You are saying "look at all the wonderful things optimizing compilers can do". And he is saying "that's cool, but it doesn't really matter".
Now I am pretty sure he wold concede immediately that there are some examples where it matters, but my experience matches what he is saying very closely.
A little bit on my background: hired in 2003 by the BBC for the express purpose of making one of their systems faster, succeeded at around 100x - 1000x with simpler code[1], hired by Apple to work on their Mac OS X performance team, similar successes with Spotlight, Mail indexing, PDF, etc. Also worked with/on Squeak. Squeak runs a bytecode interpreter, with a bunch of primitives. The interpreter is perfectly adequate for the vast majority of the system. Heck, the central page imposition routine in my BookLightning PDF page imposition program[2] is written in Objective-Smalltalk[3], or more precisely an interpreter for the language that's probably the slowest computer language implementation currently in existence. Yes it's slower than Ruby, by a wide margin. And yet BookLightning runs rings around similar programs that are based on Apple's highly optimized Quartz PDF engine. And by "rings" I mean order(s) of magnitude.
Why is BookLightning so fast? Simple: it knows that it doesn't have to unpack the individual PDF pages to impose them, and is built on top of a PDF engine that allows it to do that[4]. The benefit of an optimizing compiler in this scenario? Zilch.
At the BBC, the key insight was to move from 20+ machine distributed and SQL database backed system to a single jar event logger working in-memory with a filesystem based log[5]. How would a compiler make this transformation? And after the transformation, we probably could have run interpreted byte-code and still be fast enough, though the JIT probably helped us a little bit in not having to worry about performance of the few algorithmic parts.
As to changing a bubblesort to a mergesort, you hand-wave around that, but I think the point is that this is not a safe transformation, because the author may have specifically chosen bubblesort because he likes its characteristics for the data he knows will be encountered (or cares more about that case).
When I worked at Apple, I saw the same pattern: low-level optimizations of the type you discuss generally gained a few percent here and there, and we were happy to get them in the context of machine/system wide optimizations. However, the application-specific optimizations that were possible when you consider the whole stack and opportunities for crossing or collapsing layer boundaries were a few factors x or orders of magnitude.
And here is where the "optimizing compiler" mindset actually starts to get downright harmful for performance: in order for the compiler to be allowed to do a better job, you typically need to nail down the semantics much tighter, giving less leeway for application programmers. So you make the automatic %-optimizations that don't really matter easier by reducing or eliminating opportunities for the order-of-magnitude improvements.
And yes, I totally agree that optimizations should be user-visible and controllable, rather than automagically applied by the compiler. Again, the opportunities are much bigger for the code that matters, and the stuff that is applied automatically doesn't matter for most of the code it is applied to.
[1] http://link.springer.com/chapter/10.1007%2F978-1-4614-9299-3...
[2] http://www.metaobject.com/Products/