I see why it's useful to translate the program into instrumentation codes. This has two benefits: 1) to hide the complexities of transactional memory away from the compiler; 2) allow a degree of abstraction so that we could implement different transactional memories schemes under the hood.

If you've opened the transaction for read before a write, you can do some optimization since it was already accessed (txnOpenForWrite(a) outside the if loop) - so the 7th, 9th, 10th, and 12th lines can be erased shown on the next slide).

*write before read ^^