Index: lib/Target/ARM/ARMCodeGenPrepare.cpp
===================================================================
--- lib/Target/ARM/ARMCodeGenPrepare.cpp
+++ lib/Target/ARM/ARMCodeGenPrepare.cpp
@@ -686,6 +686,7 @@
 }
 
 void IRPromoter::Cleanup() {
+  LLVM_DEBUG(dbgs() << "ARM CGP: Cleanup..\n");
   // Some zexts will now have become redundant, along with their trunc
   // operands, so remove them
   for (auto V : *Visited) {
@@ -729,6 +730,7 @@
 }
 
 void IRPromoter::ConvertTruncs() {
+  LLVM_DEBUG(dbgs() << "ARM CGP: Converting truncs..\n");
   IRBuilder<> Builder{Ctx};
 
   for (auto *V : *Visited) {
@@ -736,12 +738,13 @@
       continue;
 
     auto *Trunc = cast<TruncInst>(V);
-    assert(LessThanTypeSize(Trunc) && "expected narrow trunc");
-
     Builder.SetInsertPoint(Trunc);
-    unsigned NumBits =
-      cast<IntegerType>(Trunc->getType())->getScalarSizeInBits();
-    ConstantInt *Mask = ConstantInt::get(Ctx, APInt::getMaxValue(NumBits));
+    IntegerType *SrcTy = cast<IntegerType>(Trunc->getOperand(0)->getType());
+    IntegerType *DestTy = cast<IntegerType>(TruncTysMap[Trunc][0]);
+
+    unsigned NumBits = DestTy->getScalarSizeInBits();
+    ConstantInt *Mask =
+      ConstantInt::get(SrcTy, APInt::getMaxValue(NumBits).getZExtValue());
     Value *Masked = Builder.CreateAnd(Trunc->getOperand(0), Mask);
 
     if (auto *I = dyn_cast<Instruction>(Masked))
@@ -783,6 +786,12 @@
         TruncTysMap[I].push_back(I->getOperand(i)->getType());
     }
   }
+  for (auto *V : Visited) {
+    if (!isa<TruncInst>(V) || Sources.count(V))
+      continue;
+    auto *Trunc = cast<TruncInst>(V);
+    TruncTysMap[Trunc].push_back(Trunc->getDestTy());
+  }
 
   // Convert adds and subs using negative immediates to equivalent instructions
   // that use positive constants.
@@ -791,14 +800,14 @@
   // Insert zext instructions between sources and their users.
   ExtendSources();
 
-  // Convert any truncs, that aren't sources, into AND masks.
-  ConvertTruncs();
-
   // Promote visited instructions, mutating their types in place. Also insert
   // DSP intrinsics, if enabled, for adds and subs which would be unsafe to
   // promote.
   PromoteTree();
 
+  // Convert any truncs, that aren't sources, into AND masks.
+  ConvertTruncs();
+
   // Insert trunc instructions for use by calls, stores etc...
   TruncateSinks();
 
Index: test/CodeGen/ARM/CGP/arm-cgp-calls.ll
===================================================================
--- test/CodeGen/ARM/CGP/arm-cgp-calls.ll
+++ test/CodeGen/ARM/CGP/arm-cgp-calls.ll
@@ -201,7 +201,7 @@
 }
 
 ; CHECK-LABEL: promote_arg_pass_to_call
-; CHECK-NOT: uxt
+; CHECK: uxtb
 define i16 @promote_arg_pass_to_call(i16 zeroext %arg1, i16 zeroext %arg2) {
   %conv = add nuw i16 %arg1, 15
   %mul = mul nuw nsw i16 %conv, 3
Index: test/CodeGen/ARM/CGP/arm-cgp-casts.ll
===================================================================
--- test/CodeGen/ARM/CGP/arm-cgp-casts.ll
+++ test/CodeGen/ARM/CGP/arm-cgp-casts.ll
@@ -437,7 +437,7 @@
 }
 
 ; CHECK-COMMON-LABEL: trunc_sink_less_than_ret
-; CHECK-COMMON-NOT: uxt
+; CHECK-COMMON: uxt{{.*}}b
 define i8 @trunc_sink_less_than_ret(i16 zeroext %a, i16 zeroext %b, i16 zeroext %c, i8 zeroext %d, i8 zeroext %e) {
 entry:
   %sub = sub nuw i16 %b, %a
@@ -632,3 +632,25 @@
   %conv5 = zext i8 %or to i32
   ret i32 %conv5
 }
+
+; CHECK-LABEL: replace_trunk_with_mask
+; CHECK: udiv
+; CHECK: mls
+; CHECK: uxtb
+define i32 @replace_trunk_with_mask(i16* %a) {
+entry:
+  %0 = load i16, i16* %a
+  %cmp = icmp eq i16 %0, 0
+  br i1 %cmp, label %cond.end, label %cond.false
+
+cond.false:
+  %1 = urem i16 535, %0
+  %.lhs.trunc = trunc i16 %1 to i8
+  %2 = udiv i8 %.lhs.trunc, 3
+  %phitmp = zext i8 %2 to i32
+  br label %cond.end
+
+cond.end:
+  %cond = phi i32 [ %phitmp, %cond.false ], [ 0, %entry ]
+  ret i32 %cond
+}