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][Offload] Introducing AMDGCN-flavored SPIR-V support for OpenMP

TL; DR Introduce AMDGCN-flavored SPIR-V support for OpenMP offloading for AMD GPUs, using Comgr to compile SPIR-V at runtime into native AMDGPU code objects. Motivation AMDGCN-flavored SPIR-V is available in Clang today as an offload target for the HIP language. Its introduction was discussed in [llvm-dev] [RFC][HIPSPV] Emitting HIP device code as SPIR-V . This RFC seeks to extend the support of this target for the OpenMP offload project on AMD GPUs. Hence, many of the changes are specific to…

[Showcase] Mapping MLIR linalg.matmul to an RTL Systolic Array

Hi everyone! I’d like to share my project, hw-mlir-lab, where I use MLIR to lower matrix multiplication (linalg.matmul) to a systolic array simulated in Verilator. The project demonstrates an end-to-end flow involving tiling, padding, bufferization, custom MLIR operations, and lowering to a C interface for RTL simulation. I’ve also packaged the environment in Docker, so you can reproduce the entire flow using make demo without manually setting up the dependencies. I hope it will be useful to…

Question on permutation maps vs masking in rank-reducing vectorization (Ref: PR #226490)

Hi everyone, I’ve been reading through recent vectorization patches to better understand the MLIR vector dialect, and I was looking at PR #226490 ([mlir][linalg] Fix rank-reducing insert_slice vectorization writing to wrong dims). The author fixed an out-of-bounds write by explicitly passing a permutation_map built from the non-dropped result dimensions, which corrects the previous behavior that incorrectly relied on an implied minor-identity map. I see this is related to the ongoing RFC about…

RFC: Rework floating-point analysis for ppc_fp128 (PPCDoubleDouble)

For the purposes of this RFC, we will assume that the layout of PPCDoubleDouble is {hi, lo}, since that reads in the same order as hi + lo. PPCDoubleDouble has been a constant pain point for analysis and optimization. It goes against many of the assumptions relied upon when making floating-point optimizations. Some of these issues include: Subnormal and normal values share overlapping ranges. PPCDoubleDouble has a subnormal epsilon, meaning values such as 1.0 + DBL_TRUE_MIN and DBL_MAX +…

A web-based matcher for checking how often an IR pattern occurs in real code

Hi, I’d like to share a research prototype and ask for feedback: MetaphIR, a small pattern language for LLVM IR, with an in-browser matcher that runs patterns against a corpus of real-world IR (an indexed snapshot of llvm-opt-benchmark-nightly). Try it: MetaphIR Matcher The “Help” page on the site introduces the syntax and semantics. This is an early prototype, so the syntax is still evolving. The matcher downloads index data on demand (around 100–600 MB for a broad query). The cache is…

Address Sanitizer and static linking

Hi all. I’m looking for a bit of clarification from something in the ASAN docs. The address sanitizer docs state: Static linking of executables is not supported. But what does that actually mean? Something is always statically linked into the excecutable, the ASAN runtime itself is statically linked - so what is it being statically linked that ASAN does not support? libc++? Thanks! 3 posts - 2 participants Read full topic

[LoadStoreVectorizer] CPU store-packing opportunities and default pipeline scheduling

Hi, While investigating adjacent store-combining opportunities on X86, we found several cases that remain after the default O2/O3 pipeline: Consecutive scalar stores: four float stores at offsets 8, 12, 16 and 20 can be combined into one <4 x float> store. Mixed vector/scalar stores: a <2 x float> store followed by two float stores at offsets 0, 8 and 12 can be combined into one <4 x float> store. A PBRT-derived chain: stores of <2 x float>, <2 x float>, float, and float at offsets 0, 8, 16 and…

[RFC] Reducing release note merge conflicts with per-change fragments

LLVM contributors currently add entries directly to each project’s release notes. Multiple PRs often add entries in the same part of a section. Even when their code changes are unrelated, edits to adjacent lines in the release notes can cause merge conflicts. Authors then have to resolve these additional documentation conflicts when updating their branches or merging their changes. During a discussion of this problem on Discord, someone mentioned blurb, which CPython uses. It stores release…

