I'm not currently the right person to review this, resigning.

First off, thanks for the discourse!

Sure, simply run the readelf commands from this commit message or something similar on the clang toolchain you built, both in your build and install directories, and perhaps with this commit reverted too.

Looks like my previous message was truncated. I think it is 'not incorrect' for the rpath to point to LLVM_LIBRARY_DIR, which is the install path at least for the nixpkgs llvm install. In this case, the prefix clang is installed to differs from the one LLVM is installed to. If the rpath isn't set for clang there, it can't find libLLVM.

If any LLVM subprojects are built separately, the LLVM build directory LLVM_LIBRARY_DIR is added to both the build and install runpaths in llvm_setup_rpath(), which is incorrect when installed.

Doesn't build for me, tried against main (8bf7f86d7966) and your latest commit to vplan (668045eb7762).

Makes sense to me, one minor suggestion. I note that this introduces a few loads/stores to the stack for fixed length vectorization, but having it work correctly is the priority, those may be possible to remove later.

Are there other parent revisions? The base isn't a valid commit and the patch doesn't apply to main nor commits from the end of January, for example CompareByComesBefore isn't in the patch context but is on main and is old.

I've been down the list of MCA output changes and agree those look good, and your reasoning seems good to me. Accepting as-is, since we think represents an improvement.

LGTM!

An alternate route to implementation landed via D139637, D139850.

Wrong action selected... :)

Thanks for the input Florian!

As @fhahn mentioned in D137159, the MemoryDependenceAnalysis infrastructure is considered deprecated, so this is not viable.

• Fix test naming.
• Also cover the zero passthru case.

Address review comments:

• Remove superflous braces.
• Prefer indexed operations in the presence of a non-undef passthrough to a dup.

LGTM, but please address the comment about poison and allow time for other reviewers to chime in.

• Update a test name (an old one crept in)

• Remove 'q' accidentally copied into the test names.

• Test zext/sext.
• Test unpacked forms (no negative tests, since dup with unpacked types is not specifically supported currently as there is currently no known way to get a DUP_MERGE_PASSTHRU from C/C++ code).
• Drop immediate tests since those are covered.

• Rebase after b4028fbc1a88

• Address review comments: Move comment down and rename BisectExtendVector => HalveAndExtendVector.

Thanks Bradley -- I'll continue this one in the near future though your input is welcomed.

LGTM with minor comments.

LGTM, Thanks for trying the path of making it generic and thanks for the input from the other reviewers!

One suggestion inline.

LGTM, but adding other reviewers for visibility. The predictable ask is to make it a target-agnostic combine: reviewers, thoughts?

Hi @dyung, thanks for the report and reproducer. Matt's away today so I've applied a revert in e1790c8c290d773cd5b1fc79f80b7a23dceb7589.

This is my first gambit, and I'm relatively new to making changes to LoopVectorize. Please don't hold back if this is the wrong approach, I'm happy to put legwork in to finding the right route to making this work.

LGTM, with one minor nit.

LGTM, one nit.

LGTM. (With the fix) I've tested this produces the expected output.

Looks good to me, with one minor suggestion.

(Full disclosure: Ben is a new colleague of mine)

I will submit first thing next week unless there are further comments.

• Rebase onto 898699831b54, which fixes the zero-extending case in visitAND.
  • A backport has been requested for this one in https://github.com/llvm/llvm-project/issues/57089.
• Use .getValueType().

• Fix copy-paste error for the legality check for the ZERO_EXTEND case.

Ihanks @RKSimon. The tests are unaffected. I've opted to follow your suggestion.

Thanks for the review comments, Simon!

New patch at https://reviews.llvm.org/D131503

• Stylistic changes to shrink the code a little.
• Add check for signext_inreg type.
• Update tests.