TheRock/compiler/amd-llvm/flang/docs/MeetingNotes/2026/2026-01-28.md

Combined Call 2026-01-28

Agenda

• Design docs and/or RFCs
  • Flang failure with thin lto on LLVM 22-RC1 - Thank you for fixing it!
    • https://github.com/llvm/llvm-project/issues/176404
  • [RFC] What to do regarding the Flang Call Notes document
    • Proposal: in the Flang repo
      • This Google doc becomes a working document for drafting the agenda, taking notes during the call
      • Post the agenda the day before the meeting, indicate to edit document directly with additional items or other changes, reply to the Slack post if that is not possible
      • flang/docs/MeetingNotes/YYYY/
        • files of finalized notes go in this directory, in a file like YYYY-MM-DD.md
      • flang/docs/MeetingNotes/README.md contains the boilerplate information, point to the GettingStarted.md document for call information to avoid tracking in two different places
  • [RFC] Support classic flang driver options in flang
    • Suggestion was to accept the alternate spellings but give a deprecation warning immediately to encourage migrating to the regular spelling
    • Time frame for full deprecation/removal?
    • Should classic Flang be treated like a totally separate compiler, like Intel vs. Cray?
  • [RFC] Automatic static promotion of large local variables in Flang
    • There is a similar issue with OpenMP privatize
    • Michael Kruse - If you do this, you run out of stack space quite quickly. So the default should be on the heap, but with a flag to specifically put arrays (even of unknown size) on the stack
    • Ted Johnson - Cray has flags for both behaviors, but there are a lot of flags to control the exact behavior (default is heap for certain use cases, stack for others)
    • Pthreads has a way to change the stack size via a function
    • What about executable stacks and how they factor in?
  • Flang-tidy
    • Meeting with authors has been deferred, will create meeting including Tarun Prabhu
  • Cross-compilation of real(kind=16)
    • PR will come out, remaining issues can be resolved via that discussion
  • [RFC] Support -fstrict-aliasing and -fno-strict-aliasing
    • Related issue: https://github.com/llvm/llvm-project/issues/171912
    • We can’t do anything reasonable with it without violating the Fortran standard, will leave it open for a little while longer but likely will not support this
  • Changes to builtin modules
    • PR was reverted due to buildbot failure involving Windows
    • Reviewers asked for big changes, working on a compromise involving putting TODOs in other runtimes
  • [RFC] Add a New Byte Type to LLVM IR
    • Among other things, “Fixes known correctness bugs in memcpy lowering and load merging/widening”. Semantics are different from integer types and may allow other memory optimizations. This is a step towards eventual removal of Undef from LLVM-IR.
    • Related group working towards a formal specification of LLVM-IR
  • `-ffp-contract=fast` Violates the Fortran Standard
    • Reply from the standards committee: https://mailman.j3-fortran.org/pipermail/j3/2026-January/015531.html
• PRs of Note
  • Tests from Cray’s CCE internal test suite in llvm-test-suite [PR]
    • Proposed directory structure with one test in linked PR
  • [Flang][FIR] Introduce FIRToCoreMLIR pass. #168703
• Issues of Note
  • gcc warning suppression (Michael Kruse)
    • https://github.com/llvm/llvm-project/pull/174918
    • Also see message in Slack #general
• FYI
  • 2026 EuroLLVM Developers' Meeting - Registration Open!
  • Call for Mentors - 2026 EuroLLVM Developers’ Meeting
    • https://discourse.llvm.org/t/call-for-mentors-2026-eurollvm-developers-meeting/89484/
    • Google Form here: EuroLLVM Mentorship Program Application
  • 2026 LLVM Community Area Team Elections - Voting is open!
    • Voting closes Feb 9
  • LLVM 22.1.0-rc2 Released!
    • The next release will be 22.1.0-rc3 on Tuesday the 10th of Feb
    • Who is writing the flang release notes?
  • Improvement of Fujitsu test suite
  • [RFC] Use pre-compiled headers to speed up LLVM build by ~1.5-2x
    • Related to what has already been done in Flang, attempt to use this approach more generally
  • Added LoopInvariantCodeMotion pass for [HL]FIR. (PR #173438)
• Other topics as time allows
  *

Details

• Consists of over 721,000 lines of code, documentation, build files, and test
• To date, over 11,703 commits have been made to Flang