2023Q2 week log

I tend to keep quite a lot of notes on the development related (sometimes at work, sometimes not) I do on a week-by-week basis, and thought it might be fun to write up the parts that were public. This may or may not be of wider interest, but it aims to be a useful aide-mémoire for my purposes at least. Weeks with few entries might be due to focusing on downstream work (or perhaps just a less productive week - I am only human!).

Week of 15th May 2023

• Corrected Clang codegen support for half FP types when the zhinx extension is available (D150777.
• Rebased and committed patches to implement MC layer support for the bfloat16 extensions (unblocked now a new PDF was posted in the riscv-bfloat16 repo). Zfbfmin, zvfbfmin, zvfbfwma. Also made a trivial typo fix to the spec.
• Looked at cleaning up the usage of report_fatal_error in the RISC-V backend and also fixed a bug encountered while looking at this. D150674.
• Participated in the ongoing discussion about adding atomics lowering to the RISC-V psABI, including a slightly altered lowering of some primitives in order to allow for forwards compatibility with "table A.7".
• Usual mix of upstream LLVM reviews, and a number of RISC-V psABI or ASM manual reviews.
• LLVM Weekly #489.
• Missed some weeks - busy with EuroLLVM etc.

Week of 17th April 2023

• Still pinging for an updated riscv-bfloat1y spec version that incorporates the fcvt.bf16.s encoding fix.
• Bumped the version of the experimental Zfa RISC-V extension supported by LLVM to 0.2 (D146834). This was very straightforward as after inspecting the spec history, it was clear there were no changes that would impact the compiler.
• Filed a couple of pull requests against the riscv-zacas repo (RISC-V Atomic Compare and Swap extension).
  • #8 made the dependency on the A extension explicit.
  • #7 attempted to explicitly reference the extension for misaligned atomics, though it seems won't be merged. I do feel uncomfortable with RISC-V extensions that can have their semantics changed by other standard extensions without this possibility being called out very explicitly. As I note in the PR, failure to appreciate this might mean that conformance tests written for zacas might fail on a system with zacas_zam. I see a slight parallel to a recent discussion about RISC-V profiles.
• Fixed the canonical ordering used for ISA naming strings in RISCVISAInfo (this will mainly affect the string stored in build attributes). This was fixed in D148615 which built on the pre-committed test case.
• A whole bunch of upstream LLVM reviews. As noted in D148315 I'm thinking we should probably relaxing the ordering rules for ISA strings in -march in order to avoid issues due to spec changes and incompatibilities between GCC and Clang.
• LLVM Weekly #485.

Week of 10th April 2023

• Some days off due to the Easter holidays, so less to report this week.
• Updated RISC-V bfloat16 patches (Zfbfmin, Zvfbfmin, Zvfbfwma), incorporating new fcvt.bf16.s encoding. Also filed an issue about the way in which the dependencies of the vector bfloat16 extensions is specified.
• Blogged about updating the Wren language benchmarks.
• Variety of upstream LLVM reviews.
• LLVM Weekly #484.

Week of 3rd April 2023

• Some days off due to the Easter holidays, so less to report this week.
• Posted MC layer (assembler/disassembler) patches for the bfloat16 extensions: Zfbfmin, Zvfbfmin, Zvfbfwma.
  • Also posted a PR to the riscv-bfloat16 spec to clarify the vector extension dependencies.
  • Pinged on my bug report about the fcvt.bf16.s encoding clash. Once this is resolved, the LLVM MC layer patches can land.
• Updated authorship information for riscv-toolchain-conventions which now has a range of contributors beyond myself.
• Usual mix of upstream LLVM reviews. This included some discussion on changing the shadow call stack register to x3 which spilled into the psABI PR where I suggested some alternate wording.
• LLVM Weekly #483.