[please ignore it][analyzer] bugfix: Reject non-symbolic iterator offsets

We found some crashes in the engine about non-symbolic iterator offsets. Syncing with upstream. github.com/llvm/llvm-project [analyzer] Reject non-symbolic iterator offsets (#232622) main ← relaxcn:fix/iterator-symbolic-offset opened 06:52AM - 16 Oct 26 UTC relaxcn +103 -28 `evalBinOp()` can return a concrete value for a constrained iterator offset. Sto…ring `getAsSymbol()` from that result leaves a null offset in the model, which is later dereferenced. Return failure for non-symbolic results…

[CfP] 2026 LLVM Memory Safety Workshop

We’re happy to announce that the detailed agenda for the LLVM Memory Safety Workshop is now available: https://llvm.swoogo.com/2026devmtg/agenda. For additional information, we’ve included here the abstracts for the talks and roundtables. Long Talks Validating Pointer Provenance with BorrowSanitizer Aliasing is central to the performance and security of systems programming, but its semantics are often underspecified. This ambiguity has made it easy for developers to inadvertently introduce…

[LLD, Mach-O] RFC: A Flexible Multipass Algorithm for Branch-Range Extension

[LLD, Mach-O] RFC: A Flexible Multi-pass Algorithm for Branch-Range Extension Summary This RFC proposes a flexible, multipass algorithm for resolving out-of-range ARM64 branches in LLD’s Mach-O port. The algorithm eliminates fixed slop reservations. It uses compact 4-byte islands for most out-of-range branches and 12-byte thunks for branches that require greater reach. Each extender type has a configurable hop limit. The planner tries extender types in order of increasing reach, using a single…

MachineScheduler: why does the heuristic unconditionally relax the pressure limit by LiveThru?

Hi all, I’d like to sanity-check my understanding of the register pressure heuristics in MachineScheduler, and get feedback on a possible refinement. Background — the code In RegPressureTracker::getUpwardPressureDelta (llvm-project/llvm/lib/CodeGen/RegisterPressure.cpp at main · llvm/llvm-project · GitHub)): unsigned Limit = RCI->getRegPressureSetLimit(PSetID); if (!LiveThruPressure.empty()) Limit += LiveThruPressure[PSetID]; My understanding of why Limit += LiveThru exists The upstream commit…

RFC: extend LIBOMPTARGET_INFO functionality to allow verbosity control

The current control for LIBOMPTARGET_INFO is only a bitmask toggling categories of debug messages with no control over verbosity. This RFC proposes adding an additional knob of control with verbosity similar to what is already done in LIBOMPTARGET_DEBUG. To achieve this I propose two different approaches: Proposal A: extend LIBOMPTARGET_INFO parsing to handle verbosity control: Verbosity level from 1-3 (described in table below) LIBOMPTARGET_INFO=all:1 (all = -1 bitmask at level 1 verbosity)…

[RFC][AArch64] HBC Optimization Based on CSSPGO

Assisted-by: AI tool (for English translation and formatting) 1. Motivation The Armv8.8-A introduced Hinted Conditional Branch instructions (BC.cond, FEAT_HBC). This instruction branches conditionally to a label at a PC-relative offset, with a hint that this branch will behave very consistently and is very unlikely to change direction. When the prediction holds, the CPU can reduce branch-predictor pressure and improve performance in hot code with highly biased branches. LLVM already provides…

RFC: Improve the cross-platform compatibility and security of the `llvm/utils/update_test_body.py` script

Regarding llvm/utils/update_test_body.py in the Elaborated tests section of the official documentation. ; RUN: rm -rf %t && split-file %s %t && cd %t Background The official documentation describes two issues with this check command: cross-platform compatibility and system security. Regarding cross-platform compatibility: The official documentation uses standard Unix commands at the beginning of the test scripts in the example file llvm/utils/update_test_body.py, which is only compatible with…

RFC: Adding `fpconvert` as a generalization of `fpext` and `fptrunc`

LLVM IR currently has two instructions for converting between floating-point types: fpext and fptrunc. The verifier only compares the bit widths of the source and destination types: fpext requires the source to be narrower, and fptrunc requires the source to be wider. This leads to a few problems: Conversions between types of the same width cannot be expressed, such as bfloat ↔ half and fp128 ↔ ppc_fp128. Some conversions are lossy in both directions. x86_fp80 has a larger exponent range than…

