Comparison of ARMCC, IAR and GCC compilers

I am working on one project that can hardly fit into 64k of microcontroller memory. And I thought it might be time to look at different compilers to choose which one can reduce the program size more aggressively.



I present to your attention a small comparison.



For the test, I made a project in Cube MX that includes USB_DEVICE and Mass Storage Class. These are pretty large libraries for testing.



The resulting main.c looks something like this:

int main() 
{
  HAL_Init();
  SystemClock_Config();
  MX_GPIO_Init();
  MX_USB_DEVICE_Init();
}

Define USBD_DEBUG_LEVEL is set to 0 so that USB stack debug messages do not require printf



Included compilers:

• IAR EWARM 8.32.1
• arm-none-eab-gcc 7-2018-q2-update (STM32 Cube IDE 1.4.2)
• ARMCC v5.06 update 7 (Keil uVision 5.32 environment)
• ARMCC v6.14.1 (Keil uVision 5.32 environment)

IAR settings:

• Optimized for size
• Run-time library NORMAL
• No low level IO (printf disabled)
• Linker optimizations enabled: Inline small routines, merge duplicate sections

GCC settings:

• Reduced runtime library --specs = nano.specs
• Optimize for size -Os
• Place functions in their own sections --ffunction-sections
• Place data in their own sections --ffdata-sections
• Discard unused sections -Wl, --gc-sections

Armcc5 settings:

• Use micro lib
• Use cross module optimization
• Optimization -O3
• One ELF section per function --split_sections

Armcc6 settings:

• Use micro lib
• Optimization image size -Oz
• One ELF section per function --split_sections

UPD

IAR has a multi-file compilation option. If you turn it on, then IAR goes ahead by a significant margin. The firmware is reduced to 12,746 bytes.



As you can see, armcc6 is slightly better than IAR. It is followed by armcc5, 4% behind, and gcc is 8% behind the leader.



It should be noted that the KEIL option "Use cross module optimization" significantly increased the compilation time, but did not nearly reduce the code size.