[LLVMbugs] [Bug 3120] Small interpreter-like testcase where LLVM fails to shine

[bugzilla-daemon at cs.uiuc.edu]

http://llvm.org/bugs/show_bug.cgi?id=3120


Bob Wilson <bob.wilson at apple.com> changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
             Status|NEW                         |RESOLVED
         Resolution|                            |FIXED




--- Comment #9 from Bob Wilson <bob.wilson at apple.com>  2009-11-25 11:46:01 ---
svn r89865 enables tail duplication for indirect branches on x86, and with that
change, the performance is much better*:

32-bit:

Switched interpreter           240
Threaded interpreter           144
Recursive interpreter          821
Closure-based interpreter     1106

64-bit:

Switched interpreter           584
Threaded interpreter           136
Recursive interpreter          733
Closure-based interpreter      508

* The 64-bit "switched interpreter" result here is bad because of a code
alignment problem.  The function is 16-byte aligned but the performance suffers
badly if it is not placed on a 32-byte aligned boundary.  I think this may be
due to the branch predictor fetching 32-byte aligned data.  If I manually
increase the alignment of that function, the runtime drops from 584 to 291,
which is better than gcc's 335 but still not as good as the 240 I measured
before.  That manual alignment change also increases the "threaded interpreter"
runtime slightly from 136 to 144.  I will file a separate PR for the alignment
issue.


-- 
Configure bugmail: http://llvm.org/bugs/userprefs.cgi?tab=email
------- You are receiving this mail because: -------
You are on the CC list for the bug.