Path: utzoo!news-server.csri.toronto.edu!rutgers!apple!amdcad!dvorak.amd.com!proton!tim
From: tim@proton.amd.com (Tim Olson)
Newsgroups: comp.arch
Subject: Re: register save
Message-ID: <1991Mar13.173856.23261@dvorak.amd.com>
Date: 13 Mar 91 17:38:56 GMT
References: <1991Mar11.192116.1974@dgbt.doc.ca> <912@spim.mips.COM> <3255@crdos1.crd.ge.COM>
Sender: usenet@dvorak.amd.com (Usenet News)
Reply-To: tim@amd.com (Tim Olson)
Organization: Advanced Micro Devices, Austin, TX
Lines: 61

In article <3255@crdos1.crd.ge.COM> davidsen@crdos1.crd.ge.com (bill davidsen) writes:
|   Well I'm not about to argue but I may pose a question, in systems
| which use register windows,  does someone have a good handle on the cost
| of doing all the register to register I see just before a call? I see
| some papers which show big gains by not moving data register to memory
| to register putting arguments on a stack, but they seem to discount the
| cost of the moves, and the reduced performance of the optimizer having
| to treat a lot of registers as temps or plain unusable. I would have to
| read some papers to clearly unstand how to measure the optimizer impact,
| so if someone has done it speak up.

The register-to-register movement that you mention is not due to a
windowed-register implementation; simply passing parameters in
registers will cause this, because a function may be called by more
than one parent, and they all must agree on what registers the
parameters must be passed in.

However, the register-to-register movement is not required all the
time; register allocators can bind local variables to the correct
outgoing parameter register through copy propagation.  For example,
the function:

----------------------------------------
void g(int, int, int);

void f(int j)
{
    int i1, i2;

    i1 = j+5;
    i2 = j-7;
    g(i2, j, i1);
}
----------------------------------------

compiles to the following, using the MetaWare C compiler for the 29K:

----------------------------------------
_f:
	sub	gr1,gr1,24
	asgeu	V_SPILL,gr1,gr126
	add	lr1,gr1,36
;5   |    int i1, i2;			<- local variables i1, i2
;6   |
;7   |    i1 = j+5;
	add	lr4,lr8,5		<- i1 is assigned local register 4
;8   |    i2 = j-7;				(3rd parameter reg)
	sub	lr2,lr8,7		<- i2 is assigned local register 2
;9   |    g(i2, j, i1);				(1st parameter reg)
	call	lr0,_g			<- so this call needs no copying,
	add	lr3,lr8,0		<- 	except incoming parameter j
;10  |}
	add	gr1,gr1,24
	nop
	jmpi	lr0
	asleu	V_FILL,lr1,gr127
----------------------------------------
--
	-- Tim Olson
	Advanced Micro Devices
	(tim@amd.com)