Ridiculous perf difference in almost-the-same F# code

sample.fs

// Version A
let private renderTriangleAlways renderer ctx v1 v2 v3 c =
    let ymin', aes = getAEs v1 v2 v3

    let mutable ymin = ymin'
    for ae1, ae2 in aes do
        let ymax = ae1.YMax
        for y = ymin to ymax - 1 do
            let minX = int (min ae1.X ae2.X)
            let maxX = int <| ceil (max ae1.X ae2.X)
            for x = minX to maxX do
                NativeInterop.NativePtr.set ctx.Context.Pixels (x + y * ctx.Width) c
            ae1.X <- ae1.X + ae1.CoeffX
            ae2.X <- ae2.X + ae2.CoeffX
        ymin <- ymax

// Version B - ~2 times slower
let private renderTriangleZBuffer renderer ctx v1 v2 v3 c =
    let ymin', aes = getAEs v1 v2 v3

    let mutable ymin = ymin'
    for ae1, ae2 in aes do
        let ymax = ae1.YMax
        for y = ymin to ymax - 1 do
            let minX = int (min ae1.X ae2.X)
            let maxX = int <| ceil (max ae1.X ae2.X)
            for x = minX to maxX do
                let idx = y * ctx.Context.Width + x
                NativeInterop.NativePtr.set ctx.Context.Pixels idx c
            ae1.X <- ae1.X + ae1.CoeffX
            ae2.X <- ae2.X + ae2.CoeffX
        ymin <- ymax