FastMath breaks elytra on OptiFine 1.12.2 F5

[surfaceflinger]

Looking straight up while flying elytra will kick player for "Invalid move player packet received" or glitch arms + legs of player making them disappear.

How to reproduce: 1. Enable Fastmath in performance settings 2. Look up while flying elytra
May be only happening on macOS because only one player reported this on discord before.

vid: https://www.youtube.com/watch?v=JruqdJO1-vs

[Nessiesson]

Can confirm it happens on Windows 10 as well. From what I've been able to determine, the bug was introduced in OptiFine_1.12.2_HD_U_F4. Another way to reproduce the bug is to stand next to a block at head height like this and fly straight up. Your X and Z coordinates are somehow set to NaN.

[AMDBartek]

I can confirm that this happens on 1.12.2 when I am using the “Fast Maths” option on my client.

Edit: happens on Windows, Mac OS and Linux.

Edit 2: also happens while playing on my Minecraft server. (Small edit: this bug seems to effect both singleplayer and multiplayer)

[TheoCGaming]

Hmm. I'll test it in singleplayer and let you know if it bugs like shown.

Edit: Using F5 as it is the only version of OF1.12 I have on hand.

EDIT 2: Confirmed for F5 and confirmed in Singleplayer. Will attempt to get F3 screenshots and send them.

EDIT 3: While testing, my game froze, waiting for recovery | Subedit: Closing game due to game saving chunks indefinitely.

Edit 4: Captured screenshot, game froze again. Needed to printscreen
Image:

[xxDark]

                        if (f < 0.0F)
                        {
                            double d9 = d7 * (double)(-MathHelper.sin(f)) * 0.04D;
                            assert !Double.isNaN(vec3d.x * d9 / d5);
                            assert !Double.isNaN(vec3d.z * d9 / d5);
                            this.motionY += d9 * 3.2D;
                            this.motionX -= vec3d.x * d9 / d5;
                            this.motionZ -= vec3d.z * d9 / d5;
                        }

These assertions fail because vec3d.x, d5 are 0.0D.
0.0 * N / 0.0 == NaN
See EntityLivingBase#travel

There is a bug somewhere in this code:

    protected final Vec3d getVectorForRotation(float yaw, float pitch)
    {
        float f = MathHelper.cos(-pitch * 0.017453292F - (float)Math.PI);
        float f1 = MathHelper.sin(-pitch * 0.017453292F - (float)Math.PI);
        float f2 = -MathHelper.cos(-yaw * 0.017453292F);
        float f3 = MathHelper.sin(-yaw * 0.017453292F);
    }

@sp614x May you look into this? Kinda want some fix on that.
We suspect that this issue is caused by rounding. Fast math causes sin/cos calls to return inaccurate values.