I'm abandoned this change. But I don't think this problem is over yet...
On 14 April 2014 19:43, Ard Biesheuvel ard.biesheuvel@linaro.org wrote:
On 14 April 2014 19:32, Leif Lindholm leif.lindholm@linaro.org wrote:
On Mon, Apr 14, 2014 at 02:03:29PM +0100, Ryan Harkin wrote:
I'll be changing uefi-build.sh soon so that it uses the bare-metal
toolchains.
I haven't pushed the change yet because I'm still testing stuff.
Once I make the change, you will need to have the bare-metal toolchain installed or you won't be able to build any more.
So let me know if this will be a problem!
This will be a problem.
Some background:
OK, so I have to confess that I've been using the wrong toolchain for
quite
some time now.
I already have the linux-gnu- toolchain installed and it's always
worked, so
I've been generally happy to keep using it. Whenever I get something
new from
ARM (UEFI, TF, ...), it usually tells me to use the bare-metal
toolchain. I
follow the instructions first time, then repeat with my "normal"
toolchain.
However, I've now come across a problem where building for an AArch64
platform
with the linux-gnu- platform toolchain doesn't boot, whereas using the bare-metal toolchain is fine.
This is a bug in EDK2 and should be fixed there. Not defaulting to building with embedded toolchains is a good thing which lets us spot things like this.
It would be helpful to have a bit of background on how the difference between bare metal and glibc toolchain causes this particular issue.
Yes, it would :-/
I haven't a clue how it causes it, I only know that it happens.
If someone wants to explicitly build with an embedded toolchain, they can do that just fine using CROSS_COMPILE/CROSS_COMPILE64.
Of course, I have never really known what the difference between the
two is.
Enter the toolchain FAQ:
https://wiki.linaro.org/WorkingGroups/ToolChain/FAQ
"The bare-metal ABI will assume a different C library (newlib for
example) - or
even no C library to the Linux ABI (which assumes glibc). Therefore, the compiler may make different function calls depending on what it
believes is
available above and beyond the Standard C library.
This effectively describes the difference between -fhosted and -ffreestanding. In other words, you can already make glibc gcc behave as bare metal gcc (in this respect) by passing a single -f flag, and gcc even has its own set of headers (e.g. stdint[-gcc].h) to use if -ffreestanding prevents it from using the ones provided by the C library.
Also the bare-metal ABI and Linux ABI for the 32-bit Instruction sets
make
different assumptions about the storage size of enums and wchar_t which
you
have to be careful of (not a complete list). And the difference between
the
32-bit and 64-bit ABIs are also numerous and subtle (the obvious
example being
pointer sizes)."
I understand this is a copy and paste job, but how would the difference between 32 bit and 64 bit ABIs relate to the difference between bare metal and glibc?
I suppose that's a question for the toolchain team.
... ooops!
I am aware of the purist toolchain engineer view on this, but in the real world, people will be using the default toolchain on their system.
Perhaps even push back a bit? I mean, what ffs is the point of, for instance, defining uint32_t as 'unsigned int' on glibc for arm32 and 'unsigned long' on bare metal for arm32. If these things are causing us grief, we should report them to the toolchain group as bugs.
I agree: why must the types be different. I understand the libraries and
function call being different, but type sizes should be fixed. How on earth do we cope with that?? But whether it's a bug or a feature is a different matter.
We must not cripple this reality by trying to conform to an idealistic fantasy.
I think the problem is deeper than this. The real world you talk about is the current reality, where we are depending on UEFI and the kernel being built with the same toolchain and where we build them ourselves.
The other real world is one where UEFI will be provided by the vendor and the kernel by a distro. Neither of which will be under our control or build environment. So depending on the toolchain to sort this stuff out isn't going to fly, surely? I'm glad it's not my problem to sort out ;-)
Imagine what happens when a BIOS vendor builds UEFI using the ARM compiler. Or, more likely: a Microsoft compiler.
So ... could you raise a defect on the issue you've seen?
Sure. Were you thinking a launchpad bug, or raising with ARM? Or even one of those JIRA things in the Landing Team space?
At the moment, the platform I am working on is using a binary build of UEFI supplied by ARM; that's the only one that works. I have no control over what toolchain that's built with, of course. This is my alternate reality.
/ Leif
Linaro-uefi mailing list Linaro-uefi@lists.linaro.org http://lists.linaro.org/mailman/listinfo/linaro-uefi