[Apanatshka]

I've seen a poster presentation at the student research competition of a PL conference where the student created a tool to do this after the fact on binaries. It looked quite promising, but I think preliminary results were single digit percentage speedup for most benchmarks. I'll have to see if I can find it again.

[kevincox]

Single digits is a huge improvement. This is single digits across billions of programs with no change on their part. In the email Google claimed single digit improvements and I'm sure they are thrilled. 1% CPU savings across Google is many, many millions of dollars a year.

If you contribute a 1% CPU savings change to GCC or LLVM you would have a very hard time being carbon-positive over your lifespan.