[clang] [Clang] Fix __builtin_clear_padding() for bitfields (PR #216063)

Victor Campos via cfe-commits cfe-commits at lists.llvm.org
Wed Sep 16 03:44:56 PDT 2026


https://github.com/vhscampos updated https://github.com/llvm/llvm-project/pull/216063

>From 53e7461cf1dcff3e645e908d375457873d2ee936 Mon Sep 17 00:00:00 2001
From: Victor Campos <victor.campos at arm.com>
Date: Wed, 5 Aug 2026 16:05:21 +0100
Subject: [PATCH 1/2] [Clang] Fix __builtin_clear_padding() for bitfields

This patch fixes the padding clearing logic for bitfields.

The padding clearing logic for bitfields is wrong for these cases:

 - Over-sized bitfields don't have their padding bits recognised as
   such.
 - In big endian mode, the occupied bits are wrongly determined because
   the different byte order is dismissed.

A big number of new tests have also been added to cover many different
scenarios that involve bitfields. Preexisting tests had their check
patterns regenerated with the updater script.

Assisted-by: Codex. Used to help debugging and to generate tests.
---
 clang/lib/AST/ASTContext.cpp                  |   78 +-
 .../CodeGen/builtin-clear-padding-codegen.c   | 4493 +++++++++-
 .../builtin-clear-padding-codegen.cpp         | 7235 +++++++++++++++--
 3 files changed, 10654 insertions(+), 1152 deletions(-)

diff --git a/clang/lib/AST/ASTContext.cpp b/clang/lib/AST/ASTContext.cpp
index ad80074723817..61936e44f6278 100644
--- a/clang/lib/AST/ASTContext.cpp
+++ b/clang/lib/AST/ASTContext.cpp
@@ -15860,9 +15860,7 @@ struct PaddingCalculator {
 
       auto FieldOffset = ASTLayout.getFieldOffset(Field->getFieldIndex());
       if (Field->isBitField()) {
-        OccuppiedIntervals.push_back(ASTContext::BitInterval{
-            StartBitOffset + FieldOffset,
-            StartBitOffset + FieldOffset + Field->getBitWidthValue()});
+        VisitBitfield(Field, StartBitOffset + FieldOffset);
       } else {
         Stack.push_back(Data{StartBitOffset + FieldOffset,
                              Field->getType().getCanonicalType(),
@@ -15871,6 +15869,80 @@ struct PaddingCalculator {
     }
   }
 
+  void VisitBitfield(const FieldDecl *Field, uint64_t StartBitOffset) {
+    assert(Field->isBitField() && !Field->isUnnamedBitField());
+    if (Field->isZeroLengthBitField())
+      return;
+
+    const uint64_t DeclaredSizeInBits = Field->getBitWidthValue();
+
+    // Handle over-sized bitfields:
+    //   unsigned char a : 12;
+    // In this case, DeclaredSizeInBits is 12, but the actually occupied bit
+    // size is 8, while the remaining 4 bits are padding.
+    const uint64_t OccupiedSizeInBits =
+        std::min(DeclaredSizeInBits,
+                 static_cast<uint64_t>(Ctx.getIntWidth(Field->getType())));
+
+    if (Ctx.getTargetInfo().isLittleEndian()) {
+      OccuppiedIntervals.push_back(
+          {StartBitOffset, StartBitOffset + OccupiedSizeInBits});
+      return;
+    }
+
+    // In big endian mode, the sequence of occupied bits traverses bytes in
+    // increasing address order, just like in little endian. However, within
+    // each byte, the traversal starts from the most significant bit. This is
+    // where it differs from little endian.
+    //
+    // If the interval contains whole bytes in the middle, then for these
+    // nothing changes, and they constitute a contiguous interval. However for
+    // the partially occupied bytes in either end, if present, their bit
+    // intervals need to be adjusted so that they count from the MSB instead.
+    //
+    // FIXME: For over-sized bitfields in BE, Clang allocates padding bits
+    // before the occupied bits. This violates the ABI rules, which say that
+    // padding should be allocated after, regardless of endianness. The current
+    // code accommodates for Clang's current behaviour though, and bumps Start
+    // forward to skip the leading padding bits.
+    const uint64_t Start =
+        StartBitOffset + DeclaredSizeInBits - OccupiedSizeInBits;
+    const uint64_t End = Start + OccupiedSizeInBits;
+    const uint64_t CharWidth = Ctx.getCharWidth();
+
+    // Special case: all the occupied bits are contained within a single byte.
+    const uint64_t ByteStart = llvm::alignDown(Start, CharWidth);
+    const uint64_t ByteEnd = llvm::alignTo(End, CharWidth);
+    if (ByteStart == ByteEnd - CharWidth) {
+      const uint64_t Length = End - Start;
+      const uint64_t Offset = Start - ByteStart;
+      OccuppiedIntervals.push_back(
+          {ByteEnd - Offset - Length, ByteEnd - Offset});
+      return;
+    }
+
+    // Compute the contiguous interval in the middle, comprised of whole bytes,
+    // if any.
+    const uint64_t MiddleIntervalStart = llvm::alignTo(Start, CharWidth);
+    const uint64_t MiddleIntervalEnd = llvm::alignDown(End, CharWidth);
+    if (MiddleIntervalStart != MiddleIntervalEnd)
+      OccuppiedIntervals.push_back({MiddleIntervalStart, MiddleIntervalEnd});
+
+    // Compute the partially occupied first byte's interval, if any, counting
+    // from the MSB.
+    if (Start != MiddleIntervalStart) {
+      const uint64_t Length = MiddleIntervalStart - Start;
+      OccuppiedIntervals.push_back({ByteStart, ByteStart + Length});
+    }
+
+    // Compute the partially occupied last byte's interval, if any, counting
+    // from the MSB.
+    if (End != MiddleIntervalEnd) {
+      const uint64_t Length = End - MiddleIntervalEnd;
+      OccuppiedIntervals.push_back({ByteEnd - Length, ByteEnd});
+    }
+  }
+
   void VisitComplex(const ComplexType *CT, uint64_t StartBitOffset) {
     QualType ElementQualType = CT->getElementType().getCanonicalType();
     auto ElementSize = Ctx.getTypeSizeInChars(ElementQualType);
diff --git a/clang/test/CodeGen/builtin-clear-padding-codegen.c b/clang/test/CodeGen/builtin-clear-padding-codegen.c
index fceb7b623ba6d..2265414fb5f2e 100644
--- a/clang/test/CodeGen/builtin-clear-padding-codegen.c
+++ b/clang/test/CodeGen/builtin-clear-padding-codegen.c
@@ -3,8 +3,8 @@
 // RUN: %clang_cc1 -std=c11 -triple=x86_64-pc-windows-msvc -emit-llvm -o - %s | FileCheck %s --check-prefix=WINDOWS
 // RUN: %clang_cc1 -std=c11 -triple=armv7a-none-eabi -emit-llvm -o - %s | FileCheck %s --check-prefix=ARM-LE
 // RUN: %clang_cc1 -std=c11 -triple=armebv7a-none-eabi -emit-llvm -o - %s | FileCheck %s --check-prefix=ARM-BE
-// RUN: %clang_cc1 -std=c11 -triple=aarch64-none-eabi -emit-llvm -o - %s | FileCheck %s --check-prefix=ARM-LE
-// RUN: %clang_cc1 -std=c11 -triple=aarch64_be-none-eabi -emit-llvm -o - %s | FileCheck %s --check-prefix=ARM-BE
+// RUN: %clang_cc1 -std=c11 -triple=aarch64-none-eabi -emit-llvm -o - %s | FileCheck %s --check-prefix=AARCH64-LE
+// RUN: %clang_cc1 -std=c11 -triple=aarch64_be-none-eabi -emit-llvm -o - %s | FileCheck %s --check-prefix=AARCH64-BE
 
 struct Empty {};
 
@@ -16,20 +16,53 @@ struct Empty {};
 // LINUX-NEXT:    [[TMP0:%.*]] = load ptr, ptr [[E_ADDR]], align 8
 // LINUX-NEXT:    ret void
 //
-// WINDOWS-LABEL: define dso_local void @testEmpty(ptr noundef %e) #0 {
-// WINDOWS-NEXT: entry:
-// WINDOWS-NEXT:   %e.addr = alloca ptr, align 8
-// WINDOWS-NEXT:   store ptr %e, ptr %e.addr, align 8
-// WINDOWS-NEXT:   %0 = load ptr, ptr %e.addr, align 8
-// WINDOWS-NEXT:   %1 = getelementptr i8, ptr %0, i32 0
-// WINDOWS-NEXT:   store i8 0, ptr %1, align 1
-// WINDOWS-NEXT:   %2 = getelementptr i8, ptr %0, i32 1
-// WINDOWS-NEXT:   store i8 0, ptr %2, align 1
-// WINDOWS-NEXT:   %3 = getelementptr i8, ptr %0, i32 2
-// WINDOWS-NEXT:   store i8 0, ptr %3, align 1
-// WINDOWS-NEXT:   %4 = getelementptr i8, ptr %0, i32 3
-// WINDOWS-NEXT:   store i8 0, ptr %4, align 1
-// WINDOWS-NEXT:   ret void
+// WINDOWS-LABEL: define dso_local void @testEmpty(
+// WINDOWS-SAME: ptr noundef [[E:%.*]]) #[[ATTR0:[0-9]+]] {
+// WINDOWS-NEXT:  [[ENTRY:.*:]]
+// WINDOWS-NEXT:    [[E_ADDR:%.*]] = alloca ptr, align 8
+// WINDOWS-NEXT:    store ptr [[E]], ptr [[E_ADDR]], align 8
+// WINDOWS-NEXT:    [[TMP0:%.*]] = load ptr, ptr [[E_ADDR]], align 8
+// WINDOWS-NEXT:    [[TMP1:%.*]] = getelementptr i8, ptr [[TMP0]], i32 0
+// WINDOWS-NEXT:    store i8 0, ptr [[TMP1]], align 1
+// WINDOWS-NEXT:    [[TMP2:%.*]] = getelementptr i8, ptr [[TMP0]], i32 1
+// WINDOWS-NEXT:    store i8 0, ptr [[TMP2]], align 1
+// WINDOWS-NEXT:    [[TMP3:%.*]] = getelementptr i8, ptr [[TMP0]], i32 2
+// WINDOWS-NEXT:    store i8 0, ptr [[TMP3]], align 1
+// WINDOWS-NEXT:    [[TMP4:%.*]] = getelementptr i8, ptr [[TMP0]], i32 3
+// WINDOWS-NEXT:    store i8 0, ptr [[TMP4]], align 1
+// WINDOWS-NEXT:    ret void
+//
+// ARM-LE-LABEL: define dso_local void @testEmpty(
+// ARM-LE-SAME: ptr noundef [[E:%.*]]) #[[ATTR0:[0-9]+]] {
+// ARM-LE-NEXT:  [[ENTRY:.*:]]
+// ARM-LE-NEXT:    [[E_ADDR:%.*]] = alloca ptr, align 4
+// ARM-LE-NEXT:    store ptr [[E]], ptr [[E_ADDR]], align 4
+// ARM-LE-NEXT:    [[TMP0:%.*]] = load ptr, ptr [[E_ADDR]], align 4
+// ARM-LE-NEXT:    ret void
+//
+// ARM-BE-LABEL: define dso_local void @testEmpty(
+// ARM-BE-SAME: ptr noundef [[E:%.*]]) #[[ATTR0:[0-9]+]] {
+// ARM-BE-NEXT:  [[ENTRY:.*:]]
+// ARM-BE-NEXT:    [[E_ADDR:%.*]] = alloca ptr, align 4
+// ARM-BE-NEXT:    store ptr [[E]], ptr [[E_ADDR]], align 4
+// ARM-BE-NEXT:    [[TMP0:%.*]] = load ptr, ptr [[E_ADDR]], align 4
+// ARM-BE-NEXT:    ret void
+//
+// AARCH64-LE-LABEL: define dso_local void @testEmpty(
+// AARCH64-LE-SAME: ptr noundef [[E:%.*]]) #[[ATTR0:[0-9]+]] {
+// AARCH64-LE-NEXT:  [[ENTRY:.*:]]
+// AARCH64-LE-NEXT:    [[E_ADDR:%.*]] = alloca ptr, align 8
+// AARCH64-LE-NEXT:    store ptr [[E]], ptr [[E_ADDR]], align 8
+// AARCH64-LE-NEXT:    [[TMP0:%.*]] = load ptr, ptr [[E_ADDR]], align 8
+// AARCH64-LE-NEXT:    ret void
+//
+// AARCH64-BE-LABEL: define dso_local void @testEmpty(
+// AARCH64-BE-SAME: ptr noundef [[E:%.*]]) #[[ATTR0:[0-9]+]] {
+// AARCH64-BE-NEXT:  [[ENTRY:.*:]]
+// AARCH64-BE-NEXT:    [[E_ADDR:%.*]] = alloca ptr, align 8
+// AARCH64-BE-NEXT:    store ptr [[E]], ptr [[E_ADDR]], align 8
+// AARCH64-BE-NEXT:    [[TMP0:%.*]] = load ptr, ptr [[E_ADDR]], align 8
+// AARCH64-BE-NEXT:    ret void
 //
 void testEmpty(struct Empty *e) {
   // Emtpy struct is empty in C in Itanium ABI, no padding
@@ -46,12 +79,45 @@ void testEmpty(struct Empty *e) {
 // LINUX-NEXT:    [[TMP0:%.*]] = load ptr, ptr [[I_ADDR]], align 8
 // LINUX-NEXT:    ret void
 //
-// WINDOWS-LABEL: define dso_local void @testPrimitiveNoPadding(ptr noundef %i) #0 {
-// WINDOWS-NEXT: entry:
-// WINDOWS-NEXT:   %i.addr = alloca ptr, align 8
-// WINDOWS-NEXT:   store ptr %i, ptr %i.addr, align 8
-// WINDOWS-NEXT:   %0 = load ptr, ptr %i.addr, align 8
-// WINDOWS-NEXT:   ret void
+// WINDOWS-LABEL: define dso_local void @testPrimitiveNoPadding(
+// WINDOWS-SAME: ptr noundef [[I:%.*]]) #[[ATTR0]] {
+// WINDOWS-NEXT:  [[ENTRY:.*:]]
+// WINDOWS-NEXT:    [[I_ADDR:%.*]] = alloca ptr, align 8
+// WINDOWS-NEXT:    store ptr [[I]], ptr [[I_ADDR]], align 8
+// WINDOWS-NEXT:    [[TMP0:%.*]] = load ptr, ptr [[I_ADDR]], align 8
+// WINDOWS-NEXT:    ret void
+//
+// ARM-LE-LABEL: define dso_local void @testPrimitiveNoPadding(
+// ARM-LE-SAME: ptr noundef [[I:%.*]]) #[[ATTR0]] {
+// ARM-LE-NEXT:  [[ENTRY:.*:]]
+// ARM-LE-NEXT:    [[I_ADDR:%.*]] = alloca ptr, align 4
+// ARM-LE-NEXT:    store ptr [[I]], ptr [[I_ADDR]], align 4
+// ARM-LE-NEXT:    [[TMP0:%.*]] = load ptr, ptr [[I_ADDR]], align 4
+// ARM-LE-NEXT:    ret void
+//
+// ARM-BE-LABEL: define dso_local void @testPrimitiveNoPadding(
+// ARM-BE-SAME: ptr noundef [[I:%.*]]) #[[ATTR0]] {
+// ARM-BE-NEXT:  [[ENTRY:.*:]]
+// ARM-BE-NEXT:    [[I_ADDR:%.*]] = alloca ptr, align 4
+// ARM-BE-NEXT:    store ptr [[I]], ptr [[I_ADDR]], align 4
+// ARM-BE-NEXT:    [[TMP0:%.*]] = load ptr, ptr [[I_ADDR]], align 4
+// ARM-BE-NEXT:    ret void
+//
+// AARCH64-LE-LABEL: define dso_local void @testPrimitiveNoPadding(
+// AARCH64-LE-SAME: ptr noundef [[I:%.*]]) #[[ATTR0]] {
+// AARCH64-LE-NEXT:  [[ENTRY:.*:]]
+// AARCH64-LE-NEXT:    [[I_ADDR:%.*]] = alloca ptr, align 8
+// AARCH64-LE-NEXT:    store ptr [[I]], ptr [[I_ADDR]], align 8
+// AARCH64-LE-NEXT:    [[TMP0:%.*]] = load ptr, ptr [[I_ADDR]], align 8
+// AARCH64-LE-NEXT:    ret void
+//
+// AARCH64-BE-LABEL: define dso_local void @testPrimitiveNoPadding(
+// AARCH64-BE-SAME: ptr noundef [[I:%.*]]) #[[ATTR0]] {
+// AARCH64-BE-NEXT:  [[ENTRY:.*:]]
+// AARCH64-BE-NEXT:    [[I_ADDR:%.*]] = alloca ptr, align 8
+// AARCH64-BE-NEXT:    store ptr [[I]], ptr [[I_ADDR]], align 8
+// AARCH64-BE-NEXT:    [[TMP0:%.*]] = load ptr, ptr [[I_ADDR]], align 8
+// AARCH64-BE-NEXT:    ret void
 //
 void testPrimitiveNoPadding(int *i) {
   // This should not clear any padding, since int has no padding.
@@ -79,12 +145,45 @@ void testPrimitiveNoPadding(int *i) {
 // LINUX-NEXT:    store i8 0, ptr [[TMP6]], align 1
 // LINUX-NEXT:    ret void
 //
-// WINDOWS-LABEL: define dso_local void @testPrimitiveLongDouble(ptr noundef %ld) #0 {
-// WINDOWS-NEXT: entry:
-// WINDOWS-NEXT:   %ld.addr = alloca ptr, align 8
-// WINDOWS-NEXT:   store ptr %ld, ptr %ld.addr, align 8
-// WINDOWS-NEXT:   %0 = load ptr, ptr %ld.addr, align 8
-// WINDOWS-NEXT:   ret void
+// WINDOWS-LABEL: define dso_local void @testPrimitiveLongDouble(
+// WINDOWS-SAME: ptr noundef [[LD:%.*]]) #[[ATTR0]] {
+// WINDOWS-NEXT:  [[ENTRY:.*:]]
+// WINDOWS-NEXT:    [[LD_ADDR:%.*]] = alloca ptr, align 8
+// WINDOWS-NEXT:    store ptr [[LD]], ptr [[LD_ADDR]], align 8
+// WINDOWS-NEXT:    [[TMP0:%.*]] = load ptr, ptr [[LD_ADDR]], align 8
+// WINDOWS-NEXT:    ret void
+//
+// ARM-LE-LABEL: define dso_local void @testPrimitiveLongDouble(
+// ARM-LE-SAME: ptr noundef [[LD:%.*]]) #[[ATTR0]] {
+// ARM-LE-NEXT:  [[ENTRY:.*:]]
+// ARM-LE-NEXT:    [[LD_ADDR:%.*]] = alloca ptr, align 4
+// ARM-LE-NEXT:    store ptr [[LD]], ptr [[LD_ADDR]], align 4
+// ARM-LE-NEXT:    [[TMP0:%.*]] = load ptr, ptr [[LD_ADDR]], align 4
+// ARM-LE-NEXT:    ret void
+//
+// ARM-BE-LABEL: define dso_local void @testPrimitiveLongDouble(
+// ARM-BE-SAME: ptr noundef [[LD:%.*]]) #[[ATTR0]] {
+// ARM-BE-NEXT:  [[ENTRY:.*:]]
+// ARM-BE-NEXT:    [[LD_ADDR:%.*]] = alloca ptr, align 4
+// ARM-BE-NEXT:    store ptr [[LD]], ptr [[LD_ADDR]], align 4
+// ARM-BE-NEXT:    [[TMP0:%.*]] = load ptr, ptr [[LD_ADDR]], align 4
+// ARM-BE-NEXT:    ret void
+//
+// AARCH64-LE-LABEL: define dso_local void @testPrimitiveLongDouble(
+// AARCH64-LE-SAME: ptr noundef [[LD:%.*]]) #[[ATTR0]] {
+// AARCH64-LE-NEXT:  [[ENTRY:.*:]]
+// AARCH64-LE-NEXT:    [[LD_ADDR:%.*]] = alloca ptr, align 8
+// AARCH64-LE-NEXT:    store ptr [[LD]], ptr [[LD_ADDR]], align 8
+// AARCH64-LE-NEXT:    [[TMP0:%.*]] = load ptr, ptr [[LD_ADDR]], align 8
+// AARCH64-LE-NEXT:    ret void
+//
+// AARCH64-BE-LABEL: define dso_local void @testPrimitiveLongDouble(
+// AARCH64-BE-SAME: ptr noundef [[LD:%.*]]) #[[ATTR0]] {
+// AARCH64-BE-NEXT:  [[ENTRY:.*:]]
+// AARCH64-BE-NEXT:    [[LD_ADDR:%.*]] = alloca ptr, align 8
+// AARCH64-BE-NEXT:    store ptr [[LD]], ptr [[LD_ADDR]], align 8
+// AARCH64-BE-NEXT:    [[TMP0:%.*]] = load ptr, ptr [[LD_ADDR]], align 8
+// AARCH64-BE-NEXT:    ret void
 //
 void testPrimitiveLongDouble(long double *ld) {
   // padding [10, 15] on x86
@@ -109,100 +208,273 @@ void testPrimitiveLongDouble(long double *ld) {
 // LINUX-NEXT:    store i8 0, ptr [[TMP6]], align 1
 // LINUX-NEXT:    ret void
 //
-// WINDOWS-LABEL: define dso_local void @testBitInt(ptr noundef %bi) #0 {
-// WINDOWS-NEXT: entry:
-// WINDOWS-NEXT:   %bi.addr = alloca ptr, align 8
-// WINDOWS-NEXT:   store ptr %bi, ptr %bi.addr, align 8
-// WINDOWS-NEXT:   %0 = load ptr, ptr %bi.addr, align 8
-// WINDOWS-NEXT:   %1 = getelementptr i8, ptr %0, i32 12
-// WINDOWS-NEXT:   %2 = load i8, ptr %1, align 4
-// WINDOWS-NEXT:   %3 = and i8 %2, 1
-// WINDOWS-NEXT:   store i8 %3, ptr %1, align 4
-// WINDOWS-NEXT:   %4 = getelementptr i8, ptr %0, i32 13
-// WINDOWS-NEXT:   store i8 0, ptr %4, align 1
-// WINDOWS-NEXT:   %5 = getelementptr i8, ptr %0, i32 14
-// WINDOWS-NEXT:   store i8 0, ptr %5, align 2
-// WINDOWS-NEXT:   %6 = getelementptr i8, ptr %0, i32 15
-// WINDOWS-NEXT:   store i8 0, ptr %6, align 1
-// WINDOWS-NEXT:   ret void
+// WINDOWS-LABEL: define dso_local void @testBitInt(
+// WINDOWS-SAME: ptr noundef [[BI:%.*]]) #[[ATTR0]] {
+// WINDOWS-NEXT:  [[ENTRY:.*:]]
+// WINDOWS-NEXT:    [[BI_ADDR:%.*]] = alloca ptr, align 8
+// WINDOWS-NEXT:    store ptr [[BI]], ptr [[BI_ADDR]], align 8
+// WINDOWS-NEXT:    [[TMP0:%.*]] = load ptr, ptr [[BI_ADDR]], align 8
+// WINDOWS-NEXT:    [[TMP1:%.*]] = getelementptr i8, ptr [[TMP0]], i32 12
+// WINDOWS-NEXT:    [[TMP2:%.*]] = load i8, ptr [[TMP1]], align 4
+// WINDOWS-NEXT:    [[TMP3:%.*]] = and i8 [[TMP2]], 1
+// WINDOWS-NEXT:    store i8 [[TMP3]], ptr [[TMP1]], align 4
+// WINDOWS-NEXT:    [[TMP4:%.*]] = getelementptr i8, ptr [[TMP0]], i32 13
+// WINDOWS-NEXT:    store i8 0, ptr [[TMP4]], align 1
+// WINDOWS-NEXT:    [[TMP5:%.*]] = getelementptr i8, ptr [[TMP0]], i32 14
+// WINDOWS-NEXT:    store i8 0, ptr [[TMP5]], align 2
+// WINDOWS-NEXT:    [[TMP6:%.*]] = getelementptr i8, ptr [[TMP0]], i32 15
+// WINDOWS-NEXT:    store i8 0, ptr [[TMP6]], align 1
+// WINDOWS-NEXT:    ret void
 //
 // ARM-LE-LABEL: define dso_local void @testBitInt(
-// ARM-LE:      [[OBJ:%.*]] = load ptr, ptr {{.*}}, align {{[0-9]+}}
-// ARM-LE-NEXT: [[BYTE12:%.*]] = getelementptr i8, ptr [[OBJ]], i32 12
-// ARM-LE-NEXT: [[VALUE:%.*]] = load i8, ptr [[BYTE12]], align 4
-// ARM-LE-NEXT: [[MASKED:%.*]] = and i8 [[VALUE]], 1
-// ARM-LE-NEXT: store i8 [[MASKED]], ptr [[BYTE12]], align 4
-// ARM-LE-NEXT: [[BYTE13:%.*]] = getelementptr i8, ptr [[OBJ]], i32 13
-// ARM-LE-NEXT: store i8 0, ptr [[BYTE13]], align 1
-// ARM-LE-NEXT: [[BYTE14:%.*]] = getelementptr i8, ptr [[OBJ]], i32 14
-// ARM-LE-NEXT: store i8 0, ptr [[BYTE14]], align 2
-// ARM-LE-NEXT: [[BYTE15:%.*]] = getelementptr i8, ptr [[OBJ]], i32 15
-// ARM-LE-NEXT: store i8 0, ptr [[BYTE15]], align 1
-// ARM-LE-NEXT: ret void
+// ARM-LE-SAME: ptr noundef [[BI:%.*]]) #[[ATTR0]] {
+// ARM-LE-NEXT:  [[ENTRY:.*:]]
+// ARM-LE-NEXT:    [[BI_ADDR:%.*]] = alloca ptr, align 4
+// ARM-LE-NEXT:    store ptr [[BI]], ptr [[BI_ADDR]], align 4
+// ARM-LE-NEXT:    [[TMP0:%.*]] = load ptr, ptr [[BI_ADDR]], align 4
+// ARM-LE-NEXT:    [[TMP1:%.*]] = getelementptr i8, ptr [[TMP0]], i32 12
+// ARM-LE-NEXT:    [[TMP2:%.*]] = load i8, ptr [[TMP1]], align 4
+// ARM-LE-NEXT:    [[TMP3:%.*]] = and i8 [[TMP2]], 1
+// ARM-LE-NEXT:    store i8 [[TMP3]], ptr [[TMP1]], align 4
+// ARM-LE-NEXT:    [[TMP4:%.*]] = getelementptr i8, ptr [[TMP0]], i32 13
+// ARM-LE-NEXT:    store i8 0, ptr [[TMP4]], align 1
+// ARM-LE-NEXT:    [[TMP5:%.*]] = getelementptr i8, ptr [[TMP0]], i32 14
+// ARM-LE-NEXT:    store i8 0, ptr [[TMP5]], align 2
+// ARM-LE-NEXT:    [[TMP6:%.*]] = getelementptr i8, ptr [[TMP0]], i32 15
+// ARM-LE-NEXT:    store i8 0, ptr [[TMP6]], align 1
+// ARM-LE-NEXT:    ret void
 //
 // ARM-BE-LABEL: define dso_local void @testBitInt(
-// ARM-BE:      [[OBJ:%.*]] = load ptr, ptr {{.*}}, align {{[0-9]+}}
-// ARM-BE-NEXT: [[BYTE0:%.*]] = getelementptr i8, ptr [[OBJ]], i32 0
-// ARM-BE-NEXT: store i8 0, ptr [[BYTE0]], align {{[0-9]+}}
-// ARM-BE-NEXT: [[BYTE1:%.*]] = getelementptr i8, ptr [[OBJ]], i32 1
-// ARM-BE-NEXT: store i8 0, ptr [[BYTE1]], align 1
-// ARM-BE-NEXT: [[BYTE2:%.*]] = getelementptr i8, ptr [[OBJ]], i32 2
-// ARM-BE-NEXT: store i8 0, ptr [[BYTE2]], align 2
-// ARM-BE-NEXT: [[BYTE3:%.*]] = getelementptr i8, ptr [[OBJ]], i32 3
-// ARM-BE-NEXT: [[VALUE:%.*]] = load i8, ptr [[BYTE3]], align 1
-// ARM-BE-NEXT: [[MASKED:%.*]] = and i8 [[VALUE]], 1
-// ARM-BE-NEXT: store i8 [[MASKED]], ptr [[BYTE3]], align 1
-// ARM-BE-NEXT: ret void
+// ARM-BE-SAME: ptr noundef [[BI:%.*]]) #[[ATTR0]] {
+// ARM-BE-NEXT:  [[ENTRY:.*:]]
+// ARM-BE-NEXT:    [[BI_ADDR:%.*]] = alloca ptr, align 4
+// ARM-BE-NEXT:    store ptr [[BI]], ptr [[BI_ADDR]], align 4
+// ARM-BE-NEXT:    [[TMP0:%.*]] = load ptr, ptr [[BI_ADDR]], align 4
+// ARM-BE-NEXT:    [[TMP1:%.*]] = getelementptr i8, ptr [[TMP0]], i32 0
+// ARM-BE-NEXT:    store i8 0, ptr [[TMP1]], align 8
+// ARM-BE-NEXT:    [[TMP2:%.*]] = getelementptr i8, ptr [[TMP0]], i32 1
+// ARM-BE-NEXT:    store i8 0, ptr [[TMP2]], align 1
+// ARM-BE-NEXT:    [[TMP3:%.*]] = getelementptr i8, ptr [[TMP0]], i32 2
+// ARM-BE-NEXT:    store i8 0, ptr [[TMP3]], align 2
+// ARM-BE-NEXT:    [[TMP4:%.*]] = getelementptr i8, ptr [[TMP0]], i32 3
+// ARM-BE-NEXT:    [[TMP5:%.*]] = load i8, ptr [[TMP4]], align 1
+// ARM-BE-NEXT:    [[TMP6:%.*]] = and i8 [[TMP5]], 1
+// ARM-BE-NEXT:    store i8 [[TMP6]], ptr [[TMP4]], align 1
+// ARM-BE-NEXT:    ret void
+//
+// AARCH64-LE-LABEL: define dso_local void @testBitInt(
+// AARCH64-LE-SAME: ptr noundef [[BI:%.*]]) #[[ATTR0]] {
+// AARCH64-LE-NEXT:  [[ENTRY:.*:]]
+// AARCH64-LE-NEXT:    [[BI_ADDR:%.*]] = alloca ptr, align 8
+// AARCH64-LE-NEXT:    store ptr [[BI]], ptr [[BI_ADDR]], align 8
+// AARCH64-LE-NEXT:    [[TMP0:%.*]] = load ptr, ptr [[BI_ADDR]], align 8
+// AARCH64-LE-NEXT:    [[TMP1:%.*]] = getelementptr i8, ptr [[TMP0]], i32 12
+// AARCH64-LE-NEXT:    [[TMP2:%.*]] = load i8, ptr [[TMP1]], align 4
+// AARCH64-LE-NEXT:    [[TMP3:%.*]] = and i8 [[TMP2]], 1
+// AARCH64-LE-NEXT:    store i8 [[TMP3]], ptr [[TMP1]], align 4
+// AARCH64-LE-NEXT:    [[TMP4:%.*]] = getelementptr i8, ptr [[TMP0]], i32 13
+// AARCH64-LE-NEXT:    store i8 0, ptr [[TMP4]], align 1
+// AARCH64-LE-NEXT:    [[TMP5:%.*]] = getelementptr i8, ptr [[TMP0]], i32 14
+// AARCH64-LE-NEXT:    store i8 0, ptr [[TMP5]], align 2
+// AARCH64-LE-NEXT:    [[TMP6:%.*]] = getelementptr i8, ptr [[TMP0]], i32 15
+// AARCH64-LE-NEXT:    store i8 0, ptr [[TMP6]], align 1
+// AARCH64-LE-NEXT:    ret void
+//
+// AARCH64-BE-LABEL: define dso_local void @testBitInt(
+// AARCH64-BE-SAME: ptr noundef [[BI:%.*]]) #[[ATTR0]] {
+// AARCH64-BE-NEXT:  [[ENTRY:.*:]]
+// AARCH64-BE-NEXT:    [[BI_ADDR:%.*]] = alloca ptr, align 8
+// AARCH64-BE-NEXT:    store ptr [[BI]], ptr [[BI_ADDR]], align 8
+// AARCH64-BE-NEXT:    [[TMP0:%.*]] = load ptr, ptr [[BI_ADDR]], align 8
+// AARCH64-BE-NEXT:    [[TMP1:%.*]] = getelementptr i8, ptr [[TMP0]], i32 0
+// AARCH64-BE-NEXT:    store i8 0, ptr [[TMP1]], align 16
+// AARCH64-BE-NEXT:    [[TMP2:%.*]] = getelementptr i8, ptr [[TMP0]], i32 1
+// AARCH64-BE-NEXT:    store i8 0, ptr [[TMP2]], align 1
+// AARCH64-BE-NEXT:    [[TMP3:%.*]] = getelementptr i8, ptr [[TMP0]], i32 2
+// AARCH64-BE-NEXT:    store i8 0, ptr [[TMP3]], align 2
+// AARCH64-BE-NEXT:    [[TMP4:%.*]] = getelementptr i8, ptr [[TMP0]], i32 3
+// AARCH64-BE-NEXT:    [[TMP5:%.*]] = load i8, ptr [[TMP4]], align 1
+// AARCH64-BE-NEXT:    [[TMP6:%.*]] = and i8 [[TMP5]], 1
+// AARCH64-BE-NEXT:    store i8 [[TMP6]], ptr [[TMP4]], align 1
+// AARCH64-BE-NEXT:    ret void
 //
 void testBitInt(_BitInt(97) *bi) {
   // Storage is widened to 128 bits; clear bits [97, 128).
   __builtin_clear_padding(bi);
 }
 
