diff --git a/llvm/include/llvm/Transforms/Utils/GlobalStatus.h b/llvm/include/llvm/Transforms/Utils/GlobalStatus.h
--- a/llvm/include/llvm/Transforms/Utils/GlobalStatus.h
+++ b/llvm/include/llvm/Transforms/Utils/GlobalStatus.h
@@ -35,6 +35,9 @@
   /// can be deleted.
   bool IsLoaded = false;
 
+  /// Number of stores to the global.
+  unsigned NumStores = 0;
+
   /// Keep track of what stores to the global look like.
   enum StoredType {
     /// There is no store to this global.  It can thus be marked constant.
diff --git a/llvm/lib/Transforms/IPO/GlobalOpt.cpp b/llvm/lib/Transforms/IPO/GlobalOpt.cpp
--- a/llvm/lib/Transforms/IPO/GlobalOpt.cpp
+++ b/llvm/lib/Transforms/IPO/GlobalOpt.cpp
@@ -1444,6 +1444,42 @@
   }
 }
 
+// For a global variable with one store, if the store dominates any loads,
+// those loads will always load the stored value (as opposed to the
+// initializer), even in the presence of recursion.
+static bool forwardStoredOnceStore(
+    GlobalVariable *GV, const StoreInst *StoredOnceStore,
+    function_ref<DominatorTree &(Function &)> LookupDomTree) {
+  const Value *StoredOnceValue = StoredOnceStore->getValueOperand();
+  // We can do this optimization for non-constants in nosync + norecurse
+  // functions, but globals used in exactly one norecurse functions are already
+  // promoted to an alloca.
+  if (!isa<Constant>(StoredOnceValue))
+    return false;
+  const Function *F = StoredOnceStore->getFunction();
+  SmallVector<LoadInst *> Loads;
+  for (User *U : GV->users()) {
+    if (auto *LI = dyn_cast<LoadInst>(U)) {
+      if (LI->getFunction() == F &&
+          LI->getType() == StoredOnceValue->getType() && LI->isSimple())
+        Loads.push_back(LI);
+    }
+  }
+  // Only compute DT if we have any loads to examine.
+  bool MadeChange = false;
+  if (!Loads.empty()) {
+    auto &DT = LookupDomTree(*const_cast<Function *>(F));
+    for (auto *LI : Loads) {
+      if (DT.dominates(StoredOnceStore, LI)) {
+        LI->replaceAllUsesWith(const_cast<Value *>(StoredOnceValue));
+        LI->eraseFromParent();
+        MadeChange = true;
+      }
+    }
+  }
+  return MadeChange;
+}
+
 /// Analyze the specified global variable and optimize
 /// it if possible.  If we make a change, return true.
 static bool
@@ -1603,6 +1639,12 @@
     if (optimizeOnceStoredGlobal(GV, StoredOnceValue, DL, GetTLI))
       return true;
 
+    // Try to forward the store to any loads. If we have more than one store, we
+    // may have a store of the initializer between StoredOnceStore and a load.
+    if (GS.NumStores == 1)
+      if (forwardStoredOnceStore(GV, GS.StoredOnceStore, LookupDomTree))
+        return true;
+
     // Otherwise, if the global was not a boolean, we can shrink it to be a
     // boolean. Skip this optimization for AS that doesn't allow an initializer.
     if (SOVConstant && GS.Ordering == AtomicOrdering::NotAtomic &&
diff --git a/llvm/lib/Transforms/Utils/GlobalStatus.cpp b/llvm/lib/Transforms/Utils/GlobalStatus.cpp
--- a/llvm/lib/Transforms/Utils/GlobalStatus.cpp
+++ b/llvm/lib/Transforms/Utils/GlobalStatus.cpp
@@ -104,6 +104,8 @@
         if (SI->isVolatile())
           return true;
 
+        ++GS.NumStores;
+
         GS.Ordering = strongerOrdering(GS.Ordering, SI->getOrdering());
 
         // If this is a direct store to the global (i.e., the global is a scalar
diff --git a/llvm/test/Transforms/GlobalOpt/shrink-global-to-bool-check-debug.ll b/llvm/test/Transforms/GlobalOpt/shrink-global-to-bool-check-debug.ll
--- a/llvm/test/Transforms/GlobalOpt/shrink-global-to-bool-check-debug.ll
+++ b/llvm/test/Transforms/GlobalOpt/shrink-global-to-bool-check-debug.ll
@@ -2,21 +2,24 @@
 
 @foo = internal global i32 0, align 4
 
-define dso_local i32 @bar() {
+define void @store() {
 entry:
   store i32 5, i32* @foo, align 4
+  ret void
+}
+
+define i32 @bar() {
+entry:
   %0 = load i32, i32* @foo, align 4
   ret i32 %0
 }
 
