CODEWITHSUNDEEP
cassemblycompiler-constructionatmegaatmelstudio
GPS
GPS

Reputation: 1382

AtmelStudio compiled code: How to optimize compiler

For the code:

uint8_t count;

ISR(TIMER1_OVF_vect, ISR_NAKED)
{
    count++;
    reti();
}

The generated assembly is:

--- F:\atmel-prj\compiler-test2\compiler-test2\Debug/.././compiler-test2.c -----
{
00000048  PUSH R1       Push register on stack 
00000049  PUSH R0       Push register on stack 
0000004A  IN R0,0x3F        In from I/O location 
0000004B  PUSH R0       Push register on stack 
0000004C  CLR R1        Clear Register 
0000004D  PUSH R24      Push register on stack 
    count++;
0000004E  LDS R24,0x0100        Load direct from data space 
00000050  SUBI R24,0xFF     Subtract immediate 
00000051  STS 0x0100,R24        Store direct to data space 
}
00000053  POP R24       Pop register from stack 
00000054  POP R0        Pop register from stack 
00000055  OUT 0x3F,R0       Out to I/O location 
00000056  POP R0        Pop register from stack 
00000057  POP R1        Pop register from stack 
00000058  RETI      Interrupt return

Now, as I see it, at least one, and up to two push-pop pairs could be eliminated:

  1. to save one push-pop, eliminate r24 by using r0 instead

    PUSH R1     Push register on stack 
PUSH R0     Push register on stack 
IN R0,0x3F      In from I/O location 
PUSH R0     Push register on stack 
CLR R1      Clear Register 
count++;
LDS R0,0x0100       Load direct from data space 
SUBI R0,0xFF        Subtract immediate 
STS 0x0100,R0       Store direct to data space 
...

  2. See that no code is using r1 for 0 value, so use only r1 for all purposes.

    PUSH R1     Push register on stack 
IN R1,0x3F      In from I/O location 
PUSH R1     Push register on stack 
CLR R1      Clear Register 
count++;
LDS R0,0x0100       Load direct from data space 
SUBI R0,0xFF        Subtract immediate 
STS 0x0100,R0       Store direct to data space 
...

Either of these save us precious bytes and microseconds.

Is there a way that I can put these or similar optimizations into the atmel studio toolchain/libraries somehow, so that my compiled code gets generated slightly better?
A lot of code surrounding interrupt and function calls, and some C to Assembly translations could be optimized a lot.

Upvotes: 0

Views: 406

Answers (1)

SKZ 81
SKZ 81

Reputation: 323

I also found out that generated assembly code could be optimized. I was not using -On option by that time. I found that the code of a called function (which takes a parameter in a register) ; copies it to the stack (to keep a clean copy as an 'automatic' C variable... which is in most case useless) AND FURTHERMORE copied back this value back to originating register !!! Whereas this register was just read before, and GCC should know it.

If you use Atmel Studio with default compiler, which is GCC, your binary code optimisation is depending on the toolchain behaviour, (and therefore, options specified on its command line).

To obtain a finest result, you can try to compile your source tree using avr-gcc directly (to setup options that Atmel Studio can't set for you).

Using another toolchain is another way to gain (theorically) finest results. Anyway, i doubt any other toolchain will give an (overall) better result than GCC does.

Upvotes: 1

Related Questions