C++: Reuse even more DataFlow::Nodes

[MathiasVP]

In C/C++ dataflow we reuse some dataflow nodes that we know are sementically equivalent to gain a small amount of performance, and to make it easier to select the right dataflow node in sources, sinks, barriers, etc. For example, consider the following snippet of IR:

r1(glval<char **>) = VariableAddress[a]          :
r2(char **)        = Load[a]                     : &:r1, m1
r3(glval<char *>)  = CopyValue                   : r2

For dataflow we have a node that represents &:r1 (i.e, the value of type char***), as well as nodes for the possible indirections of &:r1: *&:r1 (i.e., the value of type char**), **&:r1 (i.e, the value of type char*), and ***&:r1 (i.e., the value of type char).

Similarly, we have a node that represents the r2 operand (i.e., the value of type char**), as well as nodes for the possible indirections of r2: *r2 (i.e., the value of type char*), and **r2 (i.e., the value of type char).

However, some of these nodes represent the semantically same value. For example, both *&:r1 and r2 represent the char** value (i.e., the value you get when you dereference VariableAddress[a]).

The getIRRepresentationOfIndirectOperand and getIRRepresentationOfIndirectInstruction predicates are used to identify such semantically identical representations.

Currently, however, getIRRepresentationOfIndirectOperand and getIRRepresentationOfIndirectInstruction only identified that *&:r1 and r2 represented the same value. But **&:r1 and *r2, and ***&:r1 and **r2 also represent the same values. This PR extends those predicates to identify these cases.

This should (hopefully) give us a small speedup, a reduced memory footprint, and possible even remove some FPs in queries in cases where we selected a equivalent (but until now not identical!) node to the one on the actual dataflow path.

[MathiasVP]

This PR is finally good to go. There's a couple of test changes in the internal repo that I need to accept as well. I'll do that once this PR has been reviewed.

DCA looks good. I've verified that all the lost results are instances of the test added in f65fe34.

[geoffw0]

The core change looks plausible to me. I can't comment on performance but the tests show considerable deduplication. 🎉

A few comments / questions...

[geoffw0]

LGTM.

DCA shows a speedup.

I've verified that all the lost results are instances of the test added in f65fe34.

👍