 ;CHECK:      @bar
 ;CHECK-NEXT: entry:
-;CHECK-NEXT:   store i1 true, i1* @foo, align 1, !dbg ![[DbgLocStore:[0-9]+]]
 ;CHECK-NEXT:   %.b = load i1, i1* @foo, align 1, !dbg ![[DbgLocLoadSel:[0-9]+]]
 ;CHECK-NEXT:   %0 = select i1 %.b, i32 5, i32 0, !dbg ![[DbgLocLoadSel]]
 ;CHECK-NEXT:   call void @llvm.dbg.value({{.*}}), !dbg ![[DbgLocLoadSel]]
 ;CHECK-NEXT:   ret i32 %0, !dbg ![[DbgLocRet:[0-9]+]]
 
-;CHECK: ![[DbgLocStore]] = !DILocation(line: 1,
-;CHECK: ![[DbgLocLoadSel]] = !DILocation(line: 2,
-;CHECK: ![[DbgLocRet]] = !DILocation(line: 3,
+;CHECK: ![[DbgLocLoadSel]] = !DILocation(line: 3,
+;CHECK: ![[DbgLocRet]] = !DILocation(line: 4,
diff --git a/llvm/test/Transforms/GlobalOpt/shrink-global-to-bool.ll b/llvm/test/Transforms/GlobalOpt/shrink-global-to-bool.ll
--- a/llvm/test/Transforms/GlobalOpt/shrink-global-to-bool.ll
+++ b/llvm/test/Transforms/GlobalOpt/shrink-global-to-bool.ll
@@ -10,11 +10,13 @@
 ; Negative test for AS(3). Skip shrink global to bool optimization.
 ; CHECK: @lvar = internal unnamed_addr addrspace(3) global i32 undef
 
-define void @test_global_var() {
+define void @test_global_var(i1 %i) {
 ; CHECK-LABEL: @test_global_var(
 ; CHECK:    store volatile i32 10, i32* undef, align 4
 ;
 entry:
+  br i1 %i, label %bb1, label %exit
+bb1:
   store i32 10, i32* @gvar
   br label %exit
 exit:
@@ -23,13 +25,15 @@
   ret void
 }
 
-define void @test_lds_var() {
+define void @test_lds_var(i1 %i) {
 ; CHECK-LABEL: @test_lds_var(
 ; CHECK:    store i32 10, i32 addrspace(3)* @lvar, align 4
 ; CHECK:    [[LD:%.*]] = load i32, i32 addrspace(3)* @lvar, align 4
 ; CHECK:    store volatile i32 [[LD]], i32* undef, align 4
 ;
 entry:
+  br i1 %i, label %bb1, label %exit
+bb1:
   store i32 10, i32 addrspace(3)* @lvar
   br label %exit
 exit:
diff --git a/llvm/test/Transforms/GlobalOpt/stored-once-forward-value.ll b/llvm/test/Transforms/GlobalOpt/stored-once-forward-value.ll
--- a/llvm/test/Transforms/GlobalOpt/stored-once-forward-value.ll
+++ b/llvm/test/Transforms/GlobalOpt/stored-once-forward-value.ll
@@ -7,15 +7,16 @@
 @g4 = internal unnamed_addr global i32 0
 @g5 = internal unnamed_addr global i32 0
 @g6 = internal unnamed_addr global i32 0
+@g7 = internal unnamed_addr global i32 0
+@g8 = internal unnamed_addr global ptr null
+@tl = internal thread_local unnamed_addr global i32 0
 
 declare void @b()
 
 define i1 @dom_const() {
 ; CHECK-LABEL: @dom_const(
-; CHECK-NEXT:    store i1 true, ptr @g1, align 1
 ; CHECK-NEXT:    call void @b()
-; CHECK-NEXT:    [[R:%.*]] = load i1, ptr @g1, align 1
-; CHECK-NEXT:    ret i1 [[R]]
+; CHECK-NEXT:    ret i1 true
 ;
   store i1 true, ptr @g1
   call void @b()
@@ -36,16 +37,29 @@
   ret i32 %r
 }
 
+define ptr @dom_thread_local_global() {
+; CHECK-LABEL: @dom_thread_local_global(
+; CHECK-NEXT:    store ptr @tl, ptr @g3, align 8
+; CHECK-NEXT:    call void @b()
+; CHECK-NEXT:    [[R:%.*]] = load ptr, ptr @g3, align 8
+; CHECK-NEXT:    ret ptr [[R]]
+;
+  store ptr @tl, ptr @g3
+  call void @b()
+  %r = load ptr, ptr @g3
+  ret ptr %r
+}
+
 define i32 @dom_different_types() {
 ; CHECK-LABEL: @dom_different_types(
-; CHECK-NEXT:    store i1 true, ptr @g3, align 1
+; CHECK-NEXT:    store i1 true, ptr @g4, align 1
 ; CHECK-NEXT:    call void @b()
-; CHECK-NEXT:    [[R:%.*]] = load i32, ptr @g3, align 4
+; CHECK-NEXT:    [[R:%.*]] = load i32, ptr @g4, align 4
 ; CHECK-NEXT:    ret i32 [[R]]
 ;
-  store i1 true, ptr @g3
+  store i1 true, ptr @g4
   call void @b()
-  %r = load i32, ptr @g3
+  %r = load i32, ptr @g4
   ret i32 %r
 }
 
