freenode/#sicl - IRC Chatlog
Search
14:27:40
beach
So, anyway, there are two completely separate phases to my idea. Phase 1 is to compute the EDUs and phase 2 is to use those EDUs to allocate registers. So I'll put phase 1 on hold for the time being and start thinking about phase 2.
15:14:26
heisig
Solving such a linear system efficiently should not be too hard. Especially since we don't need an exact solution - an approximation would be fine.
15:14:46
beach
It is also in chapter 10 of the Cleavir documentation in SICL/Code/Cleavir/Documentation.
15:17:58
beach
In the Cleavir documentation, the equations are not linear, but I think a linear system is a fine approximation.
15:52:58
beach
... which is too bad, because I am not really that interested in that particular architecture. But that's what we have for the time being. Oh, well.
15:59:03
heisig
Hmm, x86 could actually be gone soon. They will never be able to decode more than five instructions per cycle, and that is starting to hurt.
16:09:41
beach
Bike: What semantics do we attribute to UNBOX instructions? Are the responsible for checking that the input is of one of the right types (and signaling an error otherwise) and then applying the relevant unboxing operation? If so, is there any way that the type inferencer could annotate such an instruction so that some steps could be skipped in some cases?
16:11:54
Bike
boxing and unboxing is kind of still in flux. HIR unbox definitely checked types, though.
16:13:01
beach
That's what I though. Yes, maybe. But it's tricky since more than one object type could work, and the unboxing operation would be different for each one.
16:14:25
beach
I guess it's perhaps only for the ((un)signed-byte 64) that both a fixnum and a bignum might work.
1:31:13
no-defun-allowed
Another "quirk" of the linear probing hash table is that you can see when it has to probe multiple groups. With groups of 8 and a 0.8 load, there's a 16.7% probability of a group being full (so that GETHASH has to scan a second group), but with smaller groups, the probability decreases.
1:32:08
no-defun-allowed
Oh, smaller load, not groups. I was thinking about my next sentence, which is: This probability also decreases with larger groups, e.g. the 16-element groups which can be searched with 128-bit SIMD instructions.