A fixed evaluation order may be (very slightly) less efficient than if the compiler is allowed to reorder. Consider evaluating "x + f()" when x is global and the compiler can't prove that x cannot change during execution of f. If the order of evaluation is fixed from left to right, then it must do the following:

1. load x into a register

2. spill the register holding x to the stack

3. call f

4. reload the "local" version of x from the stack

5. add

On the other hand, if the order of evaluation can be chosen by the compiler, it can do this:

1. call f

2. load x into a register

3. add

(Details depend on whether the caller or the callee saves x, but you get the idea.)

In practice, all competitive optimizing compilers are SSA-based, and thus they fix an order of evaluation early in the compilation process. I always wondered how much performance could be gained by actually delaying decisions about evaluation order.

Not sure if you're tyring to suggest that SSA-based compilers cannot do this reordering. For my example above, GCC reorders as I sketched, while Clang keeps the program order: https://godbolt.org/g/8KDvFo

I think this is done early, during lowering to a "mid-end" intermediate representation (whether SSA or not). This is the best time to do it, since the legality of this reordering is a property of the source language, which you want to forget about in the mid-end. On the other hand, in this particular example (function call vs. global variable access), I don't think there is any reason to delay this; doing the call first should never be worse than loading the global first.

As for deciding other things later, instruction scheduling and global code motion can reorder a lot of things, though calls and memory accesses block many movements.