10.0^68 gives a different result on nightly vs 1.11 after llvm-muladd pass removal

[KristofferC]

Master:

julia> 10.0^68
1.0000000000000001e68

1.11:

julia> 10.0^68
1.0e68

#55802 changed stuff related to muladd which pow uses at

julia/base/math.jl

Line 1206 in 2a06376

xylo = muladd(logxlo, y, xylo)

Putting on milestone so it isn't forgotten, not to say it is a regression (I don't know).

[Seelengrab]

Arguably, the new result is more correct, no? It's exact now, and used to have an error (within one ulp, probably?).

[giordano]

% julia +1.8 -E '10.0 ^ 68'
1.0000000000000001e68
% julia +1.9 -E '10.0 ^ 68'
1.0000000000000001e68
% julia +1.10 -E '10.0 ^ 68'
1.0e68
% julia +1.11 -E '10.0 ^ 68' 
1.0e68
% julia +nightly -E '10.0 ^ 68'
1.0000000000000001e68

Note that #49405 first appeared in v1.10, and #55802 restored the muladd behaviour outside of @simd on master.

[KristofferC]

It's exact now, and used to have an error (within one ulp, probably?).

Maybe I am dumb but isn't it the opposite?

[KristofferC]

Note that #49405 first appeared in v1.10, and #55802 restored the muladd behaviour outside of @simd on master.

I still don't want to backport a behavior change like that (PkgEval caught some new errors for example).

[Seelengrab]

Maybe I am dumb but isn't it the opposite?

You're right, I mixed up the versions 🤦

[giordano]

#55802 changed stuff related to muladd which pow uses at

julia/base/math.jl

Line 1206 in 2a06376

xylo = muladd(logxlo, y, xylo)

I think you're pointing to the wrong pow_body function, should be pow_body(x::Float64, n::Integer), not pow_body(x::Float64, n::Float64), but there's still a muladd call there:

julia/base/math.jl

Line 1275 in 2a06376

err = muladd(y, xnlo, x*ynlo)

The difference in code_llvm(Base.Math.pow_body, (Float64, Int)) is this in v1.11

L124:                                             ; preds = %L119, %L28.preheader
  %value_phi6.lcssa = phi double [ %value_phi6.ph, %L28.preheader ], [ %25, %L119 ]
  %value_phi7.lcssa = phi double [ 0.000000e+00, %L28.preheader ], [ %value_phi22, %L119 ]
  %value_phi8.lcssa = phi double [ 1.000000e+00, %L28.preheader ], [ %value_phi23, %L119 ]
  %value_phi9.lcssa = phi double [ %value_phi9.ph, %L28.preheader ], [ %22, %L119 ]
;  @ math.jl:1224 within `pow_body`
; ┌ @ float.jl:493 within `*`
   %28 = fmul contract double %value_phi7.lcssa, %value_phi9.lcssa
; └
; ┌ @ float.jl:496 within `muladd`
   %29 = fmul contract double %value_phi6.lcssa, %value_phi8.lcssa
   %30 = fadd contract double %29, %28
; └
;  @ math.jl:1225 within `pow_body`
; ┌ @ float.jl:705 within `isfinite`
; │┌ @ float.jl:492 within `-`
    %31 = fsub double %value_phi9.lcssa, %value_phi9.lcssa
    %32 = fsub double %30, %30
; └└
; ┌ @ bool.jl:38 within `&`
   %33 = fcmp uno double %31, %32
; └
; ┌ @ float.jl:496 within `muladd`
   %34 = fmul contract double %value_phi8.lcssa, %value_phi9.lcssa
; └
; ┌ @ essentials.jl:796 within `ifelse`
   %35 = select i1 %33, double -0.000000e+00, double %30
   %36 = fadd contract double %34, %35
   br label %common.ret
; └

vs this in nightly:

L124:                                             ; preds = %L119, %L28.preheader
  %value_phi2.lcssa = phi double [ %value_phi2.ph, %L28.preheader ], [ %25, %L119 ]
  %value_phi3.lcssa = phi double [ 0.000000e+00, %L28.preheader ], [ %value_phi10, %L119 ]
  %value_phi4.lcssa = phi double [ 1.000000e+00, %L28.preheader ], [ %value_phi11, %L119 ]
  %value_phi5.lcssa = phi double [ %value_phi5.ph, %L28.preheader ], [ %22, %L119 ]
;  @ math.jl:1275 within `pow_body`
; ┌ @ float.jl:493 within `*`
   %28 = fmul double %value_phi3.lcssa, %value_phi5.lcssa
; └
; ┌ @ float.jl:496 within `muladd`
   %29 = fmul contract double %value_phi2.lcssa, %value_phi4.lcssa
   %30 = fadd contract double %29, %28
; └
;  @ math.jl:1276 within `pow_body`
; ┌ @ float.jl:705 within `isfinite`
; │┌ @ float.jl:492 within `-`
    %31 = fsub double %value_phi5.lcssa, %value_phi5.lcssa
    %32 = fsub double %30, %30
; └└
; ┌ @ bool.jl:38 within `&`
   %33 = fcmp uno double %31, %32
; └
; ┌ @ float.jl:496 within `muladd`
   %34 = fmul double %value_phi4.lcssa, %value_phi5.lcssa
; └
; ┌ @ essentials.jl:790 within `ifelse`
   %35 = select i1 %33, double -0.000000e+00, double %30
   %36 = fadd contract double %34, %35
   br label %common.ret

Note the fmul which now isn't annotated as contract anymore, which was precisely the point of #55802 to fix #55785.

[KristofferC]

Note the fmul which now isn't annotated as contract anymore, which was precisely the point of #55802 to fix #55785.

You might be misunderstanding the point of this issue. I am not saying there is anything wrong with #55802. I am saying that the pow implementation might be wrong in how it relied on this bug since it now gives a "wrong" answer.

[oscardssmith]

How does the performance compare between before and after? I am slightly worried that I might have been accidentally relying on LLVM re-associating all my math.

[giordano]

You might be misunderstanding the point of this issue

What I tried to show above is that the current result has been the same since v1.8, except for the 1.10-1.11 window. Since that method has last been touched in the v1.9 time frame I don't see how that relied on a future bug.

[KristofferC]

It could have been wrong back then as well? All I know is a good result turned worse which seems like a regression no matter what values it used to give on old julia versions. The history doesn't really matter.

[mbauman]

Sure, but just because Julia grants (or granted) the ability to reassociate doesn't mean that the compiler does (or did). And it doesn't mean the implementation isn't (or wasn't) buggy.

On ARM64 (M1 Pro), I only see the +1 ULP error on v1.8; all others give 1.0e68.

These two lines are probably not meant to reassociate in the ways that muladd permits:

julia/base/math.jl

Lines 1275 to 1276 in 2a06376

	err = muladd(y, xnlo, x*ynlo)
	return ifelse(isfinite(x) & isfinite(err), muladd(x, y, err), x*y)

[oscardssmith]

in general, we don't promise <1 ULP accuracy for pow. it's nice if it works out well for integers, but that isn't a requirement