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

LLVM Dev Meeting 2026 SelectionDAG Roundtable Notes

FREEZE handling - stalled by a number of PRs – Long standing regressions preventing them landing (e.g. RISCV64 sllw/sraw nodes) – When can we just fold FREEZE(LOAD(CONSTANT_POOL)) → LOAD(CONSTANT_POOL) – DAG still CSEs FREEZE(POISON/UNDEF) nodes of the same type - can affect isel Topological Sorting – We now have the ability to enable this at command line and per-target – Not just regressions - still have crashes in some tests – Sam Elliot and Craig Topper said they’d take a look at narrowing…

LLVM Dev Meeting 2026 InstructionCost Roundtable Notes

Rough notes from the roundtable discussing how we can improve InstructionCost to better model code sequences and share more cost data with backend scheduler models. The focus was mainly on Vectorizers/VectorCombine, but other passes need better cost estimates as well. Are the TTI::getCost methods actually any good for how we use them? – No way to describe Instruction sequences, how they will combine (e.g. load+extend, compares sunk into eflags, X86addsub, value tracking) – Please can we get rid…

LLVM Dev Meeting 2026 X86 Roundtable Notes

Apologies for my rough notes, but these are the topics that arose at the round table. Gather Costs – Is this the cause of the SPEC Prefer512Bit regression in [X86] Remove TuningPrefer256Bit from x86-64-v4 · Issue #231086 · llvm/llvm-project · GitHub? – Znver4 gather cost PRs still need reviewing but this is happening on all AVX512 targets – Fix TTI isLegal check to allow AVX2 to use gathers for CodeSize builds Cleanup CodeGen – X86ISelLowering.cpp recently went over 64K LoC – Many folds/tests…

[RFC] Typedefs for evaluate::Type

With [Flang] KIND De-templatization by Meinersbur · Pull Request #216960 · llvm/llvm-project · GitHub landed, some of the type aliases do not make sense anymore, for instance using Int4 = Type<TypeCategory::Integer>; using Int8 = Type<TypeCategory::Integer>; I created [Flang][NFC] Remove Type convenience aliases by Meinersbur · Pull Request #238270 · llvm/llvm-project · GitHub to clean them up. It currently does this by deleting such aliases and using Type<TypeCategory::X> directly instead. The…

[RFC] Strengthening LLVM’s newcomer contribution workflow: fuzzer-generated issues, AI use, and reviewer support

Background LLVM’s AI Tool Use Policy already prohibits using AI to fix good first issue tasks, a restriction introduced in PR #172515: AI tools must not be used to fix GitHub issues labelled good first issue. These issues are generally not urgent, and are intended to be learning opportunities for new contributors to get familiar with the codebase. Whether you are a newcomer or not, fully automating the process of fixing this issue squanders the learning opportunity and doesn’t add much value to…

LLVM Meetup Toronto - 12 November 2026

We’re pleased to collaborating with the University of Toronto Computer Science Student Union on hosting the next Toronto LLVM Meetup! Please join us for a pair of technical talks, followed by socializing with fellow LLVM enthusiasts! Location Bahen Centre of Information Technology 40 St George St, Toronto, ON M5S 2E4, Canada · Toronto, ON Room BA3200 Agenda 6:30 - 7:00 pm Doors open, socializing 7:00 – 7:30 pm Presentation “Getting started with LLVM (it’s not that scary!)” - by Ayokunle Amodu…

LLVM 23.1.5 Released!

We are happy to announce that LLVM 23.1.5 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! Download Find sources on GitHub. A note on binaries Official binaries will not be available right away. They will be provided later. Third-party binaries will be posted to the following forum thread. Note that these binaries…

LLVM Vulkan 1.1: Drop-in middle-end passes (GVN, CSE, CFG, DCE) using Vulkan compute

Hi everyone, I’d like to share LLVM Vulkan 1.1, an open-source project exploring whether standard Vulkan compute shaders can be used as an acceleration layer for LLVM optimization passes without compromising LLVM’s host semantic authority. Project site: LLVM Vulkan — Portable bulk-compute API and scoped LLVM passes GitHub: GitHub - HECer/llvm-vulkan: Portable Vulkan bulk-compute API and scoped LLVM passes, with reproducible CPU/GPU benchmarks and release packages. · GitHub Release 1.1.0:…

Flang Round Table Notes – US LLVM Dev Meeting, Oct. 27 2026

These are not listed in any particular order of importance and may include notes from discussions that happened over the two days of the conference, not just at the roundtable Call-related topics Keeping things on the biweekly meeting agenda, even if no one comments on them at the time, is helpful for visibility of the issue and ultimately worth continuing to do The old Monday technical call still needs to be removed from the official LLVM calendar Does having the one biweekly call and the…

