roundtest.jl

function rint_llvm(x::Float64)
    Base.llvmcall("""%x = call double @llvm.rint.f64(double %0)
                  ret double %x""",Float64,(Float64,),x)
end
function trunc_llvm(x::Float64)
    Base.llvmcall("""%x = call double @llvm.trunc.f64(double %0)
                  ret double %x""",Float64,(Float64,),x)
end

# warm up llvmcall: ignore errors
rint_llvm(1.0)
rint_llvm(1.0)
trunc_llvm(1.0)
trunc_llvm(1.0)

# use "add remainder" trick
function round_rem(x::Float64)
    t = trunc_llvm(x)
    trunc_llvm(x+(x-t))
end

# use "prevfloat(0.5)" trick 
round_prev(x::Float64) = trunc_llvm(x+copysign(0.49999999999999994,x))

round_rem(1.0)
round_rem(1.0)
round_prev(1.0)
round_prev(1.0)


function test_rint_llvm(N)
    t = 1.0
    s = 0.0
    for i = 1:N
        t += eps()
        s += rint_llvm(t)
    end
    s
end
function test_round_rem(N)
    t = 1.0
    s = 0.0
    for i = 1:N
        t += eps()
        s += round_rem(t)
    end
    s
end
function test_round_prev(N)
    t = 1.0
    s = 0.0
    for i = 1:N
        t += eps()
        s += round_prev(t)
    end
    s
end


@time test_rint_llvm(100_000_000);
@time test_rint_llvm(100_000_000);

@time test_round_rem(100_000_000);
@time test_round_rem(100_000_000);

@time test_round_prev(100_000_000);
@time test_round_prev(100_000_000);

times after warmup:

julia> @time test_rint_llvm(100_000_000);
elapsed time: 0.212698866 seconds (96 bytes allocated)

julia> @time test_round_rem(100_000_000);
elapsed time: 0.536767063 seconds (96 bytes allocated)

julia> @time test_round_prev(100_000_000);
elapsed time: 0.219526082 seconds (96 bytes allocated)

julia> versioninfo()
Julia Version 0.4.0-dev+1916
Commit 11d5e8c* (2014-12-02 12:53 UTC)
Platform Info:
  System: Darwin (x86_64-apple-darwin13.4.0)
  CPU: Intel(R) Core(TM) i7-2677M CPU @ 1.80GHz
  WORD_SIZE: 64
  BLAS: libopenblas (USE64BITINT DYNAMIC_ARCH NO_AFFINITY Sandybridge)
  LAPACK: libopenblas
  LIBM: libopenlibm
  LLVM: libLLVM-3.3

Looking at the @code_native, only unvectorised (vroundsd) instructions were used.