Latency regression on first error

[jakobnissen]

The following shell commands triggers a MethodError in Julia and times how long it takes to display it: time julia --startup=no -e '1 + []'

Here are the timings on some versions of Julia (mean of three runs):

• 1.10.7: 0.41 s
• 1.11.2: 2.32 s
• Today's master: 1.29 s

Time to first error is, of course, not crucial. But it does make the REPL feel more choppy and laggy.
Using --trace-compile shows 33 uncompiled signatures, but most of them are abstract, so it may possibly be an issue with inference and/or lack of @nospecialize for error-related methods.