+// LINUX-LABEL: define dso_local void @testBitIntByteAligned(
+// LINUX-SAME: ptr noundef [[BI:%.*]]) #[[ATTR0]] {
+// LINUX-NEXT:  [[ENTRY:.*:]]
+// LINUX-NEXT:    [[BI_ADDR:%.*]] = alloca ptr, align 8
+// LINUX-NEXT:    store ptr [[BI]], ptr [[BI_ADDR]], align 8
+// LINUX-NEXT:    [[TMP0:%.*]] = load ptr, ptr [[BI_ADDR]], align 8
+// LINUX-NEXT:    [[TMP1:%.*]] = getelementptr i8, ptr [[TMP0]], i32 3
+// LINUX-NEXT:    store i8 0, ptr [[TMP1]], align 1
+// LINUX-NEXT:    ret void
+//
+// WINDOWS-LABEL: define dso_local void @testBitIntByteAligned(
+// WINDOWS-SAME: ptr noundef [[BI:%.*]]) #[[ATTR0]] {
+// WINDOWS-NEXT:  [[ENTRY:.*:]]
+// WINDOWS-NEXT:    [[BI_ADDR:%.*]] = alloca ptr, align 8
+// WINDOWS-NEXT:    store ptr [[BI]], ptr [[BI_ADDR]], align 8
+// WINDOWS-NEXT:    [[TMP0:%.*]] = load ptr, ptr [[BI_ADDR]], align 8
+// WINDOWS-NEXT:    [[TMP1:%.*]] = getelementptr i8, ptr [[TMP0]], i32 3
+// WINDOWS-NEXT:    store i8 0, ptr [[TMP1]], align 1
+// WINDOWS-NEXT:    ret void
+//
 // ARM-LE-LABEL: define dso_local void @testBitIntByteAligned(
-// ARM-LE:      [[OBJ:%.*]] = load ptr, ptr {{.*}}, align {{[0-9]+}}
-// ARM-LE-NEXT: [[BYTE3:%.*]] = getelementptr i8, ptr [[OBJ]], i32 3
-// ARM-LE-NEXT: store i8 0, ptr [[BYTE3]], align 1
-// ARM-LE-NEXT: ret void
+// ARM-LE-SAME: ptr noundef [[BI:%.*]]) #[[ATTR0]] {
+// ARM-LE-NEXT:  [[ENTRY:.*:]]
+// ARM-LE-NEXT:    [[BI_ADDR:%.*]] = alloca ptr, align 4
+// ARM-LE-NEXT:    store ptr [[BI]], ptr [[BI_ADDR]], align 4
+// ARM-LE-NEXT:    [[TMP0:%.*]] = load ptr, ptr [[BI_ADDR]], align 4
+// ARM-LE-NEXT:    [[TMP1:%.*]] = getelementptr i8, ptr [[TMP0]], i32 3
+// ARM-LE-NEXT:    store i8 0, ptr [[TMP1]], align 1
+// ARM-LE-NEXT:    ret void
 //
 // ARM-BE-LABEL: define dso_local void @testBitIntByteAligned(
-// ARM-BE:      [[OBJ:%.*]] = load ptr, ptr {{.*}}, align {{[0-9]+}}
-// ARM-BE-NEXT: [[BYTE0:%.*]] = getelementptr i8, ptr [[OBJ]], i32 0
-// ARM-BE-NEXT: store i8 0, ptr [[BYTE0]], align 4
-// ARM-BE-NEXT: ret void
+// ARM-BE-SAME: ptr noundef [[BI:%.*]]) #[[ATTR0]] {
+// ARM-BE-NEXT:  [[ENTRY:.*:]]
+// ARM-BE-NEXT:    [[BI_ADDR:%.*]] = alloca ptr, align 4
+// ARM-BE-NEXT:    store ptr [[BI]], ptr [[BI_ADDR]], align 4
+// ARM-BE-NEXT:    [[TMP0:%.*]] = load ptr, ptr [[BI_ADDR]], align 4
+// ARM-BE-NEXT:    [[TMP1:%.*]] = getelementptr i8, ptr [[TMP0]], i32 0
+// ARM-BE-NEXT:    store i8 0, ptr [[TMP1]], align 4
+// ARM-BE-NEXT:    ret void
+//
+// AARCH64-LE-LABEL: define dso_local void @testBitIntByteAligned(
+// AARCH64-LE-SAME: ptr noundef [[BI:%.*]]) #[[ATTR0]] {
+// AARCH64-LE-NEXT:  [[ENTRY:.*:]]
+// AARCH64-LE-NEXT:    [[BI_ADDR:%.*]] = alloca ptr, align 8
+// AARCH64-LE-NEXT:    store ptr [[BI]], ptr [[BI_ADDR]], align 8
+// AARCH64-LE-NEXT:    [[TMP0:%.*]] = load ptr, ptr [[BI_ADDR]], align 8
+// AARCH64-LE-NEXT:    [[TMP1:%.*]] = getelementptr i8, ptr [[TMP0]], i32 3
+// AARCH64-LE-NEXT:    store i8 0, ptr [[TMP1]], align 1
+// AARCH64-LE-NEXT:    ret void
+//
+// AARCH64-BE-LABEL: define dso_local void @testBitIntByteAligned(
+// AARCH64-BE-SAME: ptr noundef [[BI:%.*]]) #[[ATTR0]] {
+// AARCH64-BE-NEXT:  [[ENTRY:.*:]]
+// AARCH64-BE-NEXT:    [[BI_ADDR:%.*]] = alloca ptr, align 8
+// AARCH64-BE-NEXT:    store ptr [[BI]], ptr [[BI_ADDR]], align 8
+// AARCH64-BE-NEXT:    [[TMP0:%.*]] = load ptr, ptr [[BI_ADDR]], align 8
+// AARCH64-BE-NEXT:    [[TMP1:%.*]] = getelementptr i8, ptr [[TMP0]], i32 0
+// AARCH64-BE-NEXT:    store i8 0, ptr [[TMP1]], align 4
+// AARCH64-BE-NEXT:    ret void
 //
 void testBitIntByteAligned(_BitInt(24) *bi) {
   // Storage is widened to 32 bits; clear the leading or trailing byte.
   __builtin_clear_padding(bi);
 }
 
