Regression with gnu-tools-for-stm32 14.3.1+st.2

[parmi93]

We are working on a project that uses MbedTLS v4.0.0, which contains some ASM code, and when we try to compile the project with gnu-tools-for-stm32 14.3.1+st.2, the compilation fails with the following error:

[build] FAILED: [code=1] Middlewares/Third_Party/mbedtls/tf-psa-crypto/drivers/builtin/CMakeFiles/builtin.dir/src/bignum_core.c.obj 
[build] C:\Users\Parminder\AppData\Local\stm32cube\bundles\gnu-tools-for-stm32\14.3.1+st.2\bin\arm-none-eabi-gcc.exe -DDEBUG -DESC_TRL_REMOVE_DBG -DESC_USER_BUILD_CONFIG_CUSTOM=SC_vendor_platform.h -DSTM32L475xx -DTF_PSA_CRYPTO_CONFIG_FILE=\"C:/smartconnect-h300/H300/mbedtls_porting/crypto_config.h\" -DUSE_HAL_DRIVER -IC:/smartconnect-h300/Middlewares/Third_Party/mbedtls/tf-psa-crypto/drivers/builtin/src -IC:/smartconnect-h300/Middlewares/Third_Party/mbedtls/tf-psa-crypto/drivers/builtin/include -IC:/smartconnect-h300/Middlewares/Third_Party/mbedtls/tf-psa-crypto/include -IC:/smartconnect-h300/Middlewares/Third_Party/mbedtls/tf-psa-crypto/core -IC:/smartconnect-h300/Middlewares/Third_Party/mbedtls/tf-psa-crypto/drivers/everest/include -IC:/smartconnect-h300/Middlewares/Third_Party/Eseye/SMARTconnect/core -IC:/smartconnect-h300/H300 -IC:/smartconnect-h300/cmake/stm32cubemx/../../Core/Inc -IC:/smartconnect-h300/cmake/stm32cubemx/../../Drivers/STM32L4xx_HAL_Driver/Inc -IC:/smartconnect-h300/cmake/stm32cubemx/../../Drivers/STM32L4xx_HAL_Driver/Inc/Legacy -IC:/smartconnect-h300/cmake/stm32cubemx/../../Middlewares/Third_Party/FreeRTOS/Source/include -IC:/smartconnect-h300/cmake/stm32cubemx/../../Middlewares/Third_Party/FreeRTOS/Source/CMSIS_RTOS_V2 -IC:/smartconnect-h300/cmake/stm32cubemx/../../Middlewares/Third_Party/FreeRTOS/Source/portable/GCC/ARM_CM4F -IC:/smartconnect-h300/cmake/stm32cubemx/../../Drivers/CMSIS/Device/ST/STM32L4xx/Include -IC:/smartconnect-h300/cmake/stm32cubemx/../../Drivers/CMSIS/Include -mcpu=cortex-m4 -mfpu=fpv4-sp-d16 -mfloat-abi=hard -mthumb -Wall -fdata-sections -ffunction-sections -fstack-usage -fcyclomatic-complexity -mcpu=cortex-m4 -mfpu=fpv4-sp-d16 -mfloat-abi=hard -mthumb -Wall -fdata-sections -ffunction-sections -fstack-usage -fcyclomatic-complexity -O0 -g3 -std=c99 -Wall -Wextra -Wwrite-strings -Wmissing-prototypes -Wformat=2 -Wno-format-nonliteral -Wvla -Wlogical-op -Wshadow -Wformat-signedness -Wformat-overflow=2 -Wformat-truncation -O0 -g3 -Werror -Wmissing-declarations -MD -MT Middlewares/Third_Party/mbedtls/tf-psa-crypto/drivers/builtin/CMakeFiles/builtin.dir/src/bignum_core.c.obj -MF Middlewares\Third_Party\mbedtls\tf-psa-crypto\drivers\builtin\CMakeFiles\builtin.dir\src\bignum_core.c.obj.d -o Middlewares/Third_Party/mbedtls/tf-psa-crypto/drivers/builtin/CMakeFiles/builtin.dir/src/bignum_core.c.obj -c C:/smartconnect-h300/Middlewares/Third_Party/mbedtls/tf-psa-crypto/drivers/builtin/src/bignum_core.c
[build] In file included from C:/smartconnect-h300/Middlewares/Third_Party/mbedtls/tf-psa-crypto/drivers/builtin/src/bignum_core.c:8:
[build] C:/smartconnect-h300/Middlewares/Third_Party/mbedtls/tf-psa-crypto/drivers/builtin/src/bignum_core.c: In function 'mbedtls_mpi_core_mla':
[build] C:/smartconnect-h300/Middlewares/Third_Party/mbedtls/tf-psa-crypto/core/tf_psa_crypto_common.h:321:13: error: 'asm' operand has impossible constraints or there are not enough registers
[build]   321 | #define asm __asm__
[build]       |             ^~~~~~~
[build] C:/smartconnect-h300/Middlewares/Third_Party/mbedtls/tf-psa-crypto/drivers/builtin/src/bn_mul.h:785:9: note: in expansion of macro 'asm'
[build]   785 |         asm volatile (
[build]       |         ^~~
[build] C:/smartconnect-h300/Middlewares/Third_Party/mbedtls/tf-psa-crypto/drivers/builtin/src/bn_mul.h:1082:25: note: in expansion of macro 'MULADDC_X2_INIT'
[build]  1082 | #define MULADDC_X4_INIT MULADDC_X2_INIT
[build]       |                         ^~~~~~~~~~~~~~~
[build] C:/smartconnect-h300/Middlewares/Third_Party/mbedtls/tf-psa-crypto/drivers/builtin/src/bn_mul.h:1088:25: note: in expansion of macro 'MULADDC_X4_INIT'
[build]  1088 | #define MULADDC_X8_INIT MULADDC_X4_INIT
[build]       |                         ^~~~~~~~~~~~~~~
[build] C:/smartconnect-h300/Middlewares/Third_Party/mbedtls/tf-psa-crypto/drivers/builtin/src/bignum_core.c:481:9: note: in expansion of macro 'MULADDC_X8_INIT'
[build]   481 |         MULADDC_X8_INIT
[build]       |         ^~~~~~~~~~~~~~~