@@ -53,57 +67,70 @@
 ; CHECK-LABEL: @no_dom(
 ; CHECK-NEXT:    br i1 [[I:%.*]], label [[BB1:%.*]], label [[END:%.*]]
 ; CHECK:       bb1:
-; CHECK-NEXT:    store i1 true, ptr @g4, align 1
+; CHECK-NEXT:    store i1 true, ptr @g5, align 1
 ; CHECK-NEXT:    br label [[END]]
 ; CHECK:       end:
 ; CHECK-NEXT:    call void @b()
-; CHECK-NEXT:    [[R:%.*]] = load i1, ptr @g4, align 1
+; CHECK-NEXT:    [[R:%.*]] = load i1, ptr @g5, align 1
 ; CHECK-NEXT:    ret i1 [[R]]
 ;
   br i1 %i, label %bb1, label %end
 bb1:
-  store i1 true, ptr @g4
+  store i1 true, ptr @g5
   br label %end
 end:
   call void @b()
-  %r = load i1, ptr @g4
+  %r = load i1, ptr @g5
   ret i1 %r
 }
 
 define i1 @dom_multiple_function_loads() {
 ; CHECK-LABEL: @dom_multiple_function_loads(
-; CHECK-NEXT:    store i1 true, ptr @g5, align 1
+; CHECK-NEXT:    store i1 true, ptr @g6, align 1
 ; CHECK-NEXT:    call void @b()
-; CHECK-NEXT:    [[R:%.*]] = load i1, ptr @g5, align 1
-; CHECK-NEXT:    ret i1 [[R]]
+; CHECK-NEXT:    ret i1 true
 ;
-  store i1 true, ptr @g5
+  store i1 true, ptr @g6
   call void @b()
-  %r = load i1, ptr @g5
+  %r = load i1, ptr @g6
   ret i1 %r
 }
 
 define i1 @other() {
 ; CHECK-LABEL: @other(
 ; CHECK-NEXT:    call void @b()
-; CHECK-NEXT:    [[R:%.*]] = load i1, ptr @g5, align 1
+; CHECK-NEXT:    [[R:%.*]] = load i1, ptr @g6, align 1
 ; CHECK-NEXT:    ret i1 [[R]]
 ;
   call void @b()
-  %r = load i1, ptr @g5
+  %r = load i1, ptr @g6
   ret i1 %r
 }
 
 define i1 @dom_volatile() {
 ; CHECK-LABEL: @dom_volatile(
-; CHECK-NEXT:    store i1 true, ptr @g6, align 1
+; CHECK-NEXT:    store i1 true, ptr @g7, align 1
 ; CHECK-NEXT:    call void @b()
-; CHECK-NEXT:    [[R:%.*]] = load volatile i1, ptr @g6, align 1
+; CHECK-NEXT:    [[R:%.*]] = load volatile i1, ptr @g7, align 1
 ; CHECK-NEXT:    ret i1 [[R]]
 ;
-  store i1 true, ptr @g6
+  store i1 true, ptr @g7
   call void @b()
-  %r = load volatile i1, ptr @g6
+  %r = load volatile i1, ptr @g7
   ret i1 %r
 }
 
+define i1 @dom_store_const_and_initializer() {
+; CHECK-LABEL: @dom_store_const_and_initializer(
+; CHECK-NEXT:    store i1 true, ptr @g8, align 1
+; CHECK-NEXT:    store i1 false, ptr @g8, align 1
+; CHECK-NEXT:    call void @b()
+; CHECK-NEXT:    [[R:%.*]] = load i1, ptr @g8, align 1
+; CHECK-NEXT:    ret i1 [[R]]
+;
+  store i1 true, ptr @g8
+  store i1 false, ptr @g8
+  call void @b()
+  %r = load i1, ptr @g8
+  ret i1 %r
+}
diff --git a/llvm/test/Transforms/PhaseOrdering/recompute-globalsaa.ll b/llvm/test/Transforms/PhaseOrdering/recompute-globalsaa.ll
--- a/llvm/test/Transforms/PhaseOrdering/recompute-globalsaa.ll
+++ b/llvm/test/Transforms/PhaseOrdering/recompute-globalsaa.ll
@@ -9,7 +9,6 @@
 define i32 @main() {
 ; CHECK-LABEL: @main(
 ; CHECK-NEXT:  entry:
-; CHECK-NEXT:    store i1 true, i1* @a, align 4
 ; CHECK-NEXT:    [[TMP0:%.*]] = load i32*, i32** @e, align 8
 ; CHECK-NEXT:    store i32 0, i32* [[TMP0]], align 4
 ; CHECK-NEXT:    store i32* null, i32** @e, align 8