rustbuild: Build libstd with ThinLTO #45529
Closed
Conversation
This file contains bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
This commit moves the standard library to get compiled with multiple codegen units and ThinLTO like the compiler itself. This I would hope is the last major step towards closing out rust-lang#45320
|
cc #45400, which enabled this for rustc |
11 tasks
kennytm
added
the
S-waiting-on-review
|
@bors r+ |
|
📌 Commit 578feb5 has been approved by |
Mark-Simulacrum
added
S-waiting-on-bors
|
⌛ Testing commit 578feb5 with merge e462ee9c3d8c56966579d0c9c861617e9f162f86... |
|
💔 Test failed - status-travis |
|
Cannot link Errors from mipsel-unknown-linux-gnuErrors from mips64el-unknown-linux-gnuabi64 |
kennytm
added
S-waiting-on-author
|
I've been trying to debug this locally and figure out what's going on. I wasn't able to reproduce until I was able to reproduce. Then I found out it's nondeterministic. That's... oh boy. |
|
Ok so... Investigation so far: First I wanted to rule out a "trivial bug" in ThinLTO or LLVM. Let's start with this source code: #![feature(thread_local)]
#[thread_local]
static mut FOO: u32 = 3;
fn foo() -> &'static u32 {
unsafe { &*(&A as *const _) }
}
pub mod a {
pub fn bar() -> u32 {
*super::foo()
}
}Compile that with: Next up we search through the output files for our thread local Ok so far so good. One IR file imports one IR file exports. ThinLTO inlined the constant/function call across module boundaries. Despite all that we retain Next up let's take a look at the object files: Ok still looking good. One object defines a So that rules out to me at least any casual oversight. Let's take a look at the test case in question. As I mentioned above it's nondeterministic, but I managed to catch a failure. A failure right now leaves a bunch of I rigged up a script to decode Intersting! LLVM IR checks out as one module is importing a Aha! Sure enough one object file has Next up is to figure out why if the LLVM module looks correct the object file looks incorrect. That... smells like an LLVM bug? A data race? Who knows! |
|
Dohoho the plot thickens! Turns out in a successful build there's not two object files. The one that has the "bad reference" didn't have the object file in the successful build, presumably optimized out? Desires for deterministic builds aside, that raised an interesting question. I compiled the program above for And so aha! Looks like the mipsel backend is deterministically not handling |
|
And it keeps going... Note that if we don't pass It looks like this is also related to Smells a lot like an LLVM bug... |
|
I've opened an upstream LLVM bug, and I don't think there's much we can do about this in the meantime, so here's to praying that gets a response in the next few days. |
|
@alexcrichton Would changing MIPS to use 1 CGU without ThinLTO workaround the bug? The TLS problem not only affects this PR but also #45187 and #45472. |
Interesting. As far as I know our builds are pretty deterministic. At least the number of object files should always be the same. |
|
That's some awesome debugging there, @alexcrichton, like reading a detective story I also think we should just default to one CGU on the affected platforms. |
|
@kennytm oh I had no idea that this was affecting other PRs! I'll implement a fix pronto. @michaelwoerister yeah I'm not entirely certain where the nondeterminism is coming from. It's probably coming from something in ThinLTO, I may do a pass to make sure we're at least feeding things to LLVM in a deterministic order. |
alexcrichton
added a commit
to alexcrichton/rust
that referenced
this pull request
Oct 31, 2017
Discovered in rust-lang#45529 it looks like cross-module TLS imports aren't quite working today, especially with `hidden` visibility which mostly comes up with multiple codegen units. As a result this completely disables compiling with ThinLTO and multiple codegen units on MIPS when bootstrapping. cc rust-lang#45654, the tracking issue for this
bors
added a commit
that referenced
this pull request
Oct 31, 2017
rustbuild: Don't build with ThinLTO on MIPS Discovered in #45529 it looks like cross-module TLS imports aren't quite working today, especially with `hidden` visibility which mostly comes up with multiple codegen units. As a result this completely disables compiling with ThinLTO and multiple codegen units on MIPS when bootstrapping. cc #45654, the tracking issue for this
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Add this suggestion to a batch that can be applied as a single commit.
This suggestion is invalid because no changes were made to the code.
Suggestions cannot be applied while the pull request is closed.
Suggestions cannot be applied while viewing a subset of changes.
Only one suggestion per line can be applied in a batch.
Add this suggestion to a batch that can be applied as a single commit.
Applying suggestions on deleted lines is not supported.
You must change the existing code in this line in order to create a valid suggestion.
Outdated suggestions cannot be applied.
This suggestion has been applied or marked resolved.
Suggestions cannot be applied from pending reviews.
Suggestions cannot be applied on multi-line comments.
Suggestions cannot be applied while the pull request is queued to merge.
Suggestion cannot be applied right now. Please check back later.
This commit moves the standard library to get compiled with multiple codegen
units and ThinLTO like the compiler itself. This I would hope is the last major
step towards closing out #45320