Flags from -O3 often flow down into -O2 as they are proven generally beneficial.

That said, I don't think -O3 has the problems it once did.

For cases where -O2 is too slow to compile, dropping a single nasty TU down to -O1 is often beneficial. -O0 is usually not useful - while faster for tiny TUs, -O1 is still pretty fast for them, and for anything larger, the increased binary size bloat of -O0 is likely to kill your link time compared to -O1's slimness.

Also debuggability matters. GCC's `-O2` is quite debuggable once you learn how to work past the possibility of hitting an <optimized out> (going up a frame or dereferencing a casted register is often all you need); this is unlike Clang, which every time I check still gives up entirely.

The real argument is -O1 vs -O2 (since -O1 is a major improvement over -O0 and -O3 is a negligible improvement over -O2) ... I suppose originally I defaulted to -O2 because that's what's generally used by distributions, which compile rarely but run the code often. This differs from development ... but does mean you're staying on the best-tested path (hitting an ICE is pretty common as it is); also, defaulting to -O2 means you know when one of your TUs hits the nasty slowness.

While mostly obsolete now, I have also heard of cases where 32-bit x86 inline asm has difficulty fulfilling constraints under register pressure at low optimization levels.

In my experience a team of 200 developers will see 1 compiler bug affect them every 10 years. This isn't scientific, but it is a good rule of thumb and may put the above in perspective.

You generally avoid O3 because it's slower. Slower to compile, and slower to run. Aggressively unrolling loops and larger inlining windows bloat code size to the degree it impacts icache.

The optimization levels aren't "how fast do you want to code to go", they're "how aggressive do you want the optimizer to be." The most aggressive optimizations are largely unproven and left in O3 until they are generally useful, at which point they move to O2.