Memset and Memcpy are not safe on uint8 if unaligned bytes.

[tbandtg]

I’ve encountered an issue with the current implementation of memset and memcpy in the standard library on the pico. When the SCB_CCR_UNALIGN_TRP flag is enabled to trap unaligned accesses, both functions may internally perform word-sized operations on unaligned addresses—even when operating on uint8_t arrays. This results in a UsageFault exception on CPUs that do not support unaligned accesses. While this is allowed by the C standard, it is problematic for platforms enforcing strict alignment. To ensure compatibility with systems that enable UNALIGN_TRP, these functions need to handle unaligned pointers correctly, for example, by using byte-wise operations when the pointer is not word-aligned.

example code.

SCB->CCR |= ( SCB_CCR_DIV_0_TRP_Msk | SCB_CCR_UNALIGN_TRP_Msk | SCB_CCR_BFHFNMIGN_Msk);
uint8_t buffer[10];
memset(&buffer[1], 0, 4);
memset(&buffer[2], 0, 4);
DEBUG_PRINT("%d , buffer1 %d, buffer2 %d\n", buffer, &buffer[1], &buffer[2]);

[tbandtg]

add_compile_options(-mno-unaligned-access)

Added to cmake did not fix the issue.

[lurch]

I guess this must be on RP2350? Looking at the RP2350 datasheet says that CCR.UNALIGN_TRP is RW, but the RP2040 datasheet says that CCR.UNALIGN_TRP is RO and hardwired to 1. (With the latter being one of the unfortunate examples of where the Reset column in the auto-generated register tables doesn't know what the real reset value is. See also https://github.com/raspberrypi/pico-feedback/issues/241 )

[tbandtg]

IT is but if it is hardwired to 1 on the 2040 wouldnt this have caused alot of problems? I found this in btstack first and then ran the above test.

[tbandtg]

At the end of the day memset and memcpy should be safe to do on uint8 arrays even with the trap enabled.

[kilograham]

I'm leaving this around in 2.2.0 as we might replace the memset implementations, but we are picking them up from the compiler on RP2350 so this must be a common problem outside of pico-sdk

[danielecht]

I've actually run into this on RP2040. Misaligned memset cause hardfaults there. However I have only seen that happen in my own code, never in the CYW-Driver

[kilograham]

on RP2040 the SDK memset implementation (which actually calls the bootrom memset) should generally be used unless you are doing something ohter than a regular SDK CMake build perhaps. (the bootrom memset never does mis-aligned reads/writes)

[danielecht]

What do you mean by

the bootrom memset never does mis-aligned reads/writes

If my software calls memset with a unaligned address should it just result in a build error? I've seen hardfaults when calling memset on packed structs with misaligned fields.

I'm using the SDKs CMake.

I can try to provide a minimal example to produce a hardfault.

[kilograham]

Please do (the RP2040 bootrom version of memcpy/memset will do bytewise operation if the memory is not aligned)

[tbandtg]

Pls give an example on how to force it to link the memcpy and memset from the bootrom , does this also work on the rp2350.

[kilograham]

On RP2040 memcpy/memset should use the bootrom versions if you use a regular CMake SDK build On RP2350 they use the C library versions, because these are much more efficient than the ones for Arm Cortex M0+, however see #2340