Dear Alexander Holler,
In message 4D3AD19B.6030102@ahsoftware.de you wrote:
because I've recently seen some other places where volatile is used to access registers without using read?() or write?() and many people seem to start using 4.5.1, I want to post this warning using a descriptive subject.
The keyword volatile might not have any effect when reading when gcc 4.5.0 or 4.5.1 is used.
Actuelly this warning should eventually be extended to carefully check results when using any of the currently available 4.5.x versions of GCC.
The appropriate bug seems to be that one:
This is probably only one part of the problem. Thereare a number of other bugs in thes ecompiler versions, for example this one:
http://gcc.gnu.org/bugzilla/show_bug.cgi?id=44392
Unfortunately there appears to be little interest within the GCC folks to fix porolems that affect only such lunatic fringe groups as the embedded folks, so I don't expect to see fixes any time soon. Even bigger projects like the Linux Foundation driven Poky (resp. Yocto project) capitulated and stopped using -Os, see for example here: http://thread.gmane.org/gmane.linux.embedded.poky/2311/focus=2565
Yes, "-Os" is default setting for U-Boot - changing it to (for example) -O2 will generate significantly larger code and even break a number of boards.
It appears there are a number of other issues with GCC 4.5.x that affect us. If you are using any such tool chain, and see strange results or unexpected breakages, it is probably a very good idea to try another (older, working) tool chain first, before spending any tme on debugging.
You might either change such code to use readb(), readw() or readl(), or you have to use either an older version of gcc or a version >= 4.5.2 (it is fixed in 4.5.2).
We will take care to provide working (at least with most of the compiler versions) I/O accessors (including sufficient memory barriers), and all code that uses the old volatile pointer approach to access device registers should be converted to use proper I/O acessors.
A patch for write?() and read?() is currently in the u-boot-arm-repository (but not in the master and not in 2010.12):
http://lists.denx.de/pipermail/u-boot/2011-January/084885.html
Thanks for raising this issue again.
Best regards,
Wolfgang Denk