Using 'gcc -Os -fomit-frame-pointer -march=core2 -mtune=core2' for

unsigned short mul_high_c(unsigned short a, unsigned short b)
{
    return (unsigned)(a * b) >> 16;
}

unsigned short mul_high_asm(unsigned short a, unsigned short b)
{
    unsigned short res;
    asm("mulw %w2" : "=d"(res),"+a"(a) : "rm"(b));
    return res;
}

I get

_mul_high_c:
	subl	$12, %esp
	movzwl	20(%esp), %eax
	movzwl	16(%esp), %edx
	addl	$12, %esp
	imull	%edx, %eax
	shrl	$16, %eax
	ret
_mul_high_asm:
	subl	$12, %esp
	movl	16(%esp), %eax
	mulw 20(%esp)
	addl	$12, %esp
	movl	%edx, %eax
	ret

mulw puts its outputs in dx:ax, and dx contains (dx:ax)>>16, so the shift is avoided.

Ignoring the weird Darwin stack adjustment code, the version with mulw is somewhat shorter and avoids a movzwl. I'm not sure what the performance difference is; mulw is listed in Agner's tables as fairly low latency, but requires a length changing prefix for memory.

This type of operation is useful in fixed-point math, such as embedded audio codecs or arithmetic coders.