[llvm-bugs] [Bug 26301] New: [x86] suboptimal codegen for vector with string of set bits (-1)

[via llvm-bugs]

https://llvm.org/bugs/show_bug.cgi?id=26301

            Bug ID: 26301
           Summary: [x86] suboptimal codegen for vector with string of set
                    bits (-1)
           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

I noticed this while looking at vmaskmov codegen because that (unnecessarily)
uses vector masks with elements of all ones or zeros:

define <4 x i32> @high_64_ones() {
  ret <4 x i32><i32 0, i32 0, i32 -1, i32 -1>
}

define <2 x i64> @high_64_ones_alt() {
  ret <2 x i64><i64 0, i64 -1>
}

$  ./llc high_ones.ll  -o -

_high_64_ones:                          ## @high_64_ones
    movaps    LCPI0_0(%rip), %xmm0    ## xmm0 = [0,0,4294967295,4294967295]
    retq

_high_64_ones_alt:                      ## @high_64_ones_alt
    movq    $-1, %rax
    movd    %rax, %xmm0
    pslldq    $8, %xmm0               ## xmm0 =
zero,zero,zero,zero,zero,zero,zero,zero,xmm0[0,1,2,3,4,5,6,7]
    retq

-----------------------------------------------------------------------------

Some might argue that the 1st case is fine; don't burden the vector units
because loads (and memory space?) are free on big Intel core systems. But I
think the 2nd would be better as:

  pcmpeqd %xmm0, %xmm0  // splat 1 bits all the way across
  pslldq  $8, %xmm0     // shift in the zeros

...to avoid the move from integer to SSE register. And that's the codegen we
should produce by default for both cases.

-- 
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/20160125/e67aaa60/attachment.html>