[aneshlya]

Description

Tests and benchmarks for the gather/scatter work (roadmap 2.2), ahead of rewriting ImproveMemoryOps and GatherCoalescePass. No compiler changes.

• Lit tests pinning today's codegen (gather-index-width, gather-scale, 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 and gives the ideal output in an Ideal: comment, so the change that reaches it updates the checks.

• 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.

• 23 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) and shapes (How to work with interleaved RGBA data #2943, Gather/Scatter can be avoided if Source and Destination are of same type. #2778).

• 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.

• 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.

Bugs found along the way and filed separately: #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. The new and changed lit tests pass on LLVM 20, 21, 22, 23 and trunk.
• The full functional suite passes with the guard pages on sse4-i32x4, avx2-i32x8, avx2-i32x16, avx512skx-x16 and avx512skx-x64, with --jit and with --addressing=64.
• The tests of the last commit were run on sse4-i32x4, avx2-i32x8, avx2-i32x16 and avx1-i32x8 (gather-alias-1/2/3).

Related Issue

• Linked to relevant issue(s): 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