+// LINUX-LABEL: define dso_local void @testVolatileBitInt(
+// LINUX-SAME: ptr noundef [[BI:%.*]]) #[[ATTR0]] {
+// LINUX-NEXT:  [[ENTRY:.*:]]
+// LINUX-NEXT:    [[BI_ADDR:%.*]] = alloca ptr, align 8
+// LINUX-NEXT:    store ptr [[BI]], ptr [[BI_ADDR]], align 8
+// LINUX-NEXT:    [[TMP0:%.*]] = load ptr, ptr [[BI_ADDR]], align 8
+// LINUX-NEXT:    [[TMP1:%.*]] = getelementptr i8, ptr [[TMP0]], i32 12
+// LINUX-NEXT:    [[TMP2:%.*]] = load i8, ptr [[TMP1]], align 4
+// LINUX-NEXT:    [[TMP3:%.*]] = and i8 [[TMP2]], 1
+// LINUX-NEXT:    store i8 [[TMP3]], ptr [[TMP1]], align 4
+// LINUX-NEXT:    [[TMP4:%.*]] = getelementptr i8, ptr [[TMP0]], i32 13
+// LINUX-NEXT:    store i8 0, ptr [[TMP4]], align 1
+// LINUX-NEXT:    [[TMP5:%.*]] = getelementptr i8, ptr [[TMP0]], i32 14
+// LINUX-NEXT:    store i8 0, ptr [[TMP5]], align 2
+// LINUX-NEXT:    [[TMP6:%.*]] = getelementptr i8, ptr [[TMP0]], i32 15
+// LINUX-NEXT:    store i8 0, ptr [[TMP6]], align 1
+// LINUX-NEXT:    ret void
+//
+// WINDOWS-LABEL: define dso_local void @testVolatileBitInt(
+// WINDOWS-SAME: ptr noundef [[BI:%.*]]) #[[ATTR0]] {
+// WINDOWS-NEXT:  [[ENTRY:.*:]]
+// WINDOWS-NEXT:    [[BI_ADDR:%.*]] = alloca ptr, align 8
+// WINDOWS-NEXT:    store ptr [[BI]], ptr [[BI_ADDR]], align 8
+// WINDOWS-NEXT:    [[TMP0:%.*]] = load ptr, ptr [[BI_ADDR]], align 8
+// WINDOWS-NEXT:    [[TMP1:%.*]] = getelementptr i8, ptr [[TMP0]], i32 12
+// WINDOWS-NEXT:    [[TMP2:%.*]] = load i8, ptr [[TMP1]], align 4
+// WINDOWS-NEXT:    [[TMP3:%.*]] = and i8 [[TMP2]], 1
+// WINDOWS-NEXT:    store i8 [[TMP3]], ptr [[TMP1]], align 4
+// WINDOWS-NEXT:    [[TMP4:%.*]] = getelementptr i8, ptr [[TMP0]], i32 13
+// WINDOWS-NEXT:    store i8 0, ptr [[TMP4]], align 1
+// WINDOWS-NEXT:    [[TMP5:%.*]] = getelementptr i8, ptr [[TMP0]], i32 14
+// WINDOWS-NEXT:    store i8 0, ptr [[TMP5]], align 2
+// WINDOWS-NEXT:    [[TMP6:%.*]] = getelementptr i8, ptr [[TMP0]], i32 15
+// WINDOWS-NEXT:    store i8 0, ptr [[TMP6]], align 1
+// WINDOWS-NEXT:    ret void
+//
 // ARM-LE-LABEL: define dso_local void @testVolatileBitInt(
-// ARM-LE:      [[OBJ:%.*]] = load ptr, ptr {{.*}}, align {{[0-9]+}}
-// ARM-LE-NEXT: [[BYTE12:%.*]] = getelementptr i8, ptr [[OBJ]], i32 12
-// ARM-LE-NEXT: [[VALUE:%.*]] = load i8, ptr [[BYTE12]], align 4
-// ARM-LE-NEXT: [[MASKED:%.*]] = and i8 [[VALUE]], 1
-// ARM-LE-NEXT: store i8 [[MASKED]], ptr [[BYTE12]], align 4
-// ARM-LE-NEXT: [[BYTE13:%.*]] = getelementptr i8, ptr [[OBJ]], i32 13
-// ARM-LE-NEXT: store i8 0, ptr [[BYTE13]], align 1
-// ARM-LE-NEXT: [[BYTE14:%.*]] = getelementptr i8, ptr [[OBJ]], i32 14
-// ARM-LE-NEXT: store i8 0, ptr [[BYTE14]], align 2
-// ARM-LE-NEXT: [[BYTE15:%.*]] = getelementptr i8, ptr [[OBJ]], i32 15
-// ARM-LE-NEXT: store i8 0, ptr [[BYTE15]], align 1
-// ARM-LE-NEXT: ret void
+// ARM-LE-SAME: ptr noundef [[BI:%.*]]) #[[ATTR0]] {
+// ARM-LE-NEXT:  [[ENTRY:.*:]]
+// ARM-LE-NEXT:    [[BI_ADDR:%.*]] = alloca ptr, align 4
+// ARM-LE-NEXT:    store ptr [[BI]], ptr [[BI_ADDR]], align 4
+// ARM-LE-NEXT:    [[TMP0:%.*]] = load ptr, ptr [[BI_ADDR]], align 4
+// ARM-LE-NEXT:    [[TMP1:%.*]] = getelementptr i8, ptr [[TMP0]], i32 12
+// ARM-LE-NEXT:    [[TMP2:%.*]] = load i8, ptr [[TMP1]], align 4
+// ARM-LE-NEXT:    [[TMP3:%.*]] = and i8 [[TMP2]], 1
+// ARM-LE-NEXT:    store i8 [[TMP3]], ptr [[TMP1]], align 4
+// ARM-LE-NEXT:    [[TMP4:%.*]] = getelementptr i8, ptr [[TMP0]], i32 13
+// ARM-LE-NEXT:    store i8 0, ptr [[TMP4]], align 1
+// ARM-LE-NEXT:    [[TMP5:%.*]] = getelementptr i8, ptr [[TMP0]], i32 14
+// ARM-LE-NEXT:    store i8 0, ptr [[TMP5]], align 2
+// ARM-LE-NEXT:    [[TMP6:%.*]] = getelementptr i8, ptr [[TMP0]], i32 15
+// ARM-LE-NEXT:    store i8 0, ptr [[TMP6]], align 1
+// ARM-LE-NEXT:    ret void
 //
 // ARM-BE-LABEL: define dso_local void @testVolatileBitInt(
-// ARM-BE:      [[OBJ:%.*]] = load ptr, ptr {{.*}}, align {{[0-9]+}}
-// ARM-BE-NEXT: [[BYTE0:%.*]] = getelementptr i8, ptr [[OBJ]], i32 0
-// ARM-BE-NEXT: store i8 0, ptr [[BYTE0]], align {{[0-9]+}}
-// ARM-BE-NEXT: [[BYTE1:%.*]] = getelementptr i8, ptr [[OBJ]], i32 1
-// ARM-BE-NEXT: store i8 0, ptr [[BYTE1]], align 1
-// ARM-BE-NEXT: [[BYTE2:%.*]] = getelementptr i8, ptr [[OBJ]], i32 2
-// ARM-BE-NEXT: store i8 0, ptr [[BYTE2]], align 2
-// ARM-BE-NEXT: [[BYTE3:%.*]] = getelementptr i8, ptr [[OBJ]], i32 3
-// ARM-BE-NEXT: [[VALUE:%.*]] = load i8, ptr [[BYTE3]], align 1
-// ARM-BE-NEXT: [[MASKED:%.*]] = and i8 [[VALUE]], 1
-// ARM-BE-NEXT: store i8 [[MASKED]], ptr [[BYTE3]], align 1
-// ARM-BE-NEXT: ret void
+// ARM-BE-SAME: ptr noundef [[BI:%.*]]) #[[ATTR0]] {
+// ARM-BE-NEXT:  [[ENTRY:.*:]]
+// ARM-BE-NEXT:    [[BI_ADDR:%.*]] = alloca ptr, align 4
+// ARM-BE-NEXT:    store ptr [[BI]], ptr [[BI_ADDR]], align 4
+// ARM-BE-NEXT:    [[TMP0:%.*]] = load ptr, ptr [[BI_ADDR]], align 4
+// ARM-BE-NEXT:    [[TMP1:%.*]] = getelementptr i8, ptr [[TMP0]], i32 0
+// ARM-BE-NEXT:    store i8 0, ptr [[TMP1]], align 8
+// ARM-BE-NEXT:    [[TMP2:%.*]] = getelementptr i8, ptr [[TMP0]], i32 1
+// ARM-BE-NEXT:    store i8 0, ptr [[TMP2]], align 1
+// ARM-BE-NEXT:    [[TMP3:%.*]] = getelementptr i8, ptr [[TMP0]], i32 2
+// ARM-BE-NEXT:    store i8 0, ptr [[TMP3]], align 2
+// ARM-BE-NEXT:    [[TMP4:%.*]] = getelementptr i8, ptr [[TMP0]], i32 3
+// ARM-BE-NEXT:    [[TMP5:%.*]] = load i8, ptr [[TMP4]], align 1
+// ARM-BE-NEXT:    [[TMP6:%.*]] = and i8 [[TMP5]], 1
+// ARM-BE-NEXT:    store i8 [[TMP6]], ptr [[TMP4]], align 1
+// ARM-BE-NEXT:    ret void
+//
+// AARCH64-LE-LABEL: define dso_local void @testVolatileBitInt(
+// AARCH64-LE-SAME: ptr noundef [[BI:%.*]]) #[[ATTR0]] {
+// AARCH64-LE-NEXT:  [[ENTRY:.*:]]
+// AARCH64-LE-NEXT:    [[BI_ADDR:%.*]] = alloca ptr, align 8
+// AARCH64-LE-NEXT:    store ptr [[BI]], ptr [[BI_ADDR]], align 8
+// AARCH64-LE-NEXT:    [[TMP0:%.*]] = load ptr, ptr [[BI_ADDR]], align 8
+// AARCH64-LE-NEXT:    [[TMP1:%.*]] = getelementptr i8, ptr [[TMP0]], i32 12
+// AARCH64-LE-NEXT:    [[TMP2:%.*]] = load i8, ptr [[TMP1]], align 4
+// AARCH64-LE-NEXT:    [[TMP3:%.*]] = and i8 [[TMP2]], 1
+// AARCH64-LE-NEXT:    store i8 [[TMP3]], ptr [[TMP1]], align 4
+// AARCH64-LE-NEXT:    [[TMP4:%.*]] = getelementptr i8, ptr [[TMP0]], i32 13
+// AARCH64-LE-NEXT:    store i8 0, ptr [[TMP4]], align 1
+// AARCH64-LE-NEXT:    [[TMP5:%.*]] = getelementptr i8, ptr [[TMP0]], i32 14
+// AARCH64-LE-NEXT:    store i8 0, ptr [[TMP5]], align 2
+// AARCH64-LE-NEXT:    [[TMP6:%.*]] = getelementptr i8, ptr [[TMP0]], i32 15
+// AARCH64-LE-NEXT:    store i8 0, ptr [[TMP6]], align 1
+// AARCH64-LE-NEXT:    ret void
+//
+// AARCH64-BE-LABEL: define dso_local void @testVolatileBitInt(
+// AARCH64-BE-SAME: ptr noundef [[BI:%.*]]) #[[ATTR0]] {
+// AARCH64-BE-NEXT:  [[ENTRY:.*:]]
+// AARCH64-BE-NEXT:    [[BI_ADDR:%.*]] = alloca ptr, align 8
+// AARCH64-BE-NEXT:    store ptr [[BI]], ptr [[BI_ADDR]], align 8
+// AARCH64-BE-NEXT:    [[TMP0:%.*]] = load ptr, ptr [[BI_ADDR]], align 8
+// AARCH64-BE-NEXT:    [[TMP1:%.*]] = getelementptr i8, ptr [[TMP0]], i32 0
+// AARCH64-BE-NEXT:    store i8 0, ptr [[TMP1]], align 16
+// AARCH64-BE-NEXT:    [[TMP2:%.*]] = getelementptr i8, ptr [[TMP0]], i32 1
+// AARCH64-BE-NEXT:    store i8 0, ptr [[TMP2]], align 1
+// AARCH64-BE-NEXT:    [[TMP3:%.*]] = getelementptr i8, ptr [[TMP0]], i32 2
+// AARCH64-BE-NEXT:    store i8 0, ptr [[TMP3]], align 2
+// AARCH64-BE-NEXT:    [[TMP4:%.*]] = getelementptr i8, ptr [[TMP0]], i32 3
+// AARCH64-BE-NEXT:    [[TMP5:%.*]] = load i8, ptr [[TMP4]], align 1
+// AARCH64-BE-NEXT:    [[TMP6:%.*]] = and i8 [[TMP5]], 1
+// AARCH64-BE-NEXT:    store i8 [[TMP6]], ptr [[TMP4]], align 1
+// AARCH64-BE-NEXT:    ret void
 //
 void testVolatileBitInt(volatile _BitInt(97) *bi) {
   __builtin_clear_padding(bi);
@@ -241,12 +513,45 @@ void testVolatileBitInt(volatile _BitInt(97) *bi) {
 // LINUX-NEXT:    store i8 0, ptr [[TMP12]], align 1
 // LINUX-NEXT:    ret void
 //
-// WINDOWS-LABEL: define dso_local void @testPrimitiveComplexLongDouble(ptr noundef %c) #0 {
-// WINDOWS-NEXT: entry:
-// WINDOWS-NEXT:   %c.addr = alloca ptr, align 8
-// WINDOWS-NEXT:   store ptr %c, ptr %c.addr, align 8
-// WINDOWS-NEXT:   %0 = load ptr, ptr %c.addr, align 8
-// WINDOWS-NEXT:   ret void
+// WINDOWS-LABEL: define dso_local void @testPrimitiveComplexLongDouble(
+// WINDOWS-SAME: ptr noundef [[C:%.*]]) #[[ATTR0]] {
+// WINDOWS-NEXT:  [[ENTRY:.*:]]
+// WINDOWS-NEXT:    [[C_ADDR:%.*]] = alloca ptr, align 8
+// WINDOWS-NEXT:    store ptr [[C]], ptr [[C_ADDR]], align 8
+// WINDOWS-NEXT:    [[TMP0:%.*]] = load ptr, ptr [[C_ADDR]], align 8
+// WINDOWS-NEXT:    ret void
+//
+// ARM-LE-LABEL: define dso_local void @testPrimitiveComplexLongDouble(
+// ARM-LE-SAME: ptr noundef [[C:%.*]]) #[[ATTR0]] {
+// ARM-LE-NEXT:  [[ENTRY:.*:]]
+// ARM-LE-NEXT:    [[C_ADDR:%.*]] = alloca ptr, align 4
+// ARM-LE-NEXT:    store ptr [[C]], ptr [[C_ADDR]], align 4
+// ARM-LE-NEXT:    [[TMP0:%.*]] = load ptr, ptr [[C_ADDR]], align 4
+// ARM-LE-NEXT:    ret void
+//
+// ARM-BE-LABEL: define dso_local void @testPrimitiveComplexLongDouble(
+// ARM-BE-SAME: ptr noundef [[C:%.*]]) #[[ATTR0]] {
+// ARM-BE-NEXT:  [[ENTRY:.*:]]
+// ARM-BE-NEXT:    [[C_ADDR:%.*]] = alloca ptr, align 4
+// ARM-BE-NEXT:    store ptr [[C]], ptr [[C_ADDR]], align 4
+// ARM-BE-NEXT:    [[TMP0:%.*]] = load ptr, ptr [[C_ADDR]], align 4
+// ARM-BE-NEXT:    ret void
+//
+// AARCH64-LE-LABEL: define dso_local void @testPrimitiveComplexLongDouble(
+// AARCH64-LE-SAME: ptr noundef [[C:%.*]]) #[[ATTR0]] {
+// AARCH64-LE-NEXT:  [[ENTRY:.*:]]
+// AARCH64-LE-NEXT:    [[C_ADDR:%.*]] = alloca ptr, align 8
+// AARCH64-LE-NEXT:    store ptr [[C]], ptr [[C_ADDR]], align 8
+// AARCH64-LE-NEXT:    [[TMP0:%.*]] = load ptr, ptr [[C_ADDR]], align 8
+// AARCH64-LE-NEXT:    ret void
+//
+// AARCH64-BE-LABEL: define dso_local void @testPrimitiveComplexLongDouble(
+// AARCH64-BE-SAME: ptr noundef [[C:%.*]]) #[[ATTR0]] {
+// AARCH64-BE-NEXT:  [[ENTRY:.*:]]
+// AARCH64-BE-NEXT:    [[C_ADDR:%.*]] = alloca ptr, align 8
+// AARCH64-BE-NEXT:    store ptr [[C]], ptr [[C_ADDR]], align 8
+// AARCH64-BE-NEXT:    [[TMP0:%.*]] = load ptr, ptr [[C_ADDR]], align 8
+// AARCH64-BE-NEXT:    ret void
 //
 void testPrimitiveComplexLongDouble(_Complex long double *c) {
   // padding [10, 15] and [26, 31] on x86
@@ -266,12 +571,45 @@ union U1 {
 // LINUX-NEXT:    [[TMP0:%.*]] = load ptr, ptr [[U_ADDR]], align 8
 // LINUX-NEXT:    ret void
 //
-// WINDOWS-LABEL: define dso_local void @testUnionDifferentLength(ptr noundef %u) #0 {
-// WINDOWS-NEXT: entry:
-// WINDOWS-NEXT:   %u.addr = alloca ptr, align 8
-// WINDOWS-NEXT:   store ptr %u, ptr %u.addr, align 8
-// WINDOWS-NEXT:   %0 = load ptr, ptr %u.addr, align 8
-// WINDOWS-NEXT:   ret void
+// WINDOWS-LABEL: define dso_local void @testUnionDifferentLength(
+// WINDOWS-SAME: ptr noundef [[U:%.*]]) #[[ATTR0]] {
+// WINDOWS-NEXT:  [[ENTRY:.*:]]
+// WINDOWS-NEXT:    [[U_ADDR:%.*]] = alloca ptr, align 8
+// WINDOWS-NEXT:    store ptr [[U]], ptr [[U_ADDR]], align 8
+// WINDOWS-NEXT:    [[TMP0:%.*]] = load ptr, ptr [[U_ADDR]], align 8
+// WINDOWS-NEXT:    ret void
+//
+// ARM-LE-LABEL: define dso_local void @testUnionDifferentLength(
+// ARM-LE-SAME: ptr noundef [[U:%.*]]) #[[ATTR0]] {
+// ARM-LE-NEXT:  [[ENTRY:.*:]]
+// ARM-LE-NEXT:    [[U_ADDR:%.*]] = alloca ptr, align 4
+// ARM-LE-NEXT:    store ptr [[U]], ptr [[U_ADDR]], align 4
+// ARM-LE-NEXT:    [[TMP0:%.*]] = load ptr, ptr [[U_ADDR]], align 4
+// ARM-LE-NEXT:    ret void
+//
+// ARM-BE-LABEL: define dso_local void @testUnionDifferentLength(
+// ARM-BE-SAME: ptr noundef [[U:%.*]]) #[[ATTR0]] {
+// ARM-BE-NEXT:  [[ENTRY:.*:]]
+// ARM-BE-NEXT:    [[U_ADDR:%.*]] = alloca ptr, align 4
+// ARM-BE-NEXT:    store ptr [[U]], ptr [[U_ADDR]], align 4
+// ARM-BE-NEXT:    [[TMP0:%.*]] = load ptr, ptr [[U_ADDR]], align 4
+// ARM-BE-NEXT:    ret void
+//
+// AARCH64-LE-LABEL: define dso_local void @testUnionDifferentLength(
+// AARCH64-LE-SAME: ptr noundef [[U:%.*]]) #[[ATTR0]] {
+// AARCH64-LE-NEXT:  [[ENTRY:.*:]]
+// AARCH64-LE-NEXT:    [[U_ADDR:%.*]] = alloca ptr, align 8
+// AARCH64-LE-NEXT:    store ptr [[U]], ptr [[U_ADDR]], align 8
+// AARCH64-LE-NEXT:    [[TMP0:%.*]] = load ptr, ptr [[U_ADDR]], align 8
+// AARCH64-LE-NEXT:    ret void
+//
+// AARCH64-BE-LABEL: define dso_local void @testUnionDifferentLength(
+// AARCH64-BE-SAME: ptr noundef [[U:%.*]]) #[[ATTR0]] {
+// AARCH64-BE-NEXT:  [[ENTRY:.*:]]
+// AARCH64-BE-NEXT:    [[U_ADDR:%.*]] = alloca ptr, align 8
+// AARCH64-BE-NEXT:    store ptr [[U]], ptr [[U_ADDR]], align 8
+// AARCH64-BE-NEXT:    [[TMP0:%.*]] = load ptr, ptr [[U_ADDR]], align 8
+// AARCH64-BE-NEXT:    ret void
 //
 void testUnionDifferentLength(union U1 *u) {
   // This should not clear the object representation bits of the non-active member.
@@ -309,26 +647,115 @@ union U2 {
 // LINUX-NEXT:    store i8 0, ptr [[TMP7]], align 1
 // LINUX-NEXT:    ret void
 //
-// WINDOWS-LABEL: define dso_local void @testUnionTailPaddingOfLongestMember(ptr noundef %u) #0 {
-// WINDOWS-NEXT: entry:
-// WINDOWS-NEXT:   %u.addr = alloca ptr, align 8
-// WINDOWS-NEXT:   store ptr %u, ptr %u.addr, align 8
-// WINDOWS-NEXT:   %0 = load ptr, ptr %u.addr, align 8
-// WINDOWS-NEXT:   %1 = getelementptr i8, ptr %0, i32 1
-// WINDOWS-NEXT:   store i8 0, ptr %1, align 1
-// WINDOWS-NEXT:   %2 = getelementptr i8, ptr %0, i32 2
-// WINDOWS-NEXT:   store i8 0, ptr %2, align 2
-// WINDOWS-NEXT:   %3 = getelementptr i8, ptr %0, i32 3
-// WINDOWS-NEXT:   store i8 0, ptr %3, align 1
-// WINDOWS-NEXT:   %4 = getelementptr i8, ptr %0, i32 4
-// WINDOWS-NEXT:   store i8 0, ptr %4, align 4
-// WINDOWS-NEXT:   %5 = getelementptr i8, ptr %0, i32 5
-// WINDOWS-NEXT:   store i8 0, ptr %5, align 1
-// WINDOWS-NEXT:   %6 = getelementptr i8, ptr %0, i32 6
-// WINDOWS-NEXT:   store i8 0, ptr %6, align 2
-// WINDOWS-NEXT:   %7 = getelementptr i8, ptr %0, i32 7
-// WINDOWS-NEXT:   store i8 0, ptr %7, align 1
-// WINDOWS-NEXT:   ret void
+// WINDOWS-LABEL: define dso_local void @testUnionTailPaddingOfLongestMember(
+// WINDOWS-SAME: ptr noundef [[U:%.*]]) #[[ATTR0]] {
+// WINDOWS-NEXT:  [[ENTRY:.*:]]
+// WINDOWS-NEXT:    [[U_ADDR:%.*]] = alloca ptr, align 8
+// WINDOWS-NEXT:    store ptr [[U]], ptr [[U_ADDR]], align 8
+// WINDOWS-NEXT:    [[TMP0:%.*]] = load ptr, ptr [[U_ADDR]], align 8
+// WINDOWS-NEXT:    [[TMP1:%.*]] = getelementptr i8, ptr [[TMP0]], i32 1
+// WINDOWS-NEXT:    store i8 0, ptr [[TMP1]], align 1
+// WINDOWS-NEXT:    [[TMP2:%.*]] = getelementptr i8, ptr [[TMP0]], i32 2
+// WINDOWS-NEXT:    store i8 0, ptr [[TMP2]], align 2
+// WINDOWS-NEXT:    [[TMP3:%.*]] = getelementptr i8, ptr [[TMP0]], i32 3
+// WINDOWS-NEXT:    store i8 0, ptr [[TMP3]], align 1
+// WINDOWS-NEXT:    [[TMP4:%.*]] = getelementptr i8, ptr [[TMP0]], i32 4
+// WINDOWS-NEXT:    store i8 0, ptr [[TMP4]], align 4
+// WINDOWS-NEXT:    [[TMP5:%.*]] = getelementptr i8, ptr [[TMP0]], i32 5
+// WINDOWS-NEXT:    store i8 0, ptr [[TMP5]], align 1
+// WINDOWS-NEXT:    [[TMP6:%.*]] = getelementptr i8, ptr [[TMP0]], i32 6
+// WINDOWS-NEXT:    store i8 0, ptr [[TMP6]], align 2
+// WINDOWS-NEXT:    [[TMP7:%.*]] = getelementptr i8, ptr [[TMP0]], i32 7
+// WINDOWS-NEXT:    store i8 0, ptr [[TMP7]], align 1
+// WINDOWS-NEXT:    ret void
+//
+// ARM-LE-LABEL: define dso_local void @testUnionTailPaddingOfLongestMember(
+// ARM-LE-SAME: ptr noundef [[U:%.*]]) #[[ATTR0]] {
+// ARM-LE-NEXT:  [[ENTRY:.*:]]
+// ARM-LE-NEXT:    [[U_ADDR:%.*]] = alloca ptr, align 4
+// ARM-LE-NEXT:    store ptr [[U]], ptr [[U_ADDR]], align 4
+// ARM-LE-NEXT:    [[TMP0:%.*]] = load ptr, ptr [[U_ADDR]], align 4
+// ARM-LE-NEXT:    [[TMP1:%.*]] = getelementptr i8, ptr [[TMP0]], i32 1
+// ARM-LE-NEXT:    store i8 0, ptr [[TMP1]], align 1
+// ARM-LE-NEXT:    [[TMP2:%.*]] = getelementptr i8, ptr [[TMP0]], i32 2
+// ARM-LE-NEXT:    store i8 0, ptr [[TMP2]], align 2
+// ARM-LE-NEXT:    [[TMP3:%.*]] = getelementptr i8, ptr [[TMP0]], i32 3
+// ARM-LE-NEXT:    store i8 0, ptr [[TMP3]], align 1
+// ARM-LE-NEXT:    [[TMP4:%.*]] = getelementptr i8, ptr [[TMP0]], i32 4
+// ARM-LE-NEXT:    store i8 0, ptr [[TMP4]], align 4
+// ARM-LE-NEXT:    [[TMP5:%.*]] = getelementptr i8, ptr [[TMP0]], i32 5
+// ARM-LE-NEXT:    store i8 0, ptr [[TMP5]], align 1
+// ARM-LE-NEXT:    [[TMP6:%.*]] = getelementptr i8, ptr [[TMP0]], i32 6
+// ARM-LE-NEXT:    store i8 0, ptr [[TMP6]], align 2
+// ARM-LE-NEXT:    [[TMP7:%.*]] = getelementptr i8, ptr [[TMP0]], i32 7
+// ARM-LE-NEXT:    store i8 0, ptr [[TMP7]], align 1
+// ARM-LE-NEXT:    ret void
+//
+// ARM-BE-LABEL: define dso_local void @testUnionTailPaddingOfLongestMember(
+// ARM-BE-SAME: ptr noundef [[U:%.*]]) #[[ATTR0]] {
+// ARM-BE-NEXT:  [[ENTRY:.*:]]
+// ARM-BE-NEXT:    [[U_ADDR:%.*]] = alloca ptr, align 4
+// ARM-BE-NEXT:    store ptr [[U]], ptr [[U_ADDR]], align 4
+// ARM-BE-NEXT:    [[TMP0:%.*]] = load ptr, ptr [[U_ADDR]], align 4
+// ARM-BE-NEXT:    [[TMP1:%.*]] = getelementptr i8, ptr [[TMP0]], i32 1
+// ARM-BE-NEXT:    store i8 0, ptr [[TMP1]], align 1
+// ARM-BE-NEXT:    [[TMP2:%.*]] = getelementptr i8, ptr [[TMP0]], i32 2
+// ARM-BE-NEXT:    store i8 0, ptr [[TMP2]], align 2
+// ARM-BE-NEXT:    [[TMP3:%.*]] = getelementptr i8, ptr [[TMP0]], i32 3
+// ARM-BE-NEXT:    store i8 0, ptr [[TMP3]], align 1
+// ARM-BE-NEXT:    [[TMP4:%.*]] = getelementptr i8, ptr [[TMP0]], i32 4
+// ARM-BE-NEXT:    store i8 0, ptr [[TMP4]], align 4
+// ARM-BE-NEXT:    [[TMP5:%.*]] = getelementptr i8, ptr [[TMP0]], i32 5
+// ARM-BE-NEXT:    store i8 0, ptr [[TMP5]], align 1
+// ARM-BE-NEXT:    [[TMP6:%.*]] = getelementptr i8, ptr [[TMP0]], i32 6
+// ARM-BE-NEXT:    store i8 0, ptr [[TMP6]], align 2
+// ARM-BE-NEXT:    [[TMP7:%.*]] = getelementptr i8, ptr [[TMP0]], i32 7
+// ARM-BE-NEXT:    store i8 0, ptr [[TMP7]], align 1
+// ARM-BE-NEXT:    ret void
+//
+// AARCH64-LE-LABEL: define dso_local void @testUnionTailPaddingOfLongestMember(
+// AARCH64-LE-SAME: ptr noundef [[U:%.*]]) #[[ATTR0]] {
+// AARCH64-LE-NEXT:  [[ENTRY:.*:]]
+// AARCH64-LE-NEXT:    [[U_ADDR:%.*]] = alloca ptr, align 8
+// AARCH64-LE-NEXT:    store ptr [[U]], ptr [[U_ADDR]], align 8
+// AARCH64-LE-NEXT:    [[TMP0:%.*]] = load ptr, ptr [[U_ADDR]], align 8
+// AARCH64-LE-NEXT:    [[TMP1:%.*]] = getelementptr i8, ptr [[TMP0]], i32 1
+// AARCH64-LE-NEXT:    store i8 0, ptr [[TMP1]], align 1
+// AARCH64-LE-NEXT:    [[TMP2:%.*]] = getelementptr i8, ptr [[TMP0]], i32 2
+// AARCH64-LE-NEXT:    store i8 0, ptr [[TMP2]], align 2
+// AARCH64-LE-NEXT:    [[TMP3:%.*]] = getelementptr i8, ptr [[TMP0]], i32 3
+// AARCH64-LE-NEXT:    store i8 0, ptr [[TMP3]], align 1
+// AARCH64-LE-NEXT:    [[TMP4:%.*]] = getelementptr i8, ptr [[TMP0]], i32 4
+// AARCH64-LE-NEXT:    store i8 0, ptr [[TMP4]], align 4
+// AARCH64-LE-NEXT:    [[TMP5:%.*]] = getelementptr i8, ptr [[TMP0]], i32 5
+// AARCH64-LE-NEXT:    store i8 0, ptr [[TMP5]], align 1
+// AARCH64-LE-NEXT:    [[TMP6:%.*]] = getelementptr i8, ptr [[TMP0]], i32 6
+// AARCH64-LE-NEXT:    store i8 0, ptr [[TMP6]], align 2
+// AARCH64-LE-NEXT:    [[TMP7:%.*]] = getelementptr i8, ptr [[TMP0]], i32 7
+// AARCH64-LE-NEXT:    store i8 0, ptr [[TMP7]], align 1
+// AARCH64-LE-NEXT:    ret void
+//
+// AARCH64-BE-LABEL: define dso_local void @testUnionTailPaddingOfLongestMember(
+// AARCH64-BE-SAME: ptr noundef [[U:%.*]]) #[[ATTR0]] {
+// AARCH64-BE-NEXT:  [[ENTRY:.*:]]
+// AARCH64-BE-NEXT:    [[U_ADDR:%.*]] = alloca ptr, align 8
+// AARCH64-BE-NEXT:    store ptr [[U]], ptr [[U_ADDR]], align 8
+// AARCH64-BE-NEXT:    [[TMP0:%.*]] = load ptr, ptr [[U_ADDR]], align 8
+// AARCH64-BE-NEXT:    [[TMP1:%.*]] = getelementptr i8, ptr [[TMP0]], i32 1
+// AARCH64-BE-NEXT:    store i8 0, ptr [[TMP1]], align 1
+// AARCH64-BE-NEXT:    [[TMP2:%.*]] = getelementptr i8, ptr [[TMP0]], i32 2
+// AARCH64-BE-NEXT:    store i8 0, ptr [[TMP2]], align 2
+// AARCH64-BE-NEXT:    [[TMP3:%.*]] = getelementptr i8, ptr [[TMP0]], i32 3
+// AARCH64-BE-NEXT:    store i8 0, ptr [[TMP3]], align 1
+// AARCH64-BE-NEXT:    [[TMP4:%.*]] = getelementptr i8, ptr [[TMP0]], i32 4
+// AARCH64-BE-NEXT:    store i8 0, ptr [[TMP4]], align 4
+// AARCH64-BE-NEXT:    [[TMP5:%.*]] = getelementptr i8, ptr [[TMP0]], i32 5
+// AARCH64-BE-NEXT:    store i8 0, ptr [[TMP5]], align 1
+// AARCH64-BE-NEXT:    [[TMP6:%.*]] = getelementptr i8, ptr [[TMP0]], i32 6
+// AARCH64-BE-NEXT:    store i8 0, ptr [[TMP6]], align 2
+// AARCH64-BE-NEXT:    [[TMP7:%.*]] = getelementptr i8, ptr [[TMP0]], i32 7
+// AARCH64-BE-NEXT:    store i8 0, ptr [[TMP7]], align 1
+// AARCH64-BE-NEXT:    ret void
 //
 void testUnionTailPaddingOfLongestMember(union U2 *u) {
   // This should clear the tail padding of the longest member.
@@ -381,26 +808,115 @@ struct __attribute__((aligned(4))) Baz {
 // LINUX-NEXT:    store i8 0, ptr [[TMP7]], align 1
 // LINUX-NEXT:    ret void
 //
-// WINDOWS-LABEL: define dso_local void @testStructPaddingInBetweenMembers(ptr noundef %baz) #0 {
-// WINDOWS-NEXT: entry:
-// WINDOWS-NEXT:   %baz.addr = alloca ptr, align 8
-// WINDOWS-NEXT:   store ptr %baz, ptr %baz.addr, align 8
-// WINDOWS-NEXT:   %0 = load ptr, ptr %baz.addr, align 8
-// WINDOWS-NEXT:   %1 = getelementptr i8, ptr %0, i32 1
-// WINDOWS-NEXT:   store i8 0, ptr %1, align 1
-// WINDOWS-NEXT:   %2 = getelementptr i8, ptr %0, i32 3
-// WINDOWS-NEXT:   store i8 0, ptr %2, align 1
-// WINDOWS-NEXT:   %3 = getelementptr i8, ptr %0, i32 5
-// WINDOWS-NEXT:   store i8 0, ptr %3, align 1
-// WINDOWS-NEXT:   %4 = getelementptr i8, ptr %0, i32 6
-// WINDOWS-NEXT:   store i8 0, ptr %4, align 2
-// WINDOWS-NEXT:   %5 = getelementptr i8, ptr %0, i32 7
-// WINDOWS-NEXT:   store i8 0, ptr %5, align 1
-// WINDOWS-NEXT:   %6 = getelementptr i8, ptr %0, i32 9
-// WINDOWS-NEXT:   store i8 0, ptr %6, align 1
-// WINDOWS-NEXT:   %7 = getelementptr i8, ptr %0, i32 11
-// WINDOWS-NEXT:   store i8 0, ptr %7, align 1
-// WINDOWS-NEXT:   ret void
+// WINDOWS-LABEL: define dso_local void @testStructPaddingInBetweenMembers(
+// WINDOWS-SAME: ptr noundef [[BAZ:%.*]]) #[[ATTR0]] {
+// WINDOWS-NEXT:  [[ENTRY:.*:]]
+// WINDOWS-NEXT:    [[BAZ_ADDR:%.*]] = alloca ptr, align 8
+// WINDOWS-NEXT:    store ptr [[BAZ]], ptr [[BAZ_ADDR]], align 8
