L
Hi, I’m working on an undergraduate dissertation using LLVM to optimise event-processing loops. My current method learns which handler usually follows another and creates a checked direct path between them. If the prediction is wrong, execution falls back to normal dispatch. The transformed program also uses ordinary PGO. So far, I’ve tested this on small synthetic benchmarks. It helps in some cases, but changes in event order can cause slowdowns. I’d appreciate advice on related LLVM work or…
L
This is a follow-up to the previous RFC on minimal nl_langinfo and static UTF-8 locale support. While landing the internal locale plumbing (#211470) and nl_langinfo (#243540), Michael Jones and I got talking about whether LLVM-libc should support changing the locale at runtime at all. This RFC captures the gist of that discussion, and we want to solicit opinions to figure out what matters to people. The idea is to keep the default state of the libc fixed at C.UTF-8 (or whatever single default…
L
Hi all, I’m fairly new to the LLVM community, so please let me know if I’m going about this the wrong way. I am developing a tool that tests MLIR automatically. It has found about 100 bugs that are not reported yet. Following HowToSubmitABug.md, and after reading the discussions in Fuzzing MLIR Infrastructure and Seeking advice on reporting MLIR fuzzer findings, I’d like to agree on how to report them. I filed 15 earlier findings one at a time over the past month, before I knew about the…
L
I propose putting clang-tidy lit tests under clang-tools-extra/test/clang-tidy/checkers/<module>/<check>/. New checks would use this layout, and existing checks could move incrementally. Motivation When I’m working on a clang-tidy check, I want one command that rebuilds what changed and runs that check’s tests. For modernize-use-auto, the existing Ninja target runs all 165 tests in modernize. I can run the seven use-auto tests directly with lit, which uses the clang-tidy binary already in the…
L
I found bugs related to the assembler and inline assembly using a fuzzer and automated testing. Given the recent discussion about fuzzing and the increasing number of issues, I would like to share my findings and seek advice on how to report issues, rather than just submit many issues. I have already submitted some of them. I am sorry that I had not noticed the recent discussion before submitting them. To begin with, I would especially appreciate advice on which kinds of cases are worth…
L
Hello community In the beginning of 2026, I joined the LLVM community with the dream of becoming a compiler engineer. Since then, I’ve been mentored by many people in the community, and there are many things I’ve told people I would do. Even just last week, I exchanged contacts with people at a LLVM meetup. I think it is not a manner to suddenly disappear, so I am writing this post. Last weekend, my father was diagnosed with stage 4 stomach cancer. So, I’m heading to South Korea to take care of…
L
This RFC proposes replacing LLDB’s optional libxml2 dependency with a minimal XML 1.0 parser which lives in the mono repo. Motivation libxml2 is not really optional in LLDB. It is required for gdb-remote target.xml, memory maps, library lists and dSYM plists files. If LLDB is built without libxml2, register info, loaded module lists and source remapping don’t work. On Windows it prevents testing lldb-server. libxml2 is not installed by default on Windows. Installing it at desk is awkward…
L
[libc++] Updates Unicode tables to Unicode 18 · llvm/llvm-project@f7ffa95 and [libc++] Updates Unicode tables to Unicode 18 · llvm/llvm-project@753c9ed conflict with each other… Sourced from [libc++] Updates Unicode tables to Unicode 18 by clin1234 · Pull Request #199111 · llvm/llvm-project · GitHub 1 post - 1 participant Read full topic
L
Summary This RFC proposes making LLVM’s BTF emitter usable on any ELF target behind an opt-in -gbtf flag. GCC already uses the same flag spelling. With the flag, the compiler emits .BTF and .BTF.ext describing the types, functions, and global variables of the translation unit. Without the flag, behavior is unchanged. The current patch is #183929. This thread covers the compiler side only. It does not need a linker change: the kernel already deduplicates BTF after the link (in resolve_btfids),…
L
Hi, To fix profiler hangs in Firefox’s TSan builds [1], we’re adding a TSan-only interposer of libc’s syscall() that brackets SYS_futex with __sanitizer_syscall_pre_futex / __sanitizer_syscall_post_futex [2]. There is also a proposal for Rust’s std to call the same hooks around its own futex wait [3]. If both happen, a Rust futex wait would run: std: __sanitizer_syscall_pre_futex(...) interposer: __sanitizer_syscall_pre_futex(...) syscall(SYS_futex, ...) interposer:…
