[OpenMPOpt] deduplicateRuntimeCalls(): avoid traditional map lookup pitfall
authorRoman Lebedev <lebedev.ri@gmail.com>
Thu, 16 Apr 2020 16:53:17 +0000 (19:53 +0300)
committerRoman Lebedev <lebedev.ri@gmail.com>
Thu, 16 Apr 2020 16:54:02 +0000 (19:54 +0300)
Summary:
This roughly halves time spent in that pass,
while unsurprisingly significantly reducing total memory usage.

This makes sense because most functions won't use any openmp functions..

old
```
   0.2329 (  0.5%)   0.0409 (  0.9%)   0.2738 (  0.5%)   0.2736 (  0.5%)  OpenMP specific optimizations
```
```
total runtime: 63.32s.
bytes allocated in total (ignoring deallocations): 8.34GB (131.70MB/s)
calls to allocation functions: 14526259 (229410/s)
temporary memory allocations: 3335760 (52680/s)
peak heap memory consumption: 324.36MB
peak RSS (including heaptrack overhead): 5.39GB
total memory leaked: 289.93MB
```

new
```
   0.1457 (  0.3%)   0.0276 (  0.6%)   0.1732 (  0.3%)   0.1731 (  0.3%)  OpenMP specific optimizations
```
```
total runtime: 55.01s.
bytes allocated in total (ignoring deallocations): 6.70GB (121.89MB/s)
calls to allocation functions: 14268205 (259398/s)
temporary memory allocations: 3225355 (58637/s)
peak heap memory consumption: 324.09MB
peak RSS (including heaptrack overhead): 5.39GB
total memory leaked: 289.87MB
```

diff
```
total runtime: -8.31s.
bytes allocated in total (ignoring deallocations): -1.63GB (196.58MB/s)
calls to allocation functions: -258054 (31034/s)
temporary memory allocations: -110405 (13277/s)
peak heap memory consumption: -262.36KB
peak RSS (including heaptrack overhead): 0B
total memory leaked: -61.45KB
```

Reviewers: jdoerfert, hfinkel

Reviewed By: jdoerfert

Subscribers: yaxunl, hiraditya, guansong, llvm-commits

Tags: #llvm

Differential Revision: https://reviews.llvm.org/D78299

llvm/lib/Transforms/IPO/OpenMPOpt.cpp

index 8b45350..d0e0dde 100644 (file)
@@ -248,7 +248,11 @@ private:
   /// \p ReplVal if given.
   bool deduplicateRuntimeCalls(Function &F, RuntimeFunctionInfo &RFI,
                                Value *ReplVal = nullptr) {
-    auto &Uses = RFI.UsesMap[&F];
+    auto UsesIt = RFI.UsesMap.find(&F);
+    if (UsesIt == RFI.UsesMap.end())
+      return false;
+
+    auto &Uses = UsesIt->getSecond();
     if (Uses.size() + (ReplVal != nullptr) < 2)
       return false;