+// WINDOWS-NEXT:    [[TMP0:%.*]] = load ptr, ptr [[BAZ_ADDR]], align 8
+// WINDOWS-NEXT:    [[TMP1:%.*]] = getelementptr i8, ptr [[TMP0]], i32 1
+// WINDOWS-NEXT:    store i8 0, ptr [[TMP1]], align 1
+// WINDOWS-NEXT:    [[TMP2:%.*]] = getelementptr i8, ptr [[TMP0]], i32 3
+// WINDOWS-NEXT:    store i8 0, ptr [[TMP2]], align 1
+// WINDOWS-NEXT:    [[TMP3:%.*]] = getelementptr i8, ptr [[TMP0]], i32 5
+// WINDOWS-NEXT:    store i8 0, ptr [[TMP3]], align 1
+// WINDOWS-NEXT:    [[TMP4:%.*]] = getelementptr i8, ptr [[TMP0]], i32 6
+// WINDOWS-NEXT:    store i8 0, ptr [[TMP4]], align 2
+// WINDOWS-NEXT:    [[TMP5:%.*]] = getelementptr i8, ptr [[TMP0]], i32 7
+// WINDOWS-NEXT:    store i8 0, ptr [[TMP5]], align 1
+// WINDOWS-NEXT:    [[TMP6:%.*]] = getelementptr i8, ptr [[TMP0]], i32 9
+// WINDOWS-NEXT:    store i8 0, ptr [[TMP6]], align 1
+// WINDOWS-NEXT:    [[TMP7:%.*]] = getelementptr i8, ptr [[TMP0]], i32 11
+// WINDOWS-NEXT:    store i8 0, ptr [[TMP7]], align 1
+// WINDOWS-NEXT:    ret void
+//
+// ARM-LE-LABEL: define dso_local void @testStructPaddingInBetweenMembers(
+// ARM-LE-SAME: ptr noundef [[BAZ:%.*]]) #[[ATTR0]] {
+// ARM-LE-NEXT:  [[ENTRY:.*:]]
+// ARM-LE-NEXT:    [[BAZ_ADDR:%.*]] = alloca ptr, align 4
+// ARM-LE-NEXT:    store ptr [[BAZ]], ptr [[BAZ_ADDR]], align 4
+// ARM-LE-NEXT:    [[TMP0:%.*]] = load ptr, ptr [[BAZ_ADDR]], align 4
+// ARM-LE-NEXT:    [[TMP1:%.*]] = getelementptr i8, ptr [[TMP0]], i32 1
+// ARM-LE-NEXT:    store i8 0, ptr [[TMP1]], align 1
+// ARM-LE-NEXT:    [[TMP2:%.*]] = getelementptr i8, ptr [[TMP0]], i32 3
+// ARM-LE-NEXT:    store i8 0, ptr [[TMP2]], align 1
+// ARM-LE-NEXT:    [[TMP3:%.*]] = getelementptr i8, ptr [[TMP0]], i32 5
+// ARM-LE-NEXT:    store i8 0, ptr [[TMP3]], align 1
+// ARM-LE-NEXT:    [[TMP4:%.*]] = getelementptr i8, ptr [[TMP0]], i32 6
+// ARM-LE-NEXT:    store i8 0, ptr [[TMP4]], align 2
+// ARM-LE-NEXT:    [[TMP5:%.*]] = getelementptr i8, ptr [[TMP0]], i32 7
+// ARM-LE-NEXT:    store i8 0, ptr [[TMP5]], align 1
+// ARM-LE-NEXT:    [[TMP6:%.*]] = getelementptr i8, ptr [[TMP0]], i32 9
+// ARM-LE-NEXT:    store i8 0, ptr [[TMP6]], align 1
+// ARM-LE-NEXT:    [[TMP7:%.*]] = getelementptr i8, ptr [[TMP0]], i32 11
+// ARM-LE-NEXT:    store i8 0, ptr [[TMP7]], align 1
+// ARM-LE-NEXT:    ret void
+//
+// ARM-BE-LABEL: define dso_local void @testStructPaddingInBetweenMembers(
+// ARM-BE-SAME: ptr noundef [[BAZ:%.*]]) #[[ATTR0]] {
+// ARM-BE-NEXT:  [[ENTRY:.*:]]
+// ARM-BE-NEXT:    [[BAZ_ADDR:%.*]] = alloca ptr, align 4
+// ARM-BE-NEXT:    store ptr [[BAZ]], ptr [[BAZ_ADDR]], align 4
+// ARM-BE-NEXT:    [[TMP0:%.*]] = load ptr, ptr [[BAZ_ADDR]], align 4
+// ARM-BE-NEXT:    [[TMP1:%.*]] = getelementptr i8, ptr [[TMP0]], i32 1
+// ARM-BE-NEXT:    store i8 0, ptr [[TMP1]], align 1
+// ARM-BE-NEXT:    [[TMP2:%.*]] = getelementptr i8, ptr [[TMP0]], i32 3
+// ARM-BE-NEXT:    store i8 0, ptr [[TMP2]], align 1
+// ARM-BE-NEXT:    [[TMP3:%.*]] = getelementptr i8, ptr [[TMP0]], i32 5
+// ARM-BE-NEXT:    store i8 0, ptr [[TMP3]], align 1
+// ARM-BE-NEXT:    [[TMP4:%.*]] = getelementptr i8, ptr [[TMP0]], i32 6
+// ARM-BE-NEXT:    store i8 0, ptr [[TMP4]], align 2
+// ARM-BE-NEXT:    [[TMP5:%.*]] = getelementptr i8, ptr [[TMP0]], i32 7
+// ARM-BE-NEXT:    store i8 0, ptr [[TMP5]], align 1
+// ARM-BE-NEXT:    [[TMP6:%.*]] = getelementptr i8, ptr [[TMP0]], i32 9
+// ARM-BE-NEXT:    store i8 0, ptr [[TMP6]], align 1
+// ARM-BE-NEXT:    [[TMP7:%.*]] = getelementptr i8, ptr [[TMP0]], i32 11
+// ARM-BE-NEXT:    store i8 0, ptr [[TMP7]], align 1
+// ARM-BE-NEXT:    ret void
+//
+// AARCH64-LE-LABEL: define dso_local void @testStructPaddingInBetweenMembers(
+// AARCH64-LE-SAME: ptr noundef [[BAZ:%.*]]) #[[ATTR0]] {
+// AARCH64-LE-NEXT:  [[ENTRY:.*:]]
+// AARCH64-LE-NEXT:    [[BAZ_ADDR:%.*]] = alloca ptr, align 8
+// AARCH64-LE-NEXT:    store ptr [[BAZ]], ptr [[BAZ_ADDR]], align 8
+// AARCH64-LE-NEXT:    [[TMP0:%.*]] = load ptr, ptr [[BAZ_ADDR]], align 8
+// AARCH64-LE-NEXT:    [[TMP1:%.*]] = getelementptr i8, ptr [[TMP0]], i32 1
+// AARCH64-LE-NEXT:    store i8 0, ptr [[TMP1]], align 1
+// AARCH64-LE-NEXT:    [[TMP2:%.*]] = getelementptr i8, ptr [[TMP0]], i32 3
+// AARCH64-LE-NEXT:    store i8 0, ptr [[TMP2]], align 1
+// AARCH64-LE-NEXT:    [[TMP3:%.*]] = getelementptr i8, ptr [[TMP0]], i32 5
+// AARCH64-LE-NEXT:    store i8 0, ptr [[TMP3]], align 1
+// AARCH64-LE-NEXT:    [[TMP4:%.*]] = getelementptr i8, ptr [[TMP0]], i32 6
+// AARCH64-LE-NEXT:    store i8 0, ptr [[TMP4]], align 2
+// AARCH64-LE-NEXT:    [[TMP5:%.*]] = getelementptr i8, ptr [[TMP0]], i32 7
+// AARCH64-LE-NEXT:    store i8 0, ptr [[TMP5]], align 1
+// AARCH64-LE-NEXT:    [[TMP6:%.*]] = getelementptr i8, ptr [[TMP0]], i32 9
+// AARCH64-LE-NEXT:    store i8 0, ptr [[TMP6]], align 1
+// AARCH64-LE-NEXT:    [[TMP7:%.*]] = getelementptr i8, ptr [[TMP0]], i32 11
+// AARCH64-LE-NEXT:    store i8 0, ptr [[TMP7]], align 1
+// AARCH64-LE-NEXT:    ret void
+//
+// AARCH64-BE-LABEL: define dso_local void @testStructPaddingInBetweenMembers(
+// AARCH64-BE-SAME: ptr noundef [[BAZ:%.*]]) #[[ATTR0]] {
+// AARCH64-BE-NEXT:  [[ENTRY:.*:]]
+// AARCH64-BE-NEXT:    [[BAZ_ADDR:%.*]] = alloca ptr, align 8
+// AARCH64-BE-NEXT:    store ptr [[BAZ]], ptr [[BAZ_ADDR]], align 8
+// AARCH64-BE-NEXT:    [[TMP0:%.*]] = load ptr, ptr [[BAZ_ADDR]], align 8
+// AARCH64-BE-NEXT:    [[TMP1:%.*]] = getelementptr i8, ptr [[TMP0]], i32 1
+// AARCH64-BE-NEXT:    store i8 0, ptr [[TMP1]], align 1
+// AARCH64-BE-NEXT:    [[TMP2:%.*]] = getelementptr i8, ptr [[TMP0]], i32 3
+// AARCH64-BE-NEXT:    store i8 0, ptr [[TMP2]], align 1
+// AARCH64-BE-NEXT:    [[TMP3:%.*]] = getelementptr i8, ptr [[TMP0]], i32 5
+// AARCH64-BE-NEXT:    store i8 0, ptr [[TMP3]], align 1
+// AARCH64-BE-NEXT:    [[TMP4:%.*]] = getelementptr i8, ptr [[TMP0]], i32 6
+// AARCH64-BE-NEXT:    store i8 0, ptr [[TMP4]], align 2
+// AARCH64-BE-NEXT:    [[TMP5:%.*]] = getelementptr i8, ptr [[TMP0]], i32 7
+// AARCH64-BE-NEXT:    store i8 0, ptr [[TMP5]], align 1
+// AARCH64-BE-NEXT:    [[TMP6:%.*]] = getelementptr i8, ptr [[TMP0]], i32 9
+// AARCH64-BE-NEXT:    store i8 0, ptr [[TMP6]], align 1
+// AARCH64-BE-NEXT:    [[TMP7:%.*]] = getelementptr i8, ptr [[TMP0]], i32 11
+// AARCH64-BE-NEXT:    store i8 0, ptr [[TMP7]], align 1
+// AARCH64-BE-NEXT:    ret void
 //
 void testStructPaddingInBetweenMembers(struct Baz *baz) {
   // this should clear all the padding in between various members
@@ -429,26 +945,115 @@ void testStructPaddingInBetweenMembers(struct Baz *baz) {
 // LINUX-NEXT:    store i8 0, ptr [[TMP7]], align 1
 // LINUX-NEXT:    ret void
 //
-// WINDOWS-LABEL: define dso_local void @testStructVolatile(ptr noundef %baz) #0 {
-// WINDOWS-NEXT: entry:
-// WINDOWS-NEXT:   %baz.addr = alloca ptr, align 8
-// WINDOWS-NEXT:   store ptr %baz, ptr %baz.addr, align 8
-// WINDOWS-NEXT:   %0 = load ptr, ptr %baz.addr, align 8
-// WINDOWS-NEXT:   %1 = getelementptr i8, ptr %0, i32 1
-// WINDOWS-NEXT:   store i8 0, ptr %1, align 1
-// WINDOWS-NEXT:   %2 = getelementptr i8, ptr %0, i32 3
-// WINDOWS-NEXT:   store i8 0, ptr %2, align 1
-// WINDOWS-NEXT:   %3 = getelementptr i8, ptr %0, i32 5
-// WINDOWS-NEXT:   store i8 0, ptr %3, align 1
-// WINDOWS-NEXT:   %4 = getelementptr i8, ptr %0, i32 6
-// WINDOWS-NEXT:   store i8 0, ptr %4, align 2
-// WINDOWS-NEXT:   %5 = getelementptr i8, ptr %0, i32 7
-// WINDOWS-NEXT:   store i8 0, ptr %5, align 1
-// WINDOWS-NEXT:   %6 = getelementptr i8, ptr %0, i32 9
-// WINDOWS-NEXT:   store i8 0, ptr %6, align 1
-// WINDOWS-NEXT:   %7 = getelementptr i8, ptr %0, i32 11
-// WINDOWS-NEXT:   store i8 0, ptr %7, align 1
-// WINDOWS-NEXT:   ret void
+// WINDOWS-LABEL: define dso_local void @testStructVolatile(
+// WINDOWS-SAME: ptr noundef [[BAZ:%.*]]) #[[ATTR0]] {
+// WINDOWS-NEXT:  [[ENTRY:.*:]]
+// WINDOWS-NEXT:    [[BAZ_ADDR:%.*]] = alloca ptr, align 8
+// WINDOWS-NEXT:    store ptr [[BAZ]], ptr [[BAZ_ADDR]], align 8
+// WINDOWS-NEXT:    [[TMP0:%.*]] = load ptr, ptr [[BAZ_ADDR]], align 8
+// WINDOWS-NEXT:    [[TMP1:%.*]] = getelementptr i8, ptr [[TMP0]], i32 1
+// WINDOWS-NEXT:    store i8 0, ptr [[TMP1]], align 1
+// WINDOWS-NEXT:    [[TMP2:%.*]] = getelementptr i8, ptr [[TMP0]], i32 3
+// WINDOWS-NEXT:    store i8 0, ptr [[TMP2]], align 1
+// WINDOWS-NEXT:    [[TMP3:%.*]] = getelementptr i8, ptr [[TMP0]], i32 5
+// WINDOWS-NEXT:    store i8 0, ptr [[TMP3]], align 1
+// WINDOWS-NEXT:    [[TMP4:%.*]] = getelementptr i8, ptr [[TMP0]], i32 6
+// WINDOWS-NEXT:    store i8 0, ptr [[TMP4]], align 2
+// WINDOWS-NEXT:    [[TMP5:%.*]] = getelementptr i8, ptr [[TMP0]], i32 7
+// WINDOWS-NEXT:    store i8 0, ptr [[TMP5]], align 1
+// WINDOWS-NEXT:    [[TMP6:%.*]] = getelementptr i8, ptr [[TMP0]], i32 9
+// WINDOWS-NEXT:    store i8 0, ptr [[TMP6]], align 1
+// WINDOWS-NEXT:    [[TMP7:%.*]] = getelementptr i8, ptr [[TMP0]], i32 11
+// WINDOWS-NEXT:    store i8 0, ptr [[TMP7]], align 1
+// WINDOWS-NEXT:    ret void
+//
+// ARM-LE-LABEL: define dso_local void @testStructVolatile(
+// ARM-LE-SAME: ptr noundef [[BAZ:%.*]]) #[[ATTR0]] {
+// ARM-LE-NEXT:  [[ENTRY:.*:]]
+// ARM-LE-NEXT:    [[BAZ_ADDR:%.*]] = alloca ptr, align 4
+// ARM-LE-NEXT:    store ptr [[BAZ]], ptr [[BAZ_ADDR]], align 4
+// ARM-LE-NEXT:    [[TMP0:%.*]] = load ptr, ptr [[BAZ_ADDR]], align 4
+// ARM-LE-NEXT:    [[TMP1:%.*]] = getelementptr i8, ptr [[TMP0]], i32 1
+// ARM-LE-NEXT:    store i8 0, ptr [[TMP1]], align 1
+// ARM-LE-NEXT:    [[TMP2:%.*]] = getelementptr i8, ptr [[TMP0]], i32 3
+// ARM-LE-NEXT:    store i8 0, ptr [[TMP2]], align 1
+// ARM-LE-NEXT:    [[TMP3:%.*]] = getelementptr i8, ptr [[TMP0]], i32 5
+// ARM-LE-NEXT:    store i8 0, ptr [[TMP3]], align 1
+// ARM-LE-NEXT:    [[TMP4:%.*]] = getelementptr i8, ptr [[TMP0]], i32 6
+// ARM-LE-NEXT:    store i8 0, ptr [[TMP4]], align 2
+// ARM-LE-NEXT:    [[TMP5:%.*]] = getelementptr i8, ptr [[TMP0]], i32 7
+// ARM-LE-NEXT:    store i8 0, ptr [[TMP5]], align 1
+// ARM-LE-NEXT:    [[TMP6:%.*]] = getelementptr i8, ptr [[TMP0]], i32 9
+// ARM-LE-NEXT:    store i8 0, ptr [[TMP6]], align 1
+// ARM-LE-NEXT:    [[TMP7:%.*]] = getelementptr i8, ptr [[TMP0]], i32 11
+// ARM-LE-NEXT:    store i8 0, ptr [[TMP7]], align 1
+// ARM-LE-NEXT:    ret void
+//
+// ARM-BE-LABEL: define dso_local void @testStructVolatile(
+// ARM-BE-SAME: ptr noundef [[BAZ:%.*]]) #[[ATTR0]] {
+// ARM-BE-NEXT:  [[ENTRY:.*:]]
+// ARM-BE-NEXT:    [[BAZ_ADDR:%.*]] = alloca ptr, align 4
+// ARM-BE-NEXT:    store ptr [[BAZ]], ptr [[BAZ_ADDR]], align 4
+// ARM-BE-NEXT:    [[TMP0:%.*]] = load ptr, ptr [[BAZ_ADDR]], align 4
+// ARM-BE-NEXT:    [[TMP1:%.*]] = getelementptr i8, ptr [[TMP0]], i32 1
+// ARM-BE-NEXT:    store i8 0, ptr [[TMP1]], align 1
+// ARM-BE-NEXT:    [[TMP2:%.*]] = getelementptr i8, ptr [[TMP0]], i32 3
+// ARM-BE-NEXT:    store i8 0, ptr [[TMP2]], align 1
+// ARM-BE-NEXT:    [[TMP3:%.*]] = getelementptr i8, ptr [[TMP0]], i32 5
+// ARM-BE-NEXT:    store i8 0, ptr [[TMP3]], align 1
+// ARM-BE-NEXT:    [[TMP4:%.*]] = getelementptr i8, ptr [[TMP0]], i32 6
+// ARM-BE-NEXT:    store i8 0, ptr [[TMP4]], align 2
+// ARM-BE-NEXT:    [[TMP5:%.*]] = getelementptr i8, ptr [[TMP0]], i32 7
+// ARM-BE-NEXT:    store i8 0, ptr [[TMP5]], align 1
+// ARM-BE-NEXT:    [[TMP6:%.*]] = getelementptr i8, ptr [[TMP0]], i32 9
+// ARM-BE-NEXT:    store i8 0, ptr [[TMP6]], align 1
+// ARM-BE-NEXT:    [[TMP7:%.*]] = getelementptr i8, ptr [[TMP0]], i32 11
+// ARM-BE-NEXT:    store i8 0, ptr [[TMP7]], align 1
+// ARM-BE-NEXT:    ret void
+//
+// AARCH64-LE-LABEL: define dso_local void @testStructVolatile(
+// AARCH64-LE-SAME: ptr noundef [[BAZ:%.*]]) #[[ATTR0]] {
+// AARCH64-LE-NEXT:  [[ENTRY:.*:]]
+// AARCH64-LE-NEXT:    [[BAZ_ADDR:%.*]] = alloca ptr, align 8
+// AARCH64-LE-NEXT:    store ptr [[BAZ]], ptr [[BAZ_ADDR]], align 8
+// AARCH64-LE-NEXT:    [[TMP0:%.*]] = load ptr, ptr [[BAZ_ADDR]], align 8
+// AARCH64-LE-NEXT:    [[TMP1:%.*]] = getelementptr i8, ptr [[TMP0]], i32 1
+// AARCH64-LE-NEXT:    store i8 0, ptr [[TMP1]], align 1
+// AARCH64-LE-NEXT:    [[TMP2:%.*]] = getelementptr i8, ptr [[TMP0]], i32 3
+// AARCH64-LE-NEXT:    store i8 0, ptr [[TMP2]], align 1
+// AARCH64-LE-NEXT:    [[TMP3:%.*]] = getelementptr i8, ptr [[TMP0]], i32 5
+// AARCH64-LE-NEXT:    store i8 0, ptr [[TMP3]], align 1
+// AARCH64-LE-NEXT:    [[TMP4:%.*]] = getelementptr i8, ptr [[TMP0]], i32 6
+// AARCH64-LE-NEXT:    store i8 0, ptr [[TMP4]], align 2
+// AARCH64-LE-NEXT:    [[TMP5:%.*]] = getelementptr i8, ptr [[TMP0]], i32 7
+// AARCH64-LE-NEXT:    store i8 0, ptr [[TMP5]], align 1
+// AARCH64-LE-NEXT:    [[TMP6:%.*]] = getelementptr i8, ptr [[TMP0]], i32 9
+// AARCH64-LE-NEXT:    store i8 0, ptr [[TMP6]], align 1
+// AARCH64-LE-NEXT:    [[TMP7:%.*]] = getelementptr i8, ptr [[TMP0]], i32 11
+// AARCH64-LE-NEXT:    store i8 0, ptr [[TMP7]], align 1
+// AARCH64-LE-NEXT:    ret void
+//
+// AARCH64-BE-LABEL: define dso_local void @testStructVolatile(
+// AARCH64-BE-SAME: ptr noundef [[BAZ:%.*]]) #[[ATTR0]] {
+// AARCH64-BE-NEXT:  [[ENTRY:.*:]]
+// AARCH64-BE-NEXT:    [[BAZ_ADDR:%.*]] = alloca ptr, align 8
+// AARCH64-BE-NEXT:    store ptr [[BAZ]], ptr [[BAZ_ADDR]], align 8
+// AARCH64-BE-NEXT:    [[TMP0:%.*]] = load ptr, ptr [[BAZ_ADDR]], align 8
+// AARCH64-BE-NEXT:    [[TMP1:%.*]] = getelementptr i8, ptr [[TMP0]], i32 1
+// AARCH64-BE-NEXT:    store i8 0, ptr [[TMP1]], align 1
+// AARCH64-BE-NEXT:    [[TMP2:%.*]] = getelementptr i8, ptr [[TMP0]], i32 3
+// AARCH64-BE-NEXT:    store i8 0, ptr [[TMP2]], align 1
+// AARCH64-BE-NEXT:    [[TMP3:%.*]] = getelementptr i8, ptr [[TMP0]], i32 5
+// AARCH64-BE-NEXT:    store i8 0, ptr [[TMP3]], align 1
+// AARCH64-BE-NEXT:    [[TMP4:%.*]] = getelementptr i8, ptr [[TMP0]], i32 6
+// AARCH64-BE-NEXT:    store i8 0, ptr [[TMP4]], align 2
+// AARCH64-BE-NEXT:    [[TMP5:%.*]] = getelementptr i8, ptr [[TMP0]], i32 7
+// AARCH64-BE-NEXT:    store i8 0, ptr [[TMP5]], align 1
+// AARCH64-BE-NEXT:    [[TMP6:%.*]] = getelementptr i8, ptr [[TMP0]], i32 9
+// AARCH64-BE-NEXT:    store i8 0, ptr [[TMP6]], align 1
+// AARCH64-BE-NEXT:    [[TMP7:%.*]] = getelementptr i8, ptr [[TMP0]], i32 11
+// AARCH64-BE-NEXT:    store i8 0, ptr [[TMP7]], align 1
+// AARCH64-BE-NEXT:    ret void
 //
 void testStructVolatile(volatile struct Baz *baz) {
   // this should clear all the padding in between various members
@@ -513,26 +1118,147 @@ struct S3 {
 // LINUX-NEXT:    store i8 0, ptr [[TMP21]], align 1
 // LINUX-NEXT:    ret void
 //
-// WINDOWS-LABEL: define dso_local void @testStructWithLongDouble(ptr noundef %s) #0 {
-// WINDOWS-NEXT: entry:
-// WINDOWS-NEXT:   %s.addr = alloca ptr, align 8
-// WINDOWS-NEXT:   store ptr %s, ptr %s.addr, align 8
-// WINDOWS-NEXT:   %0 = load ptr, ptr %s.addr, align 8
-// WINDOWS-NEXT:   %1 = getelementptr i8, ptr %0, i32 9
-// WINDOWS-NEXT:   store i8 0, ptr %1, align 1
-// WINDOWS-NEXT:   %2 = getelementptr i8, ptr %0, i32 10
-// WINDOWS-NEXT:   store i8 0, ptr %2, align 2
-// WINDOWS-NEXT:   %3 = getelementptr i8, ptr %0, i32 11
-// WINDOWS-NEXT:   store i8 0, ptr %3, align 1
-// WINDOWS-NEXT:   %4 = getelementptr i8, ptr %0, i32 12
-// WINDOWS-NEXT:   store i8 0, ptr %4, align 4
-// WINDOWS-NEXT:   %5 = getelementptr i8, ptr %0, i32 13
-// WINDOWS-NEXT:   store i8 0, ptr %5, align 1
-// WINDOWS-NEXT:   %6 = getelementptr i8, ptr %0, i32 14
-// WINDOWS-NEXT:   store i8 0, ptr %6, align 2
-// WINDOWS-NEXT:   %7 = getelementptr i8, ptr %0, i32 15
-// WINDOWS-NEXT:   store i8 0, ptr %7, align 1
-// WINDOWS-NEXT:   ret void
+// WINDOWS-LABEL: define dso_local void @testStructWithLongDouble(
+// WINDOWS-SAME: ptr noundef [[S:%.*]]) #[[ATTR0]] {
+// WINDOWS-NEXT:  [[ENTRY:.*:]]
+// WINDOWS-NEXT:    [[S_ADDR:%.*]] = alloca ptr, align 8
+// WINDOWS-NEXT:    store ptr [[S]], ptr [[S_ADDR]], align 8
+// WINDOWS-NEXT:    [[TMP0:%.*]] = load ptr, ptr [[S_ADDR]], align 8
+// WINDOWS-NEXT:    [[TMP1:%.*]] = getelementptr i8, ptr [[TMP0]], i32 9
+// WINDOWS-NEXT:    store i8 0, ptr [[TMP1]], align 1
+// WINDOWS-NEXT:    [[TMP2:%.*]] = getelementptr i8, ptr [[TMP0]], i32 10
+// WINDOWS-NEXT:    store i8 0, ptr [[TMP2]], align 2
+// WINDOWS-NEXT:    [[TMP3:%.*]] = getelementptr i8, ptr [[TMP0]], i32 11
+// WINDOWS-NEXT:    store i8 0, ptr [[TMP3]], align 1
+// WINDOWS-NEXT:    [[TMP4:%.*]] = getelementptr i8, ptr [[TMP0]], i32 12
+// WINDOWS-NEXT:    store i8 0, ptr [[TMP4]], align 4
+// WINDOWS-NEXT:    [[TMP5:%.*]] = getelementptr i8, ptr [[TMP0]], i32 13
+// WINDOWS-NEXT:    store i8 0, ptr [[TMP5]], align 1
+// WINDOWS-NEXT:    [[TMP6:%.*]] = getelementptr i8, ptr [[TMP0]], i32 14
+// WINDOWS-NEXT:    store i8 0, ptr [[TMP6]], align 2
+// WINDOWS-NEXT:    [[TMP7:%.*]] = getelementptr i8, ptr [[TMP0]], i32 15
+// WINDOWS-NEXT:    store i8 0, ptr [[TMP7]], align 1
+// WINDOWS-NEXT:    ret void
+//
+// ARM-LE-LABEL: define dso_local void @testStructWithLongDouble(
+// ARM-LE-SAME: ptr noundef [[S:%.*]]) #[[ATTR0]] {
+// ARM-LE-NEXT:  [[ENTRY:.*:]]
+// ARM-LE-NEXT:    [[S_ADDR:%.*]] = alloca ptr, align 4
+// ARM-LE-NEXT:    store ptr [[S]], ptr [[S_ADDR]], align 4
+// ARM-LE-NEXT:    [[TMP0:%.*]] = load ptr, ptr [[S_ADDR]], align 4
+// ARM-LE-NEXT:    [[TMP1:%.*]] = getelementptr i8, ptr [[TMP0]], i32 9
+// ARM-LE-NEXT:    store i8 0, ptr [[TMP1]], align 1
+// ARM-LE-NEXT:    [[TMP2:%.*]] = getelementptr i8, ptr [[TMP0]], i32 10
+// ARM-LE-NEXT:    store i8 0, ptr [[TMP2]], align 2
+// ARM-LE-NEXT:    [[TMP3:%.*]] = getelementptr i8, ptr [[TMP0]], i32 11
+// ARM-LE-NEXT:    store i8 0, ptr [[TMP3]], align 1
+// ARM-LE-NEXT:    [[TMP4:%.*]] = getelementptr i8, ptr [[TMP0]], i32 12
+// ARM-LE-NEXT:    store i8 0, ptr [[TMP4]], align 4
+// ARM-LE-NEXT:    [[TMP5:%.*]] = getelementptr i8, ptr [[TMP0]], i32 13
+// ARM-LE-NEXT:    store i8 0, ptr [[TMP5]], align 1
+// ARM-LE-NEXT:    [[TMP6:%.*]] = getelementptr i8, ptr [[TMP0]], i32 14
+// ARM-LE-NEXT:    store i8 0, ptr [[TMP6]], align 2
+// ARM-LE-NEXT:    [[TMP7:%.*]] = getelementptr i8, ptr [[TMP0]], i32 15
+// ARM-LE-NEXT:    store i8 0, ptr [[TMP7]], align 1
+// ARM-LE-NEXT:    ret void
+//
+// ARM-BE-LABEL: define dso_local void @testStructWithLongDouble(
+// ARM-BE-SAME: ptr noundef [[S:%.*]]) #[[ATTR0]] {
+// ARM-BE-NEXT:  [[ENTRY:.*:]]
+// ARM-BE-NEXT:    [[S_ADDR:%.*]] = alloca ptr, align 4
+// ARM-BE-NEXT:    store ptr [[S]], ptr [[S_ADDR]], align 4
+// ARM-BE-NEXT:    [[TMP0:%.*]] = load ptr, ptr [[S_ADDR]], align 4
+// ARM-BE-NEXT:    [[TMP1:%.*]] = getelementptr i8, ptr [[TMP0]], i32 9
+// ARM-BE-NEXT:    store i8 0, ptr [[TMP1]], align 1
+// ARM-BE-NEXT:    [[TMP2:%.*]] = getelementptr i8, ptr [[TMP0]], i32 10
+// ARM-BE-NEXT:    store i8 0, ptr [[TMP2]], align 2
+// ARM-BE-NEXT:    [[TMP3:%.*]] = getelementptr i8, ptr [[TMP0]], i32 11
+// ARM-BE-NEXT:    store i8 0, ptr [[TMP3]], align 1
+// ARM-BE-NEXT:    [[TMP4:%.*]] = getelementptr i8, ptr [[TMP0]], i32 12
+// ARM-BE-NEXT:    store i8 0, ptr [[TMP4]], align 4
+// ARM-BE-NEXT:    [[TMP5:%.*]] = getelementptr i8, ptr [[TMP0]], i32 13
+// ARM-BE-NEXT:    store i8 0, ptr [[TMP5]], align 1
+// ARM-BE-NEXT:    [[TMP6:%.*]] = getelementptr i8, ptr [[TMP0]], i32 14
+// ARM-BE-NEXT:    store i8 0, ptr [[TMP6]], align 2
+// ARM-BE-NEXT:    [[TMP7:%.*]] = getelementptr i8, ptr [[TMP0]], i32 15
+// ARM-BE-NEXT:    store i8 0, ptr [[TMP7]], align 1
+// ARM-BE-NEXT:    ret void
+//
+// AARCH64-LE-LABEL: define dso_local void @testStructWithLongDouble(
+// AARCH64-LE-SAME: ptr noundef [[S:%.*]]) #[[ATTR0]] {
+// AARCH64-LE-NEXT:  [[ENTRY:.*:]]
+// AARCH64-LE-NEXT:    [[S_ADDR:%.*]] = alloca ptr, align 8
+// AARCH64-LE-NEXT:    store ptr [[S]], ptr [[S_ADDR]], align 8
+// AARCH64-LE-NEXT:    [[TMP0:%.*]] = load ptr, ptr [[S_ADDR]], align 8
+// AARCH64-LE-NEXT:    [[TMP1:%.*]] = getelementptr i8, ptr [[TMP0]], i32 17
+// AARCH64-LE-NEXT:    store i8 0, ptr [[TMP1]], align 1
+// AARCH64-LE-NEXT:    [[TMP2:%.*]] = getelementptr i8, ptr [[TMP0]], i32 18
+// AARCH64-LE-NEXT:    store i8 0, ptr [[TMP2]], align 2
+// AARCH64-LE-NEXT:    [[TMP3:%.*]] = getelementptr i8, ptr [[TMP0]], i32 19
+// AARCH64-LE-NEXT:    store i8 0, ptr [[TMP3]], align 1
+// AARCH64-LE-NEXT:    [[TMP4:%.*]] = getelementptr i8, ptr [[TMP0]], i32 20
+// AARCH64-LE-NEXT:    store i8 0, ptr [[TMP4]], align 4
+// AARCH64-LE-NEXT:    [[TMP5:%.*]] = getelementptr i8, ptr [[TMP0]], i32 21
+// AARCH64-LE-NEXT:    store i8 0, ptr [[TMP5]], align 1
+// AARCH64-LE-NEXT:    [[TMP6:%.*]] = getelementptr i8, ptr [[TMP0]], i32 22
+// AARCH64-LE-NEXT:    store i8 0, ptr [[TMP6]], align 2
+// AARCH64-LE-NEXT:    [[TMP7:%.*]] = getelementptr i8, ptr [[TMP0]], i32 23
+// AARCH64-LE-NEXT:    store i8 0, ptr [[TMP7]], align 1
+// AARCH64-LE-NEXT:    [[TMP8:%.*]] = getelementptr i8, ptr [[TMP0]], i32 24
+// AARCH64-LE-NEXT:    store i8 0, ptr [[TMP8]], align 8
+// AARCH64-LE-NEXT:    [[TMP9:%.*]] = getelementptr i8, ptr [[TMP0]], i32 25
+// AARCH64-LE-NEXT:    store i8 0, ptr [[TMP9]], align 1
+// AARCH64-LE-NEXT:    [[TMP10:%.*]] = getelementptr i8, ptr [[TMP0]], i32 26
+// AARCH64-LE-NEXT:    store i8 0, ptr [[TMP10]], align 2
+// AARCH64-LE-NEXT:    [[TMP11:%.*]] = getelementptr i8, ptr [[TMP0]], i32 27
+// AARCH64-LE-NEXT:    store i8 0, ptr [[TMP11]], align 1
+// AARCH64-LE-NEXT:    [[TMP12:%.*]] = getelementptr i8, ptr [[TMP0]], i32 28
+// AARCH64-LE-NEXT:    store i8 0, ptr [[TMP12]], align 4
+// AARCH64-LE-NEXT:    [[TMP13:%.*]] = getelementptr i8, ptr [[TMP0]], i32 29
+// AARCH64-LE-NEXT:    store i8 0, ptr [[TMP13]], align 1
+// AARCH64-LE-NEXT:    [[TMP14:%.*]] = getelementptr i8, ptr [[TMP0]], i32 30
+// AARCH64-LE-NEXT:    store i8 0, ptr [[TMP14]], align 2
+// AARCH64-LE-NEXT:    [[TMP15:%.*]] = getelementptr i8, ptr [[TMP0]], i32 31
+// AARCH64-LE-NEXT:    store i8 0, ptr [[TMP15]], align 1
+// AARCH64-LE-NEXT:    ret void
+//
+// AARCH64-BE-LABEL: define dso_local void @testStructWithLongDouble(
+// AARCH64-BE-SAME: ptr noundef [[S:%.*]]) #[[ATTR0]] {
+// AARCH64-BE-NEXT:  [[ENTRY:.*:]]
+// AARCH64-BE-NEXT:    [[S_ADDR:%.*]] = alloca ptr, align 8
+// AARCH64-BE-NEXT:    store ptr [[S]], ptr [[S_ADDR]], align 8
+// AARCH64-BE-NEXT:    [[TMP0:%.*]] = load ptr, ptr [[S_ADDR]], align 8
