)]}'
{
  "commit": "5392aced2a4dd4cb03079f8b376a7a6f2023e235",
  "tree": "f5ad210f9aeff4486268f7014c68e8d936bc6110",
  "parents": [
    "6a685d46af0d0b903d63c9ea2a767f3e89d67fc0"
  ],
  "author": {
    "name": "Evan Brown CA",
    "email": "ezb@google.com",
    "time": "Thu Nov 05 15:08:38 2026 -0800"
  },
  "committer": {
    "name": "Copybara-Service",
    "email": "copybara-worker@google.com",
    "time": "Thu Nov 05 15:25:53 2026 -0800"
  },
  "message": "Use increment_size() in CommonFields::set_full_soo().\n\nReplace `inline_data_.set_size(1)` with ABSL_SWISSTABLE_ASSERT(empty()) and `inline_data_.increment_size()` in CommonFields::set_full_soo().\n\nEvery caller of set_full_soo() (find_or_prepare_insert_soo(), Copy(), and Rehash()) invokes it on an empty SOO table where the upper 40 size bits of `data_` are already zero. HashtableInlineData::set_size(1) computes `(data_ \u0026 kMetadataMask) | (uint64_t{1} \u003c\u003c kSizeShift)`, forcing compilers to mask off the size field before setting bit 24 (`andl $0xffffff` + `orq $0x1000000` on x86-64, or a 4-instruction load-mask-or-store sequence when updating memory). Using increment_size() (`data_ +\u003d kSizeOneNoMetadata`) reduces register updates to a single `addq $0x1000000` on x86-64 (or `add x, x, #0x1000, lsl #12` on AArch64) and folds memory updates into a single read-modify-write addq without introducing narrow byte stores that could stall subsequent 64-bit loads of `data_`.\n\nPiperOrigin-RevId: 1015022955\n",
  "tree_diff": [
    {
      "type": "modify",
      "old_id": "536a767ffbb7f08ddd649664ffde01c0fe9363d7",
      "old_mode": 33188,
      "old_path": "absl/container/internal/raw_hash_set.h",
      "new_id": "9f7cd56867f141e0df166584c64de26f028d2b77",
      "new_mode": 33188,
      "new_path": "absl/container/internal/raw_hash_set.h"
    }
  ]
}
