Improve generated debug info to avoid assertion in SROA #13762
Conversation
|
Turns out it just runs through the memory cap trying to optimize the system image ;). |
|
So the "LLVM bug" is that they didn't buy you more memory ;-p Thanks for the fix. I can confirm that this indeed fixes the bootstrap failure on linux too. |
|
Not quite. The problem is that bugpoint limits executions of the optimizer to 400MB, the machine has plenty of RAM. |
I remember you said "LLVM does not keep such counts". Is this a recent improvement (that we could use for benchmarking) or is it just running it in a separate process and let the OS take care of it (e.g. with |
|
Yes, it's using |
|
Hmm, looks like we have a non-deterministic test failure here :(. Will see if I can reproduce. |
| @@ -4569,7 +4569,7 @@ static Function *emit_function(jl_lambda_info_t *lam) | |||
| s->name, // Variable name | |||
| topfile, // File | |||
| toplineno == -1 ? 0 : toplineno, // Line (for now, use lineno of the function) | |||
| julia_type_to_di(varinfo.value.typ, ctx.dbuilder, specsig), // Variable type | |||
| julia_type_to_di(varinfo.value.typ, ctx.dbuilder, varinfo.value.isboxed), // Variable type | |||
There was a problem hiding this comment.
isboxed isn't really well defined for varinfo.value, but regardless, it hasn't been assigned yet here. i'm not certain that this is what you want here anyways. I think you want to check after emitting the function whether a memloc needed to be created for this variable, or whether it was allocated in .value and whether varinfo.value.typ is a known type or not.
There was a problem hiding this comment.
I'm considering always passing false here and handing this at the callsite.
|
Ok, now passing |
|
Does this fix #13754? Is it okay to merge? |
|
This does fix #13754 for me. |
Improve generated debug info to avoid assertion in SROA
This is a very simple start on emitting more thorough debug info. I was planning (and still am) to work on this as part of a larger push to improve our generated debug info, but at the moment SROA is crashing on 3.7 (#13754), because our debug info is not quite valid. Interestingly, I can still see a crash if I run the system image that results from this through bugpoint, the optimizer still crashes, so we might be hitting another LLVM bug here. Bootstrap goes through fine with this though.