+// AARCH64-BE-NEXT:    [[TMP1:%.*]] = getelementptr i8, ptr [[TMP0]], i32 17
+// AARCH64-BE-NEXT:    store i8 0, ptr [[TMP1]], align 1
+// AARCH64-BE-NEXT:    [[TMP2:%.*]] = getelementptr i8, ptr [[TMP0]], i32 18
+// AARCH64-BE-NEXT:    store i8 0, ptr [[TMP2]], align 2
+// AARCH64-BE-NEXT:    [[TMP3:%.*]] = getelementptr i8, ptr [[TMP0]], i32 19
+// AARCH64-BE-NEXT:    store i8 0, ptr [[TMP3]], align 1
+// AARCH64-BE-NEXT:    [[TMP4:%.*]] = getelementptr i8, ptr [[TMP0]], i32 20
+// AARCH64-BE-NEXT:    store i8 0, ptr [[TMP4]], align 4
+// AARCH64-BE-NEXT:    [[TMP5:%.*]] = getelementptr i8, ptr [[TMP0]], i32 21
+// AARCH64-BE-NEXT:    store i8 0, ptr [[TMP5]], align 1
+// AARCH64-BE-NEXT:    [[TMP6:%.*]] = getelementptr i8, ptr [[TMP0]], i32 22
+// AARCH64-BE-NEXT:    store i8 0, ptr [[TMP6]], align 2
+// AARCH64-BE-NEXT:    [[TMP7:%.*]] = getelementptr i8, ptr [[TMP0]], i32 23
+// AARCH64-BE-NEXT:    store i8 0, ptr [[TMP7]], align 1
+// AARCH64-BE-NEXT:    [[TMP8:%.*]] = getelementptr i8, ptr [[TMP0]], i32 24
+// AARCH64-BE-NEXT:    store i8 0, ptr [[TMP8]], align 8
+// AARCH64-BE-NEXT:    [[TMP9:%.*]] = getelementptr i8, ptr [[TMP0]], i32 25
+// AARCH64-BE-NEXT:    store i8 0, ptr [[TMP9]], align 1
+// AARCH64-BE-NEXT:    [[TMP10:%.*]] = getelementptr i8, ptr [[TMP0]], i32 26
+// AARCH64-BE-NEXT:    store i8 0, ptr [[TMP10]], align 2
+// AARCH64-BE-NEXT:    [[TMP11:%.*]] = getelementptr i8, ptr [[TMP0]], i32 27
+// AARCH64-BE-NEXT:    store i8 0, ptr [[TMP11]], align 1
+// AARCH64-BE-NEXT:    [[TMP12:%.*]] = getelementptr i8, ptr [[TMP0]], i32 28
+// AARCH64-BE-NEXT:    store i8 0, ptr [[TMP12]], align 4
+// AARCH64-BE-NEXT:    [[TMP13:%.*]] = getelementptr i8, ptr [[TMP0]], i32 29
+// AARCH64-BE-NEXT:    store i8 0, ptr [[TMP13]], align 1
+// AARCH64-BE-NEXT:    [[TMP14:%.*]] = getelementptr i8, ptr [[TMP0]], i32 30
+// AARCH64-BE-NEXT:    store i8 0, ptr [[TMP14]], align 2
+// AARCH64-BE-NEXT:    [[TMP15:%.*]] = getelementptr i8, ptr [[TMP0]], i32 31
+// AARCH64-BE-NEXT:    store i8 0, ptr [[TMP15]], align 1
+// AARCH64-BE-NEXT:    ret void
 //
 void testStructWithLongDouble(struct S3 *s) {
   // "long double data[0-9]", PAD [10-15], "b", PAD [17-31]
@@ -559,20 +1285,348 @@ struct S11 {
 // LINUX-NEXT:    store i8 [[TMP3]], ptr [[TMP1]], align 1
 // LINUX-NEXT:    ret void
 //
-// WINDOWS-LABEL: define dso_local void @testBitFields(ptr noundef %s) #0 {
-// WINDOWS-NEXT: entry:
-// WINDOWS-NEXT:   %s.addr = alloca ptr, align 8
-// WINDOWS-NEXT:   store ptr %s, ptr %s.addr, align 8
-// WINDOWS-NEXT:   %0 = load ptr, ptr %s.addr, align 8
-// WINDOWS-NEXT:   %1 = getelementptr i8, ptr %0, i32 0
-// WINDOWS-NEXT:   %2 = load i8, ptr %1, align 1
-// WINDOWS-NEXT:   %3 = and i8 %2, 31
-// WINDOWS-NEXT:   store i8 %3, ptr %1, align 1
-// WINDOWS-NEXT:   ret void
+// WINDOWS-LABEL: define dso_local void @testBitFields(
+// WINDOWS-SAME: ptr noundef [[S:%.*]]) #[[ATTR0]] {
+// WINDOWS-NEXT:  [[ENTRY:.*:]]
+// WINDOWS-NEXT:    [[S_ADDR:%.*]] = alloca ptr, align 8
+// WINDOWS-NEXT:    store ptr [[S]], ptr [[S_ADDR]], align 8
+// WINDOWS-NEXT:    [[TMP0:%.*]] = load ptr, ptr [[S_ADDR]], align 8
+// WINDOWS-NEXT:    [[TMP1:%.*]] = getelementptr i8, ptr [[TMP0]], i32 0
+// WINDOWS-NEXT:    [[TMP2:%.*]] = load i8, ptr [[TMP1]], align 1
+// WINDOWS-NEXT:    [[TMP3:%.*]] = and i8 [[TMP2]], 31
+// WINDOWS-NEXT:    store i8 [[TMP3]], ptr [[TMP1]], align 1
+// WINDOWS-NEXT:    ret void
+//
+// ARM-LE-LABEL: define dso_local void @testBitFields(
+// ARM-LE-SAME: ptr noundef [[S:%.*]]) #[[ATTR0]] {
+// ARM-LE-NEXT:  [[ENTRY:.*:]]
+// ARM-LE-NEXT:    [[S_ADDR:%.*]] = alloca ptr, align 4
+// ARM-LE-NEXT:    store ptr [[S]], ptr [[S_ADDR]], align 4
+// ARM-LE-NEXT:    [[TMP0:%.*]] = load ptr, ptr [[S_ADDR]], align 4
+// ARM-LE-NEXT:    [[TMP1:%.*]] = getelementptr i8, ptr [[TMP0]], i32 0
+// ARM-LE-NEXT:    [[TMP2:%.*]] = load i8, ptr [[TMP1]], align 1
+// ARM-LE-NEXT:    [[TMP3:%.*]] = and i8 [[TMP2]], 31
+// ARM-LE-NEXT:    store i8 [[TMP3]], ptr [[TMP1]], align 1
+// ARM-LE-NEXT:    ret void
+//
+// ARM-BE-LABEL: define dso_local void @testBitFields(
+// ARM-BE-SAME: ptr noundef [[S:%.*]]) #[[ATTR0]] {
+// ARM-BE-NEXT:  [[ENTRY:.*:]]
+// ARM-BE-NEXT:    [[S_ADDR:%.*]] = alloca ptr, align 4
+// ARM-BE-NEXT:    store ptr [[S]], ptr [[S_ADDR]], align 4
+// ARM-BE-NEXT:    [[TMP0:%.*]] = load ptr, ptr [[S_ADDR]], align 4
+// ARM-BE-NEXT:    [[TMP1:%.*]] = getelementptr i8, ptr [[TMP0]], i32 0
+// ARM-BE-NEXT:    [[TMP2:%.*]] = load i8, ptr [[TMP1]], align 1
+// ARM-BE-NEXT:    [[TMP3:%.*]] = and i8 [[TMP2]], -8
+// ARM-BE-NEXT:    store i8 [[TMP3]], ptr [[TMP1]], align 1
+// ARM-BE-NEXT:    ret void
+//
+// AARCH64-LE-LABEL: define dso_local void @testBitFields(
+// AARCH64-LE-SAME: ptr noundef [[S:%.*]]) #[[ATTR0]] {
+// AARCH64-LE-NEXT:  [[ENTRY:.*:]]
+// AARCH64-LE-NEXT:    [[S_ADDR:%.*]] = alloca ptr, align 8
+// AARCH64-LE-NEXT:    store ptr [[S]], ptr [[S_ADDR]], align 8
+// AARCH64-LE-NEXT:    [[TMP0:%.*]] = load ptr, ptr [[S_ADDR]], align 8
+// AARCH64-LE-NEXT:    [[TMP1:%.*]] = getelementptr i8, ptr [[TMP0]], i32 0
+// AARCH64-LE-NEXT:    [[TMP2:%.*]] = load i8, ptr [[TMP1]], align 1
+// AARCH64-LE-NEXT:    [[TMP3:%.*]] = and i8 [[TMP2]], 31
+// AARCH64-LE-NEXT:    store i8 [[TMP3]], ptr [[TMP1]], align 1
+// AARCH64-LE-NEXT:    ret void
+//
+// AARCH64-BE-LABEL: define dso_local void @testBitFields(
+// AARCH64-BE-SAME: ptr noundef [[S:%.*]]) #[[ATTR0]] {
+// AARCH64-BE-NEXT:  [[ENTRY:.*:]]
+// AARCH64-BE-NEXT:    [[S_ADDR:%.*]] = alloca ptr, align 8
+// AARCH64-BE-NEXT:    store ptr [[S]], ptr [[S_ADDR]], align 8
+// AARCH64-BE-NEXT:    [[TMP0:%.*]] = load ptr, ptr [[S_ADDR]], align 8
+// AARCH64-BE-NEXT:    [[TMP1:%.*]] = getelementptr i8, ptr [[TMP0]], i32 0
+// AARCH64-BE-NEXT:    [[TMP2:%.*]] = load i8, ptr [[TMP1]], align 1
+// AARCH64-BE-NEXT:    [[TMP3:%.*]] = and i8 [[TMP2]], -8
+// AARCH64-BE-NEXT:    store i8 [[TMP3]], ptr [[TMP1]], align 1
+// AARCH64-BE-NEXT:    ret void
 //
 void testBitFields(struct S11 *s) {
   // "b1" [0-2], "b2" [3-4], PAD [5-7], "b3" [8-13], "b4" [14-15]
-  // to clear 5-7, we should AND 0b00011111 (31)
+  // Little endian: preserve byte 0 with 0x1f. Big endian: use 0xf8.
+  __builtin_clear_padding(s);
+}
+
+
+struct VolatileBitfield {
+  volatile unsigned char a : 5;
+  int b : 7;
+  volatile int c : 27;
+};
+
+// LINUX-LABEL: define dso_local void @testVolatileBitfields(
+// LINUX-SAME: ptr noundef [[S:%.*]]) #[[ATTR0]] {
+// LINUX-NEXT:  [[ENTRY:.*:]]
+// LINUX-NEXT:    [[S_ADDR:%.*]] = alloca ptr, align 8
+// LINUX-NEXT:    store ptr [[S]], ptr [[S_ADDR]], align 8
+// LINUX-NEXT:    [[TMP0:%.*]] = load ptr, ptr [[S_ADDR]], align 8
+// LINUX-NEXT:    [[TMP1:%.*]] = getelementptr i8, ptr [[TMP0]], i32 1
+// LINUX-NEXT:    [[TMP2:%.*]] = load i8, ptr [[TMP1]], align 1
+// LINUX-NEXT:    [[TMP3:%.*]] = and i8 [[TMP2]], 15
+// LINUX-NEXT:    store i8 [[TMP3]], ptr [[TMP1]], align 1
+// LINUX-NEXT:    [[TMP4:%.*]] = getelementptr i8, ptr [[TMP0]], i32 2
+// LINUX-NEXT:    store i8 0, ptr [[TMP4]], align 2
+// LINUX-NEXT:    [[TMP5:%.*]] = getelementptr i8, ptr [[TMP0]], i32 3
+// LINUX-NEXT:    store i8 0, ptr [[TMP5]], align 1
+// LINUX-NEXT:    [[TMP6:%.*]] = getelementptr i8, ptr [[TMP0]], i32 7
+// LINUX-NEXT:    [[TMP7:%.*]] = load i8, ptr [[TMP6]], align 1
+// LINUX-NEXT:    [[TMP8:%.*]] = and i8 [[TMP7]], 7
+// LINUX-NEXT:    store i8 [[TMP8]], ptr [[TMP6]], align 1
+// LINUX-NEXT:    ret void
+//
+// WINDOWS-LABEL: define dso_local void @testVolatileBitfields(
+// WINDOWS-SAME: ptr noundef [[S:%.*]]) #[[ATTR0]] {
+// WINDOWS-NEXT:  [[ENTRY:.*:]]
+// WINDOWS-NEXT:    [[S_ADDR:%.*]] = alloca ptr, align 8
+// WINDOWS-NEXT:    store ptr [[S]], ptr [[S_ADDR]], align 8
+// WINDOWS-NEXT:    [[TMP0:%.*]] = load ptr, ptr [[S_ADDR]], align 8
+// WINDOWS-NEXT:    [[TMP1:%.*]] = getelementptr i8, ptr [[TMP0]], i32 0
+// WINDOWS-NEXT:    [[TMP2:%.*]] = load i8, ptr [[TMP1]], align 4
+// WINDOWS-NEXT:    [[TMP3:%.*]] = and i8 [[TMP2]], 31
+// WINDOWS-NEXT:    store i8 [[TMP3]], ptr [[TMP1]], align 4
+// WINDOWS-NEXT:    [[TMP4:%.*]] = getelementptr i8, ptr [[TMP0]], i32 1
+// WINDOWS-NEXT:    store i8 0, ptr [[TMP4]], align 1
+// WINDOWS-NEXT:    [[TMP5:%.*]] = getelementptr i8, ptr [[TMP0]], i32 2
+// WINDOWS-NEXT:    store i8 0, ptr [[TMP5]], align 2
+// WINDOWS-NEXT:    [[TMP6:%.*]] = getelementptr i8, ptr [[TMP0]], i32 3
+// WINDOWS-NEXT:    store i8 0, ptr [[TMP6]], align 1
+// WINDOWS-NEXT:    [[TMP7:%.*]] = getelementptr i8, ptr [[TMP0]], i32 4
+// WINDOWS-NEXT:    [[TMP8:%.*]] = load i8, ptr [[TMP7]], align 4
+// WINDOWS-NEXT:    [[TMP9:%.*]] = and i8 [[TMP8]], 127
+// WINDOWS-NEXT:    store i8 [[TMP9]], ptr [[TMP7]], align 4
+// WINDOWS-NEXT:    [[TMP10:%.*]] = getelementptr i8, ptr [[TMP0]], i32 5
+// WINDOWS-NEXT:    store i8 0, ptr [[TMP10]], align 1
+// WINDOWS-NEXT:    [[TMP11:%.*]] = getelementptr i8, ptr [[TMP0]], i32 6
+// WINDOWS-NEXT:    store i8 0, ptr [[TMP11]], align 2
+// WINDOWS-NEXT:    [[TMP12:%.*]] = getelementptr i8, ptr [[TMP0]], i32 7
+// WINDOWS-NEXT:    store i8 0, ptr [[TMP12]], align 1
+// WINDOWS-NEXT:    [[TMP13:%.*]] = getelementptr i8, ptr [[TMP0]], i32 11
+// WINDOWS-NEXT:    [[TMP14:%.*]] = load i8, ptr [[TMP13]], align 1
+// WINDOWS-NEXT:    [[TMP15:%.*]] = and i8 [[TMP14]], 7
+// WINDOWS-NEXT:    store i8 [[TMP15]], ptr [[TMP13]], align 1
+// WINDOWS-NEXT:    ret void
+//
+// ARM-LE-LABEL: define dso_local void @testVolatileBitfields(
+// ARM-LE-SAME: ptr noundef [[S:%.*]]) #[[ATTR0]] {
+// ARM-LE-NEXT:  [[ENTRY:.*:]]
+// ARM-LE-NEXT:    [[S_ADDR:%.*]] = alloca ptr, align 4
+// ARM-LE-NEXT:    store ptr [[S]], ptr [[S_ADDR]], align 4
+// ARM-LE-NEXT:    [[TMP0:%.*]] = load ptr, ptr [[S_ADDR]], align 4
+// ARM-LE-NEXT:    [[TMP1:%.*]] = getelementptr i8, ptr [[TMP0]], i32 1
+// ARM-LE-NEXT:    [[TMP2:%.*]] = load i8, ptr [[TMP1]], align 1
+// ARM-LE-NEXT:    [[TMP3:%.*]] = and i8 [[TMP2]], 15
+// ARM-LE-NEXT:    store i8 [[TMP3]], ptr [[TMP1]], align 1
+// ARM-LE-NEXT:    [[TMP4:%.*]] = getelementptr i8, ptr [[TMP0]], i32 2
+// ARM-LE-NEXT:    store i8 0, ptr [[TMP4]], align 2
+// ARM-LE-NEXT:    [[TMP5:%.*]] = getelementptr i8, ptr [[TMP0]], i32 3
+// ARM-LE-NEXT:    store i8 0, ptr [[TMP5]], align 1
+// ARM-LE-NEXT:    [[TMP6:%.*]] = getelementptr i8, ptr [[TMP0]], i32 7
+// ARM-LE-NEXT:    [[TMP7:%.*]] = load i8, ptr [[TMP6]], align 1
+// ARM-LE-NEXT:    [[TMP8:%.*]] = and i8 [[TMP7]], 7
+// ARM-LE-NEXT:    store i8 [[TMP8]], ptr [[TMP6]], align 1
+// ARM-LE-NEXT:    ret void
+//
+// ARM-BE-LABEL: define dso_local void @testVolatileBitfields(
+// ARM-BE-SAME: ptr noundef [[S:%.*]]) #[[ATTR0]] {
+// ARM-BE-NEXT:  [[ENTRY:.*:]]
+// ARM-BE-NEXT:    [[S_ADDR:%.*]] = alloca ptr, align 4
+// ARM-BE-NEXT:    store ptr [[S]], ptr [[S_ADDR]], align 4
+// ARM-BE-NEXT:    [[TMP0:%.*]] = load ptr, ptr [[S_ADDR]], align 4
+// ARM-BE-NEXT:    [[TMP1:%.*]] = getelementptr i8, ptr [[TMP0]], i32 1
+// ARM-BE-NEXT:    [[TMP2:%.*]] = load i8, ptr [[TMP1]], align 1
+// ARM-BE-NEXT:    [[TMP3:%.*]] = and i8 [[TMP2]], -16
+// ARM-BE-NEXT:    store i8 [[TMP3]], ptr [[TMP1]], align 1
+// ARM-BE-NEXT:    [[TMP4:%.*]] = getelementptr i8, ptr [[TMP0]], i32 2
+// ARM-BE-NEXT:    store i8 0, ptr [[TMP4]], align 2
+// ARM-BE-NEXT:    [[TMP5:%.*]] = getelementptr i8, ptr [[TMP0]], i32 3
+// ARM-BE-NEXT:    store i8 0, ptr [[TMP5]], align 1
+// ARM-BE-NEXT:    [[TMP6:%.*]] = getelementptr i8, ptr [[TMP0]], i32 7
+// ARM-BE-NEXT:    [[TMP7:%.*]] = load i8, ptr [[TMP6]], align 1
+// ARM-BE-NEXT:    [[TMP8:%.*]] = and i8 [[TMP7]], -32
+// ARM-BE-NEXT:    store i8 [[TMP8]], ptr [[TMP6]], align 1
+// ARM-BE-NEXT:    ret void
+//
+// AARCH64-LE-LABEL: define dso_local void @testVolatileBitfields(
+// AARCH64-LE-SAME: ptr noundef [[S:%.*]]) #[[ATTR0]] {
+// AARCH64-LE-NEXT:  [[ENTRY:.*:]]
+// AARCH64-LE-NEXT:    [[S_ADDR:%.*]] = alloca ptr, align 8
+// AARCH64-LE-NEXT:    store ptr [[S]], ptr [[S_ADDR]], align 8
+// AARCH64-LE-NEXT:    [[TMP0:%.*]] = load ptr, ptr [[S_ADDR]], align 8
+// AARCH64-LE-NEXT:    [[TMP1:%.*]] = getelementptr i8, ptr [[TMP0]], i32 1
+// AARCH64-LE-NEXT:    [[TMP2:%.*]] = load i8, ptr [[TMP1]], align 1
+// AARCH64-LE-NEXT:    [[TMP3:%.*]] = and i8 [[TMP2]], 15
+// AARCH64-LE-NEXT:    store i8 [[TMP3]], ptr [[TMP1]], align 1
+// AARCH64-LE-NEXT:    [[TMP4:%.*]] = getelementptr i8, ptr [[TMP0]], i32 2
+// AARCH64-LE-NEXT:    store i8 0, ptr [[TMP4]], align 2
+// AARCH64-LE-NEXT:    [[TMP5:%.*]] = getelementptr i8, ptr [[TMP0]], i32 3
+// AARCH64-LE-NEXT:    store i8 0, ptr [[TMP5]], align 1
+// AARCH64-LE-NEXT:    [[TMP6:%.*]] = getelementptr i8, ptr [[TMP0]], i32 7
+// AARCH64-LE-NEXT:    [[TMP7:%.*]] = load i8, ptr [[TMP6]], align 1
+// AARCH64-LE-NEXT:    [[TMP8:%.*]] = and i8 [[TMP7]], 7
+// AARCH64-LE-NEXT:    store i8 [[TMP8]], ptr [[TMP6]], align 1
+// AARCH64-LE-NEXT:    ret void
+//
+// AARCH64-BE-LABEL: define dso_local void @testVolatileBitfields(
+// AARCH64-BE-SAME: ptr noundef [[S:%.*]]) #[[ATTR0]] {
+// AARCH64-BE-NEXT:  [[ENTRY:.*:]]
+// AARCH64-BE-NEXT:    [[S_ADDR:%.*]] = alloca ptr, align 8
+// AARCH64-BE-NEXT:    store ptr [[S]], ptr [[S_ADDR]], align 8
+// AARCH64-BE-NEXT:    [[TMP0:%.*]] = load ptr, ptr [[S_ADDR]], align 8
+// AARCH64-BE-NEXT:    [[TMP1:%.*]] = getelementptr i8, ptr [[TMP0]], i32 1
+// AARCH64-BE-NEXT:    [[TMP2:%.*]] = load i8, ptr [[TMP1]], align 1
+// AARCH64-BE-NEXT:    [[TMP3:%.*]] = and i8 [[TMP2]], -16
+// AARCH64-BE-NEXT:    store i8 [[TMP3]], ptr [[TMP1]], align 1
+// AARCH64-BE-NEXT:    [[TMP4:%.*]] = getelementptr i8, ptr [[TMP0]], i32 2
+// AARCH64-BE-NEXT:    store i8 0, ptr [[TMP4]], align 2
+// AARCH64-BE-NEXT:    [[TMP5:%.*]] = getelementptr i8, ptr [[TMP0]], i32 3
+// AARCH64-BE-NEXT:    store i8 0, ptr [[TMP5]], align 1
+// AARCH64-BE-NEXT:    [[TMP6:%.*]] = getelementptr i8, ptr [[TMP0]], i32 7
+// AARCH64-BE-NEXT:    [[TMP7:%.*]] = load i8, ptr [[TMP6]], align 1
+// AARCH64-BE-NEXT:    [[TMP8:%.*]] = and i8 [[TMP7]], -32
+// AARCH64-BE-NEXT:    store i8 [[TMP8]], ptr [[TMP6]], align 1
+// AARCH64-BE-NEXT:    ret void
+//
+void testVolatileBitfields(struct VolatileBitfield *s) {
+  __builtin_clear_padding(s);
+}
+
+// LINUX-LABEL: define dso_local void @testVolatileBitfieldsVolatileStruct(
+// LINUX-SAME: ptr noundef [[S:%.*]]) #[[ATTR0]] {
+// LINUX-NEXT:  [[ENTRY:.*:]]
+// LINUX-NEXT:    [[S_ADDR:%.*]] = alloca ptr, align 8
+// LINUX-NEXT:    store ptr [[S]], ptr [[S_ADDR]], align 8
+// LINUX-NEXT:    [[TMP0:%.*]] = load ptr, ptr [[S_ADDR]], align 8
+// LINUX-NEXT:    [[TMP1:%.*]] = getelementptr i8, ptr [[TMP0]], i32 1
+// LINUX-NEXT:    [[TMP2:%.*]] = load i8, ptr [[TMP1]], align 1
+// LINUX-NEXT:    [[TMP3:%.*]] = and i8 [[TMP2]], 15
+// LINUX-NEXT:    store i8 [[TMP3]], ptr [[TMP1]], align 1
+// LINUX-NEXT:    [[TMP4:%.*]] = getelementptr i8, ptr [[TMP0]], i32 2
+// LINUX-NEXT:    store i8 0, ptr [[TMP4]], align 2
+// LINUX-NEXT:    [[TMP5:%.*]] = getelementptr i8, ptr [[TMP0]], i32 3
+// LINUX-NEXT:    store i8 0, ptr [[TMP5]], align 1
+// LINUX-NEXT:    [[TMP6:%.*]] = getelementptr i8, ptr [[TMP0]], i32 7
+// LINUX-NEXT:    [[TMP7:%.*]] = load i8, ptr [[TMP6]], align 1
+// LINUX-NEXT:    [[TMP8:%.*]] = and i8 [[TMP7]], 7
+// LINUX-NEXT:    store i8 [[TMP8]], ptr [[TMP6]], align 1
+// LINUX-NEXT:    ret void
+//
+// WINDOWS-LABEL: define dso_local void @testVolatileBitfieldsVolatileStruct(
+// WINDOWS-SAME: ptr noundef [[S:%.*]]) #[[ATTR0]] {
+// WINDOWS-NEXT:  [[ENTRY:.*:]]
+// WINDOWS-NEXT:    [[S_ADDR:%.*]] = alloca ptr, align 8
+// WINDOWS-NEXT:    store ptr [[S]], ptr [[S_ADDR]], align 8
+// WINDOWS-NEXT:    [[TMP0:%.*]] = load ptr, ptr [[S_ADDR]], align 8
+// WINDOWS-NEXT:    [[TMP1:%.*]] = getelementptr i8, ptr [[TMP0]], i32 0
+// WINDOWS-NEXT:    [[TMP2:%.*]] = load i8, ptr [[TMP1]], align 4
+// WINDOWS-NEXT:    [[TMP3:%.*]] = and i8 [[TMP2]], 31
+// WINDOWS-NEXT:    store i8 [[TMP3]], ptr [[TMP1]], align 4
+// WINDOWS-NEXT:    [[TMP4:%.*]] = getelementptr i8, ptr [[TMP0]], i32 1
+// WINDOWS-NEXT:    store i8 0, ptr [[TMP4]], align 1
+// WINDOWS-NEXT:    [[TMP5:%.*]] = getelementptr i8, ptr [[TMP0]], i32 2
+// WINDOWS-NEXT:    store i8 0, ptr [[TMP5]], align 2
+// WINDOWS-NEXT:    [[TMP6:%.*]] = getelementptr i8, ptr [[TMP0]], i32 3
+// WINDOWS-NEXT:    store i8 0, ptr [[TMP6]], align 1
+// WINDOWS-NEXT:    [[TMP7:%.*]] = getelementptr i8, ptr [[TMP0]], i32 4
+// WINDOWS-NEXT:    [[TMP8:%.*]] = load i8, ptr [[TMP7]], align 4
+// WINDOWS-NEXT:    [[TMP9:%.*]] = and i8 [[TMP8]], 127
+// WINDOWS-NEXT:    store i8 [[TMP9]], ptr [[TMP7]], align 4
+// WINDOWS-NEXT:    [[TMP10:%.*]] = getelementptr i8, ptr [[TMP0]], i32 5
+// WINDOWS-NEXT:    store i8 0, ptr [[TMP10]], align 1
+// WINDOWS-NEXT:    [[TMP11:%.*]] = getelementptr i8, ptr [[TMP0]], i32 6
+// WINDOWS-NEXT:    store i8 0, ptr [[TMP11]], align 2
+// WINDOWS-NEXT:    [[TMP12:%.*]] = getelementptr i8, ptr [[TMP0]], i32 7
+// WINDOWS-NEXT:    store i8 0, ptr [[TMP12]], align 1
+// WINDOWS-NEXT:    [[TMP13:%.*]] = getelementptr i8, ptr [[TMP0]], i32 11
+// WINDOWS-NEXT:    [[TMP14:%.*]] = load i8, ptr [[TMP13]], align 1
+// WINDOWS-NEXT:    [[TMP15:%.*]] = and i8 [[TMP14]], 7
+// WINDOWS-NEXT:    store i8 [[TMP15]], ptr [[TMP13]], align 1
+// WINDOWS-NEXT:    ret void
+//
+// ARM-LE-LABEL: define dso_local void @testVolatileBitfieldsVolatileStruct(
+// ARM-LE-SAME: ptr noundef [[S:%.*]]) #[[ATTR0]] {
+// ARM-LE-NEXT:  [[ENTRY:.*:]]
+// ARM-LE-NEXT:    [[S_ADDR:%.*]] = alloca ptr, align 4
+// ARM-LE-NEXT:    store ptr [[S]], ptr [[S_ADDR]], align 4
+// ARM-LE-NEXT:    [[TMP0:%.*]] = load ptr, ptr [[S_ADDR]], align 4
+// ARM-LE-NEXT:    [[TMP1:%.*]] = getelementptr i8, ptr [[TMP0]], i32 1
+// ARM-LE-NEXT:    [[TMP2:%.*]] = load i8, ptr [[TMP1]], align 1
+// ARM-LE-NEXT:    [[TMP3:%.*]] = and i8 [[TMP2]], 15
+// ARM-LE-NEXT:    store i8 [[TMP3]], ptr [[TMP1]], align 1
+// ARM-LE-NEXT:    [[TMP4:%.*]] = getelementptr i8, ptr [[TMP0]], i32 2
+// ARM-LE-NEXT:    store i8 0, ptr [[TMP4]], align 2
+// ARM-LE-NEXT:    [[TMP5:%.*]] = getelementptr i8, ptr [[TMP0]], i32 3
+// ARM-LE-NEXT:    store i8 0, ptr [[TMP5]], align 1
+// ARM-LE-NEXT:    [[TMP6:%.*]] = getelementptr i8, ptr [[TMP0]], i32 7
+// ARM-LE-NEXT:    [[TMP7:%.*]] = load i8, ptr [[TMP6]], align 1
+// ARM-LE-NEXT:    [[TMP8:%.*]] = and i8 [[TMP7]], 7
+// ARM-LE-NEXT:    store i8 [[TMP8]], ptr [[TMP6]], align 1
+// ARM-LE-NEXT:    ret void
+//
+// ARM-BE-LABEL: define dso_local void @testVolatileBitfieldsVolatileStruct(
+// ARM-BE-SAME: ptr noundef [[S:%.*]]) #[[ATTR0]] {
+// ARM-BE-NEXT:  [[ENTRY:.*:]]
+// ARM-BE-NEXT:    [[S_ADDR:%.*]] = alloca ptr, align 4
+// ARM-BE-NEXT:    store ptr [[S]], ptr [[S_ADDR]], align 4
+// ARM-BE-NEXT:    [[TMP0:%.*]] = load ptr, ptr [[S_ADDR]], align 4
+// ARM-BE-NEXT:    [[TMP1:%.*]] = getelementptr i8, ptr [[TMP0]], i32 1
+// ARM-BE-NEXT:    [[TMP2:%.*]] = load i8, ptr [[TMP1]], align 1
+// ARM-BE-NEXT:    [[TMP3:%.*]] = and i8 [[TMP2]], -16
+// ARM-BE-NEXT:    store i8 [[TMP3]], ptr [[TMP1]], align 1
+// ARM-BE-NEXT:    [[TMP4:%.*]] = getelementptr i8, ptr [[TMP0]], i32 2
+// ARM-BE-NEXT:    store i8 0, ptr [[TMP4]], align 2
+// ARM-BE-NEXT:    [[TMP5:%.*]] = getelementptr i8, ptr [[TMP0]], i32 3
+// ARM-BE-NEXT:    store i8 0, ptr [[TMP5]], align 1
+// ARM-BE-NEXT:    [[TMP6:%.*]] = getelementptr i8, ptr [[TMP0]], i32 7
+// ARM-BE-NEXT:    [[TMP7:%.*]] = load i8, ptr [[TMP6]], align 1
+// ARM-BE-NEXT:    [[TMP8:%.*]] = and i8 [[TMP7]], -32
+// ARM-BE-NEXT:    store i8 [[TMP8]], ptr [[TMP6]], align 1
+// ARM-BE-NEXT:    ret void
+//
+// AARCH64-LE-LABEL: define dso_local void @testVolatileBitfieldsVolatileStruct(
+// AARCH64-LE-SAME: ptr noundef [[S:%.*]]) #[[ATTR0]] {
+// AARCH64-LE-NEXT:  [[ENTRY:.*:]]
+// AARCH64-LE-NEXT:    [[S_ADDR:%.*]] = alloca ptr, align 8
+// AARCH64-LE-NEXT:    store ptr [[S]], ptr [[S_ADDR]], align 8
+// AARCH64-LE-NEXT:    [[TMP0:%.*]] = load ptr, ptr [[S_ADDR]], align 8
+// AARCH64-LE-NEXT:    [[TMP1:%.*]] = getelementptr i8, ptr [[TMP0]], i32 1
+// AARCH64-LE-NEXT:    [[TMP2:%.*]] = load i8, ptr [[TMP1]], align 1
+// AARCH64-LE-NEXT:    [[TMP3:%.*]] = and i8 [[TMP2]], 15
+// AARCH64-LE-NEXT:    store i8 [[TMP3]], ptr [[TMP1]], align 1
+// AARCH64-LE-NEXT:    [[TMP4:%.*]] = getelementptr i8, ptr [[TMP0]], i32 2
+// AARCH64-LE-NEXT:    store i8 0, ptr [[TMP4]], align 2
+// AARCH64-LE-NEXT:    [[TMP5:%.*]] = getelementptr i8, ptr [[TMP0]], i32 3
+// AARCH64-LE-NEXT:    store i8 0, ptr [[TMP5]], align 1
+// AARCH64-LE-NEXT:    [[TMP6:%.*]] = getelementptr i8, ptr [[TMP0]], i32 7
+// AARCH64-LE-NEXT:    [[TMP7:%.*]] = load i8, ptr [[TMP6]], align 1
+// AARCH64-LE-NEXT:    [[TMP8:%.*]] = and i8 [[TMP7]], 7
+// AARCH64-LE-NEXT:    store i8 [[TMP8]], ptr [[TMP6]], align 1
+// AARCH64-LE-NEXT:    ret void
+//
+// AARCH64-BE-LABEL: define dso_local void @testVolatileBitfieldsVolatileStruct(
+// AARCH64-BE-SAME: ptr noundef [[S:%.*]]) #[[ATTR0]] {
+// AARCH64-BE-NEXT:  [[ENTRY:.*:]]
+// AARCH64-BE-NEXT:    [[S_ADDR:%.*]] = alloca ptr, align 8
+// AARCH64-BE-NEXT:    store ptr [[S]], ptr [[S_ADDR]], align 8
+// AARCH64-BE-NEXT:    [[TMP0:%.*]] = load ptr, ptr [[S_ADDR]], align 8
+// AARCH64-BE-NEXT:    [[TMP1:%.*]] = getelementptr i8, ptr [[TMP0]], i32 1
+// AARCH64-BE-NEXT:    [[TMP2:%.*]] = load i8, ptr [[TMP1]], align 1
+// AARCH64-BE-NEXT:    [[TMP3:%.*]] = and i8 [[TMP2]], -16
+// AARCH64-BE-NEXT:    store i8 [[TMP3]], ptr [[TMP1]], align 1
+// AARCH64-BE-NEXT:    [[TMP4:%.*]] = getelementptr i8, ptr [[TMP0]], i32 2
+// AARCH64-BE-NEXT:    store i8 0, ptr [[TMP4]], align 2
+// AARCH64-BE-NEXT:    [[TMP5:%.*]] = getelementptr i8, ptr [[TMP0]], i32 3
+// AARCH64-BE-NEXT:    store i8 0, ptr [[TMP5]], align 1
+// AARCH64-BE-NEXT:    [[TMP6:%.*]] = getelementptr i8, ptr [[TMP0]], i32 7
+// AARCH64-BE-NEXT:    [[TMP7:%.*]] = load i8, ptr [[TMP6]], align 1
+// AARCH64-BE-NEXT:    [[TMP8:%.*]] = and i8 [[TMP7]], -32
+// AARCH64-BE-NEXT:    store i8 [[TMP8]], ptr [[TMP6]], align 1
+// AARCH64-BE-NEXT:    ret void
+//
+void testVolatileBitfieldsVolatileStruct(volatile struct VolatileBitfield *s) {
   __builtin_clear_padding(s);
 }
 
