LLVM 15 compatibility fixes #99512

nikic · 2022-07-20T14:30:56Z

These are LLVM 15 compatibility fixes split out from #99464. There are three changes here:

Emit elementtype attribtue for ldrex/strex intrinsics. This is requires as part of the opaque pointers migration.
Make more tests compatible with opaque pointers. These are either new or aren't run on x86.
Remove a test for #[rustc_allocator]. Since codegen: use new {re,de,}allocator annotations in llvm #99574 there are more requirement on the function signature. I dropped the test entirely, since we already test the effect of the attribute elsewhere.
The main change: When a worker thread emits an error, wait for other threads to finish before unwinding the main thread and exiting. Otherwise workers may end up using globals for which destructors have already been run. This was probably never quite correct, but became an active problem with LLVM 15, because it started using global dtors in critical places, as part of ManagedStatic removal.

Fixes #99432 (and probably also #95679).

compiler/rustc_codegen_ssa/src/back/write.rs

nikic · 2022-07-27T11:13:19Z

I ended up completely rewriting the fix for the global dtor issue. The new implementation drops the AbortCodegenOnDrop wrapper, and instead integrates the functionality into OngoingCodegen (or rather a separate Coordinator struct, to limit drop scope). This ensures that all panics during codegen will result in a CodegenAborted message + thread join. Previously, panics generated as part of join_codegen() were not handled, which is really the root issue here. Similarly, WorkerFatalError handling is switched to do the same thing as CodegenAborted.

I think this makes for a nicer and more consistent solution. The main casualty are various asserts that no longer hold, as codegen can now be aborted at more different points.

These intrinsics (and a few more, but there are the only ones exposed by stdarch) require an elementtype attribute in LLVM 15.

Replace the separate AbortCodegenOnDrop guard by integrating this functionality into OngoingCodegen (or rather, the Coordinator part of it). This ensures that we send a CodegenAborted message and wait for workers to finish even if the panic occurs outside codegen_crate() (e.g. inside join_codegen()). This requires some minor changes to the handling of CodegenAborted, as it can now occur when the main thread is LLVMing rather than Codegenning.

This means that codegen_aborted may be set when new codegen requests arrive, so drop some related assertions. The new work will simply be ignored.

This attribute now does more than just place noalias on the return, and has specific requirements for the signature. Drop the test entirely, as we already check __rust_alloc attributes in other codegen tests.

cuviper · 2022-07-28T22:48:10Z

LGTM! I'm blocking rollup to be careful about the thread change, although I don't know if CI hits such errors anyway.

@bors r+ rollup=never

bors · 2022-07-28T22:48:12Z

📌 Commit 1433b29 has been approved by cuviper

It is now in the queue for this repository.

bors · 2022-07-29T01:35:23Z

⌛ Testing commit 1433b29 with merge 9de7474...

bors · 2022-07-29T04:23:01Z

☀️ Test successful - checks-actions
Approved by: cuviper
Pushing 9de7474 to master...

rust-timer · 2022-07-29T05:39:00Z

Finished benchmarking commit (9de7474): comparison url.

Instruction count

This benchmark run did not return any relevant results for this metric.

Max RSS (memory usage)

Results

Primary benchmarks: no relevant changes found
Secondary benchmarks: mixed results

	mean¹	max	count²
Regressions 😿 (primary)	N/A	N/A	0
Regressions 😿 (secondary)	2.1%	2.1%	1
Improvements 🎉 (primary)	N/A	N/A	0
Improvements 🎉 (secondary)	-2.6%	-2.6%	1
All 😿🎉 (primary)	N/A	N/A	0

Cycles

Results

Primary benchmarks: 😿 relevant regression found
Secondary benchmarks: no relevant changes found

	mean¹	max	count²
Regressions 😿 (primary)	9.2%	9.2%	1
Regressions 😿 (secondary)	N/A	N/A	0
Improvements 🎉 (primary)	N/A	N/A	0
Improvements 🎉 (secondary)	N/A	N/A	0
All 😿🎉 (primary)	9.2%	9.2%	1

If you disagree with this performance assessment, please file an issue in rust-lang/rustc-perf.