L
LLVM’s long relicensing effort is nearly done: more than 99.5% of all code ever contributed to LLVM is now covered by the Apache-2.0 WITH LLVM-exception license, has left the code base, or is too small to be covered by copyright. We think that’s a good moment to call the effort complete. This post explains why the change matters, how we got here, and what happens next. Why a single license helps One license means less guesswork for everyone who builds on LLVM. Today, anyone shipping LLVM in a…
L
I think will be good idea to switch from YAML-based regular expression matcher to script one. This will allow to properly handle such cases like [lld][ELF], [clang][analyzer] or Clang Static Analyzer, [mlir][linalg] without adding redundant lld, clang, mlir labels respectively. 1 post - 1 participant Read full topic
L
I have been using the GPU dialect as a target in my lowering and it works great, thanks! I’ve run into an issue when using the asynchronous forms of operations like gpu.launch and gpu.alloc, especially when interfacing with foreign/external clients that want to manage their own execution resources. The problem is that there’s no consistent, first principles way to say: Where asynchronous GPU work is supposed to run, and What asynchronous work depends on Instead, the default device is global…
L
(This proposal would land after: RFC: extend LIBOMPTARGET_INFO functionality to allow verbosity control and proposes formatting for the new summary lines added to DataTransfer in it) The current output for LIBOMPTARGET_INFO can be hard to parse due to the output being unformatted, namely lacking whitespace making it hard to easily and quickly parse. As an initial effort to address this I propose to add whitespace indentation on the DumpTable, KernelArgs and DataTransfer output categories:…
L
For as long as I can remember, LLDB has given each debugger its own script interpreter object and session dictionary. Under the hood, all debuggers run in the same Python interpreter and share its modules, sys and the globals of the lldb module. This is just a workaround, and there are plenty of times this illusion breaks. A concrete example of that is when two debuggers import a module with the same name (from different directories) and the last import wins. Python added subinterpreters in…
L
We’ll meet as normal tomorrow, Thu 19th November at 4pm GMT / 8am PST. If you have more topics for the agenda please drop them in the doc. 1 post - 1 participant Read full topic
L
I noticed the following differences in Clang behavior against GCC’s Clang doesn’t allow taking the address of elements Clang doesn’t allow increment/decrement on vectors, though in the table of supported vector operations it’s marked as supported No __builtin_shuffle It’s possible to mix vectors of different signedness or number of elements, while gcc doc forbids it (but gcc accepts…) Semantics of logical operations significantly differ from gcc’s. Apart from the first item, I haven’t been able…
L
November 2026 Portland (Oregon, USA) LLVM social, Thanksgiving edition. Date&time: November 24th, 2026, 6 PM. Location: BG Food Cartel , 4250 SW Rose Biggi Ave, Beaverton, OR 97005. Since this instance would fall on Thanksgiving (Thursday), we are meeting on Tuesday instead. The food cart lot is right next to a MAX station (Beaverton Central), but also has dedicated parking. The venue has an indoor seating area and a bar. Looks like the location has been a hit, but we are open to other…
L
I propose to deprecate lto::Config::PassPluginFilenames. LTO hosts should load pass plugins themselves and pass them in lto::Config::LoadedPassPlugins. This post explains the motivation, the interface contract and what changes in-tree. The scope is the lto::LTO API with its backend entry points, and dynamically loaded plugins. Statically linked extensions and the legacy LTO API (which has no plugin support) are not affected. Background lto::Config offers two ways to provide plugins today. With…
L
The context of this RFC is that I would like to propose to stronger stance on machine-generated output based on my own long-standing experience with esoteric reasoning and motivation. I have a long history of authoring code and math that I cannot explain sufficiently well, even prior to LLVM: there is some complicated, esoteric, non-verbal reasoning process that goes on in my head, that I’m not able to put down sufficiently well in communication to others. It’s always straightforward to explain…
L
