Entry point publishing in Truffle OptimizedCallTarget, on AArch64

[Andrew Dinn]

On 14/06/17 17:09, Andrew Dinn wrote:
> Correction:
> 
> Looking at the details of the test failures it appears that many of them
> are caused by an error that might be related to your code.
> 
> Many of the tests are manifesting exceptions under method
> AArch64MacroAssembler.patchJumpTarget in a call to
> ConditionFlag.fromEncoding. Here is the offending code:
> 
>         int instruction = getInt(branch);
>         int branchOffset = jumpTarget - branch;
>         PatchLabelKind type = PatchLabelKind.fromEncoding(instruction);
>         switch (type) {
>             case BRANCH_CONDITIONALLY:
>                 ConditionFlag cf =
> ConditionFlag.fromEncoding(instruction >>>
> PatchLabelKind.INFORMATION_OFFSET);
>                 super.b(cf, branchOffset, branch);
>                 break;
>             case BRANCH_UNCONDITIONALLY:
>                 super.b(branchOffset, branch);
>         . . .
> 
> So, there is something in the format of a generated branch that the
> assembler does not understand. If you are generating a TBZ that is
> susceptible to patching -- even if it is not actually installed -- then
> that might account for why I am seeing these specific unit test
> failures. I'll take a deeper look to see if I can diagnose whether the
> offending instruction is a TBZ and where it is being generated from.

This is certainly the problem that is causing some of the unit tests to
fail (I don't yet know about all of them).

The implementation of tbz/nz is using the same encoding as cbz/nz to
encode patch details into the instruction slot for a branch with an as
yet unbound label. This is use later y the jump patch routine to
construct the required jump instruction. So, with the current
implementation of tbz that fails for two reasons. The patch info doesn't
include the uimm6 argument needed to provide the test bit position. It
also means that the patch code tries to patch the jump as a cbz/nz
rather than a tbz/nz.

I'll see if I can come up with a correct implementation and then see
what that does to the failing unit test count. This probably requires a
new type of patch info encoding (TEST_BRANCH_CONDITIONALLY.encoding)
with a new layout for the other data that includes the uimm6 but
position value. I'll let you know when I have a suitable patch.

regards,


Andrew Dinn
-----------
Senior Principal Software Engineer
Red Hat UK Ltd
Registered in England and Wales under Company Registration No. 03798903
Directors: Michael Cunningham, Michael ("Mike") O'Neill, Eric Shander