From mboxrd@z Thu Jan  1 00:00:00 1970
To: 9fans@cse.psu.edu
Date: Fri, 22 Feb 2008 10:02:23 +0000
From: "Douglas A. Gwyn" <DAGwyn@null.net>
Message-ID: <47BDEB14.B9DCA12D@null.net>
Content-Type: text/plain; charset=us-ascii
Content-Transfer-Encoding: 7bit
References: <5d375e920802171343p323a2708m9f193ae24d2c14e5@mail.gmail.com>,
	<47BC116F.3060906@proweb.co.uk>
Subject: Re: [9fans] Non-stack-based calling conventions
Topicbox-Message-UUID: 5ec17a8e-ead3-11e9-9d60-3106f5b1d025

maht wrote:
> For the C version there's all sorts of pushing and popping return
> addresses and arguments on to the stack
> float
> sqr(x float)
> {
>     return x * x;
> }
> float
> hyp(x float, y float)
> {
>     return sqrt(sqr(x) + sqr(y));
> }

[Assuming that the code is corrected and the types made compatible
with the argument to sqrt.]
The extra overhead occurs only because of the support for full
recursion, with private instance data.  Many C compilers dating
back to around 1980 will in-line the calls to sqr, so the effect is:
hyp:	mul	r2,r2,r2
	mul	r3,r3,r3
	add	r2,r3,r2
	jmp	sqrt
It's hard to imagine anything more efficient (except when hyp
itself can be in-lined.)  Details of course depend on the machine
architecture and linkage convention; the above is modeled loosely
after the M*CORE, which is very nice in this respect.