Hi, My team at Intel is hiring a full-time GPU Compiler Engineer in Santa Clara, CA (hybrid: part on site, part remote). We build the LLVM-based compilers behind Intel GPUs, used for AI, HPC and data-center workloads. What you’d work on Compilers for OpenCL, OpenMP, C for Metal and SYCL / Data Parallel C++ New compiler features, and code generation and optimization for Intel GPU architectures Finding and fixing performance bottlenecks, together with the hardware and architecture teams Tooling…
L
Recently I met a situation where an operand of an inline asm call can be proven to be in the shared address space, but it was not inferred. I tried adding the following code to `collectFlatAddressExpressions`, and it worked for my case: else if (auto *CB = dyn_cast<CallBase>(&I)) { if (isa<InlineAsm>(CB->getCalledOperand())) for (Value *Arg : CB->args()) { if (Arg->getType()->isPointerTy()) PushPtrOperand(Arg); } I’d like to know why inline asm (or call instructions) is not handled here. Is it…
L
Going through another iteration of LLM policy discussions late at night, I thought we should step up to the near-philosophical level and identify what are the values of the LLVM project/community. This would allow us to argue whether any proposed policy changes align with those values or not, and I suspect at least some disagreement may be resolved in a more principled way. I see two categories The community / people involved. We have an amazing group of people who are respectful, competent,…
L
TL;DR: Please revert this policy. It is premature before we holistically rethink how LLVM will operate as AI becomes more prevalent in software development. -– [RFC] Forbid AI-generated communication was closed with the statement that the project council had approved it. This policy was already being enforced before it was written: Comment llvm/llvm-project Comment by ftynse - [mlir][LLVM] Include LLVM dialect partition headers instead of LLVMDialect.h users/zero9178/llvm-dialect-partioning ←…
L
We are happy to announce that LLVM 23.1.6 is now released! This includes the main LLVM project, and its subprojects including clang, lld, libc++, and MLIR. Huge thanks to everyone that contributed, reviewed, provided support and in any other way contributed to this release! Please note the subproject tarballs have been removed after RFC: Do “something” with the subproject tarballs in the release page Download Find sources on GitHub. A note on binaries Official binaries will not be available…
L
Welcome to the 79th edition of the MLIR (Multi-Level Intermediate Representation) Newsletter! An active month with 574 commits in the MLIR tree. Major themes : performance improvements to the pass infrastructure and greedy rewriter (6x pipeline speedup in parallel execution, 84x on deeply nested IR), a new “Logical Sharding” mechanism for ODS that halves compile-time memory, vector dialect masking unification underway, TOSA elementwise ops now lower to linalg.elementwise, and a new DLTI scoped…
L
Hi, i find that when dealing with huge variants, compilers become very slow quickly. I run this with time clang++ -c -ftemplate-depth=10000 -fbracket-depth=10000 cc-slow-absl.cpp -o tmp.o and i takes about 45 secs to finish. Is there’s any secret sauce to make this work better? Or is std::variant designed to be slow for compilation in this case? Gist gist.github.com/crlf0710 https://gist.github.com/crlf0710/8e14b927eaeaeed97a721cee8c93962d cc-slow-absl.cpp #include <absl/types/variant.h>…
L
Suggestion to improve InstCombine PR review process InstCombine receives a lot of PRs, and not all of them bring any significant real-world benefit. Some are useful, while others are just small optimizations without much practical impact. Since GitHub automatically assigns code owners as reviewers, a large number of these PRs end up in the maintainers’ review queues. This makes it difficult for them to focus on PRs that actually need attention. Possible solution Maybe we can consider changing…
L
Hello, Following the fuzzer-found issue #237126 (crash layer fixed by #208881, verifier failure remaining), I posted a root-cause analysis on the issue. The reproducible part turned out to be an instance of what I believe is a general design question about remove-dead-values, so I would like to discuss it here before writing the fix. What happens In a private function with no callers, nothing seeds the backward liveness analysis (return/terminators carry no values), so every value in the body…
L
I’d like to propose declaring LLVM’s optimization remark kinds in TableGen and generating one typed builder per kind, so that a remark’s name, outcome, message and arguments are stated once and checked when the compiler is built. This is the first half of the RFC I posted in October, split out as Florian Hahn suggested there. A prototype exists, nothing is proposed for landing, and I’m after a judgment on whether the shape is viable. A follow-up RFC proposes to carry the declared remarks into…