[aneshlya]

Description

Tests and benchmarks for the gather/scatter work, ahead of rewriting ImproveMemoryOps and GatherCoalescePass (#3990). No compiler changes.

• Lit tests pinning today's codegen (gather-index-width, gather-scale, gather-factored-offsets, gather-to-load, coalesce-gathers, masked-memops-known-mask, inline-helper-gather, perf-warn-inline-location, pipeline-order, plus the reproducers of gather coalescing doesn't happen for simple program #304, Perform full vector stores when sequentially setting uniform values in multiples of the native vector width #2720, Gather/Scatter can be avoided if Source and Destination are of same type. #2778, Codegen for varying bytes #2901, Unexpected scatter performance warnings using soa #2308, How to process H,W,C effectively? #2919 and How to work with interleaved RGBA data #2943). Each case pins today's output. Correct but suboptimal output carries an issue-linked TODO comment stating the condition under which the check may change, so the change that implements the improvement updates the check and removes the TODO.

• Lit tests for the transformations no test noticed. I switched off each transformation of the two passes, one at a time:

  • Six changed real code without failing any lit test: scale extraction, GetMoreConst, global-array bases, 2- and 4-wide coalescing loads, and all-off masked store removal.
  • Seven more were caught by only one test.

  The new tests pin each of them. They also cover --wrap-signed-int, --opt=force-aligned-memory, 32-bit x86, x64 coalescing, barriers in the coalescing window and 32-bit offset narrowing. Every pin was checked to pass on today's compiler and to fail with its transformation switched off. memory-access-fallback.ll checks that general varying addresses and unequal insert-chain bases keep the general fallback; memory-access-failed-match.ll (XFAIL) checks that a rejected base recovery leaves the IR unchanged, which step 1 of Refactor ImproveMemoryOps around shared, read-only address analysis #3990 requires.

• Guard pages in the functional test harness. The input arrays (AOT and JIT) now sit next to an inaccessible page, so an access past the end or before the start of an input faults instead of passing silently. A new rule key, // rule: run|skip on addressing=32|64, selects tests by addressing mode.

• 24 functional tests for footprint (inactive lanes and tails next to the guard pages), index semantics (8- to 64-bit signed and unsigned indices, byte offsets of 2^31 or more), conflicts (duplicate scatter indices, loads across aliasing stores, an atomic and two bases) and shapes (How to work with interleaved RGBA data #2943, Gather/Scatter can be avoided if Source and Destination are of same type. #2778, records read through inline helpers).

• Gather/scatter microbenchmarks: 01_trivial/18_gather_scatter and 03_complex/01..05. They show in the not-gated section of the dashboard until they have a stable history.

• XFAIL reproducers for the known address gaps that remain after Unsigned 32-bit gather index is scaled with nsw before zero-extension (poison for offsets above INT32_MAX) #3985. Guard-page runtime tests (address-gap-*.ispc): a masked access whose lane 0 is off and whose signed offset overflows, offsets that wrap at 2^32 between lanes, a signed constant split without nsw, a struct member at 2 GB through a varying pointer, a nested-index sum of two signed-i32 byte offsets that reaches 2^31 without any integer-to-pointer cast (the counterexample of the Refactor ImproveMemoryOps around shared, read-only address analysis #3990 design doc), and the integer-to-pointer provenance loss of Memory builtins lose pointer provenance and capture info; redundant gathers and loads survive -O2 on all targets #3983. IR tests: new sign extending an unsigned count, an unfolded +2^31 displacement that is not safe to truncate, nneg copied to split or rescaled operands (address-gap-nneg-*.ll), and an explicit int64 sum converted to a pointer. Each XFAIL test runs only the configurations that fail today; address-gaps-ok.ispc, address-nested-sum.ispc and address-provenance.ispc run the passing ones as regression gates. The kernels and drivers are shared in address-gaps.inc and address-provenance.inc.

Bugs found along the way and filed separately: #3983 (integer-to-pointer casts lose provenance), #3984 (GatherCoalescePass and negative offsets with --addressing=32), #3985 (nsw on unsigned index scaling), #3986 (crash after an undeclared symbol in a template).

