I've checked a couple of months ago, but only on one relatively small project (~6k lines); the compilation time was reduced by 50%.

The 30% faster was before support for split dwarf was enabled for llvm, I now expect to be faster or as fast as cranelift for debug.

30% faster is significantly faster

You can be 1% faster and be 'significantly' faster. 'Significantly' just tells you about variance of your samples. What matters is if it's usefully faster.

I think they meant "significant" as in "of importance", not "statistically significant".

I've been trying to catch myself claiming something is "significantly faster" and instead say it's "substantially faster" for this reason. As far as I know, "substantial" doesn't have the same kind of formal definition, and still captures the informal point I was trying to make.

I wish that we (collectively) did a better job of reporting statistically significant performance comparisons, and in that light using "substantially" is a reminder to myself to think about whether it would be hard to go the extra mile and collect the data to actually report significance properly.

Depends, 30% faster codegen is a good thing to build from but codegen is only a fraction of the actual compilation process.

It is significantly faster, just not enough to make a difference when the baseline is so slow.