[llvm] [memprof] Create a call stack table (PR #86575)

[Kazu Hirata via llvm-commits]

https://github.com/kazutakahirata created https://github.com/llvm/llvm-project/pull/86575

We are in the process of reducing the size of the indexed memprof file
by creating a hash table of call stacks so that IndexedAllocationInfo
can refer to a call stack with the hash of the call stack.  The idea
is analogus to IdToFrame, a hash table to map a FrameId to its
corresponding Frame.

This patch prepares for that future by keeping call stacks in a hash
table so that we can drop CallStack from IndexedAllocationInfo.

We will eventually serialize the hash table into an indexed profile
file just as we serialize MemProfFrameData, a hash table of Frames
indexed by FrameId.

This patch does not add a use of the hash table yet.


>From f90660ceb69a673185e8f8381ed9f33b61678d0f Mon Sep 17 00:00:00 2001
From: Kazu Hirata <kazu at google.com>
Date: Mon, 25 Mar 2024 12:41:38 -0700
Subject: [PATCH] [memprof] Create a call stack table

We are in the process of reducing the size of the indexed memprof file
by creating a hash table of call stacks so that IndexedAllocationInfo
can refer to a call stack with the hash of the call stack.  The idea
is analogus to IdToFrame, a hash table to map a FrameId to its
corresponding Frame.

This patch prepares for that future by keeping call stacks in a hash
table so that we can drop CallStack from IndexedAllocationInfo.

We will eventually serialize the hash table into an indexed profile
file just as we serialize MemProfFrameData, a hash table of Frames
indexed by FrameId.

This patch does not add a use of the hash table yet.
---
 llvm/include/llvm/ProfileData/RawMemProfReader.h | 2 ++
 llvm/lib/ProfileData/RawMemProfReader.cpp        | 1 +
 2 files changed, 3 insertions(+)

diff --git a/llvm/include/llvm/ProfileData/RawMemProfReader.h b/llvm/include/llvm/ProfileData/RawMemProfReader.h
index 6aa5caec65f791..8b8c5c9f242369 100644
--- a/llvm/include/llvm/ProfileData/RawMemProfReader.h
+++ b/llvm/include/llvm/ProfileData/RawMemProfReader.h
@@ -97,6 +97,8 @@ class MemProfReader {
   }
   // A mapping from FrameId (a hash of the contents) to the frame.
   llvm::DenseMap<FrameId, Frame> IdToFrame;
+  // A mapping from CallStackId (a hash of the contents) to the call stack.
+  llvm::DenseMap<CallStackId, SmallVector<FrameId>> CallStackIdToCallStack;
   // A mapping from function GUID, hash of the canonical function symbol to the
   // memprof profile data for that function, i.e allocation and callsite info.
   llvm::MapVector<GlobalValue::GUID, IndexedMemProfRecord> FunctionProfileData;
diff --git a/llvm/lib/ProfileData/RawMemProfReader.cpp b/llvm/lib/ProfileData/RawMemProfReader.cpp
index 5dc1ff8978154c..6e294849170b2f 100644
--- a/llvm/lib/ProfileData/RawMemProfReader.cpp
+++ b/llvm/lib/ProfileData/RawMemProfReader.cpp
@@ -447,6 +447,7 @@ Error RawMemProfReader::mapRawProfileToRecords() {
     }
 
     CallStackId CSId = hashCallStack(Callstack);
+    CallStackIdToCallStack.insert({CSId, Callstack});
 
     // We attach the memprof record to each function bottom-up including the
     // first non-inline frame.