[RFC] [Clang] Use column-major matrix SSA in every language mode

Context RFC: RFC: Represent HLSL Matrix Layout in the Clang AST The Problem #230638 exposed an issue in C\C++ that HLSL recently solved. Clang currently allows the selected memory layout to leak into the representation and interpretation of matrix values. As a result, every operation must determine whether its vector operand is a row-major memory image or a logical matrix value. That ambiguity is already is a concrete problems in CIR. The discussion in #230600 identified exactly where this…

[RFC] Error-checking APIs for memory allocation and management

Goal Define an API for allocating/managing memory that: Makes it hard to forget checking for allocation failures Validates (in debug mode) the presence of these checks at runtime Prevents accidental leaks, e.g. when a multi-step allocation fails For this RFC, I’m considering both simple allocation calls (new, make_unique) as well as container APIs that allocate memory implicitly. For containers, I am not considering objects whose move/copy operations require allocating memory (and can thus…

[RFC] Add widening multiply libcalls to RuntimeLibcalls

While working on llvm-z80, a downstream backend for the Z80, I added this lowering: [Z80] Multiply 16-bit operands to a 32-bit product in one call · llvm-z80/llvm-z80@2cffd11 · GitHub It looks useful for other targets that lower multiplication to library calls as well, so I would like to propose it upstream. Summary Add runtime libcall kinds for widening multiplication, and use them to optimize multiplies in SelectionDAG and GlobalISel where possible. Motivation On targets without a native…

Agentic AI Performance Engineering at Aire

Hello everyone, I am excited to share this opportunity, come help us shape the future of agentic performance engineering. Aire is a well-funded AI neolab, spun up by a world-class team and building an autonomous AI Research Factory. The performance, efficiency and velocity of our software stack is of the highest importance to make our products successful. We are looking to fill multiple positions across the high-performance software stack and especially on Kernels, Compilers and Tools. Join us…

LLVM Weekly - #667, October 12th 2026

LLVM Weekly - #667, October 12th 2026 If you prefer, you can read the canonical version of this issue at https://llvmweekly.org/issue/667. Welcome to the six hundred and sixty-seventh issue of LLVM Weekly, a weekly newsletter (published every Monday) covering developments in LLVM, Clang, and related projects. LLVM Weekly is brought to you by Alex Bradbury. Subscribe to future issues at https://llvmweekly.org and pass it on to anyone else you think may be interested. Please send any tips or…

Growing an LLDB Language Plugin API

I have in the past commented in various threads that I would like to make the LLDB Swift language plugin an actual plugin that can be loaded into any LLDB. I want to give an update on this effort, lay out my plan to get there, and encourage others to participate. My end goal is to have a semi-stable plugin API for languages that sits somewhere in the middle between the SBAPI and the lldb_private API: It would have a limited compatibility window, for example 3 major llvm.org releases, to allow…

LLVM 2026 dev conf debug-info round-table

Is there interest in a debug-info round-table at the 2026 US LLVM dev conference coming up? The deadline to request a time slot is Oct 16, and I’d be happy to register a time and run it. Does anyone have day/time preferences? Most of the debug-info specific talks are on the 27th, so I was thinking the 28th made sense. Also/alternatively, will anyone be around on days before or after the conference and want to meet to discuss? I will be there a few days before, and the day after. 6 posts - 4…

2026 LLVM Developers Meeting - Registration Closes October 16th

The 2026 US LLVM Developers’ Meeting is one of the largest compiler conferences in the world, bringing the LLVM Project community together each October in the heart of Silicon Valley. Organized by the LLVM Foundation and community volunteers, it unites developers and users of LLVM, MLIR, Clang, and related subprojects from across the globe. So much of what makes open source work happens in person. Meeting the people behind the patches, putting faces to usernames, and having unplanned…

[RFC] C++26 compatibilty for llvm::Twine concatenation

Overview In P2591 “Concatenation of strings and string views”, new concatenation overloads are added between std::string and std::string_view: template<class charT, class traits, class Allocator> constexpr basic_string<charT, traits, Allocator> operator+(const basic_string<charT, traits, Allocator>& lhs, type_identity_t<basic_string_view<charT, traits>> rhs); template<class charT, class traits, class Allocator> constexpr basic_string<charT, traits, Allocator> operator+(basic_string<charT,…