However, if we switch the toolchain to gnu-tools-for-stm32 13.3.1+st.9 (via the VSCode Bundle Manager UI), it compiles without any problems.

It looks like that this is the piece of code that is causing the issue 

https://github.com/Mbed-TLS/TF-PSA-Crypto/blob/76920edddcad00ac41b248e12d937b845df7bedb/drivers/builtin/src/bn_mul.h#L785-L818, but I'm not sure if this is really a issue with the toolchain or an issue of MbedTLS.

        asm volatile (

            /* - Make sure loop is 4-byte aligned to avoid stalls
             *   upon repeated non-word aligned instructions in
             *   some microarchitectures.
             * - Don't use ldm with post-increment or back-to-back
             *   loads with post-increment and same address register
             *   to avoid stalls on some microarchitectures.
             * - Bunch loads and stores to reduce latency on some
             *   microarchitectures. E.g., on Cortex-M4, the first
             *   in a series of load/store operations has latency
             *   2 cycles, while subsequent loads/stores are single-cycle. */
#define MULADDC_X2_CORE                                           \
           ".p2align  2                                   \n\t"   \
            "ldr      %[a0], [%[in]],  #+8                \n\t"   \
            "ldr      %[b0], [%[acc]], #+8                \n\t"   \
            "ldr      %[a1], [%[in],  #-4]                \n\t"   \
            "ldr      %[b1], [%[acc], #-4]                \n\t"   \
            "umaal    %[b0], %[carry], %[scalar], %[a0]   \n\t"   \
            "umaal    %[b1], %[carry], %[scalar], %[a1]   \n\t"   \
            "str      %[b0], [%[acc], #-8]                \n\t"   \
            "str      %[b1], [%[acc], #-4]                \n\t"

#define MULADDC_X2_STOP                                      \
            : [a0]     "=&r" (tmp_a0),                       \
              [b0]     "=&r" (tmp_b0),                       \
              [a1]     "=&r" (tmp_a1),                       \
              [b1]     "=&r" (tmp_b1),                       \
              [in]     "+r"  (s),                            \
              [acc]    "+r"  (d),                            \
              [carry]  "+l"  (c)                             \
            : [scalar] "r"   (b)                             \
            : "memory"                                       \
        );   

[Julien D · Accepted solution]

This issue relates to https://github.com/Mbed-TLS/mbedtls/issues/9875 and https://gcc.gnu.org/bugzilla/show_bug.cgi?id=115042. The latter having been fixed in the mainstream branch, it will be integrated into one of the upcoming toolchain releases.

In the meantime, you can ensure that an optimization level of -O1 or higher is set for this source file.