Clean up MemReq and associated code

[leekillough]

This is cleanup I did when exhaustively reviewing PR #205 .

In compressed instructions where rs1 = rd, make the decoding code assign to both rs1 and rd in the same statement, to avoid mistakes and to prevent creating a dependency on whether rs1 or rd is initialized first (some of the previous code assigned to one and then used it in a later statement to assign to the other, which is confusing -- this makes it clear that both are being assigned the same value at once).

Rename make_lsq_hash() to LSQHash(), keeping in line with CamelCase.

Clean up MemReq class to only have a constructor, and not a .Set() function to change it after it's created (it still has an assignment operator= which is used in RevMemOp::setMemReq()). Split the constructor across multiple lines, one member per line, for readability. Directly initialize MemReq objects using this constructor, instead of default-initializing a MemReq value and using .Set() afterwards.

Since they are common operations, add LSQHash() and LSQHashPair() functions to MemReq. LSQHash() returns the namespace-scope LSQHash() on selected MemReq members, and LSQHashPair() returns {LSQHash(), *this}, a pair consisting of the hash and the request, for storing in std::unordered_map.

Make data members of MemReq private, only visible to MemReq and friend classes RevTracer, RevMem and RevProc.

Instead of MarkLoadComplete() being a member functor of MemReq which expects a MemReq argument, change MarkLoadComplete() to be a member function of MemReq which takes no arguments and calls a MarkLoadCompleteFunc functor with *this as its argument, so that req.MarkLoadComplete(req) can simply be called as req.MarkLoadComplete().

Use std::function<void(const MemReq&)> consistently throughout the code. Previously std::function<void(MemReq)> was used in some places. While it may work because of conversions, it is not as clean.

Using the new helper functions, replace the common idiom R->LSQueue->insert({make_lsq_hash(Inst.rd, RevRegClass::RegGPR, F->GetHartToExecID()), req}); with the shorter R->LSQueue->insert(req.LSQHashPair());

Add std::move() around MemReq objects the last time that they are used locally, as a minor optimization (prevents having to create deep copies of dynamically-allocated objects such as lambda captures -- allows them to be shallowly moved instead).

Make the underlying type for RevReg explicitly uint16_t, since uint16_t is the type used for register numbers in a lot of APIs.

Templatize the MemReq constructor and the LSQHash(), DependencySet() and DependencyClear() functions, so that they accept either RevReg enum class or an integer register number. Replace hard-coded 10 with RevReg::a0 and 0 with RevReg::zero.

In many .cpp files, using namespace SST::RevCPU; was used to pull in symbols, but the rest of the file still defined functions in global namespace unless they were qualified names (using ::). Change all of those files to use namespace SST::RevCPU{ ... } instead of using namespace SST::RevCPU;, so that everything in the file is defined in the SST::RevCPU namespace.

[kpgriesser]

Not a concern but for future consideration (or maybe this is already there). If I wanted to put in another memory agent, like a DMA, it would be nice to build it on the MemCtl/MemReq classes. The receiving agent would receive a block of data and have a custom completion. The requests could go through the same queue to the same memory port or in parallel to a separate memory port. I've been playing with this in a coprocessor memory agent sharing the same lsq. Thanks much

[leekillough]

Not a concern but for future consideration (or maybe this is already there). If I wanted to put in another memory agent, like a DMA, it would be nice to build it on the MemCtl/MemReq classes.

That seems a bigger scope than the MemReq class, which is intended to be very lightweight and is used to implement MemHierarchy and track outstanding requests. The containers (e.g. std::unordered_map) were carefully chosen and tuned. Expanding MemReq's scope would make this run much slower.

[rkabrick]

LGTM