[RFC] Add a device address sanitizer for GPUs (DASan)

TL;DR: Adds -fsanitize=daddress as a new address sanitizer for AMD GPUs. Provides fast, low-overhead address sanitization for all AMD GPUs without shadow memory. An address carries its own metadata by virtue of where it was placed in virtual memory. Overhead instead scales per-allocation rather than per-granule. Compared to existing GPU implementations, DASan is several times faster, has less memory overhead, and works on all GPUs. Measured overhead was 1.35x on HeCBench and several real faults…

Clarifying "If you have recently been granted commit access, these policies apply"

On a code review, I copy-pasta’ed “try to separate formatting or whitespace changes from functional changes, either by correcting the format first (ideally) or afterward.” from LLVM Developer Policy - LLVM The author observed (correctly) that that quote is part of an “if” block: If you have recently been granted commit access, these policies apply: … You are allowed to commit patches without approval which you think are obvious. … Also, try to separate formatting or whitespace changes from…

LLVM Dev meeting 2026: Runtimes Workshop

Here are the notes from the runtimes workshop last week. Notes are by me unless otherwise specified. I’ll try to get the presentations posted soon. If you’d like any updates to the notes feel free to email me or message me on discord. Thank you everyone who attended! LLVM + Linux (Notes by Petr Hosek and Michael Jones) LLVM/Linux if you prefer Linux implies GNU (typically), you might use some pieces that are not GNU (LLVM, musl), but you generally have to adapt to the GNU shaped hole A new door…

Sunsetting "old" LiveDebugValues / VarLocBasedLDV

Hi, Something I brought up at the debug-info roundtable at the conference, and shared here for wider feedback: I’d like to sunset and eventually delete the “old” way of representing variable locations after instruction selection. This forces some work onto target maintainers, reduces some general maintenance burden, and is necessary step for removing debug instructions after isel. Background: LLVM post-isel used to define the register-location and instruction-position of a variable assignment…

[RFC] Clearly define the meaning of ViewLikeOpInterface

I am currently working on CIR over Enzyme. About three weeks ago, I submitted a patch to wire some CIR ops to ViewLikeOpInterface (PR #230897). However, this causes MLIR’s core Local Alias Analysis (LAA) to return MustAlias for two views with different offsets, as they are based on the same underlying object. Therefore, I proposed another change to add an interface method indicating whether a view may overlap with other views in PR #231285. This allows LAA to return MayAlias and hand the…

Settings for lldb's DataFileCache, as I add new things in there

I’ve been looking at using DataFileCache for some additional use cases, and I don’t think I’m doing a great job of making the settings understandable. I wanted to see what anyone (Greg especially) might think. Today we have these settings for the DataFileCache, symbols.enable-lldb-index-cache default off symbols.lldb-index-cache-expiration-days default 7 symbols.lldb-index-cache-max-byte-size symbols.lldb-index-cache-max-percent symbols.lldb-index-cache-path These allow the user to control…

LLVM Weekly - #670, November 2nd 2026

LLVM Weekly - #670, November 2nd 2026 If you prefer, you can read the canonical version of this issue at https://llvmweekly.org/issue/670. Welcome to the six hundred and seventieth 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…

O2 vs O3 used in practice in large code bases

Hi, There are several discussion on what the optimization levels do. My question is, which optimization level is used in practice to compile medium and large code bases (i.e. thousands to millions of loc) in actual products we can use. For example, I couldn’t figure out in which optimization level the llvm project itself is compiled (the download on https://releases.llvm.org/ for example for windows). 2 posts - 2 participants Read full topic

[mlir][vector][PSA] Disallow pure reduction dims in vector.contract

Let me draw the context first. In the following example, there are two reduction dimensions: r1 is a “pure reduction” dimension (only present in lhs), r2 is a “contraction” dim (present in both lhs and rhs). %res = vector.contract { indexing_maps = [ affine_map<(r1, p, r2) -> (r1, r2)>, affine_map<(r1, p, r2) -> (p, r2)>, affine_map<(r1, p, r2) -> (p)> ], iterator_types = ["reduction", "parallel", "reduction"], kind = #vector.kind<add> } %lhs, %rhs, %acc : vector<7x3xi32>, vector<2x3xi32> into…

[RFC] Unify masking in the Vector dialect

Hi folks! This is a proposal to define how masking functions within the Vector dialect. It follows the discussion in #225244 and #226128, and it is related to #227821. Today an op can be masked by wrapping it in vector.mask, or by passing a mask operand. For transfer_read, transfer_write and gather both are legal at the same time, and nothing says what the combination means. The RFC in one sentence: An op must have exactly one mask and can be masked in only one way, either via vector.mask or…

