3 ms·

I do a lot of low-level systems programming; think rasterizers, drivers, etc. There's numerous occasions where I've needed to hand-code some sort of work-around

I do a lot of low-level systems programming; think rasterizers, drivers, etc. There's numerous occasions where I've needed to hand-code some sort of work-around, or to paper over a gap in the functionality provided by the hardware. In C/++, this isn't really a thing: the language only has so much flexibility in what in can say. When I really have to get into things, though, I end up writing assembly for the truly hot-loops. In that case, as fugly as x86/64 is, it is one of the best ISAs for exactly expressing the algorithm I have in mind. Personally, I find the most advanced NEON specs are about a decade behind the x86 vector specs. Generally, while the ARM implementations can keep up cycle-for-cycle on commodity parts for scalar, algorithmically, x86/64 blows it out of the water. I think the article cites about 5x? In the worst case, I've seen places where ARM falls entirely out of the I$ and thrashes its D$ & I$ on scalar work-arounds for gaps in its NEON set, where x86/64 manages to just squeak by completely vectorized. In those cases I see 50x (or more) delta, even though the spec sheet would imply that ARM is "just as good" as x86/64. I like Mike Abrash's quote on this (to paraphrase): "assembly isn't faster than C++; assembly is more expressive than C++".

pjmlp 9y ago
Back in the MS-DOS days, when I had lots of fun using TASM, the common way to write high performance C and other higher level languages on 8 and 16 bit computers was using inline Assembly everywhere. The typical for 80% of the application void func(args) { asm { } } Literally poor man's macro assembler. Now 40 years of optimizing compiler improvements, we have younger generations thinking that C is like writing Assembly and C compilers were always blazing fast. So when I see statements about other compilers generating slower code than C ones, I remember those days and smile.

pjmlp 9y ago

Back in the MS-DOS days, when I had lots of fun using TASM, the common way to write high performance C and other higher level languages on 8 and 16 bit computers was using inline Assembly everywhere. The typical for 80% of the application void func(args) { asm { } } Literally poor man's macro assembler. Now 40 years of optimizing compiler improvements, we have younger generations thinking that C is like writing Assembly and C compilers were always blazing fast. So when I see statements about other compilers generating slower code than C ones, I remember those days and smile.