I admit I haven't actually figured out how the first change made the given example faster.

Edit: ax.xaxis.get_ticks_position() returns "unknown" here so ax.xaxis.get_tightbbox gets called, but I have not dug any deeper...

I think this would benefit from more digging: the fact that there is no measurable time difference without sharex suggests that this method ought to be fast and the real problem for sharex is deeper.

More digging completed #26150 (comment)

I think CodeCov was telling me that we didn't have any tests with a child axes and an automatically positioned title, so I added one.

Does it actually speed things up?

If I take this example from #26150
fig_many_sharex.savefig(stream, format='svg', bbox_inches='tight')

With main I get

477 ms ± 13.3 ms per loop (mean ± std. dev. of 7 runs, 1 loop each)

with this branch

224 ms ± 6.7 ms per loop (mean ± std. dev. of 7 runs, 1 loop each)

That example has no titles though, so we are now just skipping everything. If I add a title to each subplot I get
main

511 ms ± 9.67 ms per loop (mean ± std. dev. of 7 runs, 1 loop each)

Branch

275 ms ± 6.5 ms per loop (mean ± std. dev. of 7 runs, 1 loop each)

If I add a title to each subplot and move the ticklabels to the top of each subplot, so we do actually need to calculate xaxis.get_tightbbox (but now fewer times)
main

707 ms ± 13.7 ms per loop (mean ± std. dev. of 7 runs, 1 loop each)

Branch

464 ms ± 9.57 ms per loop (mean ± std. dev. of 7 runs, 1 loop each)

We have a benchmarking test suite. I wonder if these are caught by it? @QuLogic maintains that - I actually forget where it is...

We have a benchmarking test suite

This is news to me!

We have a benchmarking test suite

This is news to me!

We do; it is here. However, we don't test bbox_inches='tight' or sharex, and unfortunately, I haven't run the benchmark in a long while now.