[RFC] Disallow index type in arith extended ops

The addui_extended, subui_extended, mulsi_extended, and mului_extended operations return a carry, borrow, or high half. The value of these results inherently depends on the bit width. The index type has no fixed width until it is lowered, but the folders compute at 64 bits. Because of this, results currently depend on pass order. For example, mului_extended 65536, 65536 : index folds its high half to 0, but lowering to SPIR-V (32-bit index) first gives 1. Index operands also trigger a…

Missing canonicalisation for tensor.splat %x[const]

I believe a canonicalisation is missing for this construct %cst = arith.constant 1.0 : f64 %dim = arith.constant 4 : index %splat = tensor.splat %cst[%dim] : tensor<?xf64> which could be canonicalised to %cst = arith.constant 1.0 : f64 %splat = tensor.splat %cst: tensor<4xf64> %casted = tensor.cast %splat : tensor<4xf64> to tensor<?xf64> Is there a reason why this is not done? If it’s just an oversight, I’m happy to open a PR 2 posts - 2 participants Read full topic

C/C++ linker error

This is probably not a llvm question, but a general C/C++ question. I’m trying to link a piece of C++ code compiled into a lib into my C project. I’ve created a minimal example without llvm and everything worked fine. This is my actual Makefile: all: pass libpass clang lexer.c parser.c list.c ast.c main.c `llvm-config --cflags --ldflags --system-libs --libs core` -lpass -L. -o kal pass: clang++ -c LLVMOptimizationPasses.cpp -o pass.o libpass: ar r libpass.a pass.o And this is the output with…

Issue with flang-commits

Hi all! I have not been receiving emails from the flang-commits list this week, and I don’t think it’s because people aren’t sending messages. We had a problem like this almost a year ago and the solution then was to clear out disk space. I am only a list moderator and don’t have this ability. I believe my message to Mailman to report the issue bounced. Any advice on how to get the list un-stuck? Thanks! 3 posts - 3 participants Read full topic

How to create a pass pipeline with the C API?

Here is what I’ve got so far: LLVMPassManagerRef FPM = NULL; ... FPM = LLVMCreateFunctionPassManagerForModule(module); ... int optim = LLVMRunFunctionPassManager(FPM, TheFunction); printf("optimized? %i\n", optim); For the example in the fource chapter of the Kaleidoscope tutorial I get: kal> def test(x) (1+2+x)*(x+(1+2)); kal> Parsed a function definition. optimized? 0 define double @test(double %x) { entry: %addtmp = fadd double 3.000000e+00, %x %addtmp1 = fadd double %x, 3.000000e+00 %multmp…

Request for Clarity around methods to preserve metadata through MLIR passes

I have a situation in which I want to ensure MLIR passes preserve a particular kind of metadata about SSA values. This includes my own passes but also upstream and other out-of-tree passes that may inadvertently drop this metadata depending on how it’s represented. The two approaches I have heard people take are: Find a way to encode the metadata in Location, which most passes are pretty good at preserving due to the requirement to populate an op with a Location at creation time. Encode the…

Clarifying the boundaries of AI use for good first issues

Recently, while reviewing a PR addressing a good first issue, I ran into some confusion: the author was a first-time contributor and disclosed their use of an LLM, for example, to learn about Clang’s code and understand the relevant context. However, I wasn’t quite sure whether these uses violated the current policy. The current AI tool policy states that AI tools must not be used to fix good first issues. However, the explanation in the same paragraph also refers to fully automating the…

[RFC] TableGen support for RegisterBankInfo v2

Note: This is a continuation of [RFC] TableGen support for RegisterBankInfo - Code Generation / Common CodeGen Infrastructure - LLVM Discussion Forums. Proposal This RFC tries to provide a way to define register bank mapping rules by matching TypedOperands, with generating time checks. Following tablegen classes are introduced: class RegBankTypeMapping<RegisterBank bank, TypedOperand type> { RegisterBank Bank = bank; TypedOperand Type = type; } class RegBankMappingRule<list<RegBankTypeMapping>…

RFC: tail_chaincc - A Maximum-Efficiency Tail Call Convention

Summary I propose to add a new calling convention (tentatively named tail_chaincc) to LLVM and Clang. This new convention is intended to be used with [[clang::musttail]] in C and C++ (and any other language that wants to use guaranteed tail calls with LLVM, e.g. become in Rust) and is designed to address several practical issues we have encountered when using tail calls. This new convention combines the callee-cleanup mechanics of tailcc with the register optimizations of preserve_none, leading…