From mboxrd@z Thu Jan  1 00:00:00 1970
Message-ID: <8f2a08f3fdef3a1a5ab430265782f80f@vitanuova.com>
From: C H Forsyth <forsyth@vitanuova.com>
Date: Fri, 14 May 2004 06:30:50 +0100
To: 9fans@cse.psu.edu
Subject: Re: [9fans] brucee's 8c on sources?
In-Reply-To: <40A435AA.3030908@chunder.com>
MIME-Version: 1.0
Content-Type: text/plain; charset="US-ASCII"
Content-Transfer-Encoding: 7bit
Topicbox-Message-UUID: 7aa06ff4-eacd-11e9-9e20-41e7f4b1d025

>>another good speedup is a simple vlong assignment,
>>which used to be MOVL; CLD; REP; MOVSL; and is now

another example is conversion between 32-bit and 64-bit values.
loading vlong as long from memory as in
	g((ulong)x); /* x vlong */
should just be loading the correct half of the vlong, shouldn't it,
but it was
	MOVL	x+-8(SP),CX
	MOVL	CX,(SP)
	MOVL	x+-4(SP),CX
	MOVL	CX,4(SP)
	CALL	,_v2ul+0(SB)
	MOVL	AX,.safe+-16(SP)
	MOVL	.safe+-16(SP),AX
	MOVL	AX,(SP)
	CALL	,g+0(SB)
and now is
	MOVL	x+-8(SP),AX
	MOVL	AX,(SP)
	CALL	,g+0(SB)
ie, it does the obvious thing

the original scheme was easy to implement and was adequate
when there were few 64-bit values in sight,
but once offsets were vlong everywhere, and Qid.path was vlong,
they began to spread fairly widely through the code.