@rustbot label: -perf-regression

the arithmetic mean of the percent change ↩ ↩²
number of relevant changes ↩ ↩²

Update to LLVM 15 For preliminary testing. Some LLVM 15 compatibility fixes were applied separately in rust-lang#99512. Release timeline: * LLVM 15 branched on Jul 26. * The final LLVM 15.0.0 release is scheduled for Sep 6. * Current nightly (1.65.0) is scheduled for Nov 3. Changes in this PR (apart from the LLVM update): * Pass `--set llvm.allow-old-toolchain` for many Docker images. LLVM 16 will require GCC >= 7.1, while LLVM 15 still allows older compilers with an option. Specify the option for builders still using GCC 5.4. rust-lang#95026 updated some of the used toolchains, but not all. * Use the `+atomics-32` target feature for thumbv6m. * Explicitly link libatomic when cross-compiling LLVM to 32-bit target. * Explicitly disable zstd support, to avoid libzstd.so dependency. New LLVM patches ([commits](https://github.com/rust-lang/llvm-project/commits/rustc/15.0-2022-08-09)): * [rust-only] Fix ICE with GCC 5.4 (nikic/llvm-project@15be58d) * [rust-only] Fix build with GCC 5.4 (nikic/llvm-project@774edc1) * ~~[rust-only] Fix build with GCC 5.2 (nikic/llvm-project@1a6069a7bb35ace1e40d566035cbf7ed2fa3b1f7)~~ * ~~[rust-only] Fix ICE with GCC 5.2 (nikic/llvm-project@493081f2909206e0ed55af68a4058a76c0ad7a64)~~ * ~~[rust-only] Fix build with GCC 5.2 (nikic/llvm-project@0fc5979d738c3a1f9510fe2d62417f7d2af37817)~~ * [backported] Addition of `+atomics` target feature (llvm/llvm-project@57bdd98). * [backported] Revert compiler-rt change that broke powerpc (llvm/llvm-project@9c68b43) * [awaiting backport] Fix RelLookupTableConverter on gnux32 (nikic/llvm-project@639388a / llvm/llvm-project#57021) Tested images: dist-x86_64-linux, armhf-gnu, arm-android, dist-s390x-linux, dist-x86_64-illumos, dist-x86_64-freebsd, wasm32, dist-x86_64-musl, dist-various-1, dist-riscv64-linux, dist-mips-linux, dist-mipsel-linux, dist-powerpc-linux, dist-aarch64-linux, dist-x86_64-apple, x86_64-msvc-1, x86_64-msvc-2, dist-various-2, dist-arm-linux Tested up to the usual ipv6 error: test-various, i686-gnu, x86_64-gnu-nopt r? `@ghost`

rustbot added the T-compiler Relevant to the compiler team, which will review and decide on the PR/issue. label Jul 20, 2022

rust-highfive assigned cuviper Jul 20, 2022

rust-highfive added the S-waiting-on-review Status: Awaiting review from the assignee but also interested parties. label Jul 20, 2022

nikic mentioned this pull request Jul 20, 2022

rustc master has a flaky failure when compiled against LLVM main #99432

Closed

nikic commented Jul 20, 2022

View reviewed changes

compiler/rustc_codegen_ssa/src/back/write.rs Outdated Show resolved Hide resolved

nikic force-pushed the llvm-15-fixes branch from aec637a to 489b1be Compare July 27, 2022 11:09

nikic force-pushed the llvm-15-fixes branch from 489b1be to d4aadac Compare July 27, 2022 11:14

nikic added 5 commits July 27, 2022 16:19

Add elementtype attributes for llvm.arm.ldrex/strex intrinsics

f653d3a

These intrinsics (and a few more, but there are the only ones exposed by stdarch) require an elementtype attribute in LLVM 15.

Update codegen test for opaque pointers

6a1f77d

Also wait on other threads if a WorkerFatalError occurs

c87e20b

This means that codegen_aborted may be set when new codegen requests arrive, so drop some related assertions. The new work will simply be ignored.

Remove outdated rustc_allocator test

e6f0e35

This attribute now does more than just place noalias on the return, and has specific requirements for the signature. Drop the test entirely, as we already check __rust_alloc attributes in other codegen tests.

nikic force-pushed the llvm-15-fixes branch from d4aadac to 92bb4a5 Compare July 27, 2022 14:19

This comment has been minimized.

Sign in to view

Make CFI tests opaque pointers compatible

1433b29

nikic force-pushed the llvm-15-fixes branch from 92bb4a5 to 1433b29 Compare July 27, 2022 14:37

bors added S-waiting-on-bors Status: Waiting on bors to run and complete tests. Bors will change the label on completion. and removed S-waiting-on-review Status: Awaiting review from the assignee but also interested parties. labels Jul 28, 2022

bors added the merged-by-bors This PR was explicitly merged by bors. label Jul 29, 2022

bors merged commit 9de7474 into rust-lang:master Jul 29, 2022

rustbot added this to the 1.64.0 milestone Jul 29, 2022

bors mentioned this pull request Jul 29, 2022

Rewrite LLVM's archive writer in Rust #97485

Merged

nikic mentioned this pull request Jul 29, 2022

Update to LLVM 15 #99464

Merged

workingjubilee mentioned this pull request Oct 13, 2024

failed to run LLVM passes: unknown pass name 'unknown-pass' #95679

Closed

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

LLVM 15 compatibility fixes #99512

LLVM 15 compatibility fixes #99512

nikic commented Jul 20, 2022 •

edited

Loading

nikic commented Jul 27, 2022

This comment has been minimized.

cuviper commented Jul 28, 2022

bors commented Jul 28, 2022

bors commented Jul 29, 2022

bors commented Jul 29, 2022

rust-timer commented Jul 29, 2022

LLVM 15 compatibility fixes #99512

LLVM 15 compatibility fixes #99512

Conversation

nikic commented Jul 20, 2022 • edited Loading

nikic commented Jul 27, 2022

This comment has been minimized.

cuviper commented Jul 28, 2022

bors commented Jul 28, 2022

bors commented Jul 29, 2022

bors commented Jul 29, 2022

rust-timer commented Jul 29, 2022

Instruction count

Max RSS (memory usage)

Cycles

Footnotes

nikic commented Jul 20, 2022 •

edited

Loading