[AVR] Jumps are off by 2

[Patryk27]

Hi,

Ever since 6859685 (or, more precisely, 84428da), the AVR backend encodes jumps across basic blocks with an extra offset of 2.

That's related to this check:

llvm-project/llvm/lib/Target/AVR/MCTargetDesc/AVRAsmBackend.cpp

Line 521 in 1919db9

if (const auto *SymA = Target.getSymA())

... which wasn't triggered for basic blocks before, but it is now (i.e. shouldForceRelocation() returns false for such jumps now).

I'm working on a pure LLVM reproducer, but the original rustc issue has been reported here - if you compare the *.elf from older toolchain and the newer toolchain, you'll see that calls are alright, but jumps have an extra offset.

[Patryk27]

Alright, adding Value -= 1; here:

llvm-project/llvm/lib/Target/AVR/MCTargetDesc/AVRAsmBackend.cpp

Line 89 in 57cd100

static void adjustRelativeBranch(unsigned Size, const MCFixup &Fixup,

... seems to solve the problem, but is it the correct solution?

[Patryk27]

Issue has been already solved, just forgot to add closes xyz to the commit message.

[benshi001]

Though 86a60e7 has been merged, the fix is incorrect, since the llvm-objdump still has problem, as I have indicated at #102936 (comment)

[benshi001]

@Patryk27 If you are busy, please assign this issue to me. I will take a time to find a better solution later this month.

[Patryk27]

Ah, right, I forgot about it - feel free to take a stab at it.

[benshi001]

@Patryk27 Your fix is good! I have made a mistake, and I have confirmed the output of both llvm-mc -show-encoding and llvm-objdump -d are correct by your patch

[Patryk27]

Cool, thanks for checking! 🙂

[benshi001]

@Patryk27 Though your fix is good, there is still an underlying issue, which I have fixed by #109124. You are appreciated to have a look.