Index: lib/CodeGen/BranchFolding.cpp
===================================================================
--- lib/CodeGen/BranchFolding.cpp
+++ lib/CodeGen/BranchFolding.cpp
@@ -2015,11 +2015,6 @@
           IsSafe = false;
           break;
         }
-
-        if (MO.isKill() && Uses.count(Reg))
-          // Kills a register that's read by the instruction at the point of
-          // insertion. Remove the kill marker.
-          MO.setIsKill(false);
       }
     }
     if (!IsSafe)
@@ -2030,12 +2025,16 @@
       break;
 
     // Remove kills from ActiveDefsSet, these registers had short live ranges.
-    for (const MachineOperand &MO : TIB->operands()) {
+    for (MachineOperand &MO : TIB->operands()) {
       if (!MO.isReg() || !MO.isUse() || !MO.isKill())
         continue;
       unsigned Reg = MO.getReg();
       if (!Reg)
         continue;
+      if (Uses.count(Reg))
+        // Kills a register that's read by the instruction at the point of
+        // insertion. Remove the kill marker.
+        MO.setIsKill(false);
       if (!AllDefsSet.count(Reg)) {
         LocalKills.push_back(Reg);
         continue;
Index: test/CodeGen/SystemZ/branch-folder-hoist.mir
===================================================================
--- /dev/null
+++ test/CodeGen/SystemZ/branch-folder-hoist.mir
@@ -0,0 +1,66 @@
+# RUN: llc -verify-machineinstrs -mtriple=s390x-ibm-linux -o - %s -run-pass=branch-folder | FileCheck %s
+
+--- |
+  target datalayout = "E-m:e-i1:8:16-i8:8:16-i64:64-f128:64-a:8:16-n32:64"
+  target triple = "s390x-ibm-linux"
+
+  @d = external dso_local global i64, align 1
+  @o = external dso_local local_unnamed_addr global i16, align 1
+  @i = external dso_local local_unnamed_addr global i16, align 1
+  @g = external dso_local local_unnamed_addr global i32, align 1
+
+  define dso_local void @f3() local_unnamed_addr {
+  entry:
+    ret void
+  }
+
+...
+---
+name:            f3
+tracksRegLiveness: true
+body:             |
+  bb.0:
+    successors: %bb.2(0x30000000), %bb.3(0x50000000)
+    liveins: $r1d, $r2l
+
+    renamable $r3d = LGRL @d :: (load 4 from got, align 8)
+    renamable $r0d = LGF renamable $r1d, 0, $noreg :: (dereferenceable load 2 from @g, align 1)
+    dead renamable $r3d = LG killed renamable $r3d, 0, $noreg :: (volatile dereferenceable load 4 from @d, align 1)
+    TMLL renamable $r0l, 65535, implicit-def $cc
+    BRC 15, 7, %bb.3, implicit killed $cc
+    J %bb.2
+
+  bb.2:
+    successors: %bb.4(0x80000000)
+    liveins: $r0d, $r1d, $r2l
+
+    $r0h = COPY $r0l
+    $r3d = COPY killed $r0d
+    renamable $r3d = LGRL @o :: (load 4 from got, align 8)
+    MVHHI killed renamable $r3d, 0, 0 :: (store 1 into @o)
+    J %bb.4
+
+  bb.3:
+    successors: %bb.4(0x80000000)
+    liveins: $r0d, $r1d, $r2l
+
+    $r0h = COPY $r0l
+    $r3d = COPY killed $r0d
+    renamable $r3d = LGRL @i :: (load 4 from got, align 8)
+    MVHHI killed renamable $r3d, 0, 0 :: (store 1 into @i)
+
+  bb.4:
+    liveins: $r1d, $r2l
+
+    CHI killed renamable $r2l, 0, implicit-def $cc
+    renamable $r2d = LGHI 5
+    Return
+...
+
+# Note: the second COPY should not have killed flag on $r0d as $r0l is used in
+# the TMLL instruction
+
+# CHECK: bb.0:
+# CHECK: $r0h = COPY $r0l
+# CHECK-NEXT: $r3d = COPY $r0d
+# CHECK-NEXT: TMLL renamable $r0l, 65535, implicit-def $cc