<alyssa>
OK, so if I replace the imov's with iand's (with the source and itself), the problem remains
<alyssa>
Dunno if that's good or bad
<alyssa>
For one it says the imov isn't 'special' as far as int ops go
<alyssa>
Maybe I should copyprop it away and stuff worrying so much
<alyssa>
If this is a genuine bug that needs more attn, it'll show up on CTS (though maybe not till GLES3)
<alyssa>
And I'll know what to look for
<alyssa>
If it's not, well, shrug :P
<alyssa>
(It's totally possible there's a hw bug or something that never got discovered since this is a bad code path :P)
* alyssa
implements copy prop
<alyssa>
MIR is suddenly becoming a real IR I guess
vstehle has joined #panfrost
<alyssa>
There, just sent off an overly long patchset with the weekend's changes. *whistles*
* alyssa
is off to do weekday stuff, evidently, adieu \o
robclark has quit [Quit: Connection closed for inactivity]
<tomeu>
alyssa: the plan isn't to run all tests and fail if they all not pass, but to maintain a list of "expectations"
<tomeu>
which are the tests that should be passing atm
<tomeu>
and when we commit code that changes the expectations, that commit should ideally update the expectations (adding the tests that now start to pass)
<tomeu>
alyssa: the crash when changing governors is a good one
<tomeu>
indeed, you should do that if you want to have the GPU running at 100%
stikonas has joined #panfrost
stikonas has quit [Remote host closed the connection]
bbrezillon has joined #panfrost
_whitelogger has joined #panfrost
stikonas has joined #panfrost
<tomeu>
alyssa: I'm seeing crashes when freeing random stuff, and running under valgrind gave me this: http://paste.debian.net/1078756/
<tomeu>
haven't checkd if it isn't a false positive
<tomeu>
that's in your indexing branch
<alyssa>
x
<alyssa>
tomeu: I'm worried such a list will be impossibly difficult to keep up to date
<alyssa>
Since when I fix a test, usually a bunch of other tests also fix that I'm not aware of (since it's annoying to rerun the entire CTS)
<alyssa>
Not sure what you mean by "a good one"..? :P
<alyssa>
That valgrind trace looks like a legitimate bug. I'll have to poke around when I get a chance (not sure when that'll be)
<tomeu>
alyssa: the second is due to ctx->temp_count being 0, but then calling ra_set_node_reg
<tomeu>
alyssa: regarding updating the test list, you can just let a run complete in your repo, and that will tell you what expectations have changed and thus what needs to be updated
<tomeu>
a big part of this is for running the entire CTS to stop being annoying :)
<HdkR>
That reminds me that I should run my code through valgrind
* HdkR
adds to task list to remember
<tomeu>
works suprisingly well on arm
<HdkR>
Did gdb fix its issue of breakpoints not working on AArch64?
<HdkR>
Where it would break and `c` just wouldn't continue
<tomeu>
don't know, I'm currently on rk3288
stikonas has quit [Remote host closed the connection]
<HdkR>
I guess probably
<HdkR>
Last time I tested was...over a year ago at this point
<HdkR>
Although last time I tested was when I was disassembling a JIT and GDB would try handling a GCC style stack frame and explode
<HdkR>
Which was awesome
<tomeu>
yeah, saw such things when debugging memory issues in browsers on arm
<tomeu>
hopefulyl stuff has been fixed since
<alyssa>
tomeu: "you can just let a run complete in your repo, and that will tell you what expectations have changed" What is this voodoo? :P
<tomeu>
well, you push code, and the pipeline will fail because of differences between expectations and reality
<tomeu>
a diff will be presented to you that can be applied to your branch
<tomeu>
the subsequent run should succeed, if there's no flip-flops
<tomeu>
panfrost-ci isn't there yet, but it's what I implemented for virgl
<alyssa>
tomeu: Like, I push to a branch on my personal repo, and then I can forcepush over with changes to the expectations list (so we don't create a million micro commits)?
<tomeu>
yep, but first I need to create a lava account for your branch
<tomeu>
and well, need to get the CI working :)
* alyssa
smiles and trusts you know what you're talking about :)