I have two pieces of code that are very similar. Both of the code are compiled with the `-O0` flag in AC5. When I look at the assembly of the first code, I found that some variables loaded into registers would not be stored into RAM back. As show in the figure below for pOut2, pBias, pB, pB2, and pA2.
But in the assembly of the other code, all variables loaded into registers would be stored in RAM. As show in the figure below.
I don't want the assembly code store the register values into RAM, because the `STR` instruction accupies clock cycle. I want the compiler to maximize the use of registers. How can I achieve this?
Well, you could start by using a higher optimization level than -O0...The ARM only has so many registers, and if your function has too many local variables, it won't have any choice other than to put some of them on the stack. (your pictures don't show enough of the code to see if this is what is happening, though.)
Thanks for your reply. What puzzled me was why the assembly code of these two pieces of code were so inconsistent. So I can't compare program performance at the -O0 option. Higher optimization does improve register utilization, and the assembly code is more consistent.
I then tried compiling again using -O0 option in AC6, at which point both pieces of code compile into assembly code that store the register values into RAM. It seems that programs compiled with AC6 are more consistent.