[llvm-bugs] [Bug 27854] New: [x86, SSE] insertps chosen when unpcklps would be better

[via llvm-bugs]

https://llvm.org/bugs/show_bug.cgi?id=27854

            Bug ID: 27854
           Summary: [x86, SSE] insertps chosen when unpcklps would be
                    better
           Product: libraries
           Version: trunk
          Hardware: PC
                OS: All
            Status: NEW
          Severity: normal
          Priority: P
         Component: Backend: X86
          Assignee: unassignedbugs at nondot.org
          Reporter: spatel+llvm at rotateright.com
                CC: llvm-bugs at lists.llvm.org
    Classification: Unclassified

Forking this off of bug 27826:

If we have a sequence of insertelements to fill in a vector:

define <4 x float> @goo(float %f0, float %f1, float %f2, float %f3) {
  %ins0 = insertelement <4 x float> undef, float %f0, i32 0
  %ins1 = insertelement <4 x float> %ins0, float %f1, i32 1
  %ins2 = insertelement <4 x float> %ins1, float %f2, i32 2
  %ins3 = insertelement <4 x float> %ins2, float %f3, i32 3
  ret <4 x float> %ins3
}

We do the optimal thing with SSE2:
$ ./llc -o - inselts.ll 
...
    unpcklps    %xmm3, %xmm1    ## xmm1 = xmm1[0],xmm3[0],xmm1[1],xmm3[1]
    unpcklps    %xmm2, %xmm0    ## xmm0 = xmm0[0],xmm2[0],xmm0[1],xmm2[1]
    unpcklps    %xmm1, %xmm0    ## xmm0 = xmm0[0],xmm1[0],xmm0[1],xmm1[1]
    retq

The first two instructions are independent, so they can execute in parallel
given enough hardware.

But given the opportunity to use ever more shuffle instructions with each ISA
extension:
$ ./llc -o - inselts.ll  -mattr=sse4.1
...
    insertps    $16, %xmm1, %xmm0 ## xmm0 = xmm0[0],xmm1[0],xmm0[2,3]
    insertps    $32, %xmm2, %xmm0 ## xmm0 = xmm0[0,1],xmm2[0],xmm0[3]
    insertps    $48, %xmm3, %xmm0 ## xmm0 = xmm0[0,1,2],xmm3[0]
    retq

We now have a sequence of 3 dependent instructions, and each of those
instructions is larger in size too.

-- 
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/20160524/e1401907/attachment-0001.html>