XIO

xWire

Posts from every feed QUEST has found across the web — blogs, news, podcasts, channels and more.

LLVM Discussion Forums - Latest topics

Latest topics

[RFC][AMDGPU] Deprecating and removing LLVM’s R600/TeraScale backend

I would like to propose removing LLVM’s R600/TeraScale support from the AMDGPU target. This proposal only affects the R600 and TeraScale code paths. Support for GCN, RDNA, CDNA and current AMDGPU targets will not be affected. Historical background R600 originated at ATI Technologies and reached the market after AMD acquired ATI, with the Radeon HD 2900 XT launched in May 2007. It implemented ATI’s first unified shader architecture for PCs and was part of the first generation of TeraScale, which…

[RFC] Extend update_cc_test_checks.py to support multi-IR tests

Authors: Vicky Nguyen (aka iamvickynguyen) and Andrzej Warzyński (aka banach-space) We would like to propose extending update_cc_test_checks.py to support tests that verify multiple intermediate representations (IRs) within a single source file. Our primary motivation is Clang’s builtin code-generation tests, which increasingly exercise both the traditional Clang CodeGen pipeline and ClangIR. These tests share the same C/C++ source code but generate checks for different IRs, including LLVM IR…

A New `ub.assume` Operation

We were discussing integer overflow / range bugs and missed optimizations at the MLIR Meetup and folks expressed interest in an upstream operation to attach integer range information to SSA values. Is there broader interest in adding such an operation to MLIR? I’d like to focus on a design where metadata is discovered with a data flow analysis as opposed to encoding it in the type system (e.g. !my_int<32, range=[0, 128]>). The latter is incompatible with existing builtin types. There are two…

[RFC][AMDGPU] Performance-driven expansion of f32 atomic RMW on gfx12

Summary I’d like to propose a discussion about performance-driven expansion of floating-point atomicrmw fadd operations on gfx12 (gfx1200/gfx1201). Measurements show that native global_atomic_add_f32 has substantially lower throughput than integer atomics and can severely degrade the performance of unrelated memory operations running concurrently on the same GPU. A compare-and-swap (CAS) loop avoids much of this interference, although native atomics remain faster for some low-contention…

[RFC][MLIR] Align CODEOWNERS with active reviewing

Hi everyone, Authors: Andrzej Warzyński and Federico Bruzzone We would like to clarify the distinction between GitHub CODEOWNERS and PR subscriptions, and propose that we use these mechanisms more consistently in MLIR. For context: LLVM’s CODEOWNERS automatically requests reviews when a PR modifies matching files or directories. Its header also explains that adding yourself means committing to review a substantial proportion of PRs in that area. LLVM’s pr-subscribers-* teams let people receive…

[RFC] Representing Fortran aliasing rules in LLVM IR: options and costs

The full design document, with background, all measurements and the kernels, is in #242549 (rendered). This post summarizes it and lists the questions we would like answered. The problem Flang encodes Fortran’s aliasing guarantees in TBAA metadata, with one TBAA tree per function. The trees mix two kinds of facts: facts that hold everywhere, such as “a descriptor field never overlaps array data” or “two different module variables never overlap”; facts that hold only during one invocation of a…

[RFC] Host LLVM formal verification tools as part of LLVM project

Summary This RFC proposes adding LLVM formal verification tools directly into the LLVM GitHub project within a dedicated, separate repository. This infrastructure will serve as a common ground-truth testbed for evaluating LLVM IR semantics, checking tool expectations, and feeding an interactive Compiler Explorer instance for formal verification. Motivation & Background This is a follow up to the LLVM Dev Meeting round table discussion. The community currently utilizes several formal…

Guidance regarding finding proper reviewers

Hi everyone, I recently submitted three PRs to MLIR as I’m working toward gaining Contributor status. One of them was reviewed quickly, which was great, but I haven’t been able to get any feedback on the other two (PR #226759 and PR #227145) despite a few polite pings. I know reviewers volunteer their time, so I want to make sure I’m routing these to the right people rather than tagging folks outside their area of expertise. So far, I’ve tried assigning people who have worked on similar code,…

