[llvm] [memprof] Omit the key length for the record table (PR #89527)

Kazu Hirata via llvm-commits llvm-commits at lists.llvm.org
Sat Apr 20 19:43:10 PDT 2024


https://github.com/kazutakahirata created https://github.com/llvm/llvm-project/pull/89527

The record table has a constant key length, so we don't need to
serialize or deserialize it for every key-data pair.  Omitting the key
length saves 0.06% of the indexed MemProf file size.

Note that it's OK to change the format because Version2 is still under
development.


>From 70bcb4f5ace5c0bdd0b5b24814f3b31cc0da01c7 Mon Sep 17 00:00:00 2001
From: Kazu Hirata <kazu at google.com>
Date: Sat, 20 Apr 2024 18:59:43 -0700
Subject: [PATCH] [memprof] Omit the key length for the record table

The record table has a constant key length, so we don't need to
serialize or deserialize it for every key-data pair.  Omitting the key
length saves 0.06% of the indexed MemProf file size.

Note that it's OK to change the format because Version2 is still under
development.
---
 llvm/include/llvm/ProfileData/MemProf.h     | 12 +++++++++---
 llvm/include/llvm/Support/OnDiskHashTable.h |  2 +-
 2 files changed, 10 insertions(+), 4 deletions(-)

diff --git a/llvm/include/llvm/ProfileData/MemProf.h b/llvm/include/llvm/ProfileData/MemProf.h
index aa6cdf198485b0..f356e3a54a3645 100644
--- a/llvm/include/llvm/ProfileData/MemProf.h
+++ b/llvm/include/llvm/ProfileData/MemProf.h
@@ -471,12 +471,16 @@ class RecordLookupTrait {
 
   hash_value_type ComputeHash(uint64_t K) { return K; }
 
-  static std::pair<offset_type, offset_type>
+  std::pair<offset_type, offset_type>
   ReadKeyDataLength(const unsigned char *&D) {
     using namespace support;
 
+    // Starting with Version2, we don't read the key length because it is a
+    // constant.
     offset_type KeyLen =
-        endian::readNext<offset_type, llvm::endianness::little>(D);
+        Version < Version2
+            ? endian::readNext<offset_type, llvm::endianness::little>(D)
+            : sizeof(uint64_t);
     offset_type DataLen =
         endian::readNext<offset_type, llvm::endianness::little>(D);
     return std::make_pair(KeyLen, DataLen);
@@ -534,7 +538,9 @@ class RecordWriterTrait {
 
     endian::Writer LE(Out, llvm::endianness::little);
     offset_type N = sizeof(K);
-    LE.write<offset_type>(N);
+    // Starting with Version2, we omit the key length because it is a constant.
+    if (Version < Version2)
+      LE.write<offset_type>(N);
     offset_type M = V.serializedSize(Version);
     LE.write<offset_type>(M);
     return std::make_pair(N, M);
diff --git a/llvm/include/llvm/Support/OnDiskHashTable.h b/llvm/include/llvm/Support/OnDiskHashTable.h
index f6b4055e74de7e..b6dbea53f3da6d 100644
--- a/llvm/include/llvm/Support/OnDiskHashTable.h
+++ b/llvm/include/llvm/Support/OnDiskHashTable.h
@@ -377,7 +377,7 @@ template <typename Info> class OnDiskChainedHashTable {
 
       // Determine the length of the key and the data.
       const std::pair<offset_type, offset_type> &L =
-          Info::ReadKeyDataLength(Items);
+          InfoPtr->ReadKeyDataLength(Items);
       offset_type ItemLen = L.first + L.second;
 
       // Compare the hashes.  If they are not the same, skip the entry entirely.



More information about the llvm-commits mailing list