I attached a JMH benchmark showcasing the problem. The results on my machine:

Benchmark                             Mode  Cnt   Score    Error  Units
FoldCompare.foldableComparesExternal  avgt    3  26,630 ± 13,493  ns/op
FoldCompare.foldableComparesInline    avgt    3  57,526 ± 10,844  ns/op
FoldCompare.unsignedCompare           avgt    3  26,499 ± 11,984  ns/op 

C2 has code to merge handwritten range checks (e.g. value >= 0 && value < 100) into one CmpU (value <u 100). See https://github.com/openjdk/jdk/blob/c3a632dca75d2fad0a60e03e7b4fc64edb1e906e/src/hotspot/share/opto/ifnode.cpp#L687.

However, this doesn't work if the range check is directly written in a loop (see the foldableComparesInline method in the benchmark). It works if the exactly same check is extracted into an extra method.

From my analysis, the problem stems from the !region->has_phi() check here https://github.com/openjdk/jdk/blob/c3a632dca75d2fad0a60e03e7b4fc64edb1e906e/src/hotspot/share/opto/ifnode.cpp#L760C56-L760C74. In the case of the extracted method, the method gets first inlined, and projections both point to a region that then directly points to another region, but is not associated with a Phi. In contrast, the manually inlined check directly points to a region with a Phi, bailing out at this check.

My understanding of the existing code isn't good enough to tell how this can be improved. I assume the Phi check is too conservative, but I also assume that it is there for a valid reason.