[llvm-bugs] [Bug 34380] New: [X86] Very bad sequences generated for building smaller vectors from bigger ones

[via llvm-bugs]

https://bugs.llvm.org/show_bug.cgi?id=34380

            Bug ID: 34380
           Summary: [X86] Very bad sequences generated for building
                    smaller vectors from bigger ones
           Product: libraries
           Version: trunk
          Hardware: All
                OS: All
            Status: NEW
          Severity: enhancement
          Priority: P
         Component: Backend: X86
          Assignee: unassignedbugs at nondot.org
          Reporter: ayman.musa at intel.com
                CC: llvm-bugs at lists.llvm.org

IR that build a new vector with elements from a bigger sized vector gets
lowered to a very bad X86 instruction sequence.

Some example:

define <8 x i16> @test1(<16 x i16> %vec) {
  %res = shufflevector <16 x i16> %vec, <16 x i16> undef, <8 x i32> <i32 14,
i32 8, i32 4, i32 12, i32 9, i32 4, i32 14, i32 15>
  ret <8 x i16> %res 
}

LLVM emits:
     vextracti128    $1, %ymm0, %xmm1 
     vpshufb .LCPI0_0(%rip), %xmm1, %xmm1 # xmm1 =
xmm1[12,13,0,1,8,9,8,9,2,3,2,3,12,13,14,15]
     vpshufhw    $0, %xmm0, %xmm0 # xmm0 = xmm0[0,1,2,3,4,4,4,4]
     vpshufd $250, %xmm0, %xmm0      # xmm0 = xmm0[2,2,3,3]
     vpblendw    $36, %xmm0, %xmm1, %xmm0 # xmm0 =
xmm1[0,1],xmm0[2],xmm1[3,4],xmm0[5],xmm1[6,7]
     vzeroupper

While it can be replaced with (perm_idxs holds the indexes
[14,8,4,12,9,4,14,1]):
     vbroadcasti128  perm_idxs(%rip), %ymm1 # ymm1 = mem[0,1,0,1]
     vpermw  %ymm0, %ymm1, %ymm0
     vzeroupper

============================================================================

define <4 x i32> @test_16xi32_to_4xi32_perm_mask9(<16 x i32> %vec) {
 ; CHECK-LABEL: test_16xi32_to_4xi32_perm_mask9
 ; CHECK:       # BB#0:
 ; CHECK-NEXT:    vmov{{.*}} .LCPI{{.*}}(%rip), %zmm{{0|1|2}}
 ; CHECK-NEXT:    vpermd %zmm0, %zmm1, %zmm0
 ; CHECK-NEXT:    retq 
   %res = shufflevector <16 x i32> %vec, <16 x i32> undef, <4 x i32> <i32 12,
i32 9, i32 4, i32 10>
   ret <4 x i32> %res 
 }

LLVM emits:
     vextracti64x4   $1, %zmm0, %ymm1
     vmovdqa .LCPI237_0(%rip), %ymm2 # ymm2 = <4,1,u,2,u,u,u,u>
     vpermd  %ymm1, %ymm2, %ymm1
     vpermq  $232, %ymm0, %ymm0      # ymm0 = ymm0[0,2,2,3]
     vpblendd    $4, %xmm0, %xmm1, %xmm0 # xmm0 = xmm1[0,1],xmm0[2],xmm1[3]
     vzeroupper

While it can be replaced with (perm_idxs holds the indexes [12,9,4,10]):
     vbroadcasti128  perm_idxs(%rip), %zmm1 # zmm1 = mem[0,1,0,1,0,1,0,1]
     vpermd %zmm0, %zmm1, %zmm0
     vzeroupper

-- 
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/20170830/54db1c81/attachment.html>