[nihalpasham]

TL;DR: zero functional change. Every commit is a code move. Same behavior, same PTX, same everything, just smaller files.

The problem

Some of our files got genuinely huge:

generated_intrinsic_targets.rs   147,261 lines  (one 69k const + one 76k #[test] fn)
importer generated.rs             94,518 lines  (ONE 51,700-line function)
resolve.rs                        42,000 lines  (559 fns in one file)
render.rs                         25,281
rvalue.rs                         15,707
...and 12 more above 4k

Why that hurts:

• rust-analyzer: a 147k-line file drags every editing session in that crate
• compile times: rustc cuts codegen units per module, and a single item can never be split, so that 51.7k-line function is one serial chunk on every build, and it sits in the crates we rebuild most
• review: a catalog bump used to diff into one monolithic file
• navigation: finding anything meant scrolling a 42k-line file

What we did

Every 4k+ line file became a directory of small modules, cut along real seams:

before                          after
resolve.rs        42,000   →    resolve/        52 files, one per intrinsic family
render.rs         25,281   →    render/         24 files, one per pipeline stage
rvalue.rs         15,707   →    rvalue/         17 files, one per MIR concern
commands.rs       11,297   →    commands/       20 files, one per subcommand
cuda-macros lib   10,310   →    lib.rs + 16     one per proc-macro
types/target/aggregate/memory/model              5 more source splits
4 big test suites                                split by feature area

And the generator now emits its three monoliths as per-family files, the same way it already sharded dialect-nvvm and cuda-device:

importer generated.rs    94,518   →   generated/                    39 shards + tests/
targets                 147,261   →   generated_intrinsic_targets/  per family, biggest 9.9k
lowering                 17,861   →   generated_intrinsics/         35 shards, biggest 4.5k

Nobody outside a split file notices: each new mod.rs re-exports the old surface, so no call site changed anywhere.

No functionality change

• 16 of 17 commits are pure moves, verified item-for-item:
  • fn-count parity per file (resolve.rs: 559 fns before, 559 after)
  • test-count parity per crate (38 / 54 / 78 / 131 suite tests unchanged, etc.)
  • the sharding commit was diffed against the old monoliths: 1,017/1,017 dispatch arms, 1,016/1,016 table entries, 549/549 lowering impls, content-identical
• The three deliberate edits, each flagged in its commit:
  1. rvalue's Aggregate arm body extracted into its own fn (own commit, prev_op order preserved)
  2. two include_str! paths in cuda-macros tests gained one ../ (tests moved a directory deeper)
  3. the generator sharding itself (dispatch becomes a per-family chain; the abi_id test goes from globally-sorted to per-family-sorted + globally-unique, safe because all lookups are linear scans)

Validation

• cargo test --workspace: 5,287 passed / 0 failed / 173 ignored (baseline 4,264; the growth is the 76k-line single test fn becoming 1,016 small tests, plus feat(codegen): forward assert-bounded result-bundle indexing #1143)
• smoketest: 219 / 219 pass, ran twice (before and after rebasing onto feat(codegen): forward assert-bounded result-bundle indexing #1143)
• cargo run -p cuda-intrinsics-gen -- check: outputs current
• fmt, clippy (--workspace --all-targets), rustdoc: all clean, zero warnings

Review guide

• Go commit-by-commit: each commit is one file's split and diffs as a move
• Heads-up: open PRs touching the pre-split files (feat(intrinsics): add ordered sparse Ampere floating MMA #1139 feat(intrinsics): add CTA barrier reductions #1138 Implement generated mbarrier_arrive_no_complete intrinsic #1101 refactor(codegen): parse targets once, one vocabulary owner #1184 feat(wgmma): add two-slot F16 partial-wait pipeline lowering #1145 Use the shared cutile-rs host crates in cuda-oxide #1102 Add borrowed DeviceBuffer subviews for typed kernel launches #1081 cargo-oxide: fmt must not build the codegen backend #1188) will need a rebase after this lands