mehdi_amini added a comment. Also the recent measurements looks like this patch was a little slower. Another thing you could provide outside of a reproducer, is the output of a run with tcmalloc and the environment variable MALLOCSTATS https://reviews.llvm.org/D27960