i386 assembly timing issue

[David Howells <dhowells@cambridge.redhat.com>]

I've attached two slightly different bits of i386 assembly that achieve the
same end, but in slightly different ways. Can some one tell me why Case 1 is
faster than Case 2? Case 1 involves an extra CALL instruction.

	* Case 1 has a little wrapper function that saves ECX and EDX before
	  calling rwsem_wake().

	* Case 2 merges the contents of the wrapper with the caller.

Case 1 is what's generated by the rw-semaphore inline assembly code as of
2.4.4-pre5. Case 2 looks like it ought to be a faster version of the same
thing.

David

###############################################################################
#
# CASE 1: registers saved in the rwsem_wake register saving stub
#
.text
	.align 16

#
# void test_up_read(struct rw_semaphore *sem)
# {
# 	up_read(sem);
# }
#
.globl test_up_read
	.type	test_up_read,@function
test_up_read:
	movl	4(%esp), %eax
	movl	$-1, %edx
	xadd	%edx,(%eax)
	js	test_up_read_contention
test_up_read_done:
	ret

#
# Register saving stub for rwsem_wake
#
.globl __rwsem_wake
__rwsem_wake:
	pushl	%edx
	pushl	%ecx
	call	rwsem_wake
	popl	%ecx
	popl	%edx
	ret

#
# Contention handler stub for up_read
#
.section .text.lock,"ax"
test_up_read_contention:
	decl	%edx
	testl	$65535,%edx
	jnz	test_up_read_done
	call	__rwsem_wake
	jmp	test_up_read_done

###############################################################################
#
# CASE 2: registers saved in the contention handler stub
#
.text
	.align 16

#
# void test_up_read(struct rw_semaphore *sem)
# {
# 	up_read(sem);
# }
#
.globl test_up_read
	.type	test_up_read,@function
test_up_read:
	movl	4(%esp), %eax
	movl	$-1, %edx
	xadd	%edx,(%eax)
	js	test_up_read_contention
test_up_read_done:
	ret

#
# Contention handler stub for up_read
#
.section .text.lock,"ax"
test_up_read_contention:
	decl	%edx
	testl	$65535,%edx
	jnz	test_up_read_done
	pushl	%edx
	pushl	%ecx
	call	__rwsem_wake
	popl	%ecx
	popl	%edx
	jmp	test_up_read_done