答复: [COMMIT] romcc 0.63 and some fixes...

YhLu YhLu at tyan.com
Fri Jun 25 19:28:00 CEST 2004


So next week, we can get the updated romcc?

-----邮件原件-----
发件人: ebiederman at lnxi.com [mailto:ebiederman at lnxi.com] 
发送时间: 2004年6月25日 17:41
收件人: YhLu
抄送: LinuxBIOS
主题: Re: [COMMIT] romcc 0.63 and some fixes...

YhLu <YhLu at tyan.com> writes:

> Eric,
> 
> The 
> printo_debug("this mainboard is only designed for 2 cpus\r\n");
> 
> causes the segment dump.

A update on this...

I took some time and tracked the problem down.  So far
I have exposed 2 problems in romcc with your code.

The first that was causing the core dump I have fixed.

The second is a little more insidious.  Basically
when call/return instructions are introduced the control
flow graph is no longer a graph.  Because the structure
changes depending on which path you take through the code.

What happens currently is I treat it like a graph and get
a conservative estimate on the control flow structure.  So no
incorrect transformations happen to the code.  However there are a
number of cases and your code was one where code that is correct
does not appear to be.  Because it appears that there are additional
paths to a basic block, when there are not.  

The result is some legitimate cases will not compile.  And likely
that some optimization opportunities are lost.  I suspect the lost
optimization opportunities are responsible for the code size growth.

Without calls and returns the dominance relationship between
instructions is transitive. Instruction X dominates instruction Y if
you must execute X before Y.  With call/return pairs the dominance
relationship is no longer quite transitive which is going to take
some delicate manipulation to let romcc understand this without
breaking it.

I am going to spend a little more time looking at this and when I am
done I am going to commit however many fixes/enhancements that I have.


Eric



More information about the coreboot mailing list