)]}'
{
  "commit": "73bf0cd3cca39e159c6f3d4fb5aaff43c69faf6a",
  "tree": "12d95b220cbeb7a56383772eb9d05fcd1c5495dd",
  "parents": [
    "213649ef72d12130827fe25f8a470bd61c7483f3"
  ],
  "author": {
    "name": "Yonatan Nachum",
    "email": "ynachum@amazon.com",
    "time": "Mon Jun 22 10:36:40 2026 +0000"
  },
  "committer": {
    "name": "Yonatan Nachum",
    "email": "ynachum@amazon.com",
    "time": "Tue Jun 23 10:55:43 2026 +0000"
  },
  "message": "efa: Optimize WR memset on WR creation\n\nSince the WQE size (64 or 128 bytes) is a runtime variable, the compiler\ncannot inline memset and emits a function call instead. This adds\nfunction call overhead and multiple instructions to the post send hot\npath.\n\nReplace with an explicit u64 zeroing loop. The compiler recognizes the\nsmall iteration count (8 or 16) and emits inlined stores, avoiding the\nfunction call and the overhead because of it.\n\nMeasured with perf stat on osu_bibw (2x p4d.24xlarge,\nFI_EFA_USE_DEVICE_RDMA\u003d0):\nBefore: 7.93B cycles (memset function call)\nAfter:  7.42B cycles (inlined loop)\n\nSigned-off-by: Yonatan Nachum \u003cynachum@amazon.com\u003e\n",
  "tree_diff": [
    {
      "type": "modify",
      "old_id": "f7dd9bb266fff965841d52f002cf12845864f13c",
      "old_mode": 33188,
      "old_path": "providers/efa/verbs.c",
      "new_id": "ef951ca43eb92d913ba0811f2ce92c6e3eee1d31",
      "new_mode": 33188,
      "new_path": "providers/efa/verbs.c"
    }
  ]
}