Author Topic: Porting projects to the i.MX RT  (Read 446 times)

Offline mark

  • Global Moderator
  • Hero Member
  • *****
  • Posts: 3188
    • View Profile
    • uTasker
Porting projects to the i.MX RT
« on: May 20, 2022, 08:07:16 PM »
Hi All

Due to the present processor shortage - especially Kinetis family parts - there is a lot of activity porting existing products to be able to run in i.MX RT parts, which tend to be more readily available.

During the transfer of code that has run correctly for years on the Kinetis (Cortex M0+, Cortex M4) there are some initial surprises when it hard faults on the Cortex M7 of the i.MX RT. The Cortex-M7 is very sensitive to alignment and I have found that the GCC compiler with high optimisation often generates code that access them with mis-aligned accesses and cause hard faults.

By reducing the optimisation of the compiler it often goes away and I also find that declaring variables that cause hard faults as volatile stops the compiler optimising and then corrects access alignment - from what I hear I expect you have also identified such.

I had a discussion with Eric Styger here (see discussion at the bottom) https://mcuoneclipse.com/2021/10/12/spilling-the-beans-volatile-qualifier/

where I believe that the use of volatile to solve such things (when they have been correctly understood) is valid. There are arguments about this being the wrong way but - as you can see - I think that it is valid in such cases and that is what I do use. You will find various such work-arounds in the code that were needed for the i.MX RT. For example, new routines
    21.12.2021 Add fnSafeGetBufLittleShort(), fnSafeGetBufLittleLong(), fnSafeGetBufBigShort(), fnSafeGetBufBigLong() {103}
    09.01.2022 Add fnSafePutBufLittleShort(), fnSafePutBufLittleLong(), fnSafePutBufBigShort(), fnSafePutBufBigLong() {104}
are used in some code and drivers to copy buffer content in a safe way.

Usually it is quite obvious that the hard faults are due to the compiler optimising accesses (trying to use 32 bit instead of 8 bit but accessing on unaligned boundaries) by looking at the disassembled code that caused the crash. Then adding the volatile keyword to the variables in question will show that it then doesn't optimise and the code works as expected.

It is a bit of a nuisance (IAR, for example, doesn't have this problem) but when you have experienced and solved it once or twice when moving existing code to the i.MX RT you will find that it is easy to recognise and solve so it is also not that serious.

Regards

Mark