diff --git a/llvm/lib/CodeGen/TypePromotion.cpp b/llvm/lib/CodeGen/TypePromotion.cpp
--- a/llvm/lib/CodeGen/TypePromotion.cpp
+++ b/llvm/lib/CodeGen/TypePromotion.cpp
@@ -102,7 +102,6 @@
 namespace {
 class IRPromoter {
   LLVMContext &Ctx;
-  IntegerType *OrigTy = nullptr;
   unsigned PromotedWidth = 0;
   SetVector<Value *> &Visited;
   SetVector<Value *> &Sources;
@@ -122,16 +121,13 @@
   void Cleanup();
 
 public:
-  IRPromoter(LLVMContext &C, IntegerType *Ty, unsigned Width,
+  IRPromoter(LLVMContext &C, unsigned Width,
              SetVector<Value *> &visited, SetVector<Value *> &sources,
              SetVector<Instruction *> &sinks,
              SmallPtrSetImpl<Instruction *> &wrap)
-      : Ctx(C), OrigTy(Ty), PromotedWidth(Width), Visited(visited),
+      : Ctx(C), PromotedWidth(Width), Visited(visited),
         Sources(sources), Sinks(sinks), SafeWrap(wrap) {
     ExtTy = IntegerType::get(Ctx, PromotedWidth);
-    assert(OrigTy->getPrimitiveSizeInBits().getFixedSize() <
-               ExtTy->getPrimitiveSizeInBits().getFixedSize() &&
-           "Original type not smaller than extended type");
   }
 
   void Mutate();
@@ -244,7 +240,7 @@
 bool TypePromotion::isSink(Value *V) {
   // TODO The truncate also isn't actually necessary because we would already
   // proved that the data value is kept within the range of the original data
-  // type.
+  // type. We currently remove any truncs inserted for handling zext sinks.
 
   // Sinks are:
   // - points where the value in the register is being observed, such as an
@@ -591,11 +587,9 @@
       continue;
     }
 
-    // Unless they produce a value that is narrower than ExtTy, we can
-    // replace the result of the zext with the input of a newly inserted
-    // trunc.
-    if (NewInsts.count(Src) && isa<TruncInst>(Src) &&
-        Src->getType() == OrigTy) {
+    // We've inserted a trunc for a zext sink, but we already know that the
+    // input is in range, negating the need for the trunc.
+    if (NewInsts.count(Src) && isa<TruncInst>(Src)) {
       auto *Trunc = cast<TruncInst>(Src);
       assert(Trunc->getOperand(0)->getType() == ExtTy &&
              "expected inserted trunc to be operating on i32");
@@ -636,9 +630,8 @@
 }
 
 void IRPromoter::Mutate() {
-  LLVM_DEBUG(dbgs() << "IR Promotion: Promoting use-def chains from "
-                    << OrigTy->getBitWidth() << " to " << PromotedWidth
-                    << "-bits\n");
+  LLVM_DEBUG(dbgs() << "IR Promotion: Promoting use-def chains to "
+                    << PromotedWidth << "-bits\n");
 
   // Cache original types of the values that will likely need truncating
   for (auto *I : Sinks) {
@@ -878,8 +871,8 @@
   if (ToPromote < 2 || (Blocks.size() == 1 && (NonFreeArgs > SafeWrap.size())))
     return false;
 
-  IRPromoter Promoter(*Ctx, cast<IntegerType>(OrigTy), PromotedWidth,
-                      CurrentVisited, Sources, Sinks, SafeWrap);
+  IRPromoter Promoter(*Ctx, PromotedWidth, CurrentVisited, Sources, Sinks,
+                      SafeWrap);
   Promoter.Mutate();
   return true;
 }
diff --git a/llvm/test/Transforms/TypePromotion/ARM/casts.ll b/llvm/test/Transforms/TypePromotion/ARM/casts.ll
--- a/llvm/test/Transforms/TypePromotion/ARM/casts.ll
+++ b/llvm/test/Transforms/TypePromotion/ARM/casts.ll
@@ -996,8 +996,8 @@
   ret i1 %tobool
 }
 
-define i32 @dont_replace_trunc_2(i16* %a, i8* %b) {
-; CHECK-LABEL: @dont_replace_trunc_2(
+define i32 @dont_return_inserted_trunc(i16* %a, i8* %b) {
+; CHECK-LABEL: @dont_return_inserted_trunc(
 ; CHECK-NEXT:  entry:
 ; CHECK-NEXT:    [[TMP0:%.*]] = load i16, i16* [[A:%.*]], align 2
 ; CHECK-NEXT:    [[TMP1:%.*]] = zext i16 [[TMP0]] to i32
@@ -1010,8 +1010,7 @@
 ; CHECK-NEXT:    [[TMP5:%.*]] = trunc i32 [[OR]] to i8
 ; CHECK-NEXT:    store i8 [[TMP5]], i8* [[B]], align 1
 ; CHECK-NEXT:    [[TMP6:%.*]] = trunc i32 [[OR]] to i8
-; CHECK-NEXT:    [[CONV5:%.*]] = zext i8 [[TMP6]] to i32
-; CHECK-NEXT:    ret i32 [[CONV5]]
+; CHECK-NEXT:    ret i32 [[OR]]
 ;
 entry:
   %0 = load i16, i16* %a, align 2
@@ -1037,10 +1036,9 @@
 ; CHECK-NEXT:    [[TMP3:%.*]] = and i32 [[TMP2]], 255
 ; CHECK-NEXT:    [[TMP4:%.*]] = udiv i32 [[TMP3]], 3
 ; CHECK-NEXT:    [[TMP5:%.*]] = trunc i32 [[TMP4]] to i8
-; CHECK-NEXT:    [[PHITMP:%.*]] = zext i8 [[TMP5]] to i32
 ; CHECK-NEXT:    br label [[COND_END]]
 ; CHECK:       cond.end:
-; CHECK-NEXT:    [[COND:%.*]] = phi i32 [ [[PHITMP]], [[COND_FALSE]] ], [ 0, [[ENTRY:%.*]] ]
+; CHECK-NEXT:    [[COND:%.*]] = phi i32 [ [[TMP4]], [[COND_FALSE]] ], [ 0, [[ENTRY:%.*]] ]
 ; CHECK-NEXT:    ret i32 [[COND]]
 ;
 entry: