diff --git a/llvm/lib/Transforms/IPO/ArgumentPromotion.cpp b/llvm/lib/Transforms/IPO/ArgumentPromotion.cpp
--- a/llvm/lib/Transforms/IPO/ArgumentPromotion.cpp
+++ b/llvm/lib/Transforms/IPO/ArgumentPromotion.cpp
@@ -48,6 +48,7 @@
 #include "llvm/Analysis/MemoryLocation.h"
 #include "llvm/Analysis/TargetLibraryInfo.h"
 #include "llvm/Analysis/TargetTransformInfo.h"
+#include "llvm/Analysis/ValueTracking.h"
 #include "llvm/IR/Argument.h"
 #include "llvm/IR/Attributes.h"
 #include "llvm/IR/BasicBlock.h"
@@ -312,10 +313,34 @@
           LoadInst *newLoad =
               IRB.CreateLoad(OrigLoad->getType(), V, V->getName() + ".val");
           newLoad->setAlignment(OrigLoad->getAlign());
-          // Transfer the AA info too.
-          AAMDNodes AAInfo;
-          OrigLoad->getAAMetadata(AAInfo);
-          newLoad->setAAMetadata(AAInfo);
+          newLoad->copyMetadata(*OrigLoad, {LLVMContext::MD_align});
+          // The more aggressive assumptions made by the other metadatas may
+          // only hold when the control flow actually reaches the load instr; so
+          // they should only be copied if the load is guaranteed to execute.
+          BasicBlock *BB = OrigLoad->getParent();
+          if (BB == &BB->getParent()->getEntryBlock()) {
+            bool IsOrigLoadGuaranteedToExecute = true;
+            for (const Instruction &II : *BB) {
+              if (!IsOrigLoadGuaranteedToExecute || &II == OrigLoad) {
+                break;
+              }
+              IsOrigLoadGuaranteedToExecute =
+                  llvm::isGuaranteedToTransferExecutionToSuccessor(&II);
+            }
+
+            if (IsOrigLoadGuaranteedToExecute) {
+              // Transfer the AA info too.
+              AAMDNodes AAInfo;
+              OrigLoad->getAAMetadata(AAInfo);
+              newLoad->setAAMetadata(AAInfo);
+              // And other metadata.
+              newLoad->copyMetadata(*OrigLoad, {LLVMContext::MD_nontemporal,
+                                                LLVMContext::MD_nonnull,
+                                                LLVMContext::MD_dereferenceable,
+                                                LLVMContext::MD_noundef,
+                                                LLVMContext::MD_range});
+            }
+          }
 
           Args.push_back(newLoad);
           ArgAttrVec.push_back(AttributeSet());
diff --git a/llvm/test/Transforms/ArgumentPromotion/metadata.ll b/llvm/test/Transforms/ArgumentPromotion/metadata.ll
new file mode 100644
--- /dev/null
+++ b/llvm/test/Transforms/ArgumentPromotion/metadata.ll
@@ -0,0 +1,105 @@
+; NOTE: Assertions have been autogenerated by utils/update_test_checks.py UTC_ARGS: --function-signature --scrub-attributes
+; RUN: opt < %s -argpromotion -S | FileCheck %s
+
+define i32 @should_copy_range(i32* %x) {
+; CHECK-LABEL: define {{[^@]+}}@should_copy_range
+; CHECK-SAME: (i32* [[X:%.*]]) {
+; CHECK-NEXT:    [[X_VAL:%.*]] = load i32, i32* [[X]], align 4, !range !0
