Quuxplusone
commented
5 years ago Thanks for the report and the suggested, LLD has tended to focus on Arm v7a for
Linux and tends to add new functionality only when people need it. Please do
keep reporting problems as you find them.
We should be able to take it from here. With luck I'll have time to do it this
week, but it may slip till after Christmas due to holidays. If you want to go
ahead and submit it yourself there are instructions at
https://llvm.org/docs/DeveloperPolicy.html#making-and-submitting-a-patch
If you do, you would probably want something like the following sequence, which
preserves the ABI required 8-byte stack alignment, and is a bit smaller.
push {r0, r1}
ldr r0, [pc, #4]
str r0, [sp, #4]
pop {r0, pc}
.word DESTINATION
Adding tests is fairly straightforward and can generally be worked out by
following the examples in lld/test/ELF , look for the arm-thunk*.s files,
particularly those that use a linker script. What we'd do is add a new .s file
for the v6-m thunks run lld and then use llvm-objdump to check that the thunks
we'd expected had been created. The FileCheck program tests the output
https://llvm.org/docs/CommandGuide/FileCheck.html
ARMv6M does not support movw/movt Thumb2 instructions.
Current LLD version v7.0 emits such instructions to generate long jump in thunks. For example, this code generation occurs when code is relocated from flash to RAM on embedded Cortex-M0/M0+ cores.
This prevents using LLD with Cortex-M0/M0+ cores when such a feature is used.
I've written a small patch to support this feature, but I'm not a LLD developer and I do not know how to write valid tests for the LLVM infrastructure, so I guess this would require more work effort to propose a valid patch:
https://gist.github.com/eblot/43552f8c01cc7d2ee4faef42454c2c83
It only support absolute addressing, not position-independent code - I'm not sure this is a useful feature on M0+ core anyway and I would not have any real world examples to test it.