On 1/24/2013 8:43 AM, H. S. Teoh wrote:
I remember being actually offended when I read about premature
optimization. I felt insulted at the idea that my hand-written code
could actually underperform what the compiler mechanically produces. Too
bad it's actually true. :-P
I later learned that my obsession with premature optimization distracted
me from actually thinking about the algorithms, so that I was spending
90% of my time optimizing the last CPU cycles out of an O(N^2) algorithm
when even a poorly-written O(log N) algorithm would have easily
outperformed my "optimized" code.
Not to mention the humiliation of discovering, when I actually profiled
my code, that the hotspots were nowhere near where I thought they were.
I had spent 90% of my time "optimizing" for a <1% gain, when a simple
algorithm fix at the real hotspot gained 20% instantly.
Don't feel bad. Everyone goes through that.
Keep in mind that people who are attracted to the dirty job of writing compiler
back ends tend to be people who are old time assembler hackers, and they know
the tricks of the trade. Naturally, they wire that knowledge into the compiler.
And they do learn new tricks all the time, and put those in, too.
It's like thinking you can buy a $10 gimmick from the auto parts store and
increase your gas mileage 10%. Ain't gonna happen, the car makers aren't stupid.