+; CHECK-NEXT:    [[TMP1:%.*]] = call i32 @f_load_range(i32 [[X_VAL]])
+; CHECK-NEXT:    ret i32 [[TMP1]]
+;
+  %1 = call i32 @f_load_range(i32* %x)
+  ret i32 %1
+}
+
+define internal i32 @f_load_range(i32* %v) {
+; CHECK-LABEL: define {{[^@]+}}@f_load_range
+; CHECK-SAME: (i32 [[V_VAL:%.*]]) {
+; CHECK-NEXT:    ret i32 [[V_VAL]]
+;
+  %1 = load i32, i32* %v, align 4, !range !0
+  ret i32 %1
+}
+
+define i32* @should_copy_nonnull(i32** %x) {
+; CHECK-LABEL: define {{[^@]+}}@should_copy_nonnull
+; CHECK-SAME: (i32** [[X:%.*]]) {
+; CHECK-NEXT:    [[X_VAL:%.*]] = load i32*, i32** [[X]], align 4, !nonnull !1
+; CHECK-NEXT:    [[TMP1:%.*]] = call i32* @f_load_nonnull(i32* [[X_VAL]])
+; CHECK-NEXT:    ret i32* [[TMP1]]
+;
+  %1 = call i32* @f_load_nonnull(i32** %x)
+  ret i32* %1
+}
+
+define internal i32* @f_load_nonnull(i32** %v) {
+; CHECK-LABEL: define {{[^@]+}}@f_load_nonnull
+; CHECK-SAME: (i32* [[V_VAL:%.*]]) {
+; CHECK-NEXT:    ret i32* [[V_VAL]]
+;
+  %1 = load i32*, i32** %v, align 4, !nonnull !1
+  ret i32* %1
+}
+
+define i32* @should_copy_dereferenceable(i32** %x) {
+; CHECK-LABEL: define {{[^@]+}}@should_copy_dereferenceable
+; CHECK-SAME: (i32** [[X:%.*]]) {
+; CHECK-NEXT:    [[X_VAL:%.*]] = load i32*, i32** [[X]], align 4, !dereferenceable !2
+; CHECK-NEXT:    [[TMP1:%.*]] = call i32* @f_load_dereferenceable(i32* [[X_VAL]])
+; CHECK-NEXT:    ret i32* [[TMP1]]
+;
+  %1 = call i32* @f_load_dereferenceable(i32** %x)
+  ret i32* %1
+}
+
+define internal i32* @f_load_dereferenceable(i32** %v) {
+; CHECK-LABEL: define {{[^@]+}}@f_load_dereferenceable
+; CHECK-SAME: (i32* [[V_VAL:%.*]]) {
+; CHECK-NEXT:    ret i32* [[V_VAL]]
+;
+  %1 = load i32*, i32** %v, align 4, !dereferenceable !2
+  ret i32* %1
+}
+
+define i32* @no_copy_if_not_always_executed(
+; CHECK-LABEL: define {{[^@]+}}@no_copy_if_not_always_executed
+; CHECK-SAME: (i1 [[C:%.*]], i32** dereferenceable(8) [[V:%.*]]) {
+; CHECK-NEXT:    [[V_VAL:%.*]] = load i32*, i32** [[V]], align 1, !align !3
+; CHECK-NEXT:    [[TMP1:%.*]] = call i32* @f_load_not_always_executed(i1 [[C]], i32* [[V_VAL]])
+; CHECK-NEXT:    ret i32* [[TMP1]]
+;
+  i1 %c,
+  i32** dereferenceable(8) %v) {
+  %1 = call i32* @f_load_not_always_executed(i1 %c, i32** dereferenceable(8) %v)
+  ret i32* %1
+}
+
+define internal i32* @f_load_not_always_executed(
+; CHECK-LABEL: define {{[^@]+}}@f_load_not_always_executed
+; CHECK-SAME: (i1 [[C:%.*]], i32* [[V_VAL:%.*]]) {
+; CHECK-NEXT:    br i1 [[C]], label [[IF:%.*]], label [[ELSE:%.*]]
+; CHECK:       if:
+; CHECK-NEXT:    ret i32* [[V_VAL]]
+; CHECK:       else:
+; CHECK-NEXT:    ret i32* null
+;
+  i1 %c,
+  i32** dereferenceable(8) %v) {
+  br i1 %c, label %if, label %else
+
+if:
+  %1 = load i32*, i32** %v, align 1, !nonnull !1, !align !3
+  ret i32* %1
+
+else:
+  ret i32* null
+}
+
+; CHECK:      !0 = !{i32 0, i32 4}
+; CHECK-NEXT: !1 = !{}
+; CHECK-NEXT: !2 = !{i64 42}
+;
+!0 = !{i32 0, i32 4}
+!1 = !{}
+!2 = !{i64 42}
+!3 = !{i64 1}