[LLVMbugs] [Bug 675] mgrid suffering 70% performance hit against native

bugzilla-daemon at cs.uiuc.edu bugzilla-daemon at cs.uiuc.edu
Sat Feb 4 17:34:11 PST 2006


sabre at nondot.org changed:

           What    |Removed                     |Added
             Status|NEW                         |RESOLVED
          Component|opt                         |Backend: PowerPC
            Product|tools                       |libraries
         Resolution|                            |FIXED
   Target Milestone|---                         |1.7
            Version|trunk                       |1.3

------- Additional Comments From sabre at nondot.org  2006-02-04 19:34 -------
After making codegen improvements and not getting better performance, Nate simg5'd the code.

It turns out not to be a codegen problem, but actually an alignment problem: the common block was 
aligned to a 4-byte boundary (not 8) causing every load in the hot loop to fault (which flushes the 
pipeline on the G5).  Fixed with this patch:

We now equal GCC's performance on mgrid.  Thanks to Nate for doing the analysis to figure out what 
was going wrong!


------- You are receiving this mail because: -------
You are on the CC list for the bug, or are watching someone who is.

More information about the llvm-bugs mailing list