Re: Revised PPC assembly implementation

From: Paul Mackerras <paulus@samba.org>
Date: Apr 25, 2005, 09:40 UTC
Message-ID: <17004.47876.414.756912@cargo.ozlabs.ibm.com>
In-Reply-To: <20050425031337.16605.qmail@science.horizon.com>

linux@horizon.com writes:

Show 7 quoted lines

> Three changes:
> - Added stack frame as per your description.
> - Found two bugs.  (Cutting & pasting too fast.)  Fixed.
> - Minor scheduling improvements.  More to come.
> 
> Which lead to three questions:
> - Is the stack set properly now?

Not quite; you are saving 20 registers, so you need a 96-byte stack frame, like this:

	stwu	%r1,-96(%r1)
	stmw	%r13,16(%r1)
	...
	lmw	%r13,16(%r1)
	addi	%r1,%r1,96
	blr

Since sha1_core is a leaf function, I suppose you could use the lr save area (do stwu %r1,-80(%r1); stmw %r13,0(%r1)) but it seems a bit dodgy.

> - Does it produce the right answer now?

Yes.

> - Is it any faster?

I did 10 repetitions of my program that calls SHA1_Update with a 4096-byte block of zeroes 256,000 times. With my version, the average time was 4.6191 seconds with a standard deviation of 0.0157. With your version, the average was 4.6063 and the standard deviation 0.0148. So I would say that your version is probably just a little faster - of the order of 0.3% faster.

Paul.