Sunday, 2022-07-10

ghostmansd[m]It striked me today that ff/pr and dm/sm/m share same predicates so I can unify some code together. Actually, I must check than only CR predicates are allowed in ff/or (plus RC1/~RC1, but this is yet another story).07:32
ghostmansd[m]Actually this works like this: allow everything but RC1/~RC1 for dm/sm/m, and allow only CRs/RC1/~RC1 for ff/pr.07:42
ghostmansd[m]So I encoded all this information into one big table.07:42
ghostmansd[m]So I ended up with 16 insertions and 113 deletions in patch. Not bad.07:43
ghostmansd[m]Also fixed an infinite recursion in expr parser I introduced. If you define md_operand routine (and we did it for handling vector * notation), you have to be damn careful.07:48
lkclff and pr are modes, they do not take arguments.10:53
lkclwait... ff yes, takes the same predicate-mode bits10:57
* lkcl checking pr10:57
lkcl11invCR-bitRc=1: pred-result CR sel10:58
lkclyes that takes the same predicate-mode bits as well10:59
lkcl01invCR-bitRc=1: ffirst CR sel10:59
lkclfor fail-first10:59
lkclThe Mode table for Arithmetic and Logical operations is laid out as follows:10:59
lkcl0-123 4description10:59
lkcl01invCR-bitRc=1: ffirst CR sel10:59
lkcl01invVLi RC1Rc=0: ffirst z/nonz11:00
lkcl11invCR-bitRc=1: pred-result CR sel11:00
lkcl11invzz RC1Rc=0: pred-result z/nonz11:00
lkclyes. interesting. they do indeed use the same so/ns/eq/ne/..../gt/ge11:01
lkclmy memory's hosed.11:03
lkcland i'm not sure i've got test_caller unit tests yet for fail-first or predicate-result so it's not stuck in my long-term memory11:04
lkclghostmansd[m], i had to rename RS in svindex to SVG to specifically avoid it getting detected and processed as a GPR11:49
lkclthe position does not change the size does not change nothing changes except the name.11:49
ghostmansd[m]lkcl, ack; will update it in binutils12:55
lkclthx. otherwise it gets treated as a register source field, which is bad12:59
lkcleven the fact that it's called "RS" confuses ISACaller and the parser12:59
ghostmansd[m]For binutils it'll also do since RS sets GPR register flag, and that one's later being checked.15:41
lkclah yeah16:20
ghostmansd[m]So, a simple numeric field within the same bit range as RS, right?17:03
lkclyes 5-bits although strictly speaking it will be <<2.17:25
lkclso numbers 0 4 8 12 16 .... 12417:25
lkclbut i'm not quite ready for that yet17:26
lkclso just a numeric field 0-31 is fine for now17:27
ghostmansd[m]Should I put it on the review on mailing list, or it's a bit early yet?19:43
ghostmansd[m]Because there's already svindex patch, but, if you feel svindex might change soon again, I can postpone it and keep it local19:43

Generated by 2.17.1 by Marius Gedminas - find it at!