(This does not apply to e.g. auto-vectorization, which replaces 'x[0] = y[0] * z[0]; x[1] = y[1] * z[1]; ...' by a single (SSE?) instructions which calculates all of that at once. Such instructions are much faster and would require dropping down to assembly without a good optimizer.)
Absolutely. Fun to read, made me smarter, but half of it could be different in any other version of gcc.
And frankly, if these are your problems, you're doing pretty well. I worked on a product that had some numerical parts, which were sometimes the limiting factor in performance, though usually not. So one day I was talking to a guy who wrote the numerical code about why regression testing was difficult and subtle, and it was revealed that our numerical results could differ from build to build. I.e., a single build of the product always produced the same results, but if you changed the floating point code, the new build might produce different (though equally accurate) results, even if semantically the program was supposed to perform the same C math operations on the same data.
Which makes perfect sense if you're using the old stack-based x87-style floating point instructions. We couldn't be... not in the 21st century on Opterons... check Makefile and gcc man page... biggest facepalm of my career.
We were able to announce a nice speedup on some workloads in the next release.