What if we had a machine that understood LLVM bitcode directly
-
-
-
We could call it JazeLLe
- Još 1 odgovor
Novi razgovor -
-
-
I actually think that would work *worse* than for more modern arches, just because you need so many more instructions to do anything useful
-
and my understanding is that "more simpler instructions" is worse in terms of synthesis than "fewer complex instructions" because it has a larger symbolic representation
- Još 2 druga odgovora
Novi razgovor -
-
-
my take is that this isn't the case: there's still a significant gap between the largest program you can legitimately synthesize and smaller actual microcontroller programs
-
that's not to say that superoptimizers wouldn't be awesome for quirky embedded architectures! they would, but you would still need a way to focus the effort on smaller program units like functions or parts of functions
Kraj razgovora
Novi razgovor -
-
-
I remember
@ricbit writing a superoptimizer for the Z80 ~20 years ago. It worked fine, but for the PCs of the era, it was too slow for anything other than small expressions.@whitequark recently used a solver to get a 8051 to do arbitrary shifts, and I presume it's a lot faster - Još 2 druga odgovora
Novi razgovor -
-
-
On such CPUs, there is often a code-size/dynamic-data-size tradeoff to be evaluated, as well as code-size/code-speed tradeoff; and the former evaluation really needs to consider the *entire* program. It's tough going. Libraries and their ABIs make such difficult, a balancing act.
-
For instance, on the 68HC11, choices have to be made between a memory-load followed by sequence of 4 shift instructions to maneouver a data-bit into the right position, versus using bitwise AND with a branch and a load-constant instruction: [more]
- Još 1 odgovor
Novi razgovor -
Čini se da učitavanje traje već neko vrijeme.
Twitter je možda preopterećen ili ima kratkotrajnih poteškoća u radu. Pokušajte ponovno ili potražite dodatne informacije u odjeljku Status Twittera.