[LLVMbugs] [Bug 3457] New: scheduling (?) severely pessimizing scimark2 montecarlo loop

[bugzilla-daemon at cs.uiuc.edu]

http://llvm.org/bugs/show_bug.cgi?id=3457

           Summary: scheduling (?) severely pessimizing scimark2 montecarlo
                    loop
           Product: libraries
           Version: 1.0
          Platform: PC
        OS/Version: All
            Status: NEW
          Keywords: code-quality
          Severity: normal
          Priority: P2
         Component: Common Code Generator Code
        AssignedTo: unassignedbugs at nondot.org
        ReportedBy: clattner at apple.com
                CC: llvmbugs at cs.uiuc.edu


Created an attachment (id=2486)
 --> (http://llvm.org/bugs/attachment.cgi?id=2486)
testcase 

The main loop in MonteCarlo_integrate is:

        %2 = tail call double @Random_nextDouble(%struct.anon* %0) nounwind    
        ; <double> [#uses=2]
        %3 = tail call double @Random_nextDouble(%struct.anon* %0) nounwind    
        ; <double> [#uses=2]
        %4 = mul double %2, %2          ; <double> [#uses=1]
        %5 = mul double %3, %3          ; <double> [#uses=1]
        %6 = add double %4, %5          ; <double> [#uses=1]
...

We compile this to:

        movl    %esi, (%esp)
        call    L_Random_nextDouble$stub
        fstpt   -56(%ebp)
        movl    %esi, (%esp)
        call    L_Random_nextDouble$stub
        fstpl   -32(%ebp)
        fldt    -56(%ebp)
        fstpl   -24(%ebp)
        movsd   -24(%ebp), %xmm0
        mulsd   %xmm0, %xmm0
...

This is horrible because the conversion from long double to double got
scheduled after the second call, so we now get fstpt/fldt instructions (80-bit
fp load/stores), which are really really really slow (10x slower than fstpl.  I
really want to see something like:

        movl    %esi, (%esp)
        call    L_Random_nextDouble$stub
        fstpl   -24(%ebp)

        movl    %esi, (%esp)
        call    L_Random_nextDouble$stub
        fstpl   -32(%ebp)

        movsd   -24(%ebp), %xmm0
        mulsd   %xmm0, %xmm0
...

One less spill, and no crazy 80-bit fp load and stores!


-- 
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.