[LLVMbugs] [Bug 21230] New: Performance regression in vector shuffle + pmaddwd

[bugzilla-daemon at llvm.org]

http://llvm.org/bugs/show_bug.cgi?id=21230

            Bug ID: 21230
           Summary: Performance regression in vector shuffle + pmaddwd
           Product: libraries
           Version: trunk
          Hardware: PC
                OS: Linux
            Status: NEW
          Severity: normal
          Priority: P
         Component: Backend: X86
          Assignee: unassignedbugs at nondot.org
          Reporter: andrew.b.adams at gmail.com
                CC: llvmbugs at cs.uiuc.edu
    Classification: Unclassified

Created attachment 13178
  --> http://llvm.org/bugs/attachment.cgi?id=13178&action=edit
ll that produces worse x86 with trunk than release 3.5

The attached ll computes 3*a + 5*b where a and b are 16-bit vectors and the
result is a 32-bit vector. It does this with a clever application of pmaddwd.
It used to produce this asm with release 3.5

movdqa    %xmm0, %xmm4
punpcklwd    %xmm2, %xmm4 
movdqa    %xmm1, %xmm5
punpcklwd    %xmm3, %xmm5  
pmaddwd    %xmm5, %xmm4
punpckhwd    %xmm2, %xmm0 
punpckhwd    %xmm3, %xmm1 
pmaddwd    %xmm0, %xmm1
movdqa    %xmm4, %xmm0

but produces this slower asm with trunk:

pshufd    $78, %xmm0, %xmm5 
punpcklwd    %xmm2, %xmm0    
pshufd    $78, %xmm1, %xmm4     
punpcklwd    %xmm3, %xmm1   
pmaddwd    %xmm1, %xmm0
pshufd    $78, %xmm2, %xmm1      
punpcklwd    %xmm1, %xmm5    
pshufd    $78, %xmm3, %xmm1       
punpcklwd    %xmm1, %xmm4   
pmaddwd    %xmm5, %xmm4
movdqa    %xmm4, %xmm1

-- 
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/20141009/d6b6e484/attachment.html>