[llvm-bugs] [Bug 25581] New: [ppc] redundant extsw and sldi instructions

[via llvm-bugs]

https://llvm.org/bugs/show_bug.cgi?id=25581

            Bug ID: 25581
           Summary: [ppc] redundant extsw and sldi instructions
           Product: libraries
           Version: trunk
          Hardware: PC
                OS: Linux
            Status: NEW
          Severity: normal
          Priority: P
         Component: Backend: PowerPC
          Assignee: unassignedbugs at nondot.org
          Reporter: carrot at google.com
                CC: llvm-bugs at lists.llvm.org
    Classification: Unclassified

The following is a hot function from 471.omnetpp

void cMessageHeap::shiftup(int from)                                            
{
    // restores heap structure (in a sub-heap)                                  
    int i,j;
    cMessage *temp;                                                             

    i=from;
    while ((j=2*i) <= n)                                                        
    {                                                                           
        if (j<n && (*h[j] > *h[j+1]))   //direction                             
            j++;
        if (*h[i] > *h[j])  //is change necessary?
        {
            temp=h[j];                                                          
            (h[j]=h[i])->heapindex=j;
            (h[i]=temp)->heapindex=i;                                           
            i=j;
        }
        else
            break;
    }
}

When compiled with options -fno-strict-aliasing -O2 -m64 -mvsx -mcpu=power8
I got:


_ZN12cMessageHeap7shiftupEi:            # @_ZN12cMessageHeap7shiftupEi
.Lfunc_begin14:
# BB#0:                                 # %entry
        lwz 7, 64(3)
        slwi 8, 4, 1 
        cmpw     8, 7 
        bgtlr 0
# BB#1:                                 # %while.body.lr.ph
        crxor 20, 20, 20
        li 5, 144
        creqv 21, 21, 21
        .align  4
.LBB14_2:                               # %while.body
                                        # =>This Inner Loop Header: Depth=1
        mr 6, 4 
        cmpw     8, 7 
        bge 0, .LBB14_9
# BB#3:                                 # %land.lhs.true
                                        #   in Loop: Header=BB14_2 Depth=1
        ori 4, 8, 1                  // A
        extsw 9, 8                   // B
        ld 7, 56(3)
        cror 1, 20, 20
        extsw 10, 4                  // C
        sldi 9, 9, 3                 // D
        sldi 11, 10, 3               // E
        ldx 10, 7, 9 
        ldx 9, 7, 11
        lxsdx 0, 10, 5
        lxsdx 1, 9, 5 
        xscmpudp 1, 0, 1 
        blt      1, .LBB14_8
        ...

In the source code note that j comes from (i*2), so the lowest bit is 0, and
then we have following conclusion:

1) j+1 equals to j|1, so instruction A actually computes j+1

2) extsw(j+1) equals to extsw(j) + 1

3) (j+1) << 3 equals to j<<3 + 8

So instructions ABCDE can be replaced with shorter code sequence

    extsw 9, 8
    sldi 9, 9, 3
    addi 11, 9, 8

-- 
You are receiving this mail because:
You are on the CC list for the bug.
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.llvm.org/pipermail/llvm-bugs/attachments/20151120/3212cece/attachment-0001.html>