diff --git a/clang/lib/CodeGen/CGStmt.cpp b/clang/lib/CodeGen/CGStmt.cpp
index 7dd82d73d6b6..e46cfb433ab8 100644
--- a/clang/lib/CodeGen/CGStmt.cpp
+++ b/clang/lib/CodeGen/CGStmt.cpp
@@ -1242,9 +1242,17 @@ void CodeGenFunction::EmitDoStmt(const DoStmt &S,
   // As long as the condition is true, iterate the loop.
   if (EmitBoolCondBranch) {
     uint64_t BackedgeCount = getProfileCount(S.getBody()) - ParentCount;
-    Builder.CreateCondBr(
+    auto *I = Builder.CreateCondBr(
         BoolCondVal, LoopBody, LoopExit.getBlock(),
         createProfileWeightsForLoop(S.getCond(), BackedgeCount));
+
+    // Key Instructions: Emit the condition and branch as separate atoms to
+    // match existing loop stepping behaviour. FIXME: We could have the branch
+    // as the backup location for the condition, which would probably be a
+    // better experience (no jumping to the brace).
+    if (auto *I = dyn_cast<llvm::Instruction>(BoolCondVal))
+      addInstToNewSourceAtom(I, nullptr);
+    addInstToNewSourceAtom(I, nullptr);
   }
 
   LoopStack.pop();
diff --git a/clang/test/KeyInstructions/do.c b/clang/test/KeyInstructions/do.c
new file mode 100644
index 000000000000..4b7a38cb3598
--- /dev/null
+++ b/clang/test/KeyInstructions/do.c
@@ -0,0 +1,33 @@
+// RUN: %clang -gkey-instructions -x c++ -std=c++17 %s -gmlt -S -emit-llvm -o - \
+// RUN: | FileCheck %s --implicit-check-not atomGroup --implicit-check-not atomRank
+
+// RUN: %clang -gkey-instructions -x c %s -gmlt -S -emit-llvm -o -  \
+// RUN: | FileCheck %s --implicit-check-not atomGroup --implicit-check-not atomRank
+
+// Perennial quesiton: should the `dec` be in its own source atom or not
+// (currently it is).
+
+// Another question - we've made the cmp and br separate source atoms for
+// now, to match existing behaviour in this case:
+// 1. do {
+// 2.   something();
+// 3. }
+// 4. while (--A);
+// Non key instruction behaviour is: 2, 4[, 3, 2, 4]+
+// The cond br is associated with the brace on line 3 and the cmp is line 4;
+// if they were in the same atom group we'd step just: 2, 3[, 2, 3]+
+// FIXME: We could arguably improve the behaviour by making them the same
+// group but having the cmp higher precedence, resulting in: 2, 4[, 2, 4]+.
+
+void a(int A) {
+// CHECK: %dec = add nsw i32 %0, -1, !dbg [[G1R2:!.*]]
+// CHECK: store i32 %dec, ptr %A.addr{{.*}}, !dbg [[G1R1:!.*]]
+// CHECK: %tobool = icmp ne i32 %dec, 0, !dbg [[G2R1:!.*]]
+// CHECK: br i1 %tobool, label %do.body, label %do.end, !dbg [[G3R1:!.*]], !llvm.loop
+    do { } while (--A);
+}
+
+// CHECK: [[G1R2]] = !DILocation({{.*}}, atomGroup: 1, atomRank: 2)
+// CHECK: [[G1R1]] = !DILocation({{.*}}, atomGroup: 1, atomRank: 1)
+// CHECK: [[G2R1]] = !DILocation({{.*}}, atomGroup: 2, atomRank: 1)
+// CHECK: [[G3R1]] = !DILocation({{.*}}, atomGroup: 3, atomRank: 1)