40x Performance Regression in v0.6 for simple polynomial function

[mbrookhart]

Searching for a performance regression in a piece of code, simplified it to this:

function main()
    f1(x) = x + 2*x^2 + 4*x^4
    f2(x) = x + 2*x^2 + 4*x^2*x^2

    @code_llvm(f1(1))
    @code_llvm(f2(1))
    A=rand(10,10)
    f1.(A)
    f2.(A)
    A=rand(10000,10000)
    @time f1.(A)
    @time f2.(A)
end

main()

Fresh compile of tag v0.5.2:

define double @julia_f1_72267(double) #0 {
top:
  %1 = call double @llvm.powi.f64(double %0, i32 4)
  %2 = fmul double %0, %0
  %3 = fmul double %2, 2.000000e+00
  %4 = fadd double %3, %0
  %5 = fmul double %1, 4.000000e+00
  %6 = fadd double %4, %5
  ret double %6
}

define double @julia_f2_72271(double) #0 {
top:
  %1 = fmul double %0, %0
  %2 = fmul double %1, 2.000000e+00
  %3 = fadd double %2, %0
  %4 = fmul double %1, 4.000000e+00
  %5 = fmul double %1, %4
  %6 = fadd double %3, %5
  ret double %6
}
  0.118596 seconds (3 allocations: 762.940 MB, 24.19% gc time)
  0.120200 seconds (3 allocations: 762.940 MB, 24.94% gc time)

Fresh compile of tag v0.6.0:

define double @julia_f1_60711(double) #0 !dbg !5 {
top:
  %1 = call double @llvm.pow.f64(double %0, double 4.000000e+00)
  %2 = fadd double %0, 4.000000e+00
  %notlhs = fcmp ord double %1, 0.000000e+00
  %notrhs = fcmp uno double %2, 0.000000e+00
  %3 = or i1 %notrhs, %notlhs
  br i1 %3, label %L12, label %if

if:                                               ; preds = %top
  call void @jl_throw(i8** inttoptr (i64 140123921213312 to i8**))
  unreachable

L12:                                              ; preds = %top
  %4 = fmul double %0, %0
  %5 = fmul double %4, 2.000000e+00
  %6 = fadd double %5, %0
  %7 = fmul double %1, 4.000000e+00
  %8 = fadd double %6, %7
  ret double %8
}

define double @julia_f2_60732(double) #0 !dbg !5 {
top:
  %1 = fmul double %0, %0
  %2 = fmul double %1, 2.000000e+00
  %3 = fadd double %2, %0
  %4 = fmul double %1, 4.000000e+00
  %5 = fmul double %1, %4
  %6 = fadd double %3, %5
  ret double %6
}
  5.082136 seconds (2 allocations: 762.940 MiB, 0.59% gc time)
  0.122529 seconds (2 allocations: 762.940 MiB, 26.56% gc time)

[yuyichao]

This is intentional due to the old version producing inaccurate result. Ref #19872

[mbrookhart]

Ha, okay. I tried the same idea in C++ and I see a similar regression between pow(x,2)*pow(x,2) and pow(x,4), this must be one place where generic programming is biting me.

Thanks!