We observed a performance regression mainly in G1 pause time for Clang-built HotSpot, compared to GCC-built HotSpot. It is because Clang does not inline as much as GCC does, even for functions with "inline" keyword. As a result, some of G1's hottest functions for copying objects and evacuating regions have fewer inlined callees with Clang, leading to longer GC pauses.

The easiest solution is to increase Clang's -inlinehint-threshold. A better long-term solution might be to identify those performance critical functions and mark them with ALWAYSINLINE macro instead of "inline" keyword.

Attached is a performance experiment with DaCapo benchmarks for increasing -inlinehint-threshold to 100000, with normalized and unnormalized graphs. This patch (GoogClang-100kInline) reduces incremental collection pause time ("YG Pause Time (ms)") by up to 12% (xalan) for Clang-built HotSpot.

Some more details below:

Sizes of libjvm.so:
GoogGcc-default: 25369007
GoogClang-default: 22946876
GoogClang-100kInline: 24681265

GCC version: 4.9
Clang version: trunk r355087
# Clang must be newer than https://reviews.llvm.org/rL344987, so it can properly respect the "inline" keyword for function templates.

Number of trials: 20
The heap sizes are configured to relative small values (1.5X to 5X of the minimum heap size of each benchmark), so that G1 can trigger many collections.
The graphs with "G1" as name prefix (e.g. "G1 YG Pause Time", "G1 Mixed Pause Time") are empty, because they rely on some hsperfdata counters that we added to G1 in our internal JDK. These counters do not exist in external OpenJDK yet.