Issue with GCC in Release mode

[ReinhardKeil]

GCC fails in Release mode, see https://github.com/Arm-Examples/cmsis-mlek-examples/actions/runs/17127702948

This problem can be most likely solved by disabling Float 16 support using the define documented under
https://github.com/ARM-software/CMSIS-DSP?tab=readme-ov-file#half-float-support

However the problem should be also reported to the GCC team as it is most likely a compiler issue.

[jthuangarm]

Hi Reinhard,

During my investigation, I found that cbuild does not use the Config.cmake file included inside the CMSIS-DSP pack to exclude the F16-files. Instead, it relies entirely on the PDSC file to decide which source files are included in the build.

Because of this, even though I set DISABLEFLOAT16 in the compiler options, the StatisticsFunctionsF16.c file (including other F16 sources) is still pulled in by cbuild. This is expected, since the PDSC explicitly lists this file as part of the component.

Workaround:
Exclude the StatisticsFunctionsF16.c file in the .PDSC file, which is not ideal.
I will raise an issue to the GCC team to fix this.

Related issue: ARM-software/CMSIS-DSP#242

[Evagan2018]

Hi, I have reported this error to the GCC team.

[ReinhardKeil]

@christophe0606 can you help with a workaround.

[christophe0606]

Hi,

It is not a compiler problem. The problem is that when we compile with -ffast-math, NaN or INF are a problem. Same problem for AC6 for instance.

On AC6 the workaround is:

components:
- component: CMSIS:DSP&Source
misc:
- for-compiler: AC6
C-CPP:
- -Wno-nan-infinity-disabled

But latest versions of CMSIS-DSP have removed the use of infinity for f16 (it had already been removed from f32).

Instead I am using f16/f32 max/min positive values.