diff --git a/llvm/lib/Transforms/InstCombine/InstructionCombining.cpp b/llvm/lib/Transforms/InstCombine/InstructionCombining.cpp
index 8e61e3d9fdb2d6..e31023607e8079 100644
--- a/llvm/lib/Transforms/InstCombine/InstructionCombining.cpp
+++ b/llvm/lib/Transforms/InstCombine/InstructionCombining.cpp
@@ -2450,12 +2450,13 @@ Instruction *InstCombiner::visitAllocSite(Instruction &MI) {
         replaceInstUsesWith(*C,
                             ConstantInt::get(Type::getInt1Ty(C->getContext()),
                                              C->isFalseWhenEqual()));
-      } else if (isa<BitCastInst>(I) || isa<GetElementPtrInst>(I) ||
-                 isa<AddrSpaceCastInst>(I)) {
-        replaceInstUsesWith(*I, UndefValue::get(I->getType()));
       } else if (auto *SI = dyn_cast<StoreInst>(I)) {
         for (auto *DII : DIIs)
           ConvertDebugDeclareToDebugValue(DII, SI, *DIB);
+      } else {
+        // Casts, GEP, or anything else: we're about to delete this instruction,
+        // so it can not have any valid uses.
+        replaceInstUsesWith(*I, UndefValue::get(I->getType()));
       }
       eraseInstFromFunction(*I);
     }
diff --git a/llvm/test/Transforms/InstCombine/builtin-object-size-ptr.ll b/llvm/test/Transforms/InstCombine/builtin-object-size-ptr.ll
index 4475e9554a5384..7747aea11ddab1 100644
--- a/llvm/test/Transforms/InstCombine/builtin-object-size-ptr.ll
+++ b/llvm/test/Transforms/InstCombine/builtin-object-size-ptr.ll
@@ -28,6 +28,48 @@ define i32 @foo() #0 {
   ret i32 %conv
 }
 
+; This used to crash while erasing instructions:
+; https://bugs.llvm.org/show_bug.cgi?id=43723
+
+define void @PR43723() {
+; CHECK-LABEL: @PR43723(
+; CHECK-NEXT:    ret void
+;
+  %tab = alloca [10 x i8], align 16
+  %t0 = bitcast [10 x i8]* %tab to i8*
+  call void @llvm.memset.p0i8.i64(i8* align 16 %t0, i8 9, i64 10, i1 false)
+  %t1 = call {}* @llvm.invariant.start.p0i8(i64 10, i8* align 16 %t0)
+  call void @llvm.invariant.end.p0i8({}* %t1, i64 10, i8* align 16 %t0)
+  ret void
+
+  uselistorder i8* %t0, { 1, 0, 2 }
+}
+
+define void @unknown_use_of_invariant_start({}** %p) {
+; CHECK-LABEL: @unknown_use_of_invariant_start(
+; CHECK-NEXT:    ret void
+;
+  %tab = alloca [10 x i8], align 16
+  %t0 = bitcast [10 x i8]* %tab to i8*
+  call void @llvm.memset.p0i8.i64(i8* align 16 %t0, i8 9, i64 10, i1 false)
+  %t1 = call {}* @llvm.invariant.start.p0i8(i64 10, i8* align 16 %t0)
+  call void @llvm.invariant.end.p0i8({}* %t1, i64 10, i8* align 16 %t0)
+  store {}* %t1, {}** %p
+  ret void
+}
+
+define {}* @minimal_invariant_start_use(i8 %x) {
+; CHECK-LABEL: @minimal_invariant_start_use(
+; CHECK-NEXT:    ret {}* undef
+;
+  %a = alloca i8
+  %i = call {}* @llvm.invariant.start.p0i8(i64 1, i8* %a)
+  ret {}* %i
+}
+
 declare void @llvm.lifetime.start.p0i8(i64, i8* nocapture) #1
 declare i64 @llvm.objectsize.i64.p0i8(i8*, i1) #2
 declare void @llvm.lifetime.end.p0i8(i64, i8* nocapture) #1
+declare void @llvm.memset.p0i8.i64(i8* nocapture writeonly, i8, i64, i1 immarg) #0
+declare {}* @llvm.invariant.start.p0i8(i64 immarg, i8* nocapture) #0
+declare void @llvm.invariant.end.p0i8({}*, i64 immarg, i8* nocapture) #0