AMD64HotSpot: emit jump to IC miss handler directly instead of emitting it inside verfied_entry for empty methods we emitted something like this if a inline cache is needed: prefix: 00: < IC check > ... 0b: jne <ic_miss_call> 11: nop ... verified_entry: 20: ret ic_miss_call: 21: jmp <entry of runtime function> when a method is deoptimized, HotSpot patches the verified_entry (0x20) with a jump to a stub that handles call-sites that has been made non-entrant. since this jump is 5 bytes long, it will overwrite ic_miss_call and blow up every caller that calls this method via the unverified entry (prefix). the fix is to emit the jump to the runtime function inside the unverfied entry: prefix: 00: < IC check > ... 0b: je <verified_entry> 11: jeq <entry of runtime function> 16: nop ... verified_entry: 20: ret

Case #1: Yes 2
Case #2: Yes 2
Case #3: Yes 1
Case #4: Yes 0
Case #5: No
Case #6: No
Case #7: Yes 6
Case #8: Yes 6
Case #9: No
Case #10: Yes 1
Case #11: Yes 6
Case #12: Yes 0
Case #13: No
Case #14: Yes 22
Case #15: Yes 1225
Case #16: Yes 178
Case #17: No
Case #18: Yes 1
Case #19: Yes 7
Case #20: Yes 2
Case #21: Yes 1
Case #22: No
Case #23: Yes 3
Case #24: Yes 1
Case #25: Yes 7
Case #26: No
Case #27: Yes 2
Case #28: Yes 4
Case #29: Yes 2
Case #30: Yes 1
Case #31: Yes 2
Case #32: Yes 20
Case #33: Yes 161
Case #34: Yes 48
Case #35: No
Case #36: Yes 218
Case #37: Yes 51
Case #38: Yes 247
Case #39: Yes 32
Case #40: Yes 31
Case #41: Yes 31
Case #42: Yes 25
Case #43: Yes 17
Case #44: Yes 2
Case #45: Yes 61
Case #46: Yes 25
Case #47: No
Case #48: No
Case #49: Yes 8
Case #50: Yes 0