add trace support for AArch64

[fhahn]

Similar to #1668 for AArch64. There is a bunch of x86-specific trace building code that needs to be ported.

[fhahn]

@egrimley shared a draft patch at #2442

[Vincent-lau]

Hi I am working on a project to add trace support for AArch64, I am wondering if you have any suggestions on what would be a good starting point?

[AssadHashmi]

Hi I am working on a project to add trace support for AArch64, I am wondering if you have any suggestions on what would be a good starting point?

Have a look at the draft patch #2442 and see if it can be updated to the latest master.

[Vincent-lau]

Hi, I am planning to submit a patch to add trace support for AArch64, but I am not yet a member of DynamoRIO, shall I submit a few small patches by forking the repo before asking to be added as the member?

[AssadHashmi]

Hi, I am planning to submit a patch to add trace support for AArch64, but I am not yet a member of DynamoRIO, shall I submit a few small patches by forking the repo before asking to be added as the member?

There's no need to fork, I've added you as a contributor. You should receive an invite message from GitHub.
Please follow the advice on contributing here https://dynamorio.org/page_code_reviews.html
Thanks for contributing, we look forward to your pull request.

[derekbruening]

Documenting a proposal that eliminates the internal control flow of fixup sequences, which break the linear assumptions of translation:

Pasting from https://github.com/DynamoRIO/dynamorio/pull/5045/files/b9ada873ecc0b86128169aae0bd7ad648c0f921b..fc70e80bad8a66b7f417c27673fa5659e96ff1ef#r697566139

One proposal is to change the i.b. comparison to use x2 instead of x0 as the temp reg holding the expected target immediate. Then the only thing missing is copying from the actual target reg into x2, and maybe that could be done locally. It would use a taken forward branch on the match which maybe is not ideal but certainly simple:

     str x2, TLS_REG2_SLOT
     mov x2, #trace_next_target
     eor x2, x2, jump_target
     cbz x2, match
     mov x2, jump_target
     b ibl_routine_regular_exit_stub
match:
     ldr x2, TLS_REG2_SLOT

Then we eliminate this special-case trace_exit_label.

(If jump_target == x2 then a different scheme needs to be used -- might require 2 spills b/c we want to put the app value in the x2 slot and have a local scratch.)

[derekbruening]

Documenting a fundamental issue with tracing building on AArch64: exclusive monitors.

DR assumes it can unlink a block for the purpose of recording where it goes next, in order to collect the set of consecutive blocks to use to build up a trace.

However, on AArch64, unlinking can perturb the app, since the exit stub itself along with all the DR code contains memory traffic which is likely to lose an exclusive monitor. So without something like -ldstex2cas (see https://dynamorio.org/page_ldstex.html), the trace recording path would likely never capture a successful exclusive monitor acquisition. Yet another problem to add to that page: -ldstex2cas (or a similar solution) is needed for core DR operation and not just instrumentation tools.