Make Alif examples work with LLVM

[ReinhardKeil]

Currently the projects in https://github.com/Arm-Examples/cmsis-mlek-examples/tree/main/Alif only work with AC6.

We should investigate what is required to make this projects work with LLVM (and optionally GCC).
The cdefault.yml configurations have primarily a copy of the AC6 settings - which is the first change that is required.

[BhupeshKharwa-AlifSemi]

Hi @ReinhardKeil ,

Do you have a tentative timeline for the availability of this.

Thank you,
Bhupesh

[jthuangarm]

The issues with the GNU compiler have been resolved. To fix them, we modified the compiler controls in cdefault.yml and uploaded the custom linker script linker_gnu_mram_custom.ld.src under ./board/AppKit-E7_M55_HE/ or AppKit-E7_M55_HP/. Theses custom linker script are specified in the clayer.yml files in each example.

Issues with the LLVM compiler are still under investigation.

Here’s a summary of what went wrong and how it was resolved for the Audio example. Similar fixes have also been applied to other examples.

GCC

1. error: incompatible types when assigning to type 'uint32x4_t' from type 'uint16x8_t'
   The compile errors with Helium (MVE) intrinsics came from GCC’s stricter vector type rules.
   To resolve: Enabling -flax-vector-conversions in cdefault.yml relaxed those checks so the Helium code built.

2. section nn_model' will not fit in region ITCM'
section .bss' will not fit in region DTCM'
section .stack VMA [2003e000,2003ffff] overlaps section .bss VMA [20001a00,20041bd7]
region ITCM' overflowed by 25024 bytes
   The link failures were due to memory placement. The nn_model section was being placed into ITCM, overflowing it by ~25 KB. At the same time .bss and the stack were mapped to DTCM and collided.
   To resolve: Resolved this by editing the GNU linker script. I added a .nn_model > MRAM output section. I also relocated general .data/.bss and heap/stack to SRAM0, keeping DTCM clear for small hot data only.

3. nullptr issue in Labels.cpp
   Labels.cpp contains pointers requiring explicit initialization at runtime, and if placed in ITCM, results in null pointers or incorrect data references.
   To resolve: Explicitly ensure label pointers are placed in the .labels_ro and .labels_data > MRAM section.

4. Cannot stop at breakpoint (BP) if .text is in ITCM
   Debugger sets breakpoints immediately after reset, but .text code must be copied from MRAM to ITCM first. Breakpoints fail if ITCM isn't initialized yet.
   To resolve: Remove the .nn_model section from point 2 and keep entire .text in MRAM (including functions like main and app_main) directly, so breakpoints work as expected.

[ReinhardKeil]

@jthuangarm thanks for fixing the GCC issues. It works for Debug but unfortunately Release results in an error.

[chentang16]

I have managed to build and run the Tensorflow_LiteRT_HelloWorld example on Alif AppKit-E7 using Arm LLVM clang v21.1.0. Here are the changes I have made to make it work:

1. The system.h file located in the DFP pack ...\AlifSemiconductor\Ensemble\2.0.2\Device\core\common\include needs to be patched, namely from the line 83 to the line 111 need to be commented out, which results in a much simplified __clang_copy_zero_init() as __STATIC_FORCEINLINE __NO_RETURN void __clang_copy_zero_init(void) { extern void _start(void) __NO_RETURN; _start(); }
   Otherwise, you got multiple linker errors "ld.lld: error: undefined symbol: copy_table_end" etc., since copy_table_end() or zero_table_start() etc. are not implemented in picoclib.
2. Add two additional linker option -nostartfiles -lcrt0 either to cdefault.yml or to csolution.yml . See https://developer.arm.com/documentation/107976/20-1-0/Get-started/Run-bare-metal-images?lang=en
3. Add an additional linker option -Wl,--entry=Reset_Handler
4. Increase the stack size to 0x00015000 and the heap size to 0x00010000 in regions_AppKit-E7-AIML.h
5. Add an additional line PROVIDE_HIDDEN ( __bothinit_array_start = . ); /* NEW */ before the line PROVIDE_HIDDEN ( __preinit_array_start = . ); and an additional line PROVIDE_HIDDEN ( __bothinit_array_end = . ); /* NEW */ before the line PROVIDE_HIDDEN ( __fini_array_start = . ); in the linker script clang_linker_script.ld.src

[chentang16]

I have managed to build and run the MLEK_Audio example on Alif AppKit-E7 using Arm LLVM clang v21.1.1 (Yes, meanwhile there is already a new toolchain release). Here are the changes I have made to make it work:

1. Apply the same change as 1), 2), 3) and 5) from the case above of the Tensorflow_LiteRT_HelloWorld example
2. Increase the stack size to 0x00002000 and the heap size to 0x00008000 in regions_AppKit-E7-AIML.h, which is the same stack/heap size as the case with AC6
3. Move the call of the function osKernelInitialize(); from the app_main() in main_kws.cpp to the line before NpuInit(); in main.c. In other words, the CMSIS RTOS kernel needs to be initialized before calling the Arm EthosU driver function ethosu_init(). Otherwise, the EthosU driver initialization will fail with the error "Failed to create global driver mutex (ethosu_driver.c:427)"

[chentang16]

I have managed to build and run the MLEK_Video example on Alif AppKit-E7 using Arm LLVM clang v21.1.1. Here are the changes I have made to make it work:

1. Perform the same changes for the MLEK_Audio example as described above
2. Change the __ROM0_BASE value in regions_AppKit-E7-AIML.h as #define __ROM0_BASE 0x80200000, since this example will run by the HP core

[chentang16]

I have managed to build and run the MLEK_AV_DualCore example on Alif AppKit-E7 using Arm LLVM clang v21.1.1. Here are the changes I have made to make it work:

1. Perform the same changes 1) and 2) mentioned for the case with the MLEK_Video example above
2. In the linker script template file clang_linker_script.ld.src for the MLEK_Keyword_Spotting project (HE core), replace all RAM0 usage with RAM1, so that MLEK_Keyword_Spotting project (HE core) and MLEK_Object_Detection (HP core) project will use two different/separate RAM regions.