moonchild

9 ms·

61.
by moonchild 3y ago
Extant cl compilers are quite primitive. Well known and uncomplicated analyses exist that can cope with this; e.g. in llvm https://godbolt.org/z/nadbG5zGs
62.
by moonchild 3y ago
Hrm—that's another case of missing isel from gcc/clang (c.f. < https://pp.ipd.kit.edu/firm/selgen >)—my implementation (< https://github.com/moon-chilled/fancy-memcmp/blob&#x2F
63.
by moonchild 3y ago
https://www.youtube.com/watch?v=Wa81OJnlsoI
64.
by moonchild 3y ago
read the garbage collection handbook - https://gchandbook.org/
65.
by moonchild 3y ago
I am extremely confused. > There are two scenarios in which a virtual thread cannot be unmounted during blocking operations because it is pinned to its carrier: > When it executes code inside a synchronized block or method Isn't
66.
by moonchild 3y ago
> upto =: 4 : 'x}.i.1+y' Fyi: for a while now, j has supported the more convenient 'direct definition' syntax. You can now write that definition like: upto =: {{ x}.i.1+y }} It will automatically make a monad or a dy
67.
by moonchild 3y ago
It could definitely be clearer, but what makes you say the manual is misleading?
68.
by moonchild 3y ago
Blind virtual transpose (as seen in numpy) is a bad idea. A principled, locality-aware version would be fine and good.
69.
by moonchild 3y ago
Regarding k vs j, I will just say that ngn and I both agree that both the k and j array models are far more coherent than bqn. And that k is slow for multidimensional arrays ('mangle your data so it's fast in lists' is a poo
70.
by moonchild 3y ago
You would have liked fortress.
71.
by moonchild 3y ago
> VM model for APL languages It's cute—but from my skimming a while ago fairly primitive. We can do much better with less effort using more general mechanisms. (Not a knock—it's a product of it's time—a lot of old compil
72.
by moonchild 3y ago
learn apl (k and j count)
73.
by moonchild 3y ago
You have to handle it manually. Remember the partial lines at the beginning/end of your chunks and merge them when their mates become available.
74.
by moonchild 3y ago
Lulu does not make 'high-quality hardcovers'.
75.
by moonchild 3y ago
The default probability of any branch, absent further information, should be 0.5, so I wouldn't expect this to do anything. You should rather direct your ire at llvm, which has __builtin_unpredictable which is explicitly supposed to d
76.
by moonchild 3y ago
Media codecs are a microcosm of routines that want more care and attention than compilers are yet able to give. Some of RAD's decompression kernels are written in assembly, for instance, and a particular unicode transcoding routine wa
77.
by moonchild 3y ago
3 cycles is throughput, not latency; latency for normal loads is not very different from desktop parts. They win for atomic ops because they don't have a fence built in; I don't understand why intel hasn't specced non-fencin
78.
by moonchild 3y ago
> store-store dependency You can make it an exchange rather than a plain store. Questionably worthwhile but you can do it. swp is only 3 cycles on firestorm https://dougallj.github.io/applecpu/firestorm-int.html
79.
by moonchild 3y ago
You don't need a fence if you can thread a data dependency through from all memory operations inside the critical section to the release.
80.
by moonchild 3y ago
You have your causal arrows flipped around. It is because unix is not graphical that the graphical interfaces bolted onto it are poor and the only good way to interact with it is via the command line. If a system were designed to be gra
81.
by moonchild 3y ago
Having written a 2-d curve renderer, what I want is parallel compute and high bandwidth; gpus deliver. (Especially with newer interfaces that support scatter-write; not sure how much penetration these have in the browser yet.) It's t
82.
by moonchild 3y ago
See also: unison - http://unison-code.github.io/
83.
by moonchild 3y ago
The whole reason science works is that we start by assuming what our eyes tell us is true, and then come up with models to explain that. Edit: plato believed the nature of reality could be discerned with pure reason. He was wrong. Scienc
84.
by moonchild 3y ago
I have the opposite question: why not write everything in java?
85.
by moonchild 3y ago
Common lisp and scheme have had proper numeric towers without this problem for decades.
86.
by moonchild 3y ago
You would need a dedicated core per tab . Partitioning tabs according to trust would be ~fine, but laborious and error-prone.
87.
by moonchild 3y ago
Rational numbers do not overflow or underflow. The only reason for having a type with this degenerate behaviour is to screw over the unsuspecting user who expects the language to protect them because they saw it using rational number rep
88.
by moonchild 3y ago
There is no reason why it could not continue to use the exact representation and operations if it wished (and that's what FatRat does). It's just an ill-conceived concession to performance.
89.
by moonchild 3y ago
It's actually an awful misfeature, because those Rats will automatically turn into floats when the rational representation gets too big: > WHAT 1/10 (Rat) > WHAT 1/100000000000000000000 (Num) (There is FatR
90.
by moonchild 3y ago
> Decimal calculations getting rounded off is a problem that is a lot easier to see and understand. If you carefully, tediously work by hand some calculations such as addition, multiplication, or division (but not, say, square roots, exp

More ›