IRC logs for #openrisc Saturday, 2012-06-09

stekernanother small step forward for llvm-or1k, u-boot for atlys compiled with clan/llvm now works (a couple of bugs got revealed when the spi flash subsystem got compiled in)07:05
stekernthe latest was strange, a stack offset of 4 got selected into "r1 | 4 <= const"07:07
stekernthat would only work if stack slots were 8 bytes, but imo I have defined them to 4 bytes everywhere07:08
stekernit's also possible to compile newlib with it now10:03
juliusb!!12:28
juliusbnice one12:28
juliusbyou thinking of submitting it upstream?12:28
juliusbstekern: hey, i'm doing a bit of thinking about the or2k isa13:31
juliusbyou reckon the cmov instructions are useful enoguh to keep?13:31
juliusband if so, what about the idea of having move if register is zero and move if register is non-zero?13:31
juliusbmnemonic of l.cmovz and l.cmovnz13:32
juliusbso i believe we're also thinking of not having a flag bit, but rather having comparisons store a 1 or a 0 in a destination register13:33
juliusbso you'd do your compare, have it in a register, and then pass it to l.cmov[n]z as rB in a rD <- [!]rB ? rA : rD;13:34
juliusbs/have it/have the compare result/13:34
juliusbi like the idea of cmov, avoiding branching over code which will just set a register's value13:35
juliusbso you might ahve two values oyu're chosing between to put into rD,  say r2 or r3, you could just do your compare then have two l.cmov instructions, for each13:36
juliusbso C code of "d = x ? y : z;" might turn into l.cmovnz rD, rY, rX; l.cmovz rD, rZ, rX;13:38
juliusb(that example doesn't use the compare thing, but you get what I mean, if it was x==64 you'd do your l.seqi rD, rX, 64, then do the same thing13:39
juliusb)13:39
stekernyeah, I think that's better than what we have now15:30
stekernregarding upstream, you'd need to make a commitment of being the maintainer longtime if they'd accept it, I'm not sure I can make such promises15:35
stekernbesides that, it's by no means ready for it ;)15:35
stekernbut it's pretty painless to keep a out of tree backend branch, usually little change, and when they do, they are good at announcing on the llvm-dev ml about it15:36
juliusbyeah, i imagine that'd be a bit of work I guess16:20
juliusbsounds cool though16:20
juliusb:)16:20
--- Log closed Sat Jun 09 17:24:23 2012
--- Log opened Sat Jun 09 19:43:20 2012
-!- Irssi: #openrisc: Total of 12 nicks [0 ops, 0 halfops, 0 voices, 12 normal]19:43
-!- Irssi: Join to #openrisc was synced in 24 secs19:43

Generated by irclog2html.py 2.15.2 by Marius Gedminas - find it at mg.pov.lt!