[RFC] Communicating Array Extents to LLVM Using llvm.assume.chained

RFC: Communicating Array Extents to LLVM Using llvm.assume.chained Madhur Amilkanthwar, Sjoerd Meijer, Sebastian Pop, Elvina Yakubova, Yafet Beyene, NVIDIA. We have discussed parts of this proposal informally and are sharing it here to get wider input from the community. Summary. Fortran guarantees that every array subscript stays within its dimension, but the guarantee is lost when Flang lowers an access to a linear address. We propose llvm.assume.chained, a pointer-forwarding intrinsic whose…

[RFC] Removing index sharding from ODS in favour of Logical Sharding

I recently added so-called “Logical Sharding” to MLIR-TableGen which allows you to split a dialect into multiple *Ops.cpp.inc and *Ops.h.inc pairs [mlir][ODS] Add so-called "Logical Sharding" to ops by zero9178 · Pull Request #239254 · llvm/llvm-project · GitHub The idea and motivation is that previous automatic index sharding only ever split the cpp file but never the header file which meant every TU resulting from that split as well as users of the header file, were including all op…

Bytecode compatibility guarantees for the Func dialect?

The Builtin bytecode versioning RFC introduced stability guarantees for Builtin attribute and type encodings. Subsequent changes, such as adding DLTI support, also preserve historical builtin.module encodings, although stability of operations were not explicitly covered by the RFC. We (Arm) are interested in providing bytecode compatibility guarantees for the TOSA dialect, based on the backwards compatibility guarantees of the specification. However, our artifacts also contain Func operations,…

LLVM Bay Area Monthly Meetup (Mon Nov 30 ‘26 6pm)

We’re planning the next LLVM Bay Area monthly meetup. Date: Monday Nov 30, 2025 6pm Event Address: Don Giovanni’s 235 Castro St, Mountain View, CA 94041 There is no need to RSVP for this event. There will still be a cash bar. We’ll be indoors; consider bringing hearing protection, cloves of garlic, silver or wooden staves, and some holy water. Attendees are required to adhere to the LLVM Community Code of Conduct . For any Code of Conduct reports, please contact the organizers, and also email…

Mapping Numerical Linear Algebra Code onto BLAS/LAPACK

Hi all, I come from numerical linear algebra, with a long-standing interest in compiler construction. Recently I started to explore MLIR, and I find it fascinating. To learn it, I prototyped a compiler for a problem I know very well from about 20 years ago: mapping numerical algorithms onto BLAS and LAPACK. Back then I reimplemented large parts of LAPACK by hand (FLENS-LAPACK, checked bit for bit against the reference LAPACK), and I always felt I was doing a compiler’s job. Building the…

[RFC] Checker-Name Arguments for [[clang::suppress]] in the Static Analyzer

We propose changing the Clang Static Analyzer (CSA) to honor the string arguments of [[clang::suppress("...")]], interpreting each argument as a checker name or checker package. [[clang::suppress("core.NullDereference")]] return p; // suppress null-dereference report [[clang::suppress("core")]] { // suppress every checker in package "core" ... } [[clang::suppress("core.DivideZero", "unix.Malloc")]] // suppress any listed name Context Today [[clang::suppress]] is all-or-nothing in CSA. A bare…

[RFC] Value Bounds need deferred conditional bounds

Currently ValueBoundsOpInterface relies on populateAndCompare calls to decide if some bounds can be used. For example, arith.mul needs to know if the operands are positive to decide which bound is the lower bound and which the upper bound. The problem is how it’s implemented: ValueBoundsConstraintSet continues processing the queue without finishing the current item, so some of the operations don’t “see” all the bounds and have to “compare” using limited information. And so the exact value…

Seeking advice on reporting MLIR fuzzer findings and contributing fixes

Hi everyone, I have recently been using a fuzzer I developed to test MLIR, and I have been following the community discussions about fuzzer-discovered issues and AI use. I would like to describe my workflow and ask for feedback on reporting issues and becoming more involved in the community. Fuzzing and issue reporting My fuzzer uses traditional fuzzing techniques and does not use an LLM during the fuzzing process. After finding a problem, I use an LLM to help reduce the reproducer, analyze…

