Does it make sense to check hold violations at synthesis stage

As we know from basics of STA, hold timing equation is generally of the form:

Hold_slack = Data_path_delay - hold_time_of_flop - clock_skew

Which can be re-organized as follows:

Hold_slack = Data_path_delay + launch_clock_delay - capture_clock_delay - hold_time_of_flop

Data_path_delay + launch_clock_delay may be combined to represent arrival of data at the capture flip-flop with respect to clock source. The same is evident from below figure.

The modified equation, then, becomes:

Hold_slack = Data_arrival_at_capture_flop - Clock_arrival_at_capture_flop - hold_time_of_flop

It is clear from above equation that hold checks generally comprise of a race condition between clock and data minus a fixed number (is hold_time_of_flop really a fixed number is a separate question).

Before clock-tree synthesis, we do not have one major aspect of this equation; i.e. we do not know the arrival times of clock at launch and capture flip-flops. Also, we do not know how much is the skew and uncommon clock path contributing to on-chip variations; thus, requiring extra margin for hold slack. Talking about logic synthesis, we do not even have placement data most of the times, so we do not even have correct data path delay estimates with us. So, fixing hold at synthesis will not help as there will anyways be requirement for hold fixes taking into account actual data and clock path delays.

But the most important reason for not fixing hold before clock-tree synthesis is that the task of hold fixing is not much complex for the tools available. It may be as simple as downsizing logic and/or adding buffers in data path where setup slack is available. So, during synthesis and logic placement, tools focus on getting setup targets met and focus on hold after clock tree is built.

Another point to note is that after data nets routing, data path delay may increase because of detouring of data nets as compared to originally estimated. So, it may be wise not to fix hold violations of very small magnitude even after clock-tree synthesis and wait for data nets routing to see how many of those violations are still there.