ar.pfs = r11
	br.ret.sptk.many rp;;
	END(mygetpid)

In System.map, syscall_via_epc and __start_gate_section have the same value, 0xe000000004a50000, so I figure the address in user space of syscall_via_epc is GATE_ADDR (0xa000000000000000 + 2*PERCPU_PAGE_SIZE)

The mygetpid() call appears to return the correct value.

Our figures show over an average of ten runs:
--- glibc getpid() takes 579 cycles, stddev 24 --- mygetpid takes 1468 cycles, stddev 286

I was also a little confused about the entry put into the ELF AUX vector -- was the value or the user-space address of syscall_via_epc meant to go into the vector?
Currently it says in include/asm-ia64/elf.h

#define ARCH_DLINFO \
do {

\
	extern int syscall_via_epc;			\
	NEW_AUX_ENT(AT_SYSINFO, syscall_via_epc);	\
} while (0)

I keep thinking this should be
extern char syscall_via_epc[];
extern char__start_gate_section[]; NEW_AUX_ENT(AT_SYSINFO, syscall_via_epc - __start_gate_section +
GATE_ADDR);

Peter C