[CodeGenPGO] Ensure that PGO correctly instruments cilk_for loops with atomic instrumentation. Fixes llvm#30

Author: neboat

clang/lib/CodeGen/CodeGenPGO.cpp

@@ -126,6 +126,9 @@ class PGOHash {
     BinaryOperatorNE,
     // The preceding values are available since PGO_HASH_V2.
 
+    // Cilk statements.  These values are also available with PGO_HASH_V1.
+    CilkForStmt,
+
     // Keep this last.  It's for the static assert that follows.
     LastHashType
   };
@@ -267,6 +270,7 @@ struct MapRegionCounters : public RecursiveASTVisitor<MapRegionCounters> {
   DEFINE_NESTABLE_TRAVERSAL(ObjCForCollectionStmt)
   DEFINE_NESTABLE_TRAVERSAL(CXXTryStmt)
   DEFINE_NESTABLE_TRAVERSAL(CXXCatchStmt)
+  DEFINE_NESTABLE_TRAVERSAL(CilkForStmt)
 
   /// Get version \p HashVersion of the PGO hash for \p S.
   PGOHash::HashType getHashType(PGOHashVersion HashVersion, const Stmt *S) {
@@ -327,6 +331,8 @@ struct MapRegionCounters : public RecursiveASTVisitor<MapRegionCounters> {
       }
       break;
     }
+    case Stmt::CilkForStmtClass:
+      return PGOHash::CilkForStmt;
     }
 
     if (HashVersion >= PGO_HASH_V2) {
@@ -744,6 +750,53 @@ struct ComputeRegionCounts : public ConstStmtVisitor<ComputeRegionCounts> {
     setCount(ParentCount + RHSCount - CurrentCount);
     RecordNextStmtCount = true;
   }
+
+  void VisitCilkForStmt(const CilkForStmt *S) {
+    RecordStmtCount(S);
+    if (S->getInit())
+      Visit(S->getInit());
+    if (S->getLimitStmt())
+      Visit(S->getLimitStmt());
+    if (S->getBeginStmt())
+      Visit(S->getBeginStmt());
+    if (S->getEndStmt())
+      Visit(S->getEndStmt());
+    if (S->getLoopVarDecl())
+      Visit(S->getLoopVarDecl());
+
+    uint64_t ParentCount = CurrentCount;
+
+    BreakContinueStack.push_back(BreakContinue());
+    // Visit the body region first. (This is basically the same as a while
+    // loop; see further comments in VisitWhileStmt.)
+    uint64_t BodyCount = setCount(PGO.getRegionCount(S));
+    CountMap[S->getBody()] = BodyCount;
+    Visit(S->getBody());
+    uint64_t BackedgeCount = CurrentCount;
+    BreakContinue BC = BreakContinueStack.pop_back_val();
+
+    // The increment is essentially part of the body but it needs to include
+    // the count for all the continue statements.
+    if (S->getInc()) {
+      uint64_t IncCount = setCount(BackedgeCount + BC.ContinueCount);
+      CountMap[S->getInc()] = IncCount;
+      Visit(S->getInc());
+    }
+
+    // ...then go back and propagate counts through the condition.
+    uint64_t CondCount =
+        setCount(ParentCount + BackedgeCount + BC.ContinueCount);
+    if (S->getInitCond()) {
+      CountMap[S->getInitCond()] = ParentCount;
+      Visit(S->getInitCond());
+    }
+    if (S->getCond()) {
+      CountMap[S->getCond()] = CondCount;
+      Visit(S->getCond());
+    }
+    setCount(BC.BreakCount + CondCount - BodyCount);
+    RecordNextStmtCount = true;
+  }
 };
 } // end anonymous namespace
 

clang/test/Cilk/cilkfor-pgo.cpp

@@ -0,0 +1,24 @@
+// Check that -fprofile-instrument generates atomic
+// instrumentation instructions inside of _Cilk_for loops.
+//
+// Credit to Brian Wheatman for the original source of this test.
+//
+// RUN: %clang_cc1 -triple x86_64-unknown-linux-gnu -fprofile-instrument=clang -fprofile-update=atomic %s -S -emit-llvm -fopencilk -ftapir=none -o - 2>&1 | FileCheck %s
+// expected-no-diagnostics
+
+int main() {
+  int sum = 0;
+  _Cilk_for(int i = 0; i < 1000000; i++) { sum += i; }
+
+  return sum;
+}
+
+// CHECK: @__profc_main = {{.*}}global [2 x i64] zeroinitializer, section "__llvm_prf_cnts"
+
+// CHECK-LABEL: define {{.*}}i32 @main()
+
+// CHECK: detach within %{{.+}}, label %[[PFOR_BODY:.+]], label %[[PFOR_INC:.+]]
+
+// CHECK: [[PFOR_BODY]]:
+// CHECK: atomicrmw add i64* getelementptr inbounds ([2 x i64], [2 x i64]* @__profc_main, i64 0, i64 1), i64 1 monotonic
+// CHECK: reattach within %{{.+}}, label %[[PFOR_INC]]