[LLVMbugs] [Bug 15872] New: some shuffles get decomposed to insert/extract sequence with avx

[bugzilla-daemon at llvm.org]

http://llvm.org/bugs/show_bug.cgi?id=15872

            Bug ID: 15872
           Summary: some shuffles get decomposed to insert/extract
                    sequence with avx
           Product: libraries
           Version: trunk
          Hardware: PC
                OS: All
            Status: NEW
          Severity: normal
          Priority: P
         Component: Backend: X86
          Assignee: unassignedbugs at nondot.org
          Reporter: sroland at vmware.com
                CC: llvmbugs at cs.uiuc.edu
    Classification: Unclassified

Some rather simple but non-native shuffles produce completely horrendous code
because they get decomposed to single element inserts/extracts.
Here's an example:

define <4 x i32> @shuf(<8 x i32> %inval, <4 x i32>* %ptr1) {
entry:
  %0 = shufflevector <8 x i32> %inval, <8 x i32> %inval, <4 x i32> <i32 0, i32
1, i32 4, i32 5>
  %1 = shufflevector <8 x i32> %inval, <8 x i32> %inval, <4 x i32> <i32 2, i32
3, i32 6, i32 7>
  store <4 x i32> %0, <4 x i32>* %ptr1, align 16
  ret <4 x i32> %1

This generates:

        vpextrd $1, %xmm0, %eax
        vmovd   %xmm0, %ecx
        vmovd   %ecx, %xmm1
        vpinsrd $1, %eax, %xmm1, %xmm2
        vextractf128    $1, %ymm0, %xmm1
        vmovd   %xmm1, %eax
        vpinsrd $2, %eax, %xmm2, %xmm2
        vpextrd $1, %xmm1, %eax
        vpinsrd $3, %eax, %xmm2, %xmm2
        vmovdqa %xmm2, (%rdi)
        vpextrd $3, %xmm0, %eax
        vpextrd $2, %xmm0, %ecx
        vmovd   %ecx, %xmm0
        vpinsrd $1, %eax, %xmm0, %xmm0
        vpextrd $2, %xmm1, %eax
        vpinsrd $2, %eax, %xmm0, %xmm0
        vpextrd $3, %xmm1, %eax
        vpinsrd $3, %eax, %xmm0, %xmm0
        vzeroupper
        ret

which is quite obviously a disaster.
(With float instead of i32 values it will produce
        vpshufd $1, %xmm0, %xmm1
        vinsertps       $16, %xmm1, %xmm0, %xmm2
        vextractf128    $1, %ymm0, %xmm1
        vinsertps       $32, %xmm1, %xmm2, %xmm2
        vpshufd $1, %xmm1, %xmm3
        vinsertps       $48, %xmm3, %xmm2, %xmm2
        vmovaps %xmm2, (%rdi)
        vpshufd $3, %xmm0, %xmm2
        vmovhlps        %xmm0, %xmm0, %xmm0
        vinsertps       $16, %xmm2, %xmm0, %xmm0
        vmovhlps        %xmm1, %xmm1, %xmm2
        vinsertps       $32, %xmm2, %xmm0, %xmm0
        vpshufd $3, %xmm1, %xmm1
        vinsertps       $48, %xmm1, %xmm0, %xmm0
        vzeroupper
        ret
which is not really much better neither.)

But this should really just be (or something similar, there's quite some
options, for instance using vshufps):
        vextractf128    $1, %ymm0, %xmm1
        vmovlhps        %xmm0, %xmm1, %xmm2
        vmovaps         %xmm2, (%rdi)
        vunpckhpd       %xmm0, %xmm1, %xmm0
        vzeroupper
        ret

This is somewhat similar to http://llvm.org/bugs/show_bug.cgi?id=12413 but that
one was fixed.

-- 
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/20130430/6370ad7b/attachment.html>