...and if the anode of the diode is already connected to chassis ground then you would have to "bounce the current off" a current mirror on the negative rail, there are schematics aplenty for doing that with ANOTHER opamp/transistor pair...!
Thank you for your answers.
I have understood the sharp I/V curve of the laser diode. I was too concentrated on the "laser" word, and I forgot about the classic diode characteristic.
Unfortunately, the laser diode is only available with the anode connected to the chassis ground. And with the goal to have 3GHz of bandwidth, I can't connect the RF path to the chassis ground (impedance mismatch).
So, I have modified the suggested constant current, to be compatible with negative voltage.
Here is my LTspice schema
And the simulation result (V3=V DAC or V SET / V(p001) = V LOAD / I(R3)=I LOAD)
My conclusion
The constant current source is working for negative power supply.
I can reduce the DAC precision to reduce cost.
I have added voltage drop in the -3VDC path, I got -2.13VDC instead of -3VDC at 40 mA. It seems ok.
I tested the new prototype and it works perfectly.
With a constant current source, I can calibrate and control my laser diode from 2.5 mW to 4.5 mW (optical power).
Thank you all for your support.