@@ -583,10 +1637,35 @@ void testBitFields(struct S11 *s) {
 // LINUX-NEXT:    [[I:%.*]] = alloca [4 x i32], align 16
 // LINUX-NEXT:    ret void
 //
-// WINDOWS-LABEL: define dso_local void @testArrayNoPadding() #0 {
-// WINDOWS-NEXT: entry:
-// WINDOWS-NEXT:   %i = alloca [4 x i32], align 4
-// WINDOWS-NEXT:   ret void
+// WINDOWS-LABEL: define dso_local void @testArrayNoPadding(
+// WINDOWS-SAME: ) #[[ATTR0]] {
+// WINDOWS-NEXT:  [[ENTRY:.*:]]
+// WINDOWS-NEXT:    [[I:%.*]] = alloca [4 x i32], align 4
+// WINDOWS-NEXT:    ret void
+//
+// ARM-LE-LABEL: define dso_local void @testArrayNoPadding(
+// ARM-LE-SAME: ) #[[ATTR0]] {
+// ARM-LE-NEXT:  [[ENTRY:.*:]]
+// ARM-LE-NEXT:    [[I:%.*]] = alloca [4 x i32], align 4
+// ARM-LE-NEXT:    ret void
+//
+// ARM-BE-LABEL: define dso_local void @testArrayNoPadding(
+// ARM-BE-SAME: ) #[[ATTR0]] {
+// ARM-BE-NEXT:  [[ENTRY:.*:]]
+// ARM-BE-NEXT:    [[I:%.*]] = alloca [4 x i32], align 4
+// ARM-BE-NEXT:    ret void
+//
+// AARCH64-LE-LABEL: define dso_local void @testArrayNoPadding(
+// AARCH64-LE-SAME: ) #[[ATTR0]] {
+// AARCH64-LE-NEXT:  [[ENTRY:.*:]]
+// AARCH64-LE-NEXT:    [[I:%.*]] = alloca [4 x i32], align 4
+// AARCH64-LE-NEXT:    ret void
+//
+// AARCH64-BE-LABEL: define dso_local void @testArrayNoPadding(
+// AARCH64-BE-SAME: ) #[[ATTR0]] {
+// AARCH64-BE-NEXT:  [[ENTRY:.*:]]
+// AARCH64-BE-NEXT:    [[I:%.*]] = alloca [4 x i32], align 4
+// AARCH64-BE-NEXT:    ret void
 //
 void testArrayNoPadding(void) {
   int i[4];
@@ -624,10 +1703,35 @@ void testArrayNoPadding(void) {
 // LINUX-NEXT:    store i8 0, ptr [[TMP11]], align 1
 // LINUX-NEXT:    ret void
 //
-// WINDOWS-LABEL: define dso_local void @testArrayLongDouble() #0 {
-// WINDOWS-NEXT: entry:
-// WINDOWS-NEXT:   %ld = alloca [2 x double], align 8
-// WINDOWS-NEXT:   ret void
+// WINDOWS-LABEL: define dso_local void @testArrayLongDouble(
+// WINDOWS-SAME: ) #[[ATTR0]] {
+// WINDOWS-NEXT:  [[ENTRY:.*:]]
+// WINDOWS-NEXT:    [[LD:%.*]] = alloca [2 x double], align 8
+// WINDOWS-NEXT:    ret void
+//
+// ARM-LE-LABEL: define dso_local void @testArrayLongDouble(
+// ARM-LE-SAME: ) #[[ATTR0]] {
+// ARM-LE-NEXT:  [[ENTRY:.*:]]
+// ARM-LE-NEXT:    [[LD:%.*]] = alloca [2 x double], align 8
+// ARM-LE-NEXT:    ret void
+//
+// ARM-BE-LABEL: define dso_local void @testArrayLongDouble(
+// ARM-BE-SAME: ) #[[ATTR0]] {
+// ARM-BE-NEXT:  [[ENTRY:.*:]]
+// ARM-BE-NEXT:    [[LD:%.*]] = alloca [2 x double], align 8
+// ARM-BE-NEXT:    ret void
+//
+// AARCH64-LE-LABEL: define dso_local void @testArrayLongDouble(
+// AARCH64-LE-SAME: ) #[[ATTR0]] {
+// AARCH64-LE-NEXT:  [[ENTRY:.*:]]
+// AARCH64-LE-NEXT:    [[LD:%.*]] = alloca [2 x fp128], align 16
+// AARCH64-LE-NEXT:    ret void
+//
+// AARCH64-BE-LABEL: define dso_local void @testArrayLongDouble(
+// AARCH64-BE-SAME: ) #[[ATTR0]] {
+// AARCH64-BE-NEXT:  [[ENTRY:.*:]]
+// AARCH64-BE-NEXT:    [[LD:%.*]] = alloca [2 x fp128], align 16
+// AARCH64-BE-NEXT:    ret void
 //
 void testArrayLongDouble() {
   // long double 0, [0-9] PAD [10-15]
@@ -666,34 +1770,155 @@ void testArrayLongDouble() {
 // LINUX-NEXT:    store i8 0, ptr [[TMP11]], align 1
 // LINUX-NEXT:    ret void
 //
-// WINDOWS-LABEL: define dso_local void @testArrayOfStruct() #0 {
-// WINDOWS-NEXT: entry:
-// WINDOWS-NEXT:   %s = alloca [2 x %struct.S_local], align 4
-// WINDOWS-NEXT:   %0 = getelementptr i8, ptr %s, i32 5
-// WINDOWS-NEXT:   store i8 0, ptr %0, align 1
-// WINDOWS-NEXT:   %1 = getelementptr i8, ptr %s, i32 6
-// WINDOWS-NEXT:   store i8 0, ptr %1, align 2
-// WINDOWS-NEXT:   %2 = getelementptr i8, ptr %s, i32 7
-// WINDOWS-NEXT:   store i8 0, ptr %2, align 1
-// WINDOWS-NEXT:   %3 = getelementptr i8, ptr %s, i32 13
-// WINDOWS-NEXT:   store i8 0, ptr %3, align 1
-// WINDOWS-NEXT:   %4 = getelementptr i8, ptr %s, i32 14
-// WINDOWS-NEXT:   store i8 0, ptr %4, align 2
-// WINDOWS-NEXT:   %5 = getelementptr i8, ptr %s, i32 15
-// WINDOWS-NEXT:   store i8 0, ptr %5, align 1
-// WINDOWS-NEXT:   %6 = getelementptr i8, ptr %s, i32 21
-// WINDOWS-NEXT:   store i8 0, ptr %6, align 1
-// WINDOWS-NEXT:   %7 = getelementptr i8, ptr %s, i32 22
-// WINDOWS-NEXT:   store i8 0, ptr %7, align 2
-// WINDOWS-NEXT:   %8 = getelementptr i8, ptr %s, i32 23
-// WINDOWS-NEXT:   store i8 0, ptr %8, align 1
-// WINDOWS-NEXT:   %9 = getelementptr i8, ptr %s, i32 29
-// WINDOWS-NEXT:   store i8 0, ptr %9, align 1
-// WINDOWS-NEXT:   %10 = getelementptr i8, ptr %s, i32 30
-// WINDOWS-NEXT:   store i8 0, ptr %10, align 2
-// WINDOWS-NEXT:   %11 = getelementptr i8, ptr %s, i32 31
-// WINDOWS-NEXT:   store i8 0, ptr %11, align 1
-// WINDOWS-NEXT:   ret void
+// WINDOWS-LABEL: define dso_local void @testArrayOfStruct(
+// WINDOWS-SAME: ) #[[ATTR0]] {
+// WINDOWS-NEXT:  [[ENTRY:.*:]]
+// WINDOWS-NEXT:    [[S:%.*]] = alloca [2 x [[STRUCT_S_LOCAL:%.*]]], align 4
+// WINDOWS-NEXT:    [[TMP0:%.*]] = getelementptr i8, ptr [[S]], i32 5
+// WINDOWS-NEXT:    store i8 0, ptr [[TMP0]], align 1
+// WINDOWS-NEXT:    [[TMP1:%.*]] = getelementptr i8, ptr [[S]], i32 6
+// WINDOWS-NEXT:    store i8 0, ptr [[TMP1]], align 2
+// WINDOWS-NEXT:    [[TMP2:%.*]] = getelementptr i8, ptr [[S]], i32 7
+// WINDOWS-NEXT:    store i8 0, ptr [[TMP2]], align 1
+// WINDOWS-NEXT:    [[TMP3:%.*]] = getelementptr i8, ptr [[S]], i32 13
+// WINDOWS-NEXT:    store i8 0, ptr [[TMP3]], align 1
+// WINDOWS-NEXT:    [[TMP4:%.*]] = getelementptr i8, ptr [[S]], i32 14
+// WINDOWS-NEXT:    store i8 0, ptr [[TMP4]], align 2
+// WINDOWS-NEXT:    [[TMP5:%.*]] = getelementptr i8, ptr [[S]], i32 15
+// WINDOWS-NEXT:    store i8 0, ptr [[TMP5]], align 1
+// WINDOWS-NEXT:    [[TMP6:%.*]] = getelementptr i8, ptr [[S]], i32 21
+// WINDOWS-NEXT:    store i8 0, ptr [[TMP6]], align 1
+// WINDOWS-NEXT:    [[TMP7:%.*]] = getelementptr i8, ptr [[S]], i32 22
+// WINDOWS-NEXT:    store i8 0, ptr [[TMP7]], align 2
+// WINDOWS-NEXT:    [[TMP8:%.*]] = getelementptr i8, ptr [[S]], i32 23
+// WINDOWS-NEXT:    store i8 0, ptr [[TMP8]], align 1
+// WINDOWS-NEXT:    [[TMP9:%.*]] = getelementptr i8, ptr [[S]], i32 29
+// WINDOWS-NEXT:    store i8 0, ptr [[TMP9]], align 1
+// WINDOWS-NEXT:    [[TMP10:%.*]] = getelementptr i8, ptr [[S]], i32 30
+// WINDOWS-NEXT:    store i8 0, ptr [[TMP10]], align 2
+// WINDOWS-NEXT:    [[TMP11:%.*]] = getelementptr i8, ptr [[S]], i32 31
+// WINDOWS-NEXT:    store i8 0, ptr [[TMP11]], align 1
+// WINDOWS-NEXT:    ret void
+//
+// ARM-LE-LABEL: define dso_local void @testArrayOfStruct(
+// ARM-LE-SAME: ) #[[ATTR0]] {
+// ARM-LE-NEXT:  [[ENTRY:.*:]]
+// ARM-LE-NEXT:    [[S:%.*]] = alloca [2 x [[STRUCT_S_LOCAL:%.*]]], align 4
+// ARM-LE-NEXT:    [[TMP0:%.*]] = getelementptr i8, ptr [[S]], i32 5
+// ARM-LE-NEXT:    store i8 0, ptr [[TMP0]], align 1
+// ARM-LE-NEXT:    [[TMP1:%.*]] = getelementptr i8, ptr [[S]], i32 6
+// ARM-LE-NEXT:    store i8 0, ptr [[TMP1]], align 2
+// ARM-LE-NEXT:    [[TMP2:%.*]] = getelementptr i8, ptr [[S]], i32 7
+// ARM-LE-NEXT:    store i8 0, ptr [[TMP2]], align 1
+// ARM-LE-NEXT:    [[TMP3:%.*]] = getelementptr i8, ptr [[S]], i32 13
+// ARM-LE-NEXT:    store i8 0, ptr [[TMP3]], align 1
+// ARM-LE-NEXT:    [[TMP4:%.*]] = getelementptr i8, ptr [[S]], i32 14
+// ARM-LE-NEXT:    store i8 0, ptr [[TMP4]], align 2
+// ARM-LE-NEXT:    [[TMP5:%.*]] = getelementptr i8, ptr [[S]], i32 15
+// ARM-LE-NEXT:    store i8 0, ptr [[TMP5]], align 1
+// ARM-LE-NEXT:    [[TMP6:%.*]] = getelementptr i8, ptr [[S]], i32 21
+// ARM-LE-NEXT:    store i8 0, ptr [[TMP6]], align 1
+// ARM-LE-NEXT:    [[TMP7:%.*]] = getelementptr i8, ptr [[S]], i32 22
+// ARM-LE-NEXT:    store i8 0, ptr [[TMP7]], align 2
+// ARM-LE-NEXT:    [[TMP8:%.*]] = getelementptr i8, ptr [[S]], i32 23
+// ARM-LE-NEXT:    store i8 0, ptr [[TMP8]], align 1
+// ARM-LE-NEXT:    [[TMP9:%.*]] = getelementptr i8, ptr [[S]], i32 29
+// ARM-LE-NEXT:    store i8 0, ptr [[TMP9]], align 1
+// ARM-LE-NEXT:    [[TMP10:%.*]] = getelementptr i8, ptr [[S]], i32 30
+// ARM-LE-NEXT:    store i8 0, ptr [[TMP10]], align 2
+// ARM-LE-NEXT:    [[TMP11:%.*]] = getelementptr i8, ptr [[S]], i32 31
+// ARM-LE-NEXT:    store i8 0, ptr [[TMP11]], align 1
+// ARM-LE-NEXT:    ret void
+//
+// ARM-BE-LABEL: define dso_local void @testArrayOfStruct(
+// ARM-BE-SAME: ) #[[ATTR0]] {
+// ARM-BE-NEXT:  [[ENTRY:.*:]]
+// ARM-BE-NEXT:    [[S:%.*]] = alloca [2 x [[STRUCT_S_LOCAL:%.*]]], align 4
+// ARM-BE-NEXT:    [[TMP0:%.*]] = getelementptr i8, ptr [[S]], i32 5
+// ARM-BE-NEXT:    store i8 0, ptr [[TMP0]], align 1
+// ARM-BE-NEXT:    [[TMP1:%.*]] = getelementptr i8, ptr [[S]], i32 6
+// ARM-BE-NEXT:    store i8 0, ptr [[TMP1]], align 2
+// ARM-BE-NEXT:    [[TMP2:%.*]] = getelementptr i8, ptr [[S]], i32 7
+// ARM-BE-NEXT:    store i8 0, ptr [[TMP2]], align 1
+// ARM-BE-NEXT:    [[TMP3:%.*]] = getelementptr i8, ptr [[S]], i32 13
+// ARM-BE-NEXT:    store i8 0, ptr [[TMP3]], align 1
+// ARM-BE-NEXT:    [[TMP4:%.*]] = getelementptr i8, ptr [[S]], i32 14
+// ARM-BE-NEXT:    store i8 0, ptr [[TMP4]], align 2
+// ARM-BE-NEXT:    [[TMP5:%.*]] = getelementptr i8, ptr [[S]], i32 15
+// ARM-BE-NEXT:    store i8 0, ptr [[TMP5]], align 1
+// ARM-BE-NEXT:    [[TMP6:%.*]] = getelementptr i8, ptr [[S]], i32 21
+// ARM-BE-NEXT:    store i8 0, ptr [[TMP6]], align 1
+// ARM-BE-NEXT:    [[TMP7:%.*]] = getelementptr i8, ptr [[S]], i32 22
+// ARM-BE-NEXT:    store i8 0, ptr [[TMP7]], align 2
+// ARM-BE-NEXT:    [[TMP8:%.*]] = getelementptr i8, ptr [[S]], i32 23
+// ARM-BE-NEXT:    store i8 0, ptr [[TMP8]], align 1
+// ARM-BE-NEXT:    [[TMP9:%.*]] = getelementptr i8, ptr [[S]], i32 29
+// ARM-BE-NEXT:    store i8 0, ptr [[TMP9]], align 1
+// ARM-BE-NEXT:    [[TMP10:%.*]] = getelementptr i8, ptr [[S]], i32 30
+// ARM-BE-NEXT:    store i8 0, ptr [[TMP10]], align 2
+// ARM-BE-NEXT:    [[TMP11:%.*]] = getelementptr i8, ptr [[S]], i32 31
+// ARM-BE-NEXT:    store i8 0, ptr [[TMP11]], align 1
+// ARM-BE-NEXT:    ret void
+//
+// AARCH64-LE-LABEL: define dso_local void @testArrayOfStruct(
+// AARCH64-LE-SAME: ) #[[ATTR0]] {
+// AARCH64-LE-NEXT:  [[ENTRY:.*:]]
+// AARCH64-LE-NEXT:    [[S:%.*]] = alloca [2 x [[STRUCT_S_LOCAL:%.*]]], align 4
+// AARCH64-LE-NEXT:    [[TMP0:%.*]] = getelementptr i8, ptr [[S]], i32 5
+// AARCH64-LE-NEXT:    store i8 0, ptr [[TMP0]], align 1
+// AARCH64-LE-NEXT:    [[TMP1:%.*]] = getelementptr i8, ptr [[S]], i32 6
+// AARCH64-LE-NEXT:    store i8 0, ptr [[TMP1]], align 2
+// AARCH64-LE-NEXT:    [[TMP2:%.*]] = getelementptr i8, ptr [[S]], i32 7
+// AARCH64-LE-NEXT:    store i8 0, ptr [[TMP2]], align 1
+// AARCH64-LE-NEXT:    [[TMP3:%.*]] = getelementptr i8, ptr [[S]], i32 13
+// AARCH64-LE-NEXT:    store i8 0, ptr [[TMP3]], align 1
+// AARCH64-LE-NEXT:    [[TMP4:%.*]] = getelementptr i8, ptr [[S]], i32 14
+// AARCH64-LE-NEXT:    store i8 0, ptr [[TMP4]], align 2
+// AARCH64-LE-NEXT:    [[TMP5:%.*]] = getelementptr i8, ptr [[S]], i32 15
+// AARCH64-LE-NEXT:    store i8 0, ptr [[TMP5]], align 1
+// AARCH64-LE-NEXT:    [[TMP6:%.*]] = getelementptr i8, ptr [[S]], i32 21
+// AARCH64-LE-NEXT:    store i8 0, ptr [[TMP6]], align 1
+// AARCH64-LE-NEXT:    [[TMP7:%.*]] = getelementptr i8, ptr [[S]], i32 22
+// AARCH64-LE-NEXT:    store i8 0, ptr [[TMP7]], align 2
+// AARCH64-LE-NEXT:    [[TMP8:%.*]] = getelementptr i8, ptr [[S]], i32 23
+// AARCH64-LE-NEXT:    store i8 0, ptr [[TMP8]], align 1
+// AARCH64-LE-NEXT:    [[TMP9:%.*]] = getelementptr i8, ptr [[S]], i32 29
+// AARCH64-LE-NEXT:    store i8 0, ptr [[TMP9]], align 1
+// AARCH64-LE-NEXT:    [[TMP10:%.*]] = getelementptr i8, ptr [[S]], i32 30
+// AARCH64-LE-NEXT:    store i8 0, ptr [[TMP10]], align 2
+// AARCH64-LE-NEXT:    [[TMP11:%.*]] = getelementptr i8, ptr [[S]], i32 31
+// AARCH64-LE-NEXT:    store i8 0, ptr [[TMP11]], align 1
+// AARCH64-LE-NEXT:    ret void
+//
+// AARCH64-BE-LABEL: define dso_local void @testArrayOfStruct(
+// AARCH64-BE-SAME: ) #[[ATTR0]] {
+// AARCH64-BE-NEXT:  [[ENTRY:.*:]]
+// AARCH64-BE-NEXT:    [[S:%.*]] = alloca [2 x [[STRUCT_S_LOCAL:%.*]]], align 4
+// AARCH64-BE-NEXT:    [[TMP0:%.*]] = getelementptr i8, ptr [[S]], i32 5
+// AARCH64-BE-NEXT:    store i8 0, ptr [[TMP0]], align 1
+// AARCH64-BE-NEXT:    [[TMP1:%.*]] = getelementptr i8, ptr [[S]], i32 6
+// AARCH64-BE-NEXT:    store i8 0, ptr [[TMP1]], align 2
+// AARCH64-BE-NEXT:    [[TMP2:%.*]] = getelementptr i8, ptr [[S]], i32 7
+// AARCH64-BE-NEXT:    store i8 0, ptr [[TMP2]], align 1
+// AARCH64-BE-NEXT:    [[TMP3:%.*]] = getelementptr i8, ptr [[S]], i32 13
+// AARCH64-BE-NEXT:    store i8 0, ptr [[TMP3]], align 1
+// AARCH64-BE-NEXT:    [[TMP4:%.*]] = getelementptr i8, ptr [[S]], i32 14
+// AARCH64-BE-NEXT:    store i8 0, ptr [[TMP4]], align 2
+// AARCH64-BE-NEXT:    [[TMP5:%.*]] = getelementptr i8, ptr [[S]], i32 15
+// AARCH64-BE-NEXT:    store i8 0, ptr [[TMP5]], align 1
+// AARCH64-BE-NEXT:    [[TMP6:%.*]] = getelementptr i8, ptr [[S]], i32 21
+// AARCH64-BE-NEXT:    store i8 0, ptr [[TMP6]], align 1
+// AARCH64-BE-NEXT:    [[TMP7:%.*]] = getelementptr i8, ptr [[S]], i32 22
+// AARCH64-BE-NEXT:    store i8 0, ptr [[TMP7]], align 2
+// AARCH64-BE-NEXT:    [[TMP8:%.*]] = getelementptr i8, ptr [[S]], i32 23
+// AARCH64-BE-NEXT:    store i8 0, ptr [[TMP8]], align 1
+// AARCH64-BE-NEXT:    [[TMP9:%.*]] = getelementptr i8, ptr [[S]], i32 29
+// AARCH64-BE-NEXT:    store i8 0, ptr [[TMP9]], align 1
+// AARCH64-BE-NEXT:    [[TMP10:%.*]] = getelementptr i8, ptr [[S]], i32 30
+// AARCH64-BE-NEXT:    store i8 0, ptr [[TMP10]], align 2
+// AARCH64-BE-NEXT:    [[TMP11:%.*]] = getelementptr i8, ptr [[S]], i32 31
+// AARCH64-BE-NEXT:    store i8 0, ptr [[TMP11]], align 1
+// AARCH64-BE-NEXT:    ret void
 //
 void testArrayOfStruct(void) {
   struct S_local {
@@ -736,20 +1961,85 @@ struct ArrOfStructsWithPadding {
 // LINUX-NEXT:    store i8 0, ptr [[TMP4]], align 1
 // LINUX-NEXT:    ret void
 //
-// WINDOWS-LABEL: define dso_local void @testArrOfStructsWithPadding(ptr noundef %arr) #0 {
-// WINDOWS-NEXT: entry:
-// WINDOWS-NEXT:   %arr.addr = alloca ptr, align 8
-// WINDOWS-NEXT:   store ptr %arr, ptr %arr.addr, align 8
-// WINDOWS-NEXT:   %0 = load ptr, ptr %arr.addr, align 8
-// WINDOWS-NEXT:   %1 = getelementptr i8, ptr %0, i32 1
-// WINDOWS-NEXT:   store i8 0, ptr %1, align 1
-// WINDOWS-NEXT:   %2 = getelementptr i8, ptr %0, i32 3
-// WINDOWS-NEXT:   store i8 0, ptr %2, align 1
-// WINDOWS-NEXT:   %3 = getelementptr i8, ptr %0, i32 5
-// WINDOWS-NEXT:   store i8 0, ptr %3, align 1
-// WINDOWS-NEXT:   %4 = getelementptr i8, ptr %0, i32 7
-// WINDOWS-NEXT:   store i8 0, ptr %4, align 1
-// WINDOWS-NEXT:   ret void
+// WINDOWS-LABEL: define dso_local void @testArrOfStructsWithPadding(
+// WINDOWS-SAME: ptr noundef [[ARR:%.*]]) #[[ATTR0]] {
+// WINDOWS-NEXT:  [[ENTRY:.*:]]
+// WINDOWS-NEXT:    [[ARR_ADDR:%.*]] = alloca ptr, align 8
+// WINDOWS-NEXT:    store ptr [[ARR]], ptr [[ARR_ADDR]], align 8
+// WINDOWS-NEXT:    [[TMP0:%.*]] = load ptr, ptr [[ARR_ADDR]], align 8
+// WINDOWS-NEXT:    [[TMP1:%.*]] = getelementptr i8, ptr [[TMP0]], i32 1
+// WINDOWS-NEXT:    store i8 0, ptr [[TMP1]], align 1
+// WINDOWS-NEXT:    [[TMP2:%.*]] = getelementptr i8, ptr [[TMP0]], i32 3
+// WINDOWS-NEXT:    store i8 0, ptr [[TMP2]], align 1
+// WINDOWS-NEXT:    [[TMP3:%.*]] = getelementptr i8, ptr [[TMP0]], i32 5
+// WINDOWS-NEXT:    store i8 0, ptr [[TMP3]], align 1
+// WINDOWS-NEXT:    [[TMP4:%.*]] = getelementptr i8, ptr [[TMP0]], i32 7
+// WINDOWS-NEXT:    store i8 0, ptr [[TMP4]], align 1
+// WINDOWS-NEXT:    ret void
+//
+// ARM-LE-LABEL: define dso_local void @testArrOfStructsWithPadding(
+// ARM-LE-SAME: ptr noundef [[ARR:%.*]]) #[[ATTR0]] {
+// ARM-LE-NEXT:  [[ENTRY:.*:]]
+// ARM-LE-NEXT:    [[ARR_ADDR:%.*]] = alloca ptr, align 4
+// ARM-LE-NEXT:    store ptr [[ARR]], ptr [[ARR_ADDR]], align 4
+// ARM-LE-NEXT:    [[TMP0:%.*]] = load ptr, ptr [[ARR_ADDR]], align 4
+// ARM-LE-NEXT:    [[TMP1:%.*]] = getelementptr i8, ptr [[TMP0]], i32 1
+// ARM-LE-NEXT:    store i8 0, ptr [[TMP1]], align 1
+// ARM-LE-NEXT:    [[TMP2:%.*]] = getelementptr i8, ptr [[TMP0]], i32 3
+// ARM-LE-NEXT:    store i8 0, ptr [[TMP2]], align 1
+// ARM-LE-NEXT:    [[TMP3:%.*]] = getelementptr i8, ptr [[TMP0]], i32 5
+// ARM-LE-NEXT:    store i8 0, ptr [[TMP3]], align 1
+// ARM-LE-NEXT:    [[TMP4:%.*]] = getelementptr i8, ptr [[TMP0]], i32 7
+// ARM-LE-NEXT:    store i8 0, ptr [[TMP4]], align 1
+// ARM-LE-NEXT:    ret void
+//
+// ARM-BE-LABEL: define dso_local void @testArrOfStructsWithPadding(
+// ARM-BE-SAME: ptr noundef [[ARR:%.*]]) #[[ATTR0]] {
+// ARM-BE-NEXT:  [[ENTRY:.*:]]
+// ARM-BE-NEXT:    [[ARR_ADDR:%.*]] = alloca ptr, align 4
+// ARM-BE-NEXT:    store ptr [[ARR]], ptr [[ARR_ADDR]], align 4
+// ARM-BE-NEXT:    [[TMP0:%.*]] = load ptr, ptr [[ARR_ADDR]], align 4
+// ARM-BE-NEXT:    [[TMP1:%.*]] = getelementptr i8, ptr [[TMP0]], i32 1
+// ARM-BE-NEXT:    store i8 0, ptr [[TMP1]], align 1
+// ARM-BE-NEXT:    [[TMP2:%.*]] = getelementptr i8, ptr [[TMP0]], i32 3
+// ARM-BE-NEXT:    store i8 0, ptr [[TMP2]], align 1
+// ARM-BE-NEXT:    [[TMP3:%.*]] = getelementptr i8, ptr [[TMP0]], i32 5
+// ARM-BE-NEXT:    store i8 0, ptr [[TMP3]], align 1
+// ARM-BE-NEXT:    [[TMP4:%.*]] = getelementptr i8, ptr [[TMP0]], i32 7
+// ARM-BE-NEXT:    store i8 0, ptr [[TMP4]], align 1
+// ARM-BE-NEXT:    ret void
+//
+// AARCH64-LE-LABEL: define dso_local void @testArrOfStructsWithPadding(
+// AARCH64-LE-SAME: ptr noundef [[ARR:%.*]]) #[[ATTR0]] {
+// AARCH64-LE-NEXT:  [[ENTRY:.*:]]
+// AARCH64-LE-NEXT:    [[ARR_ADDR:%.*]] = alloca ptr, align 8
+// AARCH64-LE-NEXT:    store ptr [[ARR]], ptr [[ARR_ADDR]], align 8
+// AARCH64-LE-NEXT:    [[TMP0:%.*]] = load ptr, ptr [[ARR_ADDR]], align 8
+// AARCH64-LE-NEXT:    [[TMP1:%.*]] = getelementptr i8, ptr [[TMP0]], i32 1
+// AARCH64-LE-NEXT:    store i8 0, ptr [[TMP1]], align 1
+// AARCH64-LE-NEXT:    [[TMP2:%.*]] = getelementptr i8, ptr [[TMP0]], i32 3
+// AARCH64-LE-NEXT:    store i8 0, ptr [[TMP2]], align 1
+// AARCH64-LE-NEXT:    [[TMP3:%.*]] = getelementptr i8, ptr [[TMP0]], i32 5
+// AARCH64-LE-NEXT:    store i8 0, ptr [[TMP3]], align 1
+// AARCH64-LE-NEXT:    [[TMP4:%.*]] = getelementptr i8, ptr [[TMP0]], i32 7
+// AARCH64-LE-NEXT:    store i8 0, ptr [[TMP4]], align 1
+// AARCH64-LE-NEXT:    ret void
+//
+// AARCH64-BE-LABEL: define dso_local void @testArrOfStructsWithPadding(
+// AARCH64-BE-SAME: ptr noundef [[ARR:%.*]]) #[[ATTR0]] {
+// AARCH64-BE-NEXT:  [[ENTRY:.*:]]
+// AARCH64-BE-NEXT:    [[ARR_ADDR:%.*]] = alloca ptr, align 8
+// AARCH64-BE-NEXT:    store ptr [[ARR]], ptr [[ARR_ADDR]], align 8
+// AARCH64-BE-NEXT:    [[TMP0:%.*]] = load ptr, ptr [[ARR_ADDR]], align 8
+// AARCH64-BE-NEXT:    [[TMP1:%.*]] = getelementptr i8, ptr [[TMP0]], i32 1
+// AARCH64-BE-NEXT:    store i8 0, ptr [[TMP1]], align 1
+// AARCH64-BE-NEXT:    [[TMP2:%.*]] = getelementptr i8, ptr [[TMP0]], i32 3
+// AARCH64-BE-NEXT:    store i8 0, ptr [[TMP2]], align 1
+// AARCH64-BE-NEXT:    [[TMP3:%.*]] = getelementptr i8, ptr [[TMP0]], i32 5
+// AARCH64-BE-NEXT:    store i8 0, ptr [[TMP3]], align 1
+// AARCH64-BE-NEXT:    [[TMP4:%.*]] = getelementptr i8, ptr [[TMP0]], i32 7
+// AARCH64-BE-NEXT:    store i8 0, ptr [[TMP4]], align 1
+// AARCH64-BE-NEXT:    ret void
 //
 void testArrOfStructsWithPadding(struct ArrOfStructsWithPadding *arr) {
   __builtin_clear_padding(arr);
@@ -767,16 +2057,65 @@ void testArrOfStructsWithPadding(struct ArrOfStructsWithPadding *arr) {
 // LINUX-NEXT:    store i8 0, ptr [[TMP2]], align 1
 // LINUX-NEXT:    ret void
 //
-// WINDOWS-LABEL: define dso_local void @testAtomic(ptr noundef %bar) #0 {
-// WINDOWS-NEXT: entry:
-// WINDOWS-NEXT:   %bar.addr = alloca ptr, align 8
-// WINDOWS-NEXT:   store ptr %bar, ptr %bar.addr, align 8
-// WINDOWS-NEXT:   %0 = load ptr, ptr %bar.addr, align 8
-// WINDOWS-NEXT:   %1 = getelementptr i8, ptr %0, i32 1
-// WINDOWS-NEXT:   store i8 0, ptr %1, align 1
-// WINDOWS-NEXT:   %2 = getelementptr i8, ptr %0, i32 3
-// WINDOWS-NEXT:   store i8 0, ptr %2, align 1
-// WINDOWS-NEXT:   ret void
+// WINDOWS-LABEL: define dso_local void @testAtomic(
+// WINDOWS-SAME: ptr noundef [[BAR:%.*]]) #[[ATTR0]] {
+// WINDOWS-NEXT:  [[ENTRY:.*:]]
+// WINDOWS-NEXT:    [[BAR_ADDR:%.*]] = alloca ptr, align 8
+// WINDOWS-NEXT:    store ptr [[BAR]], ptr [[BAR_ADDR]], align 8
+// WINDOWS-NEXT:    [[TMP0:%.*]] = load ptr, ptr [[BAR_ADDR]], align 8
+// WINDOWS-NEXT:    [[TMP1:%.*]] = getelementptr i8, ptr [[TMP0]], i32 1
+// WINDOWS-NEXT:    store i8 0, ptr [[TMP1]], align 1
+// WINDOWS-NEXT:    [[TMP2:%.*]] = getelementptr i8, ptr [[TMP0]], i32 3
+// WINDOWS-NEXT:    store i8 0, ptr [[TMP2]], align 1
+// WINDOWS-NEXT:    ret void
+//
+// ARM-LE-LABEL: define dso_local void @testAtomic(
+// ARM-LE-SAME: ptr noundef [[BAR:%.*]]) #[[ATTR0]] {
+// ARM-LE-NEXT:  [[ENTRY:.*:]]
+// ARM-LE-NEXT:    [[BAR_ADDR:%.*]] = alloca ptr, align 4
+// ARM-LE-NEXT:    store ptr [[BAR]], ptr [[BAR_ADDR]], align 4
+// ARM-LE-NEXT:    [[TMP0:%.*]] = load ptr, ptr [[BAR_ADDR]], align 4
+// ARM-LE-NEXT:    [[TMP1:%.*]] = getelementptr i8, ptr [[TMP0]], i32 1
+// ARM-LE-NEXT:    store i8 0, ptr [[TMP1]], align 1
+// ARM-LE-NEXT:    [[TMP2:%.*]] = getelementptr i8, ptr [[TMP0]], i32 3
+// ARM-LE-NEXT:    store i8 0, ptr [[TMP2]], align 1
+// ARM-LE-NEXT:    ret void
+//
+// ARM-BE-LABEL: define dso_local void @testAtomic(
+// ARM-BE-SAME: ptr noundef [[BAR:%.*]]) #[[ATTR0]] {
+// ARM-BE-NEXT:  [[ENTRY:.*:]]
+// ARM-BE-NEXT:    [[BAR_ADDR:%.*]] = alloca ptr, align 4
+// ARM-BE-NEXT:    store ptr [[BAR]], ptr [[BAR_ADDR]], align 4
+// ARM-BE-NEXT:    [[TMP0:%.*]] = load ptr, ptr [[BAR_ADDR]], align 4
+// ARM-BE-NEXT:    [[TMP1:%.*]] = getelementptr i8, ptr [[TMP0]], i32 1
+// ARM-BE-NEXT:    store i8 0, ptr [[TMP1]], align 1
+// ARM-BE-NEXT:    [[TMP2:%.*]] = getelementptr i8, ptr [[TMP0]], i32 3
+// ARM-BE-NEXT:    store i8 0, ptr [[TMP2]], align 1
+// ARM-BE-NEXT:    ret void
+//
+// AARCH64-LE-LABEL: define dso_local void @testAtomic(
+// AARCH64-LE-SAME: ptr noundef [[BAR:%.*]]) #[[ATTR0]] {
+// AARCH64-LE-NEXT:  [[ENTRY:.*:]]
+// AARCH64-LE-NEXT:    [[BAR_ADDR:%.*]] = alloca ptr, align 8
+// AARCH64-LE-NEXT:    store ptr [[BAR]], ptr [[BAR_ADDR]], align 8
+// AARCH64-LE-NEXT:    [[TMP0:%.*]] = load ptr, ptr [[BAR_ADDR]], align 8
+// AARCH64-LE-NEXT:    [[TMP1:%.*]] = getelementptr i8, ptr [[TMP0]], i32 1
+// AARCH64-LE-NEXT:    store i8 0, ptr [[TMP1]], align 1
+// AARCH64-LE-NEXT:    [[TMP2:%.*]] = getelementptr i8, ptr [[TMP0]], i32 3
+// AARCH64-LE-NEXT:    store i8 0, ptr [[TMP2]], align 1
+// AARCH64-LE-NEXT:    ret void
+//
+// AARCH64-BE-LABEL: define dso_local void @testAtomic(
+// AARCH64-BE-SAME: ptr noundef [[BAR:%.*]]) #[[ATTR0]] {
+// AARCH64-BE-NEXT:  [[ENTRY:.*:]]
+// AARCH64-BE-NEXT:    [[BAR_ADDR:%.*]] = alloca ptr, align 8
+// AARCH64-BE-NEXT:    store ptr [[BAR]], ptr [[BAR_ADDR]], align 8
+// AARCH64-BE-NEXT:    [[TMP0:%.*]] = load ptr, ptr [[BAR_ADDR]], align 8
+// AARCH64-BE-NEXT:    [[TMP1:%.*]] = getelementptr i8, ptr [[TMP0]], i32 1
+// AARCH64-BE-NEXT:    store i8 0, ptr [[TMP1]], align 1
+// AARCH64-BE-NEXT:    [[TMP2:%.*]] = getelementptr i8, ptr [[TMP0]], i32 3
+// AARCH64-BE-NEXT:    store i8 0, ptr [[TMP2]], align 1
+// AARCH64-BE-NEXT:    ret void
 //
 void testAtomic(_Atomic(struct Bar)* bar) {
   __builtin_clear_padding(bar);
@@ -801,20 +2140,85 @@ typedef long double LongDouble3Vec __attribute__((ext_vector_type(3)));
 // LINUX-NEXT:    store i8 0, ptr [[TMP4]], align 1
 // LINUX-NEXT:    ret void
 //
-// WINDOWS-LABEL: define dso_local void @testAttributedType(ptr noundef %v) #0 {
-// WINDOWS-NEXT: entry:
-// WINDOWS-NEXT:   %v.addr = alloca ptr, align 8
-// WINDOWS-NEXT:   store ptr %v, ptr %v.addr, align 8
-// WINDOWS-NEXT:   %0 = load ptr, ptr %v.addr, align 8
-// WINDOWS-NEXT:   %1 = getelementptr i8, ptr %0, i32 12
-// WINDOWS-NEXT:   store i8 0, ptr %1, align 4
-// WINDOWS-NEXT:   %2 = getelementptr i8, ptr %0, i32 13
-// WINDOWS-NEXT:   store i8 0, ptr %2, align 1
-// WINDOWS-NEXT:   %3 = getelementptr i8, ptr %0, i32 14
-// WINDOWS-NEXT:   store i8 0, ptr %3, align 2
-// WINDOWS-NEXT:   %4 = getelementptr i8, ptr %0, i32 15
-// WINDOWS-NEXT:   store i8 0, ptr %4, align 1
-// WINDOWS-NEXT:   ret void
+// WINDOWS-LABEL: define dso_local void @testAttributedType(
+// WINDOWS-SAME: ptr noundef [[V:%.*]]) #[[ATTR0]] {
+// WINDOWS-NEXT:  [[ENTRY:.*:]]
+// WINDOWS-NEXT:    [[V_ADDR:%.*]] = alloca ptr, align 8
+// WINDOWS-NEXT:    store ptr [[V]], ptr [[V_ADDR]], align 8
+// WINDOWS-NEXT:    [[TMP0:%.*]] = load ptr, ptr [[V_ADDR]], align 8
+// WINDOWS-NEXT:    [[TMP1:%.*]] = getelementptr i8, ptr [[TMP0]], i32 12
+// WINDOWS-NEXT:    store i8 0, ptr [[TMP1]], align 4
+// WINDOWS-NEXT:    [[TMP2:%.*]] = getelementptr i8, ptr [[TMP0]], i32 13
+// WINDOWS-NEXT:    store i8 0, ptr [[TMP2]], align 1
+// WINDOWS-NEXT:    [[TMP3:%.*]] = getelementptr i8, ptr [[TMP0]], i32 14
+// WINDOWS-NEXT:    store i8 0, ptr [[TMP3]], align 2
+// WINDOWS-NEXT:    [[TMP4:%.*]] = getelementptr i8, ptr [[TMP0]], i32 15
+// WINDOWS-NEXT:    store i8 0, ptr [[TMP4]], align 1
+// WINDOWS-NEXT:    ret void
+//
+// ARM-LE-LABEL: define dso_local void @testAttributedType(
+// ARM-LE-SAME: ptr noundef [[V:%.*]]) #[[ATTR0]] {
+// ARM-LE-NEXT:  [[ENTRY:.*:]]
+// ARM-LE-NEXT:    [[V_ADDR:%.*]] = alloca ptr, align 4
+// ARM-LE-NEXT:    store ptr [[V]], ptr [[V_ADDR]], align 4
+// ARM-LE-NEXT:    [[TMP0:%.*]] = load ptr, ptr [[V_ADDR]], align 4
+// ARM-LE-NEXT:    [[TMP1:%.*]] = getelementptr i8, ptr [[TMP0]], i32 12
+// ARM-LE-NEXT:    store i8 0, ptr [[TMP1]], align 4
+// ARM-LE-NEXT:    [[TMP2:%.*]] = getelementptr i8, ptr [[TMP0]], i32 13
+// ARM-LE-NEXT:    store i8 0, ptr [[TMP2]], align 1
+// ARM-LE-NEXT:    [[TMP3:%.*]] = getelementptr i8, ptr [[TMP0]], i32 14
+// ARM-LE-NEXT:    store i8 0, ptr [[TMP3]], align 2
+// ARM-LE-NEXT:    [[TMP4:%.*]] = getelementptr i8, ptr [[TMP0]], i32 15
+// ARM-LE-NEXT:    store i8 0, ptr [[TMP4]], align 1
+// ARM-LE-NEXT:    ret void
+//
+// ARM-BE-LABEL: define dso_local void @testAttributedType(
+// ARM-BE-SAME: ptr noundef [[V:%.*]]) #[[ATTR0]] {
+// ARM-BE-NEXT:  [[ENTRY:.*:]]
+// ARM-BE-NEXT:    [[V_ADDR:%.*]] = alloca ptr, align 4
+// ARM-BE-NEXT:    store ptr [[V]], ptr [[V_ADDR]], align 4
+// ARM-BE-NEXT:    [[TMP0:%.*]] = load ptr, ptr [[V_ADDR]], align 4
+// ARM-BE-NEXT:    [[TMP1:%.*]] = getelementptr i8, ptr [[TMP0]], i32 12
+// ARM-BE-NEXT:    store i8 0, ptr [[TMP1]], align 4
+// ARM-BE-NEXT:    [[TMP2:%.*]] = getelementptr i8, ptr [[TMP0]], i32 13
+// ARM-BE-NEXT:    store i8 0, ptr [[TMP2]], align 1
+// ARM-BE-NEXT:    [[TMP3:%.*]] = getelementptr i8, ptr [[TMP0]], i32 14
+// ARM-BE-NEXT:    store i8 0, ptr [[TMP3]], align 2
+// ARM-BE-NEXT:    [[TMP4:%.*]] = getelementptr i8, ptr [[TMP0]], i32 15
+// ARM-BE-NEXT:    store i8 0, ptr [[TMP4]], align 1
+// ARM-BE-NEXT:    ret void
+//
+// AARCH64-LE-LABEL: define dso_local void @testAttributedType(
+// AARCH64-LE-SAME: ptr noundef [[V:%.*]]) #[[ATTR0]] {
+// AARCH64-LE-NEXT:  [[ENTRY:.*:]]
+// AARCH64-LE-NEXT:    [[V_ADDR:%.*]] = alloca ptr, align 8
+// AARCH64-LE-NEXT:    store ptr [[V]], ptr [[V_ADDR]], align 8
+// AARCH64-LE-NEXT:    [[TMP0:%.*]] = load ptr, ptr [[V_ADDR]], align 8
+// AARCH64-LE-NEXT:    [[TMP1:%.*]] = getelementptr i8, ptr [[TMP0]], i32 12
+// AARCH64-LE-NEXT:    store i8 0, ptr [[TMP1]], align 4
+// AARCH64-LE-NEXT:    [[TMP2:%.*]] = getelementptr i8, ptr [[TMP0]], i32 13
+// AARCH64-LE-NEXT:    store i8 0, ptr [[TMP2]], align 1
+// AARCH64-LE-NEXT:    [[TMP3:%.*]] = getelementptr i8, ptr [[TMP0]], i32 14
+// AARCH64-LE-NEXT:    store i8 0, ptr [[TMP3]], align 2
+// AARCH64-LE-NEXT:    [[TMP4:%.*]] = getelementptr i8, ptr [[TMP0]], i32 15
+// AARCH64-LE-NEXT:    store i8 0, ptr [[TMP4]], align 1
+// AARCH64-LE-NEXT:    ret void
+//
+// AARCH64-BE-LABEL: define dso_local void @testAttributedType(
+// AARCH64-BE-SAME: ptr noundef [[V:%.*]]) #[[ATTR0]] {
+// AARCH64-BE-NEXT:  [[ENTRY:.*:]]
+// AARCH64-BE-NEXT:    [[V_ADDR:%.*]] = alloca ptr, align 8
+// AARCH64-BE-NEXT:    store ptr [[V]], ptr [[V_ADDR]], align 8
+// AARCH64-BE-NEXT:    [[TMP0:%.*]] = load ptr, ptr [[V_ADDR]], align 8
+// AARCH64-BE-NEXT:    [[TMP1:%.*]] = getelementptr i8, ptr [[TMP0]], i32 12
+// AARCH64-BE-NEXT:    store i8 0, ptr [[TMP1]], align 4
+// AARCH64-BE-NEXT:    [[TMP2:%.*]] = getelementptr i8, ptr [[TMP0]], i32 13
+// AARCH64-BE-NEXT:    store i8 0, ptr [[TMP2]], align 1
+// AARCH64-BE-NEXT:    [[TMP3:%.*]] = getelementptr i8, ptr [[TMP0]], i32 14
+// AARCH64-BE-NEXT:    store i8 0, ptr [[TMP3]], align 2
+// AARCH64-BE-NEXT:    [[TMP4:%.*]] = getelementptr i8, ptr [[TMP0]], i32 15
+// AARCH64-BE-NEXT:    store i8 0, ptr [[TMP4]], align 1
+// AARCH64-BE-NEXT:    ret void
 //
 void testAttributedType(Float3Vec* v) {
   __builtin_clear_padding(v);
@@ -896,28 +2300,157 @@ void testAttributedType(Float3Vec* v) {
 // LINUX-NEXT:    store i8 0, ptr [[TMP34]], align 1
 // LINUX-NEXT:    ret void
