See the top level README for information on where to find documentation
for the raspberry pi and the ARM processor inside.  Also find information
on how to load and run these programs.

This example is for the pi2, see other directories for other flavors
of raspberry pi.  This example switches back to SVC mode from HYP mode.

This is an LED blinker example.

The pi2 has two LEDs tied to gpios 35 and 47.

Being the first example I will spend a little more time describing it.
I have some other ramblings on baremetal and the gnu tools so I will
try not to duplicate that.

The primary use case for the raspberry pi is to run some version of
linux.  The three main files to do that are bootloader.bin, start.elf
and kernel.img.  The first two being GPU programs the last being ARM.
If you have no other files (dont have a config.txt) they (starg.elf
GPU code) copy the kernel.img file to 0x8000 in RAM, place some
code at address 0x0000 that is intended to prepare the ARM for booting
linux, then branch to 0x8000.  We simply replace the kernel.img file
with our own program.  I prefer to not mess with config.txt stuff
because it is not the primary use case, most pis out there do not use
this, so without is significantly better tested.  Also over time the
config.txt things you can play with come and go and change name, some
of the popular ones are undocumented and so on.  So I really dont want
to rely on that.  Simply replace the kernel.img and otherwise be stock.

vectors.s is the entry point for this program, even an application on
an operating system like linux has some assembly up front before
calling the main function.  For this processor the minimum is to to
set up the stack pointer and call the main function.  Because some
compilers add extra stuff if they see a main() funtion I use some
function name other than main() typically for embedded systems like this.
I have adopted the habit of using notmain() both to not be named main()
and to emphasize this is bare metal and not your average application.

See my ramblings on .data and .bss, I dont need/use them so the
bootstrap (little bit of assembly before calling the first C function)
does not have to prepare those segments.  I only need to setup the
stack and call the first C function in the project.

I normally would set the stack pointer at the top of ram...Well that is
a lie, normaly one would do such a thing, but with code like this that
mostly ports across a number of boards, it becomes a pain keeping track
of who has how much ram.  Instead for simple examples I set the stack
somewhere where it doesnt collide with the code, but also I dont have to
change every board.  Because I am using the 0x8000 entry point I can
set the stack at 0x8000 and it will grow down toward 0x0000, and that
is more than enough for these projects.  The way the ARM works it
subtracts then writes stuff so the first thing on the stack will really
be at 0x7FFC, you dont have to set it to 0x7FFC to avoid the code at
0x8000.

The gpio pin is setup as an output to drive the LED.  The blink rate
appears to be around a couple-three times a second, but may vary based
on your compiler and settings.  This program does not attempt to use
any other peripherals (a timer) it relies on simply wasting time in a
loop then changing the state of the LED.  If you were to use this line
of code:

for(ra=0;ra<0x1000;ra++) continue;

The optimizer will replace that with this one assignment:

ra = 0x1000;

And actually since that value isnt used again, it is dead code the
optimizer will likely remove that as well.

One way to get around this is to make the loop variable volatile, this
tells the compiler every time you use it grab it from and save it back
to its home in ram.  I prefer a different approach.  I have a simple
dummy function in assembly language, it simply returns.

.globl dummy
dummy:
    bx lr

The assembly is outside the visibility of the optimizer as would
anything basically not in the same file (llvm is a little different it
might "see" those other objects and optimize across them, gnu wont).

So by having that external function and by passing the loop variable
to it.

for(ra=0;ra<0x1000;ra++) dummy(ra);

We force the compiler to actually implement this code and run that loop
that many times.  Dont need to declare the variable volatile.  If
uncomfortable with assembly langauge you could create a dummy function
in a separately compiled file

void dummy ( void )
{
}

Which produces the same code.

00000000 <dummy>:
   0:   e12fff1e    bx  lr

Some toolchains have the ability to see across objects when they
optimize so you still have to be careful.  I prefer the assembly approach
to defeating the optimizer.

So this program sets up the gpio to drive the LED.  Uses the loop to kill
some time, changes state of the LED, repeat forever.  The blink rate
will be the same for the same program.  But compiler differences and
options can cause one build to be different from another in the blink
rate.  It is not really deterministic, thus the desire to use timers in
the examples that follow.  If you change the number the loop counts to
and re-build with the same tools, you should see a change in the
blink rate.

Note the Broadcom documentation uses addresses 0x7Exxxxxx for the
peripherals. That is we assume the GPU's address space to access those
things.  The ARM window into that is to date either at 0x20xxxxxx or
0x3Fxxxxxx depending on the specific broadcom chip for that board type
the pi2 uses 0x3Fxxxxxx so when you see 0x7Exxxxxx replace that
0x7E with 0x3F.

I normally dont leave the compiled output in the repository, but you
may need it to compare with your results to see why for example mine
works and yours doesnt, so I will leave these here for this example.