OpenAMP behaves differently for make and cmake when compiling for armv7-m

[raiden00pl]

There is no problems when building with cmake for armv7-m targets, but the same configuration doesn't work with make:

./tools/configure.sh nucleo-h745zi:nsh_cm7_rptun 
make -j
...
LD: nuttx                                                                                                                                                                               
arm-none-eabi-ld: /home/raiden00/git/RTOS/nuttx/nuttx/staging/libopenamp.a(remoteproc_virtio.o): in function `metal_io_read':
/home/raiden00/git/RTOS/nuttx/nuttx/include/metal/io.h:252: undefined reference to `__atomic_load_8'
arm-none-eabi-ld: /home/raiden00/git/RTOS/nuttx/nuttx/staging/libopenamp.a(remoteproc_virtio.o): in function `metal_io_write':
/home/raiden00/git/RTOS/nuttx/nuttx/include/metal/io.h:290: undefined reference to `__atomic_store_8'
make[1]: *** [Makefile:197: nuttx] Error 1

The problem doesn't occur for armv8-m targets:

./tools/configure.sh nrf5340-dk:rptun_cpunet
make -j
...
LD: nuttx                                                                                                                                                                             
CP: nuttx.hex
CP: nuttx.bin

[xiaoxiang781216]

it's strange that cmake and make generate the different result with the same config.

[raiden00pl]

Yes, the same config. I also replaced the generated .configs but the problem persists. So there must be a subtle difference in the build systems somewhere.

[anchao]

The cmake does not add the implementation of toolchain atomic support: #9760,
from the break log, maybe armv7-m does not support 64-bit atomic

[anchao]

I have a workaround for arch armv7-m to disable toolchain atomic:

$ git diff --cached 
diff --git a/openamp/libmetal.defs b/openamp/libmetal.defs
index d2df7b5b6a..8e906bd952 100644
--- a/openamp/libmetal.defs
+++ b/openamp/libmetal.defs
@@ -67,9 +67,14 @@ LIBMETAL_HDRS_SEDEXP := \
         s/@PROJECT_SYSTEM_UPPER@/nuttx/g; \
         s/@PROJECT_PROCESSOR_UPPER@/$(LIBMETAL_ARCH)/g; \
         s/@PROJECT_MACHINE_UPPER@/$(CONFIG_ARCH_CHIP)/g; \
+
+ifeq ($(CONFIG_ARCH_ARMV7M)$(CONFIG_ARCH_ARMV6M),y)
+LIBMETAL_HDRS_SEDEXP += \
         s/cmakedefine HAVE_STDATOMIC_H/include <nuttx\/compiler.h>/g; \
-        s/defined(HAVE_STDATOMIC_H)/defined(CONFIG_HAVE_ATOMICS)/g; \
-        s/cmakedefine/undef/g"
+        s/defined(HAVE_STDATOMIC_H)/defined(CONFIG_HAVE_ATOMICS)/g;
+endif
+
+LIBMETAL_HDRS_SEDEXP += s/cmakedefine/undef/g"
 
 ifeq ($(wildcard libmetal/.git),)
 libmetal.zip:

[raiden00pl]

This issue is not fixed.

[AndreHeinemans-NXP]

Is it possible that the cmake build uses different optimization settings? Maybe it is linked to this issue #12489.

The nrf5340-dk:rptun_cpunet builds well with make using -Os