 //
-// WINDOWS-LABEL: define dso_local void @testAttributedLongDoubleType(ptr noundef %v) #0 {
-// WINDOWS-NEXT: entry:
-// WINDOWS-NEXT:   %v.addr = alloca ptr, align 8
-// WINDOWS-NEXT:   store ptr %v, ptr %v.addr, align 8
-// WINDOWS-NEXT:   %0 = load ptr, ptr %v.addr, align 8
-// WINDOWS-NEXT:   %1 = getelementptr i8, ptr %0, i32 24
-// WINDOWS-NEXT:   store i8 0, ptr %1, align 8
-// WINDOWS-NEXT:   %2 = getelementptr i8, ptr %0, i32 25
-// WINDOWS-NEXT:   store i8 0, ptr %2, align 1
-// WINDOWS-NEXT:   %3 = getelementptr i8, ptr %0, i32 26
-// WINDOWS-NEXT:   store i8 0, ptr %3, align 2
-// WINDOWS-NEXT:   %4 = getelementptr i8, ptr %0, i32 27
-// WINDOWS-NEXT:   store i8 0, ptr %4, align 1
-// WINDOWS-NEXT:   %5 = getelementptr i8, ptr %0, i32 28
-// WINDOWS-NEXT:   store i8 0, ptr %5, align 4
-// WINDOWS-NEXT:   %6 = getelementptr i8, ptr %0, i32 29
-// WINDOWS-NEXT:   store i8 0, ptr %6, align 1
-// WINDOWS-NEXT:   %7 = getelementptr i8, ptr %0, i32 30
-// WINDOWS-NEXT:   store i8 0, ptr %7, align 2
-// WINDOWS-NEXT:   %8 = getelementptr i8, ptr %0, i32 31
-// WINDOWS-NEXT:   store i8 0, ptr %8, align 1
-// WINDOWS-NEXT:   ret void
+// WINDOWS-LABEL: define dso_local void @testAttributedLongDoubleType(
+// WINDOWS-SAME: ptr noundef [[V:%.*]]) #[[ATTR0]] {
+// WINDOWS-NEXT:  [[ENTRY:.*:]]
+// WINDOWS-NEXT:    [[V_ADDR:%.*]] = alloca ptr, align 8
+// WINDOWS-NEXT:    store ptr [[V]], ptr [[V_ADDR]], align 8
+// WINDOWS-NEXT:    [[TMP0:%.*]] = load ptr, ptr [[V_ADDR]], align 8
+// WINDOWS-NEXT:    [[TMP1:%.*]] = getelementptr i8, ptr [[TMP0]], i32 24
+// WINDOWS-NEXT:    store i8 0, ptr [[TMP1]], align 8
+// WINDOWS-NEXT:    [[TMP2:%.*]] = getelementptr i8, ptr [[TMP0]], i32 25
+// WINDOWS-NEXT:    store i8 0, ptr [[TMP2]], align 1
+// WINDOWS-NEXT:    [[TMP3:%.*]] = getelementptr i8, ptr [[TMP0]], i32 26
+// WINDOWS-NEXT:    store i8 0, ptr [[TMP3]], align 2
+// WINDOWS-NEXT:    [[TMP4:%.*]] = getelementptr i8, ptr [[TMP0]], i32 27
+// WINDOWS-NEXT:    store i8 0, ptr [[TMP4]], align 1
+// WINDOWS-NEXT:    [[TMP5:%.*]] = getelementptr i8, ptr [[TMP0]], i32 28
+// WINDOWS-NEXT:    store i8 0, ptr [[TMP5]], align 4
+// WINDOWS-NEXT:    [[TMP6:%.*]] = getelementptr i8, ptr [[TMP0]], i32 29
+// WINDOWS-NEXT:    store i8 0, ptr [[TMP6]], align 1
+// WINDOWS-NEXT:    [[TMP7:%.*]] = getelementptr i8, ptr [[TMP0]], i32 30
+// WINDOWS-NEXT:    store i8 0, ptr [[TMP7]], align 2
+// WINDOWS-NEXT:    [[TMP8:%.*]] = getelementptr i8, ptr [[TMP0]], i32 31
+// WINDOWS-NEXT:    store i8 0, ptr [[TMP8]], align 1
+// WINDOWS-NEXT:    ret void
+//
+// ARM-LE-LABEL: define dso_local void @testAttributedLongDoubleType(
+// ARM-LE-SAME: ptr noundef [[V:%.*]]) #[[ATTR0]] {
+// ARM-LE-NEXT:  [[ENTRY:.*:]]
+// ARM-LE-NEXT:    [[V_ADDR:%.*]] = alloca ptr, align 4
+// ARM-LE-NEXT:    store ptr [[V]], ptr [[V_ADDR]], align 4
+// ARM-LE-NEXT:    [[TMP0:%.*]] = load ptr, ptr [[V_ADDR]], align 4
+// ARM-LE-NEXT:    [[TMP1:%.*]] = getelementptr i8, ptr [[TMP0]], i32 24
+// ARM-LE-NEXT:    store i8 0, ptr [[TMP1]], align 8
+// ARM-LE-NEXT:    [[TMP2:%.*]] = getelementptr i8, ptr [[TMP0]], i32 25
+// ARM-LE-NEXT:    store i8 0, ptr [[TMP2]], align 1
+// ARM-LE-NEXT:    [[TMP3:%.*]] = getelementptr i8, ptr [[TMP0]], i32 26
+// ARM-LE-NEXT:    store i8 0, ptr [[TMP3]], align 2
+// ARM-LE-NEXT:    [[TMP4:%.*]] = getelementptr i8, ptr [[TMP0]], i32 27
+// ARM-LE-NEXT:    store i8 0, ptr [[TMP4]], align 1
+// ARM-LE-NEXT:    [[TMP5:%.*]] = getelementptr i8, ptr [[TMP0]], i32 28
+// ARM-LE-NEXT:    store i8 0, ptr [[TMP5]], align 4
+// ARM-LE-NEXT:    [[TMP6:%.*]] = getelementptr i8, ptr [[TMP0]], i32 29
+// ARM-LE-NEXT:    store i8 0, ptr [[TMP6]], align 1
+// ARM-LE-NEXT:    [[TMP7:%.*]] = getelementptr i8, ptr [[TMP0]], i32 30
+// ARM-LE-NEXT:    store i8 0, ptr [[TMP7]], align 2
+// ARM-LE-NEXT:    [[TMP8:%.*]] = getelementptr i8, ptr [[TMP0]], i32 31
+// ARM-LE-NEXT:    store i8 0, ptr [[TMP8]], align 1
+// ARM-LE-NEXT:    ret void
+//
+// ARM-BE-LABEL: define dso_local void @testAttributedLongDoubleType(
+// ARM-BE-SAME: ptr noundef [[V:%.*]]) #[[ATTR0]] {
+// ARM-BE-NEXT:  [[ENTRY:.*:]]
+// ARM-BE-NEXT:    [[V_ADDR:%.*]] = alloca ptr, align 4
+// ARM-BE-NEXT:    store ptr [[V]], ptr [[V_ADDR]], align 4
+// ARM-BE-NEXT:    [[TMP0:%.*]] = load ptr, ptr [[V_ADDR]], align 4
+// ARM-BE-NEXT:    [[TMP1:%.*]] = getelementptr i8, ptr [[TMP0]], i32 24
+// ARM-BE-NEXT:    store i8 0, ptr [[TMP1]], align 8
+// ARM-BE-NEXT:    [[TMP2:%.*]] = getelementptr i8, ptr [[TMP0]], i32 25
+// ARM-BE-NEXT:    store i8 0, ptr [[TMP2]], align 1
+// ARM-BE-NEXT:    [[TMP3:%.*]] = getelementptr i8, ptr [[TMP0]], i32 26
+// ARM-BE-NEXT:    store i8 0, ptr [[TMP3]], align 2
+// ARM-BE-NEXT:    [[TMP4:%.*]] = getelementptr i8, ptr [[TMP0]], i32 27
+// ARM-BE-NEXT:    store i8 0, ptr [[TMP4]], align 1
+// ARM-BE-NEXT:    [[TMP5:%.*]] = getelementptr i8, ptr [[TMP0]], i32 28
+// ARM-BE-NEXT:    store i8 0, ptr [[TMP5]], align 4
+// ARM-BE-NEXT:    [[TMP6:%.*]] = getelementptr i8, ptr [[TMP0]], i32 29
+// ARM-BE-NEXT:    store i8 0, ptr [[TMP6]], align 1
+// ARM-BE-NEXT:    [[TMP7:%.*]] = getelementptr i8, ptr [[TMP0]], i32 30
+// ARM-BE-NEXT:    store i8 0, ptr [[TMP7]], align 2
+// ARM-BE-NEXT:    [[TMP8:%.*]] = getelementptr i8, ptr [[TMP0]], i32 31
+// ARM-BE-NEXT:    store i8 0, ptr [[TMP8]], align 1
+// ARM-BE-NEXT:    ret void
+//
+// AARCH64-LE-LABEL: define dso_local void @testAttributedLongDoubleType(
+// AARCH64-LE-SAME: ptr noundef [[V:%.*]]) #[[ATTR0]] {
+// AARCH64-LE-NEXT:  [[ENTRY:.*:]]
+// AARCH64-LE-NEXT:    [[V_ADDR:%.*]] = alloca ptr, align 8
+// AARCH64-LE-NEXT:    store ptr [[V]], ptr [[V_ADDR]], align 8
+// AARCH64-LE-NEXT:    [[TMP0:%.*]] = load ptr, ptr [[V_ADDR]], align 8
+// AARCH64-LE-NEXT:    [[TMP1:%.*]] = getelementptr i8, ptr [[TMP0]], i32 48
+// AARCH64-LE-NEXT:    store i8 0, ptr [[TMP1]], align 16
+// AARCH64-LE-NEXT:    [[TMP2:%.*]] = getelementptr i8, ptr [[TMP0]], i32 49
+// AARCH64-LE-NEXT:    store i8 0, ptr [[TMP2]], align 1
+// AARCH64-LE-NEXT:    [[TMP3:%.*]] = getelementptr i8, ptr [[TMP0]], i32 50
+// AARCH64-LE-NEXT:    store i8 0, ptr [[TMP3]], align 2
+// AARCH64-LE-NEXT:    [[TMP4:%.*]] = getelementptr i8, ptr [[TMP0]], i32 51
+// AARCH64-LE-NEXT:    store i8 0, ptr [[TMP4]], align 1
+// AARCH64-LE-NEXT:    [[TMP5:%.*]] = getelementptr i8, ptr [[TMP0]], i32 52
+// AARCH64-LE-NEXT:    store i8 0, ptr [[TMP5]], align 4
+// AARCH64-LE-NEXT:    [[TMP6:%.*]] = getelementptr i8, ptr [[TMP0]], i32 53
+// AARCH64-LE-NEXT:    store i8 0, ptr [[TMP6]], align 1
+// AARCH64-LE-NEXT:    [[TMP7:%.*]] = getelementptr i8, ptr [[TMP0]], i32 54
+// AARCH64-LE-NEXT:    store i8 0, ptr [[TMP7]], align 2
+// AARCH64-LE-NEXT:    [[TMP8:%.*]] = getelementptr i8, ptr [[TMP0]], i32 55
+// AARCH64-LE-NEXT:    store i8 0, ptr [[TMP8]], align 1
+// AARCH64-LE-NEXT:    [[TMP9:%.*]] = getelementptr i8, ptr [[TMP0]], i32 56
+// AARCH64-LE-NEXT:    store i8 0, ptr [[TMP9]], align 8
+// AARCH64-LE-NEXT:    [[TMP10:%.*]] = getelementptr i8, ptr [[TMP0]], i32 57
+// AARCH64-LE-NEXT:    store i8 0, ptr [[TMP10]], align 1
+// AARCH64-LE-NEXT:    [[TMP11:%.*]] = getelementptr i8, ptr [[TMP0]], i32 58
+// AARCH64-LE-NEXT:    store i8 0, ptr [[TMP11]], align 2
+// AARCH64-LE-NEXT:    [[TMP12:%.*]] = getelementptr i8, ptr [[TMP0]], i32 59
+// AARCH64-LE-NEXT:    store i8 0, ptr [[TMP12]], align 1
+// AARCH64-LE-NEXT:    [[TMP13:%.*]] = getelementptr i8, ptr [[TMP0]], i32 60
+// AARCH64-LE-NEXT:    store i8 0, ptr [[TMP13]], align 4
+// AARCH64-LE-NEXT:    [[TMP14:%.*]] = getelementptr i8, ptr [[TMP0]], i32 61
+// AARCH64-LE-NEXT:    store i8 0, ptr [[TMP14]], align 1
+// AARCH64-LE-NEXT:    [[TMP15:%.*]] = getelementptr i8, ptr [[TMP0]], i32 62
+// AARCH64-LE-NEXT:    store i8 0, ptr [[TMP15]], align 2
+// AARCH64-LE-NEXT:    [[TMP16:%.*]] = getelementptr i8, ptr [[TMP0]], i32 63
+// AARCH64-LE-NEXT:    store i8 0, ptr [[TMP16]], align 1
+// AARCH64-LE-NEXT:    ret void
+//
+// AARCH64-BE-LABEL: define dso_local void @testAttributedLongDoubleType(
+// AARCH64-BE-SAME: ptr noundef [[V:%.*]]) #[[ATTR0]] {
+// AARCH64-BE-NEXT:  [[ENTRY:.*:]]
+// AARCH64-BE-NEXT:    [[V_ADDR:%.*]] = alloca ptr, align 8
+// AARCH64-BE-NEXT:    store ptr [[V]], ptr [[V_ADDR]], align 8
+// AARCH64-BE-NEXT:    [[TMP0:%.*]] = load ptr, ptr [[V_ADDR]], align 8
+// AARCH64-BE-NEXT:    [[TMP1:%.*]] = getelementptr i8, ptr [[TMP0]], i32 48
+// AARCH64-BE-NEXT:    store i8 0, ptr [[TMP1]], align 16
+// AARCH64-BE-NEXT:    [[TMP2:%.*]] = getelementptr i8, ptr [[TMP0]], i32 49
+// AARCH64-BE-NEXT:    store i8 0, ptr [[TMP2]], align 1
+// AARCH64-BE-NEXT:    [[TMP3:%.*]] = getelementptr i8, ptr [[TMP0]], i32 50
+// AARCH64-BE-NEXT:    store i8 0, ptr [[TMP3]], align 2
+// AARCH64-BE-NEXT:    [[TMP4:%.*]] = getelementptr i8, ptr [[TMP0]], i32 51
+// AARCH64-BE-NEXT:    store i8 0, ptr [[TMP4]], align 1
+// AARCH64-BE-NEXT:    [[TMP5:%.*]] = getelementptr i8, ptr [[TMP0]], i32 52
+// AARCH64-BE-NEXT:    store i8 0, ptr [[TMP5]], align 4
+// AARCH64-BE-NEXT:    [[TMP6:%.*]] = getelementptr i8, ptr [[TMP0]], i32 53
+// AARCH64-BE-NEXT:    store i8 0, ptr [[TMP6]], align 1
+// AARCH64-BE-NEXT:    [[TMP7:%.*]] = getelementptr i8, ptr [[TMP0]], i32 54
+// AARCH64-BE-NEXT:    store i8 0, ptr [[TMP7]], align 2
+// AARCH64-BE-NEXT:    [[TMP8:%.*]] = getelementptr i8, ptr [[TMP0]], i32 55
+// AARCH64-BE-NEXT:    store i8 0, ptr [[TMP8]], align 1
+// AARCH64-BE-NEXT:    [[TMP9:%.*]] = getelementptr i8, ptr [[TMP0]], i32 56
+// AARCH64-BE-NEXT:    store i8 0, ptr [[TMP9]], align 8
+// AARCH64-BE-NEXT:    [[TMP10:%.*]] = getelementptr i8, ptr [[TMP0]], i32 57
+// AARCH64-BE-NEXT:    store i8 0, ptr [[TMP10]], align 1
+// AARCH64-BE-NEXT:    [[TMP11:%.*]] = getelementptr i8, ptr [[TMP0]], i32 58
+// AARCH64-BE-NEXT:    store i8 0, ptr [[TMP11]], align 2
+// AARCH64-BE-NEXT:    [[TMP12:%.*]] = getelementptr i8, ptr [[TMP0]], i32 59
+// AARCH64-BE-NEXT:    store i8 0, ptr [[TMP12]], align 1
+// AARCH64-BE-NEXT:    [[TMP13:%.*]] = getelementptr i8, ptr [[TMP0]], i32 60
+// AARCH64-BE-NEXT:    store i8 0, ptr [[TMP13]], align 4
+// AARCH64-BE-NEXT:    [[TMP14:%.*]] = getelementptr i8, ptr [[TMP0]], i32 61
+// AARCH64-BE-NEXT:    store i8 0, ptr [[TMP14]], align 1
+// AARCH64-BE-NEXT:    [[TMP15:%.*]] = getelementptr i8, ptr [[TMP0]], i32 62
+// AARCH64-BE-NEXT:    store i8 0, ptr [[TMP15]], align 2
+// AARCH64-BE-NEXT:    [[TMP16:%.*]] = getelementptr i8, ptr [[TMP0]], i32 63
+// AARCH64-BE-NEXT:    store i8 0, ptr [[TMP16]], align 1
+// AARCH64-BE-NEXT:    ret void
 //
 void testAttributedLongDoubleType(LongDouble3Vec *v) {
   // long double elements occupy [0-9], [16-25], [32-41] on x86.
@@ -942,20 +2475,69 @@ struct UnnamedBitfieldSingleBit {
 // LINUX-NEXT:    store i8 [[TMP3]], ptr [[TMP1]], align 1
 // LINUX-NEXT:    ret void
 //
-// WINDOWS-LABEL: define dso_local void @testUnnamedBitfieldSingleBit(ptr noundef %s) #0 {
-// WINDOWS-NEXT: entry:
-// WINDOWS-NEXT:   %s.addr = alloca ptr, align 8
-// WINDOWS-NEXT:   store ptr %s, ptr %s.addr, align 8
-// WINDOWS-NEXT:   %0 = load ptr, ptr %s.addr, align 8
-// WINDOWS-NEXT:   %1 = getelementptr i8, ptr %0, i32 0
-// WINDOWS-NEXT:   %2 = load i8, ptr %1, align 1
-// WINDOWS-NEXT:   %3 = and i8 %2, -9
-// WINDOWS-NEXT:   store i8 %3, ptr %1, align 1
-// WINDOWS-NEXT:   ret void
+// WINDOWS-LABEL: define dso_local void @testUnnamedBitfieldSingleBit(
+// WINDOWS-SAME: ptr noundef [[S:%.*]]) #[[ATTR0]] {
+// WINDOWS-NEXT:  [[ENTRY:.*:]]
+// WINDOWS-NEXT:    [[S_ADDR:%.*]] = alloca ptr, align 8
+// WINDOWS-NEXT:    store ptr [[S]], ptr [[S_ADDR]], align 8
+// WINDOWS-NEXT:    [[TMP0:%.*]] = load ptr, ptr [[S_ADDR]], align 8
+// WINDOWS-NEXT:    [[TMP1:%.*]] = getelementptr i8, ptr [[TMP0]], i32 0
+// WINDOWS-NEXT:    [[TMP2:%.*]] = load i8, ptr [[TMP1]], align 1
+// WINDOWS-NEXT:    [[TMP3:%.*]] = and i8 [[TMP2]], -9
+// WINDOWS-NEXT:    store i8 [[TMP3]], ptr [[TMP1]], align 1
+// WINDOWS-NEXT:    ret void
+//
+// ARM-LE-LABEL: define dso_local void @testUnnamedBitfieldSingleBit(
+// ARM-LE-SAME: ptr noundef [[S:%.*]]) #[[ATTR0]] {
+// ARM-LE-NEXT:  [[ENTRY:.*:]]
+// ARM-LE-NEXT:    [[S_ADDR:%.*]] = alloca ptr, align 4
+// ARM-LE-NEXT:    store ptr [[S]], ptr [[S_ADDR]], align 4
+// ARM-LE-NEXT:    [[TMP0:%.*]] = load ptr, ptr [[S_ADDR]], align 4
+// ARM-LE-NEXT:    [[TMP1:%.*]] = getelementptr i8, ptr [[TMP0]], i32 0
+// ARM-LE-NEXT:    [[TMP2:%.*]] = load i8, ptr [[TMP1]], align 1
+// ARM-LE-NEXT:    [[TMP3:%.*]] = and i8 [[TMP2]], -9
+// ARM-LE-NEXT:    store i8 [[TMP3]], ptr [[TMP1]], align 1
+// ARM-LE-NEXT:    ret void
+//
+// ARM-BE-LABEL: define dso_local void @testUnnamedBitfieldSingleBit(
+// ARM-BE-SAME: ptr noundef [[S:%.*]]) #[[ATTR0]] {
+// ARM-BE-NEXT:  [[ENTRY:.*:]]
+// ARM-BE-NEXT:    [[S_ADDR:%.*]] = alloca ptr, align 4
+// ARM-BE-NEXT:    store ptr [[S]], ptr [[S_ADDR]], align 4
+// ARM-BE-NEXT:    [[TMP0:%.*]] = load ptr, ptr [[S_ADDR]], align 4
+// ARM-BE-NEXT:    [[TMP1:%.*]] = getelementptr i8, ptr [[TMP0]], i32 0
+// ARM-BE-NEXT:    [[TMP2:%.*]] = load i8, ptr [[TMP1]], align 1
+// ARM-BE-NEXT:    [[TMP3:%.*]] = and i8 [[TMP2]], -17
+// ARM-BE-NEXT:    store i8 [[TMP3]], ptr [[TMP1]], align 1
+// ARM-BE-NEXT:    ret void
+//
+// AARCH64-LE-LABEL: define dso_local void @testUnnamedBitfieldSingleBit(
+// AARCH64-LE-SAME: ptr noundef [[S:%.*]]) #[[ATTR0]] {
+// AARCH64-LE-NEXT:  [[ENTRY:.*:]]
+// AARCH64-LE-NEXT:    [[S_ADDR:%.*]] = alloca ptr, align 8
+// AARCH64-LE-NEXT:    store ptr [[S]], ptr [[S_ADDR]], align 8
+// AARCH64-LE-NEXT:    [[TMP0:%.*]] = load ptr, ptr [[S_ADDR]], align 8
+// AARCH64-LE-NEXT:    [[TMP1:%.*]] = getelementptr i8, ptr [[TMP0]], i32 0
+// AARCH64-LE-NEXT:    [[TMP2:%.*]] = load i8, ptr [[TMP1]], align 1
+// AARCH64-LE-NEXT:    [[TMP3:%.*]] = and i8 [[TMP2]], -9
+// AARCH64-LE-NEXT:    store i8 [[TMP3]], ptr [[TMP1]], align 1
+// AARCH64-LE-NEXT:    ret void
+//
+// AARCH64-BE-LABEL: define dso_local void @testUnnamedBitfieldSingleBit(
+// AARCH64-BE-SAME: ptr noundef [[S:%.*]]) #[[ATTR0]] {
+// AARCH64-BE-NEXT:  [[ENTRY:.*:]]
+// AARCH64-BE-NEXT:    [[S_ADDR:%.*]] = alloca ptr, align 8
+// AARCH64-BE-NEXT:    store ptr [[S]], ptr [[S_ADDR]], align 8
+// AARCH64-BE-NEXT:    [[TMP0:%.*]] = load ptr, ptr [[S_ADDR]], align 8
+// AARCH64-BE-NEXT:    [[TMP1:%.*]] = getelementptr i8, ptr [[TMP0]], i32 0
+// AARCH64-BE-NEXT:    [[TMP2:%.*]] = load i8, ptr [[TMP1]], align 1
+// AARCH64-BE-NEXT:    [[TMP3:%.*]] = and i8 [[TMP2]], -17
+// AARCH64-BE-NEXT:    store i8 [[TMP3]], ptr [[TMP1]], align 1
+// AARCH64-BE-NEXT:    ret void
 //
 void testUnnamedBitfieldSingleBit(struct UnnamedBitfieldSingleBit *s) {
-  // byte 0: a[bits 0-2], unnamed[bit 3], b[bits 4-7]
-  // bit 3 must be cleared
+  // Little endian: a[bits 0-2], unnamed[bit 3], b[bits 4-7].
+  // Clear bit 3 on little endian, or bit 4 on big endian.
   // Mask: 0b11110111 == -9
   __builtin_clear_padding(s);
 }
@@ -978,16 +2560,65 @@ struct UnnamedBitfieldMiddle {
 // LINUX-NEXT:    store i8 [[TMP3]], ptr [[TMP1]], align 1
 // LINUX-NEXT:    ret void
 //
-// WINDOWS-LABEL: define dso_local void @testUnnamedBitfieldMiddle(ptr noundef %s) #0 {
-// WINDOWS-NEXT: entry:
-// WINDOWS-NEXT:   %s.addr = alloca ptr, align 8
-// WINDOWS-NEXT:   store ptr %s, ptr %s.addr, align 8
-// WINDOWS-NEXT:   %0 = load ptr, ptr %s.addr, align 8
-// WINDOWS-NEXT:   %1 = getelementptr i8, ptr %0, i32 0
-// WINDOWS-NEXT:   %2 = load i8, ptr %1, align 1
-// WINDOWS-NEXT:   %3 = and i8 %2, -25
-// WINDOWS-NEXT:   store i8 %3, ptr %1, align 1
-// WINDOWS-NEXT:   ret void
+// WINDOWS-LABEL: define dso_local void @testUnnamedBitfieldMiddle(
+// WINDOWS-SAME: ptr noundef [[S:%.*]]) #[[ATTR0]] {
+// WINDOWS-NEXT:  [[ENTRY:.*:]]
+// WINDOWS-NEXT:    [[S_ADDR:%.*]] = alloca ptr, align 8
+// WINDOWS-NEXT:    store ptr [[S]], ptr [[S_ADDR]], align 8
+// WINDOWS-NEXT:    [[TMP0:%.*]] = load ptr, ptr [[S_ADDR]], align 8
+// WINDOWS-NEXT:    [[TMP1:%.*]] = getelementptr i8, ptr [[TMP0]], i32 0
+// WINDOWS-NEXT:    [[TMP2:%.*]] = load i8, ptr [[TMP1]], align 1
+// WINDOWS-NEXT:    [[TMP3:%.*]] = and i8 [[TMP2]], -25
+// WINDOWS-NEXT:    store i8 [[TMP3]], ptr [[TMP1]], align 1
+// WINDOWS-NEXT:    ret void
+//
+// ARM-LE-LABEL: define dso_local void @testUnnamedBitfieldMiddle(
+// ARM-LE-SAME: ptr noundef [[S:%.*]]) #[[ATTR0]] {
+// ARM-LE-NEXT:  [[ENTRY:.*:]]
+// ARM-LE-NEXT:    [[S_ADDR:%.*]] = alloca ptr, align 4
+// ARM-LE-NEXT:    store ptr [[S]], ptr [[S_ADDR]], align 4
+// ARM-LE-NEXT:    [[TMP0:%.*]] = load ptr, ptr [[S_ADDR]], align 4
+// ARM-LE-NEXT:    [[TMP1:%.*]] = getelementptr i8, ptr [[TMP0]], i32 0
+// ARM-LE-NEXT:    [[TMP2:%.*]] = load i8, ptr [[TMP1]], align 1
+// ARM-LE-NEXT:    [[TMP3:%.*]] = and i8 [[TMP2]], -25
+// ARM-LE-NEXT:    store i8 [[TMP3]], ptr [[TMP1]], align 1
+// ARM-LE-NEXT:    ret void
+//
+// ARM-BE-LABEL: define dso_local void @testUnnamedBitfieldMiddle(
+// ARM-BE-SAME: ptr noundef [[S:%.*]]) #[[ATTR0]] {
+// ARM-BE-NEXT:  [[ENTRY:.*:]]
+// ARM-BE-NEXT:    [[S_ADDR:%.*]] = alloca ptr, align 4
+// ARM-BE-NEXT:    store ptr [[S]], ptr [[S_ADDR]], align 4
+// ARM-BE-NEXT:    [[TMP0:%.*]] = load ptr, ptr [[S_ADDR]], align 4
+// ARM-BE-NEXT:    [[TMP1:%.*]] = getelementptr i8, ptr [[TMP0]], i32 0
+// ARM-BE-NEXT:    [[TMP2:%.*]] = load i8, ptr [[TMP1]], align 1
+// ARM-BE-NEXT:    [[TMP3:%.*]] = and i8 [[TMP2]], -25
+// ARM-BE-NEXT:    store i8 [[TMP3]], ptr [[TMP1]], align 1
+// ARM-BE-NEXT:    ret void
+//
+// AARCH64-LE-LABEL: define dso_local void @testUnnamedBitfieldMiddle(
+// AARCH64-LE-SAME: ptr noundef [[S:%.*]]) #[[ATTR0]] {
+// AARCH64-LE-NEXT:  [[ENTRY:.*:]]
+// AARCH64-LE-NEXT:    [[S_ADDR:%.*]] = alloca ptr, align 8
+// AARCH64-LE-NEXT:    store ptr [[S]], ptr [[S_ADDR]], align 8
+// AARCH64-LE-NEXT:    [[TMP0:%.*]] = load ptr, ptr [[S_ADDR]], align 8
+// AARCH64-LE-NEXT:    [[TMP1:%.*]] = getelementptr i8, ptr [[TMP0]], i32 0
+// AARCH64-LE-NEXT:    [[TMP2:%.*]] = load i8, ptr [[TMP1]], align 1
+// AARCH64-LE-NEXT:    [[TMP3:%.*]] = and i8 [[TMP2]], -25
+// AARCH64-LE-NEXT:    store i8 [[TMP3]], ptr [[TMP1]], align 1
+// AARCH64-LE-NEXT:    ret void
+//
+// AARCH64-BE-LABEL: define dso_local void @testUnnamedBitfieldMiddle(
+// AARCH64-BE-SAME: ptr noundef [[S:%.*]]) #[[ATTR0]] {
+// AARCH64-BE-NEXT:  [[ENTRY:.*:]]
+// AARCH64-BE-NEXT:    [[S_ADDR:%.*]] = alloca ptr, align 8
+// AARCH64-BE-NEXT:    store ptr [[S]], ptr [[S_ADDR]], align 8
+// AARCH64-BE-NEXT:    [[TMP0:%.*]] = load ptr, ptr [[S_ADDR]], align 8
+// AARCH64-BE-NEXT:    [[TMP1:%.*]] = getelementptr i8, ptr [[TMP0]], i32 0
+// AARCH64-BE-NEXT:    [[TMP2:%.*]] = load i8, ptr [[TMP1]], align 1
+// AARCH64-BE-NEXT:    [[TMP3:%.*]] = and i8 [[TMP2]], -25
+// AARCH64-BE-NEXT:    store i8 [[TMP3]], ptr [[TMP1]], align 1
+// AARCH64-BE-NEXT:    ret void
 //
 void testUnnamedBitfieldMiddle(struct UnnamedBitfieldMiddle *s) {
   // byte 0: a[0-2], unnamed[3-4], b[5-7]
@@ -1018,20 +2649,85 @@ struct UnnamedBitfieldSurrounding {
 // LINUX-NEXT:    store i8 [[TMP6]], ptr [[TMP4]], align 1
 // LINUX-NEXT:    ret void
 //
-// WINDOWS-LABEL: define dso_local void @testUnnamedBitfieldSurrounding(ptr noundef %s) #0 {
-// WINDOWS-NEXT: entry:
-// WINDOWS-NEXT:   %s.addr = alloca ptr, align 8
-// WINDOWS-NEXT:   store ptr %s, ptr %s.addr, align 8
-// WINDOWS-NEXT:   %0 = load ptr, ptr %s.addr, align 8
-// WINDOWS-NEXT:   %1 = getelementptr i8, ptr %0, i32 0
-// WINDOWS-NEXT:   %2 = load i8, ptr %1, align 1
-// WINDOWS-NEXT:   %3 = and i8 %2, -4
-// WINDOWS-NEXT:   store i8 %3, ptr %1, align 1
-// WINDOWS-NEXT:   %4 = getelementptr i8, ptr %0, i32 0
-// WINDOWS-NEXT:   %5 = load i8, ptr %4, align 1
-// WINDOWS-NEXT:   %6 = and i8 %5, 63
-// WINDOWS-NEXT:   store i8 %6, ptr %4, align 1
-// WINDOWS-NEXT:   ret void
+// WINDOWS-LABEL: define dso_local void @testUnnamedBitfieldSurrounding(
+// WINDOWS-SAME: ptr noundef [[S:%.*]]) #[[ATTR0]] {
+// WINDOWS-NEXT:  [[ENTRY:.*:]]
+// WINDOWS-NEXT:    [[S_ADDR:%.*]] = alloca ptr, align 8
+// WINDOWS-NEXT:    store ptr [[S]], ptr [[S_ADDR]], align 8
+// WINDOWS-NEXT:    [[TMP0:%.*]] = load ptr, ptr [[S_ADDR]], align 8
+// WINDOWS-NEXT:    [[TMP1:%.*]] = getelementptr i8, ptr [[TMP0]], i32 0
+// WINDOWS-NEXT:    [[TMP2:%.*]] = load i8, ptr [[TMP1]], align 1
+// WINDOWS-NEXT:    [[TMP3:%.*]] = and i8 [[TMP2]], -4
+// WINDOWS-NEXT:    store i8 [[TMP3]], ptr [[TMP1]], align 1
+// WINDOWS-NEXT:    [[TMP4:%.*]] = getelementptr i8, ptr [[TMP0]], i32 0
+// WINDOWS-NEXT:    [[TMP5:%.*]] = load i8, ptr [[TMP4]], align 1
+// WINDOWS-NEXT:    [[TMP6:%.*]] = and i8 [[TMP5]], 63
+// WINDOWS-NEXT:    store i8 [[TMP6]], ptr [[TMP4]], align 1
+// WINDOWS-NEXT:    ret void
+//
+// ARM-LE-LABEL: define dso_local void @testUnnamedBitfieldSurrounding(
+// ARM-LE-SAME: ptr noundef [[S:%.*]]) #[[ATTR0]] {
+// ARM-LE-NEXT:  [[ENTRY:.*:]]
+// ARM-LE-NEXT:    [[S_ADDR:%.*]] = alloca ptr, align 4
+// ARM-LE-NEXT:    store ptr [[S]], ptr [[S_ADDR]], align 4
+// ARM-LE-NEXT:    [[TMP0:%.*]] = load ptr, ptr [[S_ADDR]], align 4
+// ARM-LE-NEXT:    [[TMP1:%.*]] = getelementptr i8, ptr [[TMP0]], i32 0
+// ARM-LE-NEXT:    [[TMP2:%.*]] = load i8, ptr [[TMP1]], align 1
+// ARM-LE-NEXT:    [[TMP3:%.*]] = and i8 [[TMP2]], -4
+// ARM-LE-NEXT:    store i8 [[TMP3]], ptr [[TMP1]], align 1
+// ARM-LE-NEXT:    [[TMP4:%.*]] = getelementptr i8, ptr [[TMP0]], i32 0
+// ARM-LE-NEXT:    [[TMP5:%.*]] = load i8, ptr [[TMP4]], align 1
+// ARM-LE-NEXT:    [[TMP6:%.*]] = and i8 [[TMP5]], 63
+// ARM-LE-NEXT:    store i8 [[TMP6]], ptr [[TMP4]], align 1
+// ARM-LE-NEXT:    ret void
+//
+// ARM-BE-LABEL: define dso_local void @testUnnamedBitfieldSurrounding(
+// ARM-BE-SAME: ptr noundef [[S:%.*]]) #[[ATTR0]] {
+// ARM-BE-NEXT:  [[ENTRY:.*:]]
+// ARM-BE-NEXT:    [[S_ADDR:%.*]] = alloca ptr, align 4
+// ARM-BE-NEXT:    store ptr [[S]], ptr [[S_ADDR]], align 4
+// ARM-BE-NEXT:    [[TMP0:%.*]] = load ptr, ptr [[S_ADDR]], align 4
+// ARM-BE-NEXT:    [[TMP1:%.*]] = getelementptr i8, ptr [[TMP0]], i32 0
+// ARM-BE-NEXT:    [[TMP2:%.*]] = load i8, ptr [[TMP1]], align 1
+// ARM-BE-NEXT:    [[TMP3:%.*]] = and i8 [[TMP2]], -4
+// ARM-BE-NEXT:    store i8 [[TMP3]], ptr [[TMP1]], align 1
+// ARM-BE-NEXT:    [[TMP4:%.*]] = getelementptr i8, ptr [[TMP0]], i32 0
+// ARM-BE-NEXT:    [[TMP5:%.*]] = load i8, ptr [[TMP4]], align 1
+// ARM-BE-NEXT:    [[TMP6:%.*]] = and i8 [[TMP5]], 63
+// ARM-BE-NEXT:    store i8 [[TMP6]], ptr [[TMP4]], align 1
+// ARM-BE-NEXT:    ret void
+//
+// AARCH64-LE-LABEL: define dso_local void @testUnnamedBitfieldSurrounding(
+// AARCH64-LE-SAME: ptr noundef [[S:%.*]]) #[[ATTR0]] {
+// AARCH64-LE-NEXT:  [[ENTRY:.*:]]
+// AARCH64-LE-NEXT:    [[S_ADDR:%.*]] = alloca ptr, align 8
+// AARCH64-LE-NEXT:    store ptr [[S]], ptr [[S_ADDR]], align 8
+// AARCH64-LE-NEXT:    [[TMP0:%.*]] = load ptr, ptr [[S_ADDR]], align 8
+// AARCH64-LE-NEXT:    [[TMP1:%.*]] = getelementptr i8, ptr [[TMP0]], i32 0
+// AARCH64-LE-NEXT:    [[TMP2:%.*]] = load i8, ptr [[TMP1]], align 1
+// AARCH64-LE-NEXT:    [[TMP3:%.*]] = and i8 [[TMP2]], -4
+// AARCH64-LE-NEXT:    store i8 [[TMP3]], ptr [[TMP1]], align 1
+// AARCH64-LE-NEXT:    [[TMP4:%.*]] = getelementptr i8, ptr [[TMP0]], i32 0
+// AARCH64-LE-NEXT:    [[TMP5:%.*]] = load i8, ptr [[TMP4]], align 1
+// AARCH64-LE-NEXT:    [[TMP6:%.*]] = and i8 [[TMP5]], 63
+// AARCH64-LE-NEXT:    store i8 [[TMP6]], ptr [[TMP4]], align 1
+// AARCH64-LE-NEXT:    ret void
+//
+// AARCH64-BE-LABEL: define dso_local void @testUnnamedBitfieldSurrounding(
+// AARCH64-BE-SAME: ptr noundef [[S:%.*]]) #[[ATTR0]] {
+// AARCH64-BE-NEXT:  [[ENTRY:.*:]]
+// AARCH64-BE-NEXT:    [[S_ADDR:%.*]] = alloca ptr, align 8
+// AARCH64-BE-NEXT:    store ptr [[S]], ptr [[S_ADDR]], align 8
+// AARCH64-BE-NEXT:    [[TMP0:%.*]] = load ptr, ptr [[S_ADDR]], align 8
+// AARCH64-BE-NEXT:    [[TMP1:%.*]] = getelementptr i8, ptr [[TMP0]], i32 0
+// AARCH64-BE-NEXT:    [[TMP2:%.*]] = load i8, ptr [[TMP1]], align 1
+// AARCH64-BE-NEXT:    [[TMP3:%.*]] = and i8 [[TMP2]], -4
+// AARCH64-BE-NEXT:    store i8 [[TMP3]], ptr [[TMP1]], align 1
+// AARCH64-BE-NEXT:    [[TMP4:%.*]] = getelementptr i8, ptr [[TMP0]], i32 0
+// AARCH64-BE-NEXT:    [[TMP5:%.*]] = load i8, ptr [[TMP4]], align 1
+// AARCH64-BE-NEXT:    [[TMP6:%.*]] = and i8 [[TMP5]], 63
+// AARCH64-BE-NEXT:    store i8 [[TMP6]], ptr [[TMP4]], align 1
+// AARCH64-BE-NEXT:    ret void
 //
 void testUnnamedBitfieldSurrounding(struct UnnamedBitfieldSurrounding *s) {
   // byte 0: unnamed[0-1], a[2-5], unnamed[6-7]
@@ -1076,34 +2772,148 @@ struct UnnamedZeroWidthBitfield {
 // LINUX-NEXT:    store i8 0, ptr [[TMP12]], align 1
 // LINUX-NEXT:    ret void
 //
-// WINDOWS-LABEL: define dso_local void @testUnnamedZeroWidthBitfield(ptr noundef %s) #0 {
-// WINDOWS-NEXT: entry:
-// WINDOWS-NEXT:   %s.addr = alloca ptr, align 8
-// WINDOWS-NEXT:   store ptr %s, ptr %s.addr, align 8
-// WINDOWS-NEXT:   %0 = load ptr, ptr %s.addr, align 8
-// WINDOWS-NEXT:   %1 = getelementptr i8, ptr %0, i32 0
-// WINDOWS-NEXT:   %2 = load i8, ptr %1, align 4
-// WINDOWS-NEXT:   %3 = and i8 %2, 15
-// WINDOWS-NEXT:   store i8 %3, ptr %1, align 4
-// WINDOWS-NEXT:   %4 = getelementptr i8, ptr %0, i32 1
-// WINDOWS-NEXT:   store i8 0, ptr %4, align 1
-// WINDOWS-NEXT:   %5 = getelementptr i8, ptr %0, i32 2
-// WINDOWS-NEXT:   store i8 0, ptr %5, align 2
-// WINDOWS-NEXT:   %6 = getelementptr i8, ptr %0, i32 3
-// WINDOWS-NEXT:   store i8 0, ptr %6, align 1
-// WINDOWS-NEXT:   %7 = getelementptr i8, ptr %0, i32 4
-// WINDOWS-NEXT:   %8 = load i8, ptr %7, align 4
-// WINDOWS-NEXT:   %9 = and i8 %8, 15
-// WINDOWS-NEXT:   store i8 %9, ptr %7, align 4
-// WINDOWS-NEXT:   %10 = getelementptr i8, ptr %0, i32 5
-// WINDOWS-NEXT:   store i8 0, ptr %10, align 1
-// WINDOWS-NEXT:   %11 = getelementptr i8, ptr %0, i32 6
-// WINDOWS-NEXT:   store i8 0, ptr %11, align 2
-// WINDOWS-NEXT:   %12 = getelementptr i8, ptr %0, i32 7
-// WINDOWS-NEXT:   store i8 0, ptr %12, align 1
-// WINDOWS-NEXT:   ret void
+// WINDOWS-LABEL: define dso_local void @testUnnamedZeroWidthBitfield(
+// WINDOWS-SAME: ptr noundef [[S:%.*]]) #[[ATTR0]] {
+// WINDOWS-NEXT:  [[ENTRY:.*:]]
+// WINDOWS-NEXT:    [[S_ADDR:%.*]] = alloca ptr, align 8
+// WINDOWS-NEXT:    store ptr [[S]], ptr [[S_ADDR]], align 8
+// WINDOWS-NEXT:    [[TMP0:%.*]] = load ptr, ptr [[S_ADDR]], align 8
+// WINDOWS-NEXT:    [[TMP1:%.*]] = getelementptr i8, ptr [[TMP0]], i32 0
+// WINDOWS-NEXT:    [[TMP2:%.*]] = load i8, ptr [[TMP1]], align 4
+// WINDOWS-NEXT:    [[TMP3:%.*]] = and i8 [[TMP2]], 15
+// WINDOWS-NEXT:    store i8 [[TMP3]], ptr [[TMP1]], align 4
+// WINDOWS-NEXT:    [[TMP4:%.*]] = getelementptr i8, ptr [[TMP0]], i32 1
+// WINDOWS-NEXT:    store i8 0, ptr [[TMP4]], align 1
+// WINDOWS-NEXT:    [[TMP5:%.*]] = getelementptr i8, ptr [[TMP0]], i32 2
+// WINDOWS-NEXT:    store i8 0, ptr [[TMP5]], align 2
+// WINDOWS-NEXT:    [[TMP6:%.*]] = getelementptr i8, ptr [[TMP0]], i32 3
+// WINDOWS-NEXT:    store i8 0, ptr [[TMP6]], align 1
+// WINDOWS-NEXT:    [[TMP7:%.*]] = getelementptr i8, ptr [[TMP0]], i32 4
+// WINDOWS-NEXT:    [[TMP8:%.*]] = load i8, ptr [[TMP7]], align 4
+// WINDOWS-NEXT:    [[TMP9:%.*]] = and i8 [[TMP8]], 15
+// WINDOWS-NEXT:    store i8 [[TMP9]], ptr [[TMP7]], align 4
+// WINDOWS-NEXT:    [[TMP10:%.*]] = getelementptr i8, ptr [[TMP0]], i32 5
+// WINDOWS-NEXT:    store i8 0, ptr [[TMP10]], align 1
+// WINDOWS-NEXT:    [[TMP11:%.*]] = getelementptr i8, ptr [[TMP0]], i32 6
+// WINDOWS-NEXT:    store i8 0, ptr [[TMP11]], align 2
+// WINDOWS-NEXT:    [[TMP12:%.*]] = getelementptr i8, ptr [[TMP0]], i32 7
+// WINDOWS-NEXT:    store i8 0, ptr [[TMP12]], align 1
+// WINDOWS-NEXT:    ret void
+//
+// ARM-LE-LABEL: define dso_local void @testUnnamedZeroWidthBitfield(
+// ARM-LE-SAME: ptr noundef [[S:%.*]]) #[[ATTR0]] {
+// ARM-LE-NEXT:  [[ENTRY:.*:]]
+// ARM-LE-NEXT:    [[S_ADDR:%.*]] = alloca ptr, align 4
+// ARM-LE-NEXT:    store ptr [[S]], ptr [[S_ADDR]], align 4
+// ARM-LE-NEXT:    [[TMP0:%.*]] = load ptr, ptr [[S_ADDR]], align 4
+// ARM-LE-NEXT:    [[TMP1:%.*]] = getelementptr i8, ptr [[TMP0]], i32 0
+// ARM-LE-NEXT:    [[TMP2:%.*]] = load i8, ptr [[TMP1]], align 4
+// ARM-LE-NEXT:    [[TMP3:%.*]] = and i8 [[TMP2]], 15
+// ARM-LE-NEXT:    store i8 [[TMP3]], ptr [[TMP1]], align 4
+// ARM-LE-NEXT:    [[TMP4:%.*]] = getelementptr i8, ptr [[TMP0]], i32 1
+// ARM-LE-NEXT:    store i8 0, ptr [[TMP4]], align 1
+// ARM-LE-NEXT:    [[TMP5:%.*]] = getelementptr i8, ptr [[TMP0]], i32 2
+// ARM-LE-NEXT:    store i8 0, ptr [[TMP5]], align 2
+// ARM-LE-NEXT:    [[TMP6:%.*]] = getelementptr i8, ptr [[TMP0]], i32 3
+// ARM-LE-NEXT:    store i8 0, ptr [[TMP6]], align 1
+// ARM-LE-NEXT:    [[TMP7:%.*]] = getelementptr i8, ptr [[TMP0]], i32 4
+// ARM-LE-NEXT:    [[TMP8:%.*]] = load i8, ptr [[TMP7]], align 4
+// ARM-LE-NEXT:    [[TMP9:%.*]] = and i8 [[TMP8]], 15
