Hi,
Compiler I'm using is a 5.10 alpha - will try a release version if needed but suspect this isn't something new
I think the problem I'm having is close to this thread's issue: http://e2e.ti.com/support/development_tools/compiler/f/343/t/213287.aspx but I'm after a different result and maybe it's a more difficult result.
For benchmarking, we have a built-in cycle counter in the Cortex's PMU unit, and an assembly function to access this is:
_pmuGetCycleCount_:
MRC P15, #0, R0, C9, C13, #0
BX R14
Now the function call really isn't necessary since the function itself is only one line of code. I need to put two of these function calls
around a small body of code (about 5 instructions) that are accessing flash memory, and I don't want to have the error / overhead of the function call itself
(the branches) in the result.
I couldn't find any sort of directive in assembly to tell the compiler that an assembly function may be inlined.
So, I moved the function to 'C' like this:
unsigned int _pmuGetCycleCount_(void) {
asm(" mrc p15, #0, r0, c9, c13, #0");
}
and use it like this:
cyclecount[0] = _pmuGetCycleCount_();
FunctionToBenchmark();
cyclecount[1] = _pmuGetCycleCount_();
Now, when the above function is called as is the compiled code is fine and the value returned in R0 is saved in a temporary register and then finally written to memory correctly.
Adding the keyword 'inline' - the mrc instruction is inlined and there isn't any function call.
However, the compiler then writes over the result stored in R0; and doesn't save the result to memory anymore.
The last observations seems like it is a compiler issue - why would the compiler know to save the result of the function to cycle_count[0] and [1] when
not inlined, but then throw away the result when the function is inlined?
-Anthony
PS: I think it would be a lot simpler if there were generic intrinsics for the mcr, mrc instructions, not just the specific intrinsics hardcoded to interrupt enable/disable.
There are plenty of these coproessor registers; and for example this thread:
http://e2e.ti.com/support/development_tools/code_composer_studio/f/81/t/103040.aspx
Could have benefited too if the intrinsic allowed the complicated mrc ...
