[llvm-bugs] [Bug 31492] New: [PPC] slower vsx instructions generated for vmac

[via llvm-bugs]

https://llvm.org/bugs/show_bug.cgi?id=31492

            Bug ID: 31492
           Summary: [PPC] slower vsx instructions generated for vmac
           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

Created attachment 17789
  --> https://llvm.org/bugs/attachment.cgi?id=17789&action=edit
testcase

The attached test case is simplified from vmac. Compile it with options

-m64 -O2 -mvsx -mcpu=power8

LLVM generates following code for the while loop


.LBB0_2:                                # %while.body
                                        # =>This Inner Loop Header: Depth=1
        lxvd2x 0, 0, 7             // *
        lxvd2x 1, 0, 6             // *
        xxswapd  34, 0             // *
        xxswapd  35, 1             // *
        vaddudm 2, 3, 2            // *
        xxswapd  10, 34            // *
        mfvsrd 9, 34               // *
        mfvsrd 10, 10              // *
        #APP
        mulhdu 11, 10, 9
        #NO_APP
        lxvd2x 11, 7, 8
        lxvd2x 12, 0, 5
        mulld 9, 9, 10
        addi 7, 7, 64
        xxswapd  50, 11
        xxswapd  51, 12
        vaddudm 2, 19, 18
        xxswapd  13, 34
        mfvsrd 0, 34
        mfvsrd 12, 13
        mulld 10, 0, 12
        #APP
        mulhdu 12, 12, 0
        #NO_APP
        #APP
        addc 3, 9, 10
        adde 4, 11, 12
        #NO_APP
        bdnz .LBB0_2

There are two problems:

1. (kp)[i] is loop invariant, its loading can be hoisted before the loop.

2. llvm generates vsx code marked * for the expression

    get64PE((mp) + i) + (kp)[i]

   if we use simple integer load and add instructions, it will be shorter and
faster. For large input, it can be 35% faster.
   Looks like cost model problem in vectorization?

-- 
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/20161228/e125be6c/attachment.html>