Testing

• check-all passes on the current main (ded7463) with LLVM 21.1.8: 711 passed, 18 expected failures, 121 unsupported, no failures. The new and changed lit tests were also run on LLVM 20, 22, 23 and trunk in earlier revisions of this branch.
• The functional suite passes with the guard pages on avx2-i32x8 at the current main: 1668 tests, no compilation or execution failures, 7 skipped (AMX, Xe and the --addressing=64-only gather-index-overflow-2). Earlier revisions of this branch were run on sse4-i32x4, avx2-i32x8, avx2-i32x16, avx512skx-x16 and avx512skx-x64, with --jit and with --addressing=64.
• Every XFAIL build/check/run step was verified independently (lit stops at the first failure, so an XFAIL result alone does not establish every case): each fails on its own with the intended wrong result, and the passing variants pass.

Related Issue

• Linked to relevant issue(s): Refactor ImproveMemoryOps around shared, read-only address analysis #3990, Memory builtins lose pointer provenance and capture info; redundant gathers and loads survive -O2 on all targets #3983, GatherCoalescePass uses wrong addresses for negative constant offsets with --addressing=32 #3984, Unsigned 32-bit gather index is scaled with nsw before zero-extension (poison for offsets above INT32_MAX) #3985, Use index scale immediate of gather instructions #1256, gather() optmization #1531, Inefficient Gather with 16-bit indices  #1581, gather coalescing doesn't happen for simple program #304, Perform full vector stores when sequentially setting uniform values in multiples of the native vector width #2720, Gather/Scatter can be avoided if Source and Destination are of same type. #2778, Codegen for varying bytes #2901; discussions Unexpected scatter performance warnings using soa #2308, How to process H,W,C effectively? #2919, How to work with interleaved RGBA data #2943

Checklist

• Code has been formatted with clang-format (e.g., clang-format -i src/ispc.cpp)
• Git history has been squashed to meaningful commits (one commit per logical change)
• Compiler changes are covered by lit tests
• Language/stdlib changes include new functional tests for runtime behavior
• Documentation updated if needed

🤖 Generated with Claude Code

[dbabokin]

@codex review

[chatgpt-codex-connector]

Codex Review Summary

This comment shows the latest Codex review activity on this pull request.

Review	Status	Commit	Review trigger
📝 Code Review	✅ Completed 2026-11-02T18:41:45.682645Z	f69ec26	Manual request

ℹ️ About Codex in GitHub

Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you

• Open a pull request for review
• Mark a draft as ready
• Comment "@codex review" or "@codex security review".

Codex reacts with 👀 while any review is running, comments if it has suggestions, and reacts with 👍 once all reviews finish with no findings.

[chatgpt-codex-connector]

Codex Review: Didn't find any major issues. Nice work!

Reviewed commit: f69ec26442

ℹ️ About Codex in GitHub

Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you

• Open a pull request for review
• Mark a draft as ready
• Comment "@codex review".

If Codex has suggestions, it will comment; otherwise it will react with 👍.

Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".

[dbabokin]

Adversarial review of f670e9e: no actionable correctness or regression defects found.

I reviewed this from four independent angles: shared harness/argument handling and portability; functional-test semantics, bounds and oracles; lit/FileCheck and expected-failure coverage; benchmark correctness, measurement boundaries and dashboard integration.

Validation using a fresh build of this revision with LLVM 23.1.2 on macOS ARM64:

• cmake --build build-review --target check-all -j 8: 622 passed, 11 expected failures, 217 unsupported, no unexpected failures.
• All 24 new functional tests checked on neon-i32x4 and generic-i1x64, with both addressing modes: 23 passed plus the intended addressing-only skip in each 32-bit run; all 24 passed in each 64-bit run. Native JIT: 23 passed plus the intended skip.
• Full functional suite on neon-i32x4/aarch64: 1,653 passed, 15 skipped out of 1,668, zero compilation or execution failures.
• All six new benchmarks compiled and passed their C++ correctness checks on both targets: 502 cases total. Dashboard selection also matched the intended not-gated cases.
• Direct harness probes verified 64-byte alignment, valid boundary accesses and faults immediately beyond the protected boundary. Addressing-selection probes covered flag precedence, fixed-width architectures and JIT.
• Independently compiled and checked the standalone IR XFAILs to confirm they fail for their documented transformations, rather than an unrelated compilation error.

Nonblocking coverage suggestion: split independent runtime configurations into separate XFAIL tests. For example, in address-gap-nested-sum.ispc:9–18, the first failing execution prevents the --wrap-signed-int and -O0 runs from executing in CI. Separate tests would continuously exercise each configuration and produce an XPASS when one is fixed. The PR explicitly acknowledges this limitation and reports independent manual checks, so I am not treating it as a blocking defect.

Local coverage limits: Linux-only guard-page runtime reproducers, native x86/Windows execution and other LLVM versions were not independently rerun. The current PR CI is green, but the local results above do not replace that platform coverage.