Re: [microblaze-uclinux] microblaze+MMU optimization issues

[John Williams <jwilliams@xxxxxxxxxxxxxx>]

Hi Steve,

Steven Hein wrote:

> I've been trying to characterize what kind of performance
> we'll see from our Microblaze+MMU design.   As part of this,
> I've been trying to play with higher levels of optimization
> in kernel and user packages.   I've noticed that I can't use
> -O2 optimization in most user apps that use fork().  (I've
> observed this with the busybox, sysvinit, and dropbear
> packages).
>
> I'm sure I need to provide more detail on this....but what
> I have observed is when I fork() is called on an app compiled
> with -O2, the child process has garbage in many of the
> variables that it inherits from the parent.    When I run
> with -O (or in some cases -Os), then everything works fine.
> I haven't been able to reproduce the symptom with a simple
> piece of test code at this point.    Is this behavior
> something that you have seen?    (I see that most components
> of the release are not built with -O2, but rather -Os).

Ouch.  No, I've not seen this, or at least not characterised any issues 
down to this cause.

It's possible that this garbage is in variables that have been mapped to 
registers - in which case it would point to a problem in the vfork / 
clone task handling.

If the variables are memory-bound then we'd be looking for a problem 
with the MMU handling and page table twiddling duing fork.

> If so, is it something that I can set -O2 and disable a
> particular optimization?    Also, I saw a previous email
> from you that said you were testing a new compiler/toolchain--
> does this newer version resolve the problem?

We'd need to look at the assembly produced by gcc - obviously a minimal 
test case would help there.

Doing

$ mb-linux-objdump -S myapp > myapp.dump

on the two different versions would be the place to start.

If you can distill this down to a minimal test case it will make it much 
easier to track down.

> Also.....are there any plans for being able to build
> libraries that are optimized for a particular design
> (i.e. with HW multiplication or division enabled)?
> Is the source for the toolchains available  (i.e. to allow
> them to be built via crosstool)?

I've got an updated toolchain I'm just testing now that has a few 
different glibc builds for the different optimisations.

I had hoped to merge the MMU support directly into the next PetaLinux 
release but it's looking like another interim -mmu test cut might be 
more appropriate to get some of the recent work better tested.

John

___________________________
microblaze-uclinux mailing list
microblaze-uclinux@xxxxxxxxxxxxxx
Project Home Page : http://www.itee.uq.edu.au/~jwilliams/mblaze-uclinux
Mailing List Archive : http://www.itee.uq.edu.au/~listarch/microblaze-uclinux/