Add configuration for IO flushing during exit.

[pan-]

Description

This PR allows application to enable or disable the flush of standard I/O's when exit is called.
This change can save a lot of space for applications which don't use standard stream because even if NDEBUG is enabled, a lot of the I/O subsystem is stilled pulled when the exit function invoke fflush over standard streams.

By default, the streams are still flushed but the user can override this behavior in the configuration file:

{
        "target_overrides": {
                "*": {
                        "core.stdio-flush-at-exit": false
                }
        }
}

Related PRs

With #2715, this PR allows user to completely remove the IO subsystem from their application.

Gain:

The test have been conducted upon mbed-os-example-blinky compiled with NDEBUG enabled.

• Without the patch:

	GCC	ARMCC	IAR
RAM	11832	10152	7970
ROM	37008	26868	21266

• With this patch and Allow the trace output by mbed error to be conditional of NDEBUG. #2715:

	GCC	ARMCC	IAR
RAM	11288 bytes	8020 bytes	7350 bytes
ROM	16640 bytes	12686 bytes	12398 bytes

• Difference:

	GCC	ARMCC	IAR
RAM	-544 bytes (4.6%)	-2132 bytes (-21%)	-620 bytes (7.8%)
ROM	-20368 bytes (-55%)	-14686 bytes (-52.8%)	-8868 bytes (-41.7%)

[pan-]

@0xc0170 @sg- Could you review this PR ?

[0xc0170]

The footprint results and preserving the current behavior +1.

error() function shouldn't call abort() ? If it does, then this config is not needed, as when assert/error happens, abort is called, and it's not expected to flush streams. A caller (mbed_error() in this case) would flush buffers that uses. However abort() would need to be defined, currently is not (I assume that's why exit is invoked).

What do you think ?

[sg-]

I think using the assert() -> abort() makes sense not flushing the streams and reserving error() and flushing the streams for non optimized builds which can be guarded by NDEBUG for the size reduction

@0xc0170 is this what you were suggesting?

[0xc0170]

I think using the assert() -> abort() makes sense not flushing the streams and reserving error() and flushing the streams for non optimized builds which can be guarded by NDEBUG for the size reduction

Yes, assert() -> abort(), and I would apply the same for error(). As both are error conditions (should not happen), and shall invoke abort. However, this might be considered breaking, we could flush stderr (the stream that is used in mbed_error, I believe it's stderr) prior calling abort (the other pull request protects that via NDEBUG, would be active only for debug builds). The result would be the almost the same as you described.

exit() however would stay as it is this change-set would not be needed anymore.

@pan- Does that sound feasible?

[pan-]

@sg- @0xc0170 assert() -> abort() makes a lot of sense, it is the behavior defined by the C and C++ standard!

I still see two issues:

• _exit: This is the routine we override for newlib, it flush IO but it shouldn't. This routine is the stub called directly by abort or the one called after at_exit processing and IO flushing by exit.
• exit is still referenced by the c runtime (crt0.s).

[sg-]

_exit: This is the routine we override for newlib, it flush IO but it shouldn't. This routine is the stub called directly by abort or the one called after at_exit processing and IO flushing by exit.

That seems like incorrect behavior and should be an easy fix

exit is still referenced by the c runtime (crt0.s)

given this is intended for the return value of main and even if main returns there is no where to go it seems we could patch this to use _exit or is that nonsense?

[0xc0170]

@pan- Let's resolve this one today (those 2 issues remaining about _exit, exit), I'll talk to you

[0xc0170]

To summarise the discussion with @pan- :

we overwrite _exit for gcc (it's called via exit()). If we remove flushing stdout/err, its not valid for other toolchains (quick look at IAR, app should overwrite exit, they don't provide similar hook as gcc does ,thus we use exit for it). I think this is the way to go for now to have this flushing configurable via config.

The above issues about assert and abort shall be handled separately.

@sg- What do you think?

[sg-]

/morph test

[sg-]

@mbed-bot: TEST

HOST_OSES=ALL
BUILD_TOOLCHAINS=ALL
TARGETS=ALL

[mbed-bot]

Result: SUCCESS

Your command has finished executing! Here's what you wrote!

/morph test

Output

mbed Build Number: 962

All builds and test passed!

[mbed-bot]

[Build 986]
SUCCESS: Building succeeded and tests were run! Be sure to check the test results