Guidance on "Not Observable" Side Effect Modeling

I am looking for some guidance on side effect modeling with the MemoryEffectOpInterface. Can an op that internally allocates some memory, accesses the memory, then deallocates the memory be marked as “no effects”? Can an op that accesses constant memory be marked as “no effects”? E.g. an operation that reads from lookup table instead of computing a complex mathematical function. My answer to both questions would be “yes” because these effects are not externally observable. Apart from some edge…

Israel LLVM Meetup Group - November 24th 2026

Join the Israel LLVM community for an evening dedicated to the latest advancements in compilers and system architecture, hosted by Majestic-Labs.AI. We will kick off with key technical takeaways from the U.S. LLVM Developers’ Meeting, followed by an in-depth session on compiling high-level Deep Neural Network (DNN) representations to specific hardware architectures. Whether your focus is on MLIR, LLVM infrastructure, or backend code generation, this meetup offers an excellent opportunity to…

[RFC] An LLVM backend for Renesas RX, with Rust support

Hi all, I’d like to share an LLVM backend for the Renesas RX architecture that I’ve been developing, and ask whether there would be interest in upstreaming it as an experimental target. Background RX is a 32-bit CISC architecture from Renesas, widely used in industrial and consumer embedded devices (especially in Japan). Today the toolchain options for RX are Renesas’ proprietary CC-RX, IAR, and the GCC port. As far as I know, there is no publicly available LLVM backend for RX, which also means…

[RFC][Draft] P2300R10 std::execution Initial design drafts

This thread is intended to serve as a place to discuss ideas regarding the architecture, design, and potential implementation challenges of std::execution (Senders/Receivers), introduced in C++26. The title and description will be updated as we make progress toward an agreed-upon design. Currently related: Open issue llvm/llvm-project P2300R10: `std::execution` opened 10:01PM - 20 Aug 24 UTC ldionne libc++ clang:frontend c++26 wg21-paper **Link:** https://wg21.link/P2300R10 Pull request…

Starting a community Floating-Point Working Group

Our recent Clang C and C++ Language Working Group agendas have had several topics related to floating-point. It was also recognized that some of the concerns raised are cross-cutting in that they touch Clang, LLVM, llvm-libc, libc++, MLIR, and/or compiler-rt and so having the Clang language working group hosting the meetings means we sometimes miss important perspectives. Given the scale of the discussions and their cross-cutting nature, we felt it would make sense to start a dedicated…

[RFC]: Deprecate and remove `__builtin_sycl_unique_stable_name`

The __builtin_sycl_unique_stable_name token and corresponding SYCLUniqueStableNameExpr AST node were added to Clang during a previous effort to contribute an implementation of the SYCL 2020 specification. They are not, and will not be, used by the current effort. __builtin_sycl_unique_stable_name is currently used by Intel’s downstream DPC++ compiler implementations (the open source DPClang compiler, the Intel oneAPI DPC++/C++ compiler, and any compilers that derive from those). Intel is…

[RFC] Change in OpaqueValueExpr implementation

Hey Everyone, I ran into an assertion in OpaqueValueExpr while working on an issue and the reproducer had pack expansion inside a fold expression ((void)(co_yield Values),…); The issue I found was due to the unexpanded flag in OVE so my first instinct was the since OVE does not have the source as it’s own children it should not hold any syntactic data so I modified the code to bypass the assertion by only copying the semantic data of the source expr but that created an issue that in some case…

[RFC] Add -Xlinker= as an alias for -Xlinker

Hi everyone, I’d like to propose adding -Xlinker=<arg> as an alternative to -Xlinker <arg>, available in Clang, clang-cl, and Flang. This is a simple change following the rationale for the existing -Xclang= alias but llvm/llvm-project#238301 advised it would be better to start a quick discussion. My main motivation is that I need to let bazel, which constructs and understands response-files with a single string per line, pass -Xlinker args to clang-cl. Any objection? cc @MaskRay for the driver…