Can someone give me a reference to static single use (for example scientific paper).
I know SSA and SSI (I = information).

Thanks

--Bart.

_________________________________________________________________ Een eigen webpagina voor foto's en weblog in 2 tellen! http://spaces.live.com

Op Wed, 8 Nov 2006, schreef Bart van den Broek:

Can someone give me a reference to static single use (for example scientific paper). I know SSA and SSI (I = information).

http://www.freepascal.org/wiki/index.php/Single_assignment

Daniël

Bart van den Broek wrote

Can someone give me a reference to static single use (for example scientific paper). I know SSA and SSI (I = information).

http://www.google.com/search?q=static+single+use

First hit (the Google summary even contains all necessary information :)

Jonas

Is SSU easier to implement than SSA ? Aren't variables often more used than assigned (so more generated variables by SSU) ? How do the achieved optimizations compare ?

(At least I didn't find this info in the wiki, fpc or pedia).

Micha

All links you gave me lead to SSA, I have gathered quite some documents on the subject.
I found one reference in an SSI document to SSU and it was discredited as incomplete (no algorithms) and vague.
I thought you found some info i didnt see yet.

So i guess you are using SSA?

--Bart.

----Original Message Follows----
From: Micha Nelissen
Reply-To: FPC developers' list
To: FPC developers' list
Subject: Re: [fpc-devel] Request info on static single use Date: Wed, 08 Nov 2006 17:46:13 +0100

Is SSU easier to implement than SSA ? Aren't variables often more used than assigned (so more generated variables by SSU) ? How do the achieved optimizations compare ?

(At least I didn't find this info in the wiki, fpc or pedia).

Micha

No, it's a bit more work.

Aren't variables often more used than assigned (so more generated variables by SSU) ?

Yes.

How do the achieved optimizations compare ?

SSU gives somewhat better results on register-starved architectures. The change from SSA to SSU didn't make any difference for PPC for the compiler+rtl (it can give a difference for large procedures with a lot of variables which are used in different "phases" of course). It saves about 5kb on the compiler on Linux/i386 compared to SSA-only.

Jonas

Bart van den Broek wrote

All links you gave me lead to SSA,

No. The first google hit is a wikipedia article about SSA, but contains a subsection about SSU.

No, SSU. SSU simply means that you not only rename a variable when it's written, but also when it's read. I.e.,

x1 := y1
z1 := y1;

is changed into

x2 := y1
y2 := y1
z2 := y2
y3 := y2

This allows the register allocator to more often change which registers are spilled to memory and which are kept in registers (which is mainly useful if you have few registers available)

Jonas

No. The first google hit is a wikipedia article about SSA, but contains a subsection about SSU.

Ah found it (but no articles/reports that compares it with SSA).

Ah found it (but no articles/reports that compares it with SSA).

There are some bibliographies: http://www.cs.man.ac.uk/~irogers/SSA.xhtml and http://www.cs.man.ac.uk/~jsinger/ssa.html (with comments). (However, to view some articles you must have subscription.)

Will you be working on SSA and/or SSU with control blocks (e.g. using SSA in procedures where a for loop is used) before 2.2? Or will that be released only in a later version.

A synthetic example that will benefit is the CalculatePoint procedure of http://svn.freepascal.org/svn/fpc/trunk/tests/bench/shootout/src/mandelbrot.pp

Vincent

Vincent Snijders wrote

Will you be working on SSA and/or SSU with control blocks (e.g. using SSA in procedures where a for loop is used) before 2.2? Or will that be released only in a later version.

Later. But maybe Florian will work on it earlier.

Jonas

I don't plan it before 2.2. I prefer to concentrate our efforts on 2.2

....(However, to view some articles you must have subscription.)

Which is also my problem....