Greetings,

In order to reduce the binary size of my projects, I am using the following flags: CFLAGS= ... -Os -ffunction-sections -fdata-sections -flto ... LFLAGS= ... -Os -Wl,--gc-sections -flto ...

When building with those flags with gcc version 4.8.3 20140228 (release) [ARM/embedded-4_8-branch revision 208322] from Linaro, several variables (pxCurrentTCB) and functions (vTaskSwitchContext) get deleted because they are not referenced from C code, but from assembler code that is passed verbatim to the compiler, without disclosing the global references.

Not knowing all the magic options to the GCC embedded assembler, I got it to link by extracting vPortSVCHandler and xPortPendSVHandler functions into an assembler file . On my small examples that uses the STM32F4 cube libraries to blink some LEDs, I get a text size reduction from 9060 to 6740, only by adding -flto to the flags.

Best, Sign Bit

My 'porta.S' file is attached. Since the code was copied technically verbatim from port.c, the same copyright applies.

I regularly use the -ffunction-sections, -fdata-sections and --gc-sections options with and without optimisation, but have never experimented with -flto. It would be interesting to give it a try.

Compiler writers always tell me they hate inline assembler because it messes up their optimisers :o) In this case it may be possible to prevent the symbol being removed by accessing the symbol from code that will never get executed, but the compiler will not know does not get executed. I use that technique in some ports already (RX with the Renesas compiler for example). In the GCC Cortex-M port the function prvPortStartFirstTask() will never return because form that point on the scheduler is running, so code added after the call prvPortStartFirstTask() inside xPortStartScheduler() would hopefully (assuming volatile accesses) let the compiler know you don't want the symbols to be removed. In fact, looking at it know, there is already a call to prvTaskExitError() made after prvPortStartFirstTask() with a comment saying it is done purely to prevent compiler warnings about a static function not being called (the address of the function is placed on the task's stack, but the function is never called directly).

Regards.

I expect as the compilers and linkers get smarter, the value of LTO and WHOPR to just increase and help us cram more functionality in fewer bits.

Sure, we can trick the compiler to accept our "portable" code choke-full of assembler, or we can just export those pure assembler functions into a separate module. That code is clearly not portable anyway.

Best, Sign Bit