<somlo>
daveshah: if I use yosys at #9e4801cca7, my point-to-point wide axi rocket-litedram setup works fine; as soon as I move to the immediately adjacent #ca99b1ee8dc, linux hangs on boot
<somlo>
it's only the point-to-point AXI based design that exhibits the problem, and it doesn't look like stochasticity of p&r has anything to do with it (happens pretty predictably)
<somlo>
out of desperation at how little sense this makes, I'm building the latest yosys (#419ca5c2) with #ca99b1ee reverted, to see what happens
<daveshah>
The only change I can see in that PR is some new idstrings, which might have a small stochastic impact on synthesis
<somlo>
maybe we're tickling a migen, or litex/axi bug that somehow didn't manifest earlier?
<daveshah>
If you try changing an identifier name somewhere, that should have a similar permutation effect
<daveshah>
Seems unlikely, more likely an issue on the Yosys or nextpnr side
<somlo>
ok, let's see what happens when I build with #419ca5c2 and reverted #ca99b1ee (should be up to another hour or so -- as I mentioned yesterday, later yosys commits take much longer to build litex/rocket on the trellisboard)
<somlo>
and that's out of a sample of two (both times I tried it over the last two-three days, it took a LOOONG time to p&r the output of latest yosys)
<somlo>
ok, yosys done compiling, now running litex build
<somlo>
daveshah: reverting #ca99b1ee on top of current master (#419ca5c2) got me a working design
<somlo>
now, to be 100% sure, I'm building #419ca5c2 without the revert, to rule out some other interaction effects
<somlo>
would it make sense to diff the yosys json output with vs. without the "offending" commit ?
<daveshah>
I expect it would be too chaotic to diff in any meaningful way
<daveshah>
But you could have a look at the ilang after `proc`
<daveshah>
that is the point where that commit stops having any impact - so if the rtlil is the same after `proc` with and without the commit then it is some other stochastic impact of the commit rather than a bug in or triggered directly by the commit
<somlo>
seems without the revert, yosys-master generated json once again takes very long to p&r
<somlo>
specifically, *route* (placement takes about the same as it ever has)
nrossi has quit [Quit: Connection closed for inactivity]