[ARM] PR ld/20608

[Christophe Lyon <christophe dot lyon at linaro dot org>]

Hi,

In PR 20608, we have a case where a Thumb2 branch to PLT using b.w
targets the pre-PLT Thumb-ARM stub that is too far away.

Indeed, the linker is able to insert a pre-PLT Thumb->ARM stub to
handle Thumb mode calls through a PLT, but the check if a farcall stub
is needed doesn't take this into account correctly.

Depending on the branch-target distance, there is a 4 bytes windows
that is not handled properly: the closer range is handled by a direct
branch to the pre-PLT stub, while the further range is handle by a
farcall stub directly targeting the PLT. Between the 2 ranges, there
are 2 possible distances where a b.w instruction would result in a
"relocation truncated" error.

The attached patch fixes that by taking into account the actual final
distance when checking if a farcall stub is needed. I hope I updated
my comments in a clear way :-)

Regarding the testcase, I have added farcall-mixed-app2, which is a
copy of farcall-mixed-app with a new mid-range section right at the
offending frontier. I added a new testcase instead of updating this
existing one because adding b.w instructions enabled the linker to
replace bl by blx when needed, implying necessary modifications to the
expected results, and in effect removing tests for old architectures.

So now, we have farcall-mixed-app, farcall-mixed-app-v5 (the same one,
linked with --use-blx to allow the linker to replace bl with blx) and
farcall-mixed-app2 which uses b.w.

It seems this problem has been causing numerous random failures in
LLVM buildbots.

OK?

Thanks,

Christophe

Attachment: binutils-pr20608.log.txt
Description: Text document

Attachment: binutils-pr20608.patch.txt
Description: Text document