[lld] [ELF] Change Ctx::target to unique_ptr (PR #111260)

[Fangrui Song via llvm-commits]

MaskRay wrote:

> > > We also see a weird issue on Windows is that during clang test step, lld timed out: `Linking CXX executable unittests\Transforms\Scalar\ScalarTests.exe` took 36 minutes and caused the bot timed out. Your change is the only on the blamelist.
> > 
> > 
> > @zeroomega is this repeatable? We haven't seen this on our windows builds.
> > EDIT - we're not seeing any timeouts.. but I will check for excessive test times.
> 
> So I did a few test runs locally.
> 
> The test timeout (I should say lit probably hanged) on test `ELF/aarch64-cortex-a53-843419-thunk-align.s` If I run `ninja check-lld` with "-DLLVM_LIT_ARGS=-v -j64" , I have a high chance that the `check-lld` will hang. The host machine is a 16 cores 32 threads machine. However, if I reduces the -j to 16 or even 1, the test will finish with error on `lld :: ELF/aarch64-thunk-bti.s`. Comparing finished test run with the one that hang, I discovered that `ELF/aarch64-cortex-a53-843419-thunk-align.s` was not on the list of finished test when `-j64` was used. The hang issue resolved itself after this patch was reverted. Could it be some deadlock issue introduced to lld in this patch?

Thanks for the investigation. The output file  `%t2` from `test/ELF/aarch64-cortex-a53-843419-thunk-align.s` is 134MiB. It 
 `trapInstr` is non-zero (used of uninitialized memory as @smithp35 noticed), the non-zero values will be part of llvm-objdump -d output, which will be slow.

The test currently relies on `llvm-objdump -d `'s zero skipping feature (unless -z is specified). Ideally the test should be changed not to create such a huge output.

https://github.com/llvm/llvm-project/pull/111260