Fix some non-deterministic bugs. #638
Conversation
Periodically the PT hardware emits a PSB+ packet sequence. Each of these consists of: - A PSB packet. - other packets. - A PSBEnd packet. The Intel manual (Section 33.3.7) explains that the "other packets" are special and don't conform to the normal rules: > When a PSB packet is generated, it is followed by a PSBEND packet (Section > 33.4.2.18). One or more packets may be generated in between those two > packets, and these inform the decoder of the current state of the > processor. These packets, known collectively as PSB+, should be interpreted > as “status only”, since they do not imply any change of state at the time > of the PSB, nor are they associated directly with any instruction or event. > Thus, the normal binding and ordering rules that apply to these packets > outside of PSB+ can be ignored when these packets are between a PSB and > PSBEND. They inform the decoder of the state of the processor at the time > of the PSB. Our mistake was to allow target ip payloads in these "other packets" to update our decoder state, namely `self.cur_loc`. In some cases this would make it appear as though certain blocks were being re-executed in part. In turn this confused later parts of the JIT pipeline.
|
This looks fine to me, but let's put |
|
5a3d807 looks good? |
|
Works for me! Merge as-is? |
|
yup |
|
bors r+ |
|
Build failed: |
|
OK, so this is caused by new Rust. The -#![feature(once_cell)]
+#![feature(lazy_cell)]But this is scary: Investigating... |
|
The loader has placed the embedded bitcode at an unaligned address! |
|
Yikes! |
|
It has probably always been this way, but Rust is now checking for misaligned pointer derefs it seems... It's not clear to me why there is an alignment requirement here. I thought that misaligned reads were valid on x86_64, but I may be wrong? I know sparc64 would outright crash... You will probably tell me it is UB I imagine :) |
|
I think "Rust the language" defines misaligned reads/writes as UB, no matter the underlying platform. [IMHO this is the correct definition: you don't want people to write code that will explode in weird ways on different platforms!] |
|
Here's a discussion I had on the Rust Zulip about this: And here's an issue about the error message not being quite right: |
|
The failures in this PR have been addressed separately in ykjit/ykllvm#60 and #639 |
|
bors r+ |
|
Build succeeded: |
This fixes our non-deterministic failures.
The first commit is the meat, the next commit raises a question I'd like to investigate next.
I have an idea why this is happening, but don't want to address it in this same PR.
(this was hard work :( )