[PATCH] D16832: Minor performance tweaks to llvm-tblgen (and a few that might be a good idea)

Alexander Riccio via llvm-commits llvm-commits at lists.llvm.org
Tue Feb 2 17:35:22 PST 2016

ariccio created this revision.
ariccio added reviewers: rnk, qcolombet, craig.topper.
ariccio added a subscriber: llvm-commits.

This patch adds a reserve call to an expensive function (`llvm::LoadIntrinsics`), and may fix a few other low hanging performance fruit (I've put them in comments for now, so we can discuss).


As I'm sure other developers do, when I build LLVM, I build the entire project with the same config (`Debug`, `MinSizeRel`, `Release`, or `RelWithDebInfo`). However, the `Debug` config also builds llvm-tblgen in `Debug` mode. Later build steps that run llvm-tblgen then can actually be the slowest steps in the entire build. Nobody likes slow builds.



Index: C:/LLVM/llvm/utils/TableGen/CodeGenInstruction.cpp
--- C:/LLVM/llvm/utils/TableGen/CodeGenInstruction.cpp
+++ C:/LLVM/llvm/utils/TableGen/CodeGenInstruction.cpp
@@ -49,6 +49,11 @@
   unsigned MIOperandNo = 0;
   std::set<std::string> OperandNames;
+  // TODO: can we reserve space for e elements in OperandList?
+  // Profiling a debug build with ETW revealed that the emplace_back
+  // at the very bottom of this loop can take up to 227ms, largely
+  // spent in reallocation.
   for (unsigned i = 0, e = InDI->getNumArgs()+OutDI->getNumArgs(); i != e; ++i){
     Init *ArgInit;
     std::string ArgName;
Index: C:/LLVM/llvm/include/llvm/TableGen/Record.h
--- C:/LLVM/llvm/include/llvm/TableGen/Record.h
+++ C:/LLVM/llvm/include/llvm/TableGen/Record.h
@@ -1307,9 +1307,12 @@
   bool isSubClassOf(StringRef Name) const {
-    for (const auto &SCPair : SuperClasses)
+    for (const auto &SCPair : SuperClasses) {
+      // TODO: getNameInitAsString copy constructs a new std::string,
+      // yet we're only using it in this comparison. Is there a better way?
       if (SCPair.first->getNameInitAsString() == Name)
         return true;
+    }
     return false;
Index: C:/LLVM/llvm/utils/TableGen/CodeGenTarget.cpp
--- C:/LLVM/llvm/utils/TableGen/CodeGenTarget.cpp
+++ C:/LLVM/llvm/utils/TableGen/CodeGenTarget.cpp
@@ -441,12 +441,24 @@
   std::vector<Record*> I = RC.getAllDerivedDefinitions("Intrinsic");
   std::vector<CodeGenIntrinsic> Result;
+  // Allocate enough memory to hold enough for worst case scenario
+  // in the following loop. Profiling a debug build with ETW
+  // revealed that push_back in the loop could take up to 280ms,
+  // largely spent in reallocation.
+  Result.reserve(I.size());
   for (unsigned i = 0, e = I.size(); i != e; ++i) {
     bool isTarget = I[i]->getValueAsBit("isTarget");
     if (isTarget == TargetOnly)
+  // TODO: does this lambda need to accept parameters by value?
+  // Profiling a debug build with ETW revealed that each call of
+  // std::_Unguarded_partition (an internal function, called by std::sort),
+  // can spend hundreds of milliseconds inside this lambda calling
+  // ~CodeGenIntrinsic(), which is quite wasteful.
   std::sort(Result.begin(), Result.end(),
             [](CodeGenIntrinsic LHS, CodeGenIntrinsic RHS) {
               return LHS.Name < RHS.Name;

-------------- next part --------------
A non-text attachment was scrubbed...
Name: D16832.46726.patch
Type: text/x-patch
Size: 2623 bytes
Desc: not available
URL: <http://lists.llvm.org/pipermail/llvm-commits/attachments/20160203/f45a57a5/attachment.bin>

More information about the llvm-commits mailing list