Match the base pointer of llvm_elem and llvm_field.

[andreistefanescu]

Currently, saw supports overrides of the form llvm_execute_func [x_p, (llvm_elem x_p 1)];, while it does not support overrides of the form llvm_execute_func [(llvm_elem x_p 1), x_p];, because matching on llvm_elem x_p 1 fails unless x_p is already matched. This changes the handling of llvm_elem to match the base pointer. This changes the handling of llvm_field in a similar way.

[atomb]

Could you provide a bit more description of the goal of this PR?

[andreistefanescu]

@atomb I updated the description. Take a look at the tests, they are failing with the current version of saw.

[brianhuffman]

As I understand it, this patch basically makes llvm_elem and llvm_field work as patterns. So when you want to match llvm_elem n lhs against rhs, that reduces to matching lhs against rhs - offset (where offset is computed from n and the pointer type).

I've thought about the possibility of implementing this before, and this is exactly how I would have done it. The code changes look good, so I think it's ready to merge as soon as we fix the test cases (which I'm going to have a look at now).

[brianhuffman]

When I run saw manually on test0065_match_llvm_elem/test.saw, it works just fine. But my runtests.sh says it fails; I think it must be running an out-of-date saw executable.

[andreistefanescu]

@brianhuffman the tests pass in github actions, and fail in travis because of missing external abc. The test added by this pull request passes in both. Or are you seeing a different failure?

I guess it's time to remove travis entirely.

[brianhuffman]

My local failure was bogus, due to an out-of-date saw binary. Everything seems to be working fine. We can safely ignore the failing travis integration test.