I cannot emphasize this fact as much as it deserves. You cannot and you should not expect the compiler will do the job for you the best way possible no matter what the code is. I’m not saying compilers should not try to optimize code, I’m saying no compiler, however sophisticated it might be, will ever be able to do the job exactly the way you want, and in any given case, without you giving it directions.
Moreover, I have seen people relying on Matlab (or, even worse, Computer Algebra Systems like Mathematica and the likes) to do heavy numerical computations. It makes no sense, and I’m not just talking about the fact such systems will do the job orders of magnitude slower. This trend basically created a whole generation of new scientists who use numerical codes without having any idea about the numerical methods involved. I dare say they have no idea what they are doing, but they are doing it anyway. This extends to simpler things such as matmul or transpose in any Fortran compiler. For example, nobody should ever expect the compiler will be “clever” enough to create a local copy of a matrix when you want and not doing so when you don’t want it to happen. Expecting things like that from a compiler in every single case makes no sense. Use your brains instead.
Have you ever tried to compile Scilab from source? It requires a Fortran compiler and it compiles venerable Fortran “packs” such as ODEPACK. You can easily guess why. Now, I cannot tell what Matlab does under the hood because it is non-free software. But last time I checked the syntax of Matlab commands for numerical computations (and even the name of such commands) are very much close - if not identical - to the procedures in Fortran packages found in Netlib, GAMS, and elsewhere.
Software like Matlab, Scilab, Octave, etc is great for quickly trying a new idea, see if it works, and get the first convincing results. But if further development requires heavy computations you should switch to Fortran, or at least a decent compiler in whatever language you like. Even then, relying on abstractions is nonsense. Using black boxes without at least knowing what exactly they are doing - and how - is the worst practice ever, and yet it becomes more and more popular.
I’m not looking for a compiler that will do its “magic” so I don’t have to analyze the situation and optimize the code myself in the first place. I just want a compiler that will do what I tell it to do. Similarly, I’m not looking for a magic box that will solve that system of stiff differential equations in some way I barely know anything about. Even worse if we are talking about proprietary, non-free software. I just want a programming language clearly made for the task, giving me the tools to solve the problem myself instead of relying on black boxes.
As for the original topic, the question is not to evaluate the risks of relying on Fortran for mission critical codes, the question should be evaluating the risks of relying on the popular languages nowadays. Relying on such a language that obviously, definitely wasn’t made with Numerical Analysis in mind is a very, very bad practice.