Don't emit duplicate source locations

[bnoordhuis]

Requires a little extra bookkeeping but there are two good reasons to accept that overhead:

1. OP_source_loc is pretty big at 9 bytes so duplicates bloat the bytecode considerably

2. An OP_source_loc opcode can throw off the peephole optimizer if it's between two opcodes the optimizer is trying to match; removing duplicates reduces the chance of that happening

Fixes: #916

[bnoordhuis]

2. An OP_source_loc opcode can throw off the peephole optimizer if it's between two opcodes the optimizer is trying to match; removing duplicates reduces the chance of that happening

Looking at code_match() more closely, it seems smart enough to skip them. That leaves argument (1) and since they're filtered out in phase 3 anyway, maybe it's not worth the overhead.

[bnoordhuis]

I used callgrind and massif and benchmarked this change rather extensively on a big source file and...the results are confusing:

1. Total instruction count dropped by 0.001%, but
2. Peak memory usage went up 0.001% - wut?

They're marginal differences but callgrind and massif are exact and repeated runs are indeed identical.

I'm somewhat at a loss to explain (2) so I'll go ahead and close this.