Re: Drawing the line between speed and simplicity/elegance

[Coda Highland <chighland@...>]

On Sun, May 10, 2015 at 12:11 AM, Enrico Colombini <erix@erix.it> wrote:
> On 10-May-15 02:53, Sean Conner wrote:
>>
>>    Just on a lark, I wrote two programs, one in pure Lua; the other one
>> pure
>> C.  The programs calculate the Mandelbrot set (but don't output
>> anything---I
>> wanted just the pure calculations) [2].  Here are some timings:
>>
>>         Lua in PUC-Lua: 52.7s
>>         C:               2.7s (compiled with -O3)
>>         Lua in LuaJIT:   2.5s
>
>
> Sean, out of curiosity, did you also measure with -Os?
>
> I am asking this because, in a different context (a small microcontroller) I
> saw code emitted with -O3 that was slower than using -Os, despite 's'
> meaning 'size' and not 'speed'. It did strange things, such as jumping to a
> common piece of code (and back) from the middle of the part I was trying to
> optimize.
>
> (disclaimer: I did not check if the code in its entirety was actually
> faster, because I was interested in squeezing the last clock cycle out of a
> particular section)
>

Data point: OSX defaults to -Os these days. It isn't "minimize size at
all costs" (clang has -Oz for this, gcc doesn't have an equivalent
flag AFAIK) but rather "-O2 unless it would increase the binary size,
plus some extra size optimizations." The claim is that the cache
benefit outweighs the few lost optimizations you would have otherwise
had.

/s/ Adam