[llvm-bugs] [Bug 45794] New: [X86][AVX] Poor sitofp v4i64 to v4f32 conversion

[via llvm-bugs]

https://bugs.llvm.org/show_bug.cgi?id=45794

            Bug ID: 45794
           Summary: [X86][AVX] Poor sitofp v4i64 to v4f32 conversion
           Product: libraries
           Version: trunk
          Hardware: PC
                OS: Windows NT
            Status: NEW
          Severity: enhancement
          Priority: P
         Component: Backend: X86
          Assignee: unassignedbugs at nondot.org
          Reporter: llvm-dev at redking.me.uk
                CC: craig.topper at gmail.com, llvm-bugs at lists.llvm.org,
                    llvm-dev at redking.me.uk, spatel+llvm at rotateright.com

Split off from [Bug #39974]:

https://c.godbolt.org/z/-rvyDZ

#include <stdint.h>
#include <x86intrin.h>

__m128 foo128(__m128i x, __m128i y) {
    auto xf = __builtin_convertvector(x >> 48, __v2sf);
    auto yf = __builtin_convertvector(y >> 48, __v2sf);
    return (__v4sf) { xf[0], xf[1], yf[0], yf[1] };
}

__m128 foo256(__m256i x) {
    return __builtin_convertvector(x >> 48, __v4sf);
}

-g0 -O3 -march=btver2

define <4 x float> @foo128(<2 x i64> %0, <2 x i64> %1) {
  %3 = ashr <2 x i64> %0, <i64 48, i64 48>
  %4 = sitofp <2 x i64> %3 to <2 x float>
  %5 = ashr <2 x i64> %1, <i64 48, i64 48>
  %6 = sitofp <2 x i64> %5 to <2 x float>
  %7 = shufflevector <2 x float> %4, <2 x float> %6, <4 x i32> <i32 0, i32 1,
i32 2, i32 3>
  ret <4 x float> %7
}

define <4 x float> @foo256(<4 x i64> %0) {
  %2 = ashr <4 x i64> %0, <i64 48, i64 48, i64 48, i64 48>
  %3 = sitofp <4 x i64> %2 to <4 x float>
  ret <4 x float> %3
}

foo128:
        vpsrad  $16, %xmm0, %xmm0
        vpsrad  $16, %xmm1, %xmm1
        vpshufd $237, %xmm0, %xmm0      # xmm0 = xmm0[1,3,2,3]
        vpshufd $237, %xmm1, %xmm1      # xmm1 = xmm1[1,3,2,3]
        vcvtdq2ps       %xmm0, %xmm0
        vcvtdq2ps       %xmm1, %xmm1
        vmovlhps        %xmm1, %xmm0, %xmm0 # xmm0 = xmm0[0],xmm1[0]
        retq
foo256:
        vextractf128    $1, %ymm0, %xmm1
        vpsrad  $31, %xmm0, %xmm3
        vpsrad  $16, %xmm0, %xmm0
        vpsrad  $31, %xmm1, %xmm2
        vpsrad  $16, %xmm1, %xmm1
        vpshufd $245, %xmm0, %xmm0      # xmm0 = xmm0[1,1,3,3]
        vpshufd $245, %xmm1, %xmm1      # xmm1 = xmm1[1,1,3,3]
        vpblendw        $204, %xmm3, %xmm0, %xmm0 # xmm0 =
xmm0[0,1],xmm3[2,3],xmm0[4,5],xmm3[6,7]
        vpblendw        $204, %xmm2, %xmm1, %xmm1 # xmm1 =
xmm1[0,1],xmm2[2,3],xmm1[4,5],xmm2[6,7]
        vpackssdw       %xmm1, %xmm0, %xmm0
        vcvtdq2ps       %xmm0, %xmm0
        retq

Neither of these come off well and amazingly for different reasons:

foo128 - missing (in DAG or x86):

    shuffle(sitofp(x),sitofp(y)) -> sitofp(shuffle(x,y))

foo256 - all that shuffle code is mostly redundant (vpackss is probably
struggling to count signbits through bitcasts).

In both cases we should get to:

  vpsrad $16, %xmm0, %xmm0
  vpsrad $16, %xmm1, %xmm1
  vshufps ???, %xmm1, %xmm0, %xmm0
  vcvtdq2ps %xmm0, %xmm0
  retq

-- 
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/20200504/091ae0e5/attachment.html>