What are costs of disabling interrupts vs LDREX/STREX on Arm Cortex M?

Question

On ARM Cortex M, I'm aware of only two ways to achieve atomicity:

1. LDREX/STREX
2. Disable interrupts

Both can be used in a very similar way: For example, define volatile bool is_locked, and check / set it either with one of those two methods.

What are the costs in performance for each? I imagine we need to think of two things;

1. Throughput: How many cycles are spent on each, assuming no conflicts
2. Latency: For how long are interrupts delayed

Are there any other tradeoffs between the two?