LLVM codegen-equivalence tests

[graydon]

This is a performance test idea. It involves registering pairs of functions -- rust and rust, rust and C++, rust and LLVM -- and comparing the ratio of their outputs in terms of number of LLVM instructions and number of machine instructions.

I think this is a nice category of test because:

• it's mostly objective (subject to clang drift) and timing-insensitive so can be turned into a performance ratchet (see make check should use a performance ratchet #7459)
• it's easy to think of new testcases that "should be equivalent"
• it's easy to turn existing performance problems into such testcases

Some examples:

• Manually inlined things vs. types of inlining we expect LLVM to do
• Same for constant propagation
• Same for type specialization
• Comparing how we do these things vs. how C++/clang does them
• Unsafe code vs. safe code for same operations
• Match constructs vs. nested ifs
• Iteration constructs vs. manual counter loops
• Destructuring vs. field access

[graydon]

This is now active, #7623 added a framework and I've been adding new tests.