+// ARM-LE-NEXT:    store i8 [[TMP9]], ptr [[TMP7]], align 4
+// ARM-LE-NEXT:    [[TMP10:%.*]] = getelementptr i8, ptr [[TMP0]], i32 5
+// ARM-LE-NEXT:    store i8 0, ptr [[TMP10]], align 1
+// ARM-LE-NEXT:    [[TMP11:%.*]] = getelementptr i8, ptr [[TMP0]], i32 6
+// ARM-LE-NEXT:    store i8 0, ptr [[TMP11]], align 2
+// ARM-LE-NEXT:    [[TMP12:%.*]] = getelementptr i8, ptr [[TMP0]], i32 7
+// ARM-LE-NEXT:    store i8 0, ptr [[TMP12]], align 1
+// ARM-LE-NEXT:    ret void
+//
+// ARM-BE-LABEL: define dso_local void @testUnnamedZeroWidthBitfield(
+// ARM-BE-SAME: ptr noundef [[S:%.*]]) #[[ATTR0]] {
+// ARM-BE-NEXT:  [[ENTRY:.*:]]
+// ARM-BE-NEXT:    [[S_ADDR:%.*]] = alloca ptr, align 4
+// ARM-BE-NEXT:    store ptr [[S]], ptr [[S_ADDR]], align 4
+// ARM-BE-NEXT:    [[TMP0:%.*]] = load ptr, ptr [[S_ADDR]], align 4
+// ARM-BE-NEXT:    [[TMP1:%.*]] = getelementptr i8, ptr [[TMP0]], i32 0
+// ARM-BE-NEXT:    [[TMP2:%.*]] = load i8, ptr [[TMP1]], align 4
+// ARM-BE-NEXT:    [[TMP3:%.*]] = and i8 [[TMP2]], -16
+// ARM-BE-NEXT:    store i8 [[TMP3]], ptr [[TMP1]], align 4
+// ARM-BE-NEXT:    [[TMP4:%.*]] = getelementptr i8, ptr [[TMP0]], i32 1
+// ARM-BE-NEXT:    store i8 0, ptr [[TMP4]], align 1
+// ARM-BE-NEXT:    [[TMP5:%.*]] = getelementptr i8, ptr [[TMP0]], i32 2
+// ARM-BE-NEXT:    store i8 0, ptr [[TMP5]], align 2
+// ARM-BE-NEXT:    [[TMP6:%.*]] = getelementptr i8, ptr [[TMP0]], i32 3
+// ARM-BE-NEXT:    store i8 0, ptr [[TMP6]], align 1
+// ARM-BE-NEXT:    [[TMP7:%.*]] = getelementptr i8, ptr [[TMP0]], i32 4
+// ARM-BE-NEXT:    [[TMP8:%.*]] = load i8, ptr [[TMP7]], align 4
+// ARM-BE-NEXT:    [[TMP9:%.*]] = and i8 [[TMP8]], -16
+// ARM-BE-NEXT:    store i8 [[TMP9]], ptr [[TMP7]], align 4
+// ARM-BE-NEXT:    [[TMP10:%.*]] = getelementptr i8, ptr [[TMP0]], i32 5
+// ARM-BE-NEXT:    store i8 0, ptr [[TMP10]], align 1
+// ARM-BE-NEXT:    [[TMP11:%.*]] = getelementptr i8, ptr [[TMP0]], i32 6
+// ARM-BE-NEXT:    store i8 0, ptr [[TMP11]], align 2
+// ARM-BE-NEXT:    [[TMP12:%.*]] = getelementptr i8, ptr [[TMP0]], i32 7
+// ARM-BE-NEXT:    store i8 0, ptr [[TMP12]], align 1
+// ARM-BE-NEXT:    ret void
+//
+// AARCH64-LE-LABEL: define dso_local void @testUnnamedZeroWidthBitfield(
+// AARCH64-LE-SAME: ptr noundef [[S:%.*]]) #[[ATTR0]] {
+// AARCH64-LE-NEXT:  [[ENTRY:.*:]]
+// AARCH64-LE-NEXT:    [[S_ADDR:%.*]] = alloca ptr, align 8
+// AARCH64-LE-NEXT:    store ptr [[S]], ptr [[S_ADDR]], align 8
+// AARCH64-LE-NEXT:    [[TMP0:%.*]] = load ptr, ptr [[S_ADDR]], align 8
+// AARCH64-LE-NEXT:    [[TMP1:%.*]] = getelementptr i8, ptr [[TMP0]], i32 0
+// AARCH64-LE-NEXT:    [[TMP2:%.*]] = load i8, ptr [[TMP1]], align 4
+// AARCH64-LE-NEXT:    [[TMP3:%.*]] = and i8 [[TMP2]], 15
+// AARCH64-LE-NEXT:    store i8 [[TMP3]], ptr [[TMP1]], align 4
+// AARCH64-LE-NEXT:    [[TMP4:%.*]] = getelementptr i8, ptr [[TMP0]], i32 1
+// AARCH64-LE-NEXT:    store i8 0, ptr [[TMP4]], align 1
+// AARCH64-LE-NEXT:    [[TMP5:%.*]] = getelementptr i8, ptr [[TMP0]], i32 2
+// AARCH64-LE-NEXT:    store i8 0, ptr [[TMP5]], align 2
+// AARCH64-LE-NEXT:    [[TMP6:%.*]] = getelementptr i8, ptr [[TMP0]], i32 3
+// AARCH64-LE-NEXT:    store i8 0, ptr [[TMP6]], align 1
+// AARCH64-LE-NEXT:    [[TMP7:%.*]] = getelementptr i8, ptr [[TMP0]], i32 4
+// AARCH64-LE-NEXT:    [[TMP8:%.*]] = load i8, ptr [[TMP7]], align 4
+// AARCH64-LE-NEXT:    [[TMP9:%.*]] = and i8 [[TMP8]], 15
+// AARCH64-LE-NEXT:    store i8 [[TMP9]], ptr [[TMP7]], align 4
+// AARCH64-LE-NEXT:    [[TMP10:%.*]] = getelementptr i8, ptr [[TMP0]], i32 5
+// AARCH64-LE-NEXT:    store i8 0, ptr [[TMP10]], align 1
+// AARCH64-LE-NEXT:    [[TMP11:%.*]] = getelementptr i8, ptr [[TMP0]], i32 6
+// AARCH64-LE-NEXT:    store i8 0, ptr [[TMP11]], align 2
+// AARCH64-LE-NEXT:    [[TMP12:%.*]] = getelementptr i8, ptr [[TMP0]], i32 7
+// AARCH64-LE-NEXT:    store i8 0, ptr [[TMP12]], align 1
+// AARCH64-LE-NEXT:    ret void
+//
+// AARCH64-BE-LABEL: define dso_local void @testUnnamedZeroWidthBitfield(
+// AARCH64-BE-SAME: ptr noundef [[S:%.*]]) #[[ATTR0]] {
+// AARCH64-BE-NEXT:  [[ENTRY:.*:]]
+// AARCH64-BE-NEXT:    [[S_ADDR:%.*]] = alloca ptr, align 8
+// AARCH64-BE-NEXT:    store ptr [[S]], ptr [[S_ADDR]], align 8
+// AARCH64-BE-NEXT:    [[TMP0:%.*]] = load ptr, ptr [[S_ADDR]], align 8
+// AARCH64-BE-NEXT:    [[TMP1:%.*]] = getelementptr i8, ptr [[TMP0]], i32 0
+// AARCH64-BE-NEXT:    [[TMP2:%.*]] = load i8, ptr [[TMP1]], align 4
+// AARCH64-BE-NEXT:    [[TMP3:%.*]] = and i8 [[TMP2]], -16
+// AARCH64-BE-NEXT:    store i8 [[TMP3]], ptr [[TMP1]], align 4
+// AARCH64-BE-NEXT:    [[TMP4:%.*]] = getelementptr i8, ptr [[TMP0]], i32 1
+// AARCH64-BE-NEXT:    store i8 0, ptr [[TMP4]], align 1
+// AARCH64-BE-NEXT:    [[TMP5:%.*]] = getelementptr i8, ptr [[TMP0]], i32 2
+// AARCH64-BE-NEXT:    store i8 0, ptr [[TMP5]], align 2
+// AARCH64-BE-NEXT:    [[TMP6:%.*]] = getelementptr i8, ptr [[TMP0]], i32 3
+// AARCH64-BE-NEXT:    store i8 0, ptr [[TMP6]], align 1
+// AARCH64-BE-NEXT:    [[TMP7:%.*]] = getelementptr i8, ptr [[TMP0]], i32 4
+// AARCH64-BE-NEXT:    [[TMP8:%.*]] = load i8, ptr [[TMP7]], align 4
+// AARCH64-BE-NEXT:    [[TMP9:%.*]] = and i8 [[TMP8]], -16
+// AARCH64-BE-NEXT:    store i8 [[TMP9]], ptr [[TMP7]], align 4
+// AARCH64-BE-NEXT:    [[TMP10:%.*]] = getelementptr i8, ptr [[TMP0]], i32 5
+// AARCH64-BE-NEXT:    store i8 0, ptr [[TMP10]], align 1
+// AARCH64-BE-NEXT:    [[TMP11:%.*]] = getelementptr i8, ptr [[TMP0]], i32 6
+// AARCH64-BE-NEXT:    store i8 0, ptr [[TMP11]], align 2
+// AARCH64-BE-NEXT:    [[TMP12:%.*]] = getelementptr i8, ptr [[TMP0]], i32 7
+// AARCH64-BE-NEXT:    store i8 0, ptr [[TMP12]], align 1
+// AARCH64-BE-NEXT:    ret void
 //
 void testUnnamedZeroWidthBitfield(struct UnnamedZeroWidthBitfield *s) {
+  // The masks below describe little endian; big endian uses 0xf0.
   // byte 0: a[0-3], unnamed[4-7]
   // bytes 1-3: struct padding
   // bytes 4-7: b[bits 0-3], tail padding[bits 4-31]
@@ -1142,22 +2952,88 @@ struct UnnamedBitfieldMultiByte {
 // LINUX-NEXT:    store i8 [[TMP6]], ptr [[TMP4]], align 1
 // LINUX-NEXT:    ret void
 //
-// WINDOWS-LABEL: define dso_local void @testUnnamedBitfieldMultiByte(ptr noundef %s) #0 {
-// WINDOWS-NEXT: entry:
-// WINDOWS-NEXT:   %s.addr = alloca ptr, align 8
-// WINDOWS-NEXT:   store ptr %s, ptr %s.addr, align 8
-// WINDOWS-NEXT:   %0 = load ptr, ptr %s.addr, align 8
-// WINDOWS-NEXT:   %1 = getelementptr i8, ptr %0, i32 0
-// WINDOWS-NEXT:   %2 = load i8, ptr %1, align 2
-// WINDOWS-NEXT:   %3 = and i8 %2, 15
-// WINDOWS-NEXT:   store i8 %3, ptr %1, align 2
-// WINDOWS-NEXT:   %4 = getelementptr i8, ptr %0, i32 1
-// WINDOWS-NEXT:   %5 = load i8, ptr %4, align 1
-// WINDOWS-NEXT:   %6 = and i8 %5, -16
-// WINDOWS-NEXT:   store i8 %6, ptr %4, align 1
-// WINDOWS-NEXT:   ret void
+// WINDOWS-LABEL: define dso_local void @testUnnamedBitfieldMultiByte(
+// WINDOWS-SAME: ptr noundef [[S:%.*]]) #[[ATTR0]] {
+// WINDOWS-NEXT:  [[ENTRY:.*:]]
+// WINDOWS-NEXT:    [[S_ADDR:%.*]] = alloca ptr, align 8
+// WINDOWS-NEXT:    store ptr [[S]], ptr [[S_ADDR]], align 8
+// WINDOWS-NEXT:    [[TMP0:%.*]] = load ptr, ptr [[S_ADDR]], align 8
+// WINDOWS-NEXT:    [[TMP1:%.*]] = getelementptr i8, ptr [[TMP0]], i32 0
+// WINDOWS-NEXT:    [[TMP2:%.*]] = load i8, ptr [[TMP1]], align 2
+// WINDOWS-NEXT:    [[TMP3:%.*]] = and i8 [[TMP2]], 15
+// WINDOWS-NEXT:    store i8 [[TMP3]], ptr [[TMP1]], align 2
+// WINDOWS-NEXT:    [[TMP4:%.*]] = getelementptr i8, ptr [[TMP0]], i32 1
+// WINDOWS-NEXT:    [[TMP5:%.*]] = load i8, ptr [[TMP4]], align 1
+// WINDOWS-NEXT:    [[TMP6:%.*]] = and i8 [[TMP5]], -16
+// WINDOWS-NEXT:    store i8 [[TMP6]], ptr [[TMP4]], align 1
+// WINDOWS-NEXT:    ret void
+//
+// ARM-LE-LABEL: define dso_local void @testUnnamedBitfieldMultiByte(
+// ARM-LE-SAME: ptr noundef [[S:%.*]]) #[[ATTR0]] {
+// ARM-LE-NEXT:  [[ENTRY:.*:]]
+// ARM-LE-NEXT:    [[S_ADDR:%.*]] = alloca ptr, align 4
+// ARM-LE-NEXT:    store ptr [[S]], ptr [[S_ADDR]], align 4
+// ARM-LE-NEXT:    [[TMP0:%.*]] = load ptr, ptr [[S_ADDR]], align 4
+// ARM-LE-NEXT:    [[TMP1:%.*]] = getelementptr i8, ptr [[TMP0]], i32 0
+// ARM-LE-NEXT:    [[TMP2:%.*]] = load i8, ptr [[TMP1]], align 2
+// ARM-LE-NEXT:    [[TMP3:%.*]] = and i8 [[TMP2]], 15
+// ARM-LE-NEXT:    store i8 [[TMP3]], ptr [[TMP1]], align 2
+// ARM-LE-NEXT:    [[TMP4:%.*]] = getelementptr i8, ptr [[TMP0]], i32 1
+// ARM-LE-NEXT:    [[TMP5:%.*]] = load i8, ptr [[TMP4]], align 1
+// ARM-LE-NEXT:    [[TMP6:%.*]] = and i8 [[TMP5]], -16
+// ARM-LE-NEXT:    store i8 [[TMP6]], ptr [[TMP4]], align 1
+// ARM-LE-NEXT:    ret void
+//
+// ARM-BE-LABEL: define dso_local void @testUnnamedBitfieldMultiByte(
+// ARM-BE-SAME: ptr noundef [[S:%.*]]) #[[ATTR0]] {
+// ARM-BE-NEXT:  [[ENTRY:.*:]]
+// ARM-BE-NEXT:    [[S_ADDR:%.*]] = alloca ptr, align 4
+// ARM-BE-NEXT:    store ptr [[S]], ptr [[S_ADDR]], align 4
+// ARM-BE-NEXT:    [[TMP0:%.*]] = load ptr, ptr [[S_ADDR]], align 4
+// ARM-BE-NEXT:    [[TMP1:%.*]] = getelementptr i8, ptr [[TMP0]], i32 0
+// ARM-BE-NEXT:    [[TMP2:%.*]] = load i8, ptr [[TMP1]], align 2
+// ARM-BE-NEXT:    [[TMP3:%.*]] = and i8 [[TMP2]], -16
+// ARM-BE-NEXT:    store i8 [[TMP3]], ptr [[TMP1]], align 2
+// ARM-BE-NEXT:    [[TMP4:%.*]] = getelementptr i8, ptr [[TMP0]], i32 1
+// ARM-BE-NEXT:    [[TMP5:%.*]] = load i8, ptr [[TMP4]], align 1
+// ARM-BE-NEXT:    [[TMP6:%.*]] = and i8 [[TMP5]], 15
+// ARM-BE-NEXT:    store i8 [[TMP6]], ptr [[TMP4]], align 1
+// ARM-BE-NEXT:    ret void
+//
+// AARCH64-LE-LABEL: define dso_local void @testUnnamedBitfieldMultiByte(
+// AARCH64-LE-SAME: ptr noundef [[S:%.*]]) #[[ATTR0]] {
+// AARCH64-LE-NEXT:  [[ENTRY:.*:]]
+// AARCH64-LE-NEXT:    [[S_ADDR:%.*]] = alloca ptr, align 8
+// AARCH64-LE-NEXT:    store ptr [[S]], ptr [[S_ADDR]], align 8
+// AARCH64-LE-NEXT:    [[TMP0:%.*]] = load ptr, ptr [[S_ADDR]], align 8
+// AARCH64-LE-NEXT:    [[TMP1:%.*]] = getelementptr i8, ptr [[TMP0]], i32 0
+// AARCH64-LE-NEXT:    [[TMP2:%.*]] = load i8, ptr [[TMP1]], align 2
+// AARCH64-LE-NEXT:    [[TMP3:%.*]] = and i8 [[TMP2]], 15
+// AARCH64-LE-NEXT:    store i8 [[TMP3]], ptr [[TMP1]], align 2
+// AARCH64-LE-NEXT:    [[TMP4:%.*]] = getelementptr i8, ptr [[TMP0]], i32 1
+// AARCH64-LE-NEXT:    [[TMP5:%.*]] = load i8, ptr [[TMP4]], align 1
+// AARCH64-LE-NEXT:    [[TMP6:%.*]] = and i8 [[TMP5]], -16
+// AARCH64-LE-NEXT:    store i8 [[TMP6]], ptr [[TMP4]], align 1
+// AARCH64-LE-NEXT:    ret void
+//
+// AARCH64-BE-LABEL: define dso_local void @testUnnamedBitfieldMultiByte(
+// AARCH64-BE-SAME: ptr noundef [[S:%.*]]) #[[ATTR0]] {
+// AARCH64-BE-NEXT:  [[ENTRY:.*:]]
+// AARCH64-BE-NEXT:    [[S_ADDR:%.*]] = alloca ptr, align 8
+// AARCH64-BE-NEXT:    store ptr [[S]], ptr [[S_ADDR]], align 8
+// AARCH64-BE-NEXT:    [[TMP0:%.*]] = load ptr, ptr [[S_ADDR]], align 8
+// AARCH64-BE-NEXT:    [[TMP1:%.*]] = getelementptr i8, ptr [[TMP0]], i32 0
+// AARCH64-BE-NEXT:    [[TMP2:%.*]] = load i8, ptr [[TMP1]], align 2
+// AARCH64-BE-NEXT:    [[TMP3:%.*]] = and i8 [[TMP2]], -16
+// AARCH64-BE-NEXT:    store i8 [[TMP3]], ptr [[TMP1]], align 2
+// AARCH64-BE-NEXT:    [[TMP4:%.*]] = getelementptr i8, ptr [[TMP0]], i32 1
+// AARCH64-BE-NEXT:    [[TMP5:%.*]] = load i8, ptr [[TMP4]], align 1
+// AARCH64-BE-NEXT:    [[TMP6:%.*]] = and i8 [[TMP5]], 15
+// AARCH64-BE-NEXT:    store i8 [[TMP6]], ptr [[TMP4]], align 1
+// AARCH64-BE-NEXT:    ret void
 //
 void testUnnamedBitfieldMultiByte(struct UnnamedBitfieldMultiByte *s) {
+  // The masks below describe little endian; big endian swaps them.
   // 2 bytes: a[0-3], unnamed[4-11], b[12-15]
   // byte 0: clear bits 4-7
   // byte 1: clear bits 0-3
@@ -1167,3 +3043,1870 @@ void testUnnamedBitfieldMultiByte(struct UnnamedBitfieldMultiByte *s) {
   //   0b11110000 == -16
   __builtin_clear_padding(s);
 }
+
+// Complete bytes require no clearing, including a full 64-bit bitfield.
+struct BitfieldFullBytes {
+  unsigned char first : 8;
+  unsigned char second : 8;
+  unsigned short half : 16;
+  unsigned int word : 32;
+  unsigned long long wide : 64;
+};
+
+// LINUX-LABEL: define dso_local void @testBitfieldFullBytes(
+// LINUX-SAME: ptr noundef [[S:%.*]]) #[[ATTR0]] {
+// LINUX-NEXT:  [[ENTRY:.*:]]
+// LINUX-NEXT:    [[S_ADDR:%.*]] = alloca ptr, align 8
+// LINUX-NEXT:    store ptr [[S]], ptr [[S_ADDR]], align 8
+// LINUX-NEXT:    [[TMP0:%.*]] = load ptr, ptr [[S_ADDR]], align 8
+// LINUX-NEXT:    ret void
+//
+// WINDOWS-LABEL: define dso_local void @testBitfieldFullBytes(
+// WINDOWS-SAME: ptr noundef [[S:%.*]]) #[[ATTR0]] {
+// WINDOWS-NEXT:  [[ENTRY:.*:]]
+// WINDOWS-NEXT:    [[S_ADDR:%.*]] = alloca ptr, align 8
+// WINDOWS-NEXT:    store ptr [[S]], ptr [[S_ADDR]], align 8
+// WINDOWS-NEXT:    [[TMP0:%.*]] = load ptr, ptr [[S_ADDR]], align 8
+// WINDOWS-NEXT:    ret void
+//
+// ARM-LE-LABEL: define dso_local void @testBitfieldFullBytes(
+// ARM-LE-SAME: ptr noundef [[S:%.*]]) #[[ATTR0]] {
+// ARM-LE-NEXT:  [[ENTRY:.*:]]
+// ARM-LE-NEXT:    [[S_ADDR:%.*]] = alloca ptr, align 4
+// ARM-LE-NEXT:    store ptr [[S]], ptr [[S_ADDR]], align 4
+// ARM-LE-NEXT:    [[TMP0:%.*]] = load ptr, ptr [[S_ADDR]], align 4
+// ARM-LE-NEXT:    ret void
+//
+// ARM-BE-LABEL: define dso_local void @testBitfieldFullBytes(
+// ARM-BE-SAME: ptr noundef [[S:%.*]]) #[[ATTR0]] {
+// ARM-BE-NEXT:  [[ENTRY:.*:]]
+// ARM-BE-NEXT:    [[S_ADDR:%.*]] = alloca ptr, align 4
+// ARM-BE-NEXT:    store ptr [[S]], ptr [[S_ADDR]], align 4
+// ARM-BE-NEXT:    [[TMP0:%.*]] = load ptr, ptr [[S_ADDR]], align 4
+// ARM-BE-NEXT:    ret void
+//
+// AARCH64-LE-LABEL: define dso_local void @testBitfieldFullBytes(
+// AARCH64-LE-SAME: ptr noundef [[S:%.*]]) #[[ATTR0]] {
+// AARCH64-LE-NEXT:  [[ENTRY:.*:]]
+// AARCH64-LE-NEXT:    [[S_ADDR:%.*]] = alloca ptr, align 8
+// AARCH64-LE-NEXT:    store ptr [[S]], ptr [[S_ADDR]], align 8
+// AARCH64-LE-NEXT:    [[TMP0:%.*]] = load ptr, ptr [[S_ADDR]], align 8
+// AARCH64-LE-NEXT:    ret void
+//
+// AARCH64-BE-LABEL: define dso_local void @testBitfieldFullBytes(
+// AARCH64-BE-SAME: ptr noundef [[S:%.*]]) #[[ATTR0]] {
+// AARCH64-BE-NEXT:  [[ENTRY:.*:]]
+// AARCH64-BE-NEXT:    [[S_ADDR:%.*]] = alloca ptr, align 8
+// AARCH64-BE-NEXT:    store ptr [[S]], ptr [[S_ADDR]], align 8
+// AARCH64-BE-NEXT:    [[TMP0:%.*]] = load ptr, ptr [[S_ADDR]], align 8
+// AARCH64-BE-NEXT:    ret void
+//
+void testBitfieldFullBytes(struct BitfieldFullBytes *s) {
+  __builtin_clear_padding(s);
+}
+
+// The value ends exactly at a byte boundary. Preserve masks are 0xf8, 0xff
+// on little endian and 0x1f, 0xff on big endian.
+struct BitfieldEndAligned {
+  unsigned short : 3;
+  unsigned short value : 13;
+};
+
+// LINUX-LABEL: define dso_local void @testBitfieldEndAligned(
+// LINUX-SAME: ptr noundef [[S:%.*]]) #[[ATTR0]] {
+// LINUX-NEXT:  [[ENTRY:.*:]]
+// LINUX-NEXT:    [[S_ADDR:%.*]] = alloca ptr, align 8
+// LINUX-NEXT:    store ptr [[S]], ptr [[S_ADDR]], align 8
+// LINUX-NEXT:    [[TMP0:%.*]] = load ptr, ptr [[S_ADDR]], align 8
+// LINUX-NEXT:    [[TMP1:%.*]] = getelementptr i8, ptr [[TMP0]], i32 0
+// LINUX-NEXT:    [[TMP2:%.*]] = load i8, ptr [[TMP1]], align 2
+// LINUX-NEXT:    [[TMP3:%.*]] = and i8 [[TMP2]], -8
+// LINUX-NEXT:    store i8 [[TMP3]], ptr [[TMP1]], align 2
+// LINUX-NEXT:    ret void
+//
+// WINDOWS-LABEL: define dso_local void @testBitfieldEndAligned(
+// WINDOWS-SAME: ptr noundef [[S:%.*]]) #[[ATTR0]] {
+// WINDOWS-NEXT:  [[ENTRY:.*:]]
+// WINDOWS-NEXT:    [[S_ADDR:%.*]] = alloca ptr, align 8
+// WINDOWS-NEXT:    store ptr [[S]], ptr [[S_ADDR]], align 8
+// WINDOWS-NEXT:    [[TMP0:%.*]] = load ptr, ptr [[S_ADDR]], align 8
+// WINDOWS-NEXT:    [[TMP1:%.*]] = getelementptr i8, ptr [[TMP0]], i32 0
+// WINDOWS-NEXT:    [[TMP2:%.*]] = load i8, ptr [[TMP1]], align 2
+// WINDOWS-NEXT:    [[TMP3:%.*]] = and i8 [[TMP2]], -8
+// WINDOWS-NEXT:    store i8 [[TMP3]], ptr [[TMP1]], align 2
+// WINDOWS-NEXT:    ret void
+//
+// ARM-LE-LABEL: define dso_local void @testBitfieldEndAligned(
+// ARM-LE-SAME: ptr noundef [[S:%.*]]) #[[ATTR0]] {
+// ARM-LE-NEXT:  [[ENTRY:.*:]]
+// ARM-LE-NEXT:    [[S_ADDR:%.*]] = alloca ptr, align 4
+// ARM-LE-NEXT:    store ptr [[S]], ptr [[S_ADDR]], align 4
+// ARM-LE-NEXT:    [[TMP0:%.*]] = load ptr, ptr [[S_ADDR]], align 4
+// ARM-LE-NEXT:    [[TMP1:%.*]] = getelementptr i8, ptr [[TMP0]], i32 0
+// ARM-LE-NEXT:    [[TMP2:%.*]] = load i8, ptr [[TMP1]], align 2
+// ARM-LE-NEXT:    [[TMP3:%.*]] = and i8 [[TMP2]], -8
+// ARM-LE-NEXT:    store i8 [[TMP3]], ptr [[TMP1]], align 2
+// ARM-LE-NEXT:    ret void
+//
+// ARM-BE-LABEL: define dso_local void @testBitfieldEndAligned(
+// ARM-BE-SAME: ptr noundef [[S:%.*]]) #[[ATTR0]] {
+// ARM-BE-NEXT:  [[ENTRY:.*:]]
+// ARM-BE-NEXT:    [[S_ADDR:%.*]] = alloca ptr, align 4
+// ARM-BE-NEXT:    store ptr [[S]], ptr [[S_ADDR]], align 4
+// ARM-BE-NEXT:    [[TMP0:%.*]] = load ptr, ptr [[S_ADDR]], align 4
+// ARM-BE-NEXT:    [[TMP1:%.*]] = getelementptr i8, ptr [[TMP0]], i32 0
+// ARM-BE-NEXT:    [[TMP2:%.*]] = load i8, ptr [[TMP1]], align 2
+// ARM-BE-NEXT:    [[TMP3:%.*]] = and i8 [[TMP2]], 31
+// ARM-BE-NEXT:    store i8 [[TMP3]], ptr [[TMP1]], align 2
+// ARM-BE-NEXT:    ret void
+//
+// AARCH64-LE-LABEL: define dso_local void @testBitfieldEndAligned(
+// AARCH64-LE-SAME: ptr noundef [[S:%.*]]) #[[ATTR0]] {
+// AARCH64-LE-NEXT:  [[ENTRY:.*:]]
+// AARCH64-LE-NEXT:    [[S_ADDR:%.*]] = alloca ptr, align 8
+// AARCH64-LE-NEXT:    store ptr [[S]], ptr [[S_ADDR]], align 8
+// AARCH64-LE-NEXT:    [[TMP0:%.*]] = load ptr, ptr [[S_ADDR]], align 8
+// AARCH64-LE-NEXT:    [[TMP1:%.*]] = getelementptr i8, ptr [[TMP0]], i32 0
+// AARCH64-LE-NEXT:    [[TMP2:%.*]] = load i8, ptr [[TMP1]], align 2
+// AARCH64-LE-NEXT:    [[TMP3:%.*]] = and i8 [[TMP2]], -8
+// AARCH64-LE-NEXT:    store i8 [[TMP3]], ptr [[TMP1]], align 2
+// AARCH64-LE-NEXT:    ret void
+//
+// AARCH64-BE-LABEL: define dso_local void @testBitfieldEndAligned(
+// AARCH64-BE-SAME: ptr noundef [[S:%.*]]) #[[ATTR0]] {
+// AARCH64-BE-NEXT:  [[ENTRY:.*:]]
+// AARCH64-BE-NEXT:    [[S_ADDR:%.*]] = alloca ptr, align 8
+// AARCH64-BE-NEXT:    store ptr [[S]], ptr [[S_ADDR]], align 8
+// AARCH64-BE-NEXT:    [[TMP0:%.*]] = load ptr, ptr [[S_ADDR]], align 8
+// AARCH64-BE-NEXT:    [[TMP1:%.*]] = getelementptr i8, ptr [[TMP0]], i32 0
+// AARCH64-BE-NEXT:    [[TMP2:%.*]] = load i8, ptr [[TMP1]], align 2
+// AARCH64-BE-NEXT:    [[TMP3:%.*]] = and i8 [[TMP2]], 31
+// AARCH64-BE-NEXT:    store i8 [[TMP3]], ptr [[TMP1]], align 2
+// AARCH64-BE-NEXT:    ret void
+//
+void testBitfieldEndAligned(struct BitfieldEndAligned *s) {
+  __builtin_clear_padding(s);
+}
+
+// A normal field with two partial endpoints and a complete middle byte.
+// Preserve masks are 0xf8, 0xff, 0x1f, 0x00 on little endian and
+// 0x1f, 0xff, 0xf8, 0x00 on big endian.
+struct BitfieldPartialEnds {
+  unsigned int : 3;
+  unsigned int value : 18;
+  unsigned int : 11;
+};
+
+// LINUX-LABEL: define dso_local void @testBitfieldPartialEnds(
+// LINUX-SAME: ptr noundef [[S:%.*]]) #[[ATTR0]] {
+// LINUX-NEXT:  [[ENTRY:.*:]]
+// LINUX-NEXT:    [[S_ADDR:%.*]] = alloca ptr, align 8
+// LINUX-NEXT:    store ptr [[S]], ptr [[S_ADDR]], align 8
+// LINUX-NEXT:    [[TMP0:%.*]] = load ptr, ptr [[S_ADDR]], align 8
+// LINUX-NEXT:    [[TMP1:%.*]] = getelementptr i8, ptr [[TMP0]], i32 0
+// LINUX-NEXT:    [[TMP2:%.*]] = load i8, ptr [[TMP1]], align 4
+// LINUX-NEXT:    [[TMP3:%.*]] = and i8 [[TMP2]], -8
+// LINUX-NEXT:    store i8 [[TMP3]], ptr [[TMP1]], align 4
+// LINUX-NEXT:    [[TMP4:%.*]] = getelementptr i8, ptr [[TMP0]], i32 2
+// LINUX-NEXT:    [[TMP5:%.*]] = load i8, ptr [[TMP4]], align 2
+// LINUX-NEXT:    [[TMP6:%.*]] = and i8 [[TMP5]], 31
+// LINUX-NEXT:    store i8 [[TMP6]], ptr [[TMP4]], align 2
+// LINUX-NEXT:    [[TMP7:%.*]] = getelementptr i8, ptr [[TMP0]], i32 3
+// LINUX-NEXT:    store i8 0, ptr [[TMP7]], align 1
+// LINUX-NEXT:    ret void
+//
+// WINDOWS-LABEL: define dso_local void @testBitfieldPartialEnds(
+// WINDOWS-SAME: ptr noundef [[S:%.*]]) #[[ATTR0]] {
+// WINDOWS-NEXT:  [[ENTRY:.*:]]
+// WINDOWS-NEXT:    [[S_ADDR:%.*]] = alloca ptr, align 8
+// WINDOWS-NEXT:    store ptr [[S]], ptr [[S_ADDR]], align 8
+// WINDOWS-NEXT:    [[TMP0:%.*]] = load ptr, ptr [[S_ADDR]], align 8
+// WINDOWS-NEXT:    [[TMP1:%.*]] = getelementptr i8, ptr [[TMP0]], i32 0
+// WINDOWS-NEXT:    [[TMP2:%.*]] = load i8, ptr [[TMP1]], align 4
+// WINDOWS-NEXT:    [[TMP3:%.*]] = and i8 [[TMP2]], -8
+// WINDOWS-NEXT:    store i8 [[TMP3]], ptr [[TMP1]], align 4
+// WINDOWS-NEXT:    [[TMP4:%.*]] = getelementptr i8, ptr [[TMP0]], i32 2
+// WINDOWS-NEXT:    [[TMP5:%.*]] = load i8, ptr [[TMP4]], align 2
+// WINDOWS-NEXT:    [[TMP6:%.*]] = and i8 [[TMP5]], 31
+// WINDOWS-NEXT:    store i8 [[TMP6]], ptr [[TMP4]], align 2
+// WINDOWS-NEXT:    [[TMP7:%.*]] = getelementptr i8, ptr [[TMP0]], i32 3
+// WINDOWS-NEXT:    store i8 0, ptr [[TMP7]], align 1
+// WINDOWS-NEXT:    ret void
+//
+// ARM-LE-LABEL: define dso_local void @testBitfieldPartialEnds(
+// ARM-LE-SAME: ptr noundef [[S:%.*]]) #[[ATTR0]] {
+// ARM-LE-NEXT:  [[ENTRY:.*:]]
+// ARM-LE-NEXT:    [[S_ADDR:%.*]] = alloca ptr, align 4
+// ARM-LE-NEXT:    store ptr [[S]], ptr [[S_ADDR]], align 4
+// ARM-LE-NEXT:    [[TMP0:%.*]] = load ptr, ptr [[S_ADDR]], align 4
+// ARM-LE-NEXT:    [[TMP1:%.*]] = getelementptr i8, ptr [[TMP0]], i32 0
+// ARM-LE-NEXT:    [[TMP2:%.*]] = load i8, ptr [[TMP1]], align 4
+// ARM-LE-NEXT:    [[TMP3:%.*]] = and i8 [[TMP2]], -8
+// ARM-LE-NEXT:    store i8 [[TMP3]], ptr [[TMP1]], align 4
+// ARM-LE-NEXT:    [[TMP4:%.*]] = getelementptr i8, ptr [[TMP0]], i32 2
+// ARM-LE-NEXT:    [[TMP5:%.*]] = load i8, ptr [[TMP4]], align 2
+// ARM-LE-NEXT:    [[TMP6:%.*]] = and i8 [[TMP5]], 31
+// ARM-LE-NEXT:    store i8 [[TMP6]], ptr [[TMP4]], align 2
+// ARM-LE-NEXT:    [[TMP7:%.*]] = getelementptr i8, ptr [[TMP0]], i32 3
+// ARM-LE-NEXT:    store i8 0, ptr [[TMP7]], align 1
+// ARM-LE-NEXT:    ret void
+//
+// ARM-BE-LABEL: define dso_local void @testBitfieldPartialEnds(
+// ARM-BE-SAME: ptr noundef [[S:%.*]]) #[[ATTR0]] {
+// ARM-BE-NEXT:  [[ENTRY:.*:]]
+// ARM-BE-NEXT:    [[S_ADDR:%.*]] = alloca ptr, align 4
+// ARM-BE-NEXT:    store ptr [[S]], ptr [[S_ADDR]], align 4
+// ARM-BE-NEXT:    [[TMP0:%.*]] = load ptr, ptr [[S_ADDR]], align 4
+// ARM-BE-NEXT:    [[TMP1:%.*]] = getelementptr i8, ptr [[TMP0]], i32 0
+// ARM-BE-NEXT:    [[TMP2:%.*]] = load i8, ptr [[TMP1]], align 4
+// ARM-BE-NEXT:    [[TMP3:%.*]] = and i8 [[TMP2]], 31
+// ARM-BE-NEXT:    store i8 [[TMP3]], ptr [[TMP1]], align 4
+// ARM-BE-NEXT:    [[TMP4:%.*]] = getelementptr i8, ptr [[TMP0]], i32 2
+// ARM-BE-NEXT:    [[TMP5:%.*]] = load i8, ptr [[TMP4]], align 2
+// ARM-BE-NEXT:    [[TMP6:%.*]] = and i8 [[TMP5]], -8
+// ARM-BE-NEXT:    store i8 [[TMP6]], ptr [[TMP4]], align 2
+// ARM-BE-NEXT:    [[TMP7:%.*]] = getelementptr i8, ptr [[TMP0]], i32 3
+// ARM-BE-NEXT:    store i8 0, ptr [[TMP7]], align 1
+// ARM-BE-NEXT:    ret void
+//
+// AARCH64-LE-LABEL: define dso_local void @testBitfieldPartialEnds(
+// AARCH64-LE-SAME: ptr noundef [[S:%.*]]) #[[ATTR0]] {
+// AARCH64-LE-NEXT:  [[ENTRY:.*:]]
+// AARCH64-LE-NEXT:    [[S_ADDR:%.*]] = alloca ptr, align 8
+// AARCH64-LE-NEXT:    store ptr [[S]], ptr [[S_ADDR]], align 8
+// AARCH64-LE-NEXT:    [[TMP0:%.*]] = load ptr, ptr [[S_ADDR]], align 8
+// AARCH64-LE-NEXT:    [[TMP1:%.*]] = getelementptr i8, ptr [[TMP0]], i32 0
+// AARCH64-LE-NEXT:    [[TMP2:%.*]] = load i8, ptr [[TMP1]], align 4
+// AARCH64-LE-NEXT:    [[TMP3:%.*]] = and i8 [[TMP2]], -8
+// AARCH64-LE-NEXT:    store i8 [[TMP3]], ptr [[TMP1]], align 4
+// AARCH64-LE-NEXT:    [[TMP4:%.*]] = getelementptr i8, ptr [[TMP0]], i32 2
+// AARCH64-LE-NEXT:    [[TMP5:%.*]] = load i8, ptr [[TMP4]], align 2
+// AARCH64-LE-NEXT:    [[TMP6:%.*]] = and i8 [[TMP5]], 31
+// AARCH64-LE-NEXT:    store i8 [[TMP6]], ptr [[TMP4]], align 2
+// AARCH64-LE-NEXT:    [[TMP7:%.*]] = getelementptr i8, ptr [[TMP0]], i32 3
+// AARCH64-LE-NEXT:    store i8 0, ptr [[TMP7]], align 1
+// AARCH64-LE-NEXT:    ret void
+//
+// AARCH64-BE-LABEL: define dso_local void @testBitfieldPartialEnds(
+// AARCH64-BE-SAME: ptr noundef [[S:%.*]]) #[[ATTR0]] {
+// AARCH64-BE-NEXT:  [[ENTRY:.*:]]
+// AARCH64-BE-NEXT:    [[S_ADDR:%.*]] = alloca ptr, align 8
+// AARCH64-BE-NEXT:    store ptr [[S]], ptr [[S_ADDR]], align 8
+// AARCH64-BE-NEXT:    [[TMP0:%.*]] = load ptr, ptr [[S_ADDR]], align 8
+// AARCH64-BE-NEXT:    [[TMP1:%.*]] = getelementptr i8, ptr [[TMP0]], i32 0
+// AARCH64-BE-NEXT:    [[TMP2:%.*]] = load i8, ptr [[TMP1]], align 4
+// AARCH64-BE-NEXT:    [[TMP3:%.*]] = and i8 [[TMP2]], 31
+// AARCH64-BE-NEXT:    store i8 [[TMP3]], ptr [[TMP1]], align 4
+// AARCH64-BE-NEXT:    [[TMP4:%.*]] = getelementptr i8, ptr [[TMP0]], i32 2
+// AARCH64-BE-NEXT:    [[TMP5:%.*]] = load i8, ptr [[TMP4]], align 2
+// AARCH64-BE-NEXT:    [[TMP6:%.*]] = and i8 [[TMP5]], -8
+// AARCH64-BE-NEXT:    store i8 [[TMP6]], ptr [[TMP4]], align 2
+// AARCH64-BE-NEXT:    [[TMP7:%.*]] = getelementptr i8, ptr [[TMP0]], i32 3
+// AARCH64-BE-NEXT:    store i8 0, ptr [[TMP7]], align 1
+// AARCH64-BE-NEXT:    ret void
+//
+void testBitfieldPartialEnds(struct BitfieldPartialEnds *s) {
+  __builtin_clear_padding(s);
+}
+
+// An asymmetric single-bit interval catches endpoint and reflection errors.
+struct BitfieldSingleValueBit {
+  unsigned char : 2;
+  unsigned char value : 1;
+  unsigned char : 5;
+};
+
+// LINUX-LABEL: define dso_local void @testBitfieldSingleValueBit(
+// LINUX-SAME: ptr noundef [[S:%.*]]) #[[ATTR0]] {
+// LINUX-NEXT:  [[ENTRY:.*:]]
+// LINUX-NEXT:    [[S_ADDR:%.*]] = alloca ptr, align 8
+// LINUX-NEXT:    store ptr [[S]], ptr [[S_ADDR]], align 8
+// LINUX-NEXT:    [[TMP0:%.*]] = load ptr, ptr [[S_ADDR]], align 8
+// LINUX-NEXT:    [[TMP1:%.*]] = getelementptr i8, ptr [[TMP0]], i32 0
+// LINUX-NEXT:    [[TMP2:%.*]] = load i8, ptr [[TMP1]], align 1
+// LINUX-NEXT:    [[TMP3:%.*]] = and i8 [[TMP2]], -4
+// LINUX-NEXT:    store i8 [[TMP3]], ptr [[TMP1]], align 1
+// LINUX-NEXT:    [[TMP4:%.*]] = getelementptr i8, ptr [[TMP0]], i32 0
+// LINUX-NEXT:    [[TMP5:%.*]] = load i8, ptr [[TMP4]], align 1
+// LINUX-NEXT:    [[TMP6:%.*]] = and i8 [[TMP5]], 7
+// LINUX-NEXT:    store i8 [[TMP6]], ptr [[TMP4]], align 1
+// LINUX-NEXT:    ret void
+//
+// WINDOWS-LABEL: define dso_local void @testBitfieldSingleValueBit(
+// WINDOWS-SAME: ptr noundef [[S:%.*]]) #[[ATTR0]] {
+// WINDOWS-NEXT:  [[ENTRY:.*:]]
+// WINDOWS-NEXT:    [[S_ADDR:%.*]] = alloca ptr, align 8
+// WINDOWS-NEXT:    store ptr [[S]], ptr [[S_ADDR]], align 8
+// WINDOWS-NEXT:    [[TMP0:%.*]] = load ptr, ptr [[S_ADDR]], align 8
+// WINDOWS-NEXT:    [[TMP1:%.*]] = getelementptr i8, ptr [[TMP0]], i32 0
+// WINDOWS-NEXT:    [[TMP2:%.*]] = load i8, ptr [[TMP1]], align 1
+// WINDOWS-NEXT:    [[TMP3:%.*]] = and i8 [[TMP2]], -4
+// WINDOWS-NEXT:    store i8 [[TMP3]], ptr [[TMP1]], align 1
+// WINDOWS-NEXT:    [[TMP4:%.*]] = getelementptr i8, ptr [[TMP0]], i32 0
+// WINDOWS-NEXT:    [[TMP5:%.*]] = load i8, ptr [[TMP4]], align 1
+// WINDOWS-NEXT:    [[TMP6:%.*]] = and i8 [[TMP5]], 7
+// WINDOWS-NEXT:    store i8 [[TMP6]], ptr [[TMP4]], align 1
+// WINDOWS-NEXT:    ret void
+//
+// ARM-LE-LABEL: define dso_local void @testBitfieldSingleValueBit(
+// ARM-LE-SAME: ptr noundef [[S:%.*]]) #[[ATTR0]] {
+// ARM-LE-NEXT:  [[ENTRY:.*:]]
+// ARM-LE-NEXT:    [[S_ADDR:%.*]] = alloca ptr, align 4
+// ARM-LE-NEXT:    store ptr [[S]], ptr [[S_ADDR]], align 4
+// ARM-LE-NEXT:    [[TMP0:%.*]] = load ptr, ptr [[S_ADDR]], align 4
+// ARM-LE-NEXT:    [[TMP1:%.*]] = getelementptr i8, ptr [[TMP0]], i32 0
+// ARM-LE-NEXT:    [[TMP2:%.*]] = load i8, ptr [[TMP1]], align 1
+// ARM-LE-NEXT:    [[TMP3:%.*]] = and i8 [[TMP2]], -4
+// ARM-LE-NEXT:    store i8 [[TMP3]], ptr [[TMP1]], align 1
+// ARM-LE-NEXT:    [[TMP4:%.*]] = getelementptr i8, ptr [[TMP0]], i32 0
+// ARM-LE-NEXT:    [[TMP5:%.*]] = load i8, ptr [[TMP4]], align 1
+// ARM-LE-NEXT:    [[TMP6:%.*]] = and i8 [[TMP5]], 7
+// ARM-LE-NEXT:    store i8 [[TMP6]], ptr [[TMP4]], align 1
+// ARM-LE-NEXT:    ret void
+//
+// ARM-BE-LABEL: define dso_local void @testBitfieldSingleValueBit(
+// ARM-BE-SAME: ptr noundef [[S:%.*]]) #[[ATTR0]] {
+// ARM-BE-NEXT:  [[ENTRY:.*:]]
+// ARM-BE-NEXT:    [[S_ADDR:%.*]] = alloca ptr, align 4
+// ARM-BE-NEXT:    store ptr [[S]], ptr [[S_ADDR]], align 4
+// ARM-BE-NEXT:    [[TMP0:%.*]] = load ptr, ptr [[S_ADDR]], align 4
+// ARM-BE-NEXT:    [[TMP1:%.*]] = getelementptr i8, ptr [[TMP0]], i32 0
+// ARM-BE-NEXT:    [[TMP2:%.*]] = load i8, ptr [[TMP1]], align 1
+// ARM-BE-NEXT:    [[TMP3:%.*]] = and i8 [[TMP2]], -32
+// ARM-BE-NEXT:    store i8 [[TMP3]], ptr [[TMP1]], align 1
+// ARM-BE-NEXT:    [[TMP4:%.*]] = getelementptr i8, ptr [[TMP0]], i32 0
+// ARM-BE-NEXT:    [[TMP5:%.*]] = load i8, ptr [[TMP4]], align 1
+// ARM-BE-NEXT:    [[TMP6:%.*]] = and i8 [[TMP5]], 63
+// ARM-BE-NEXT:    store i8 [[TMP6]], ptr [[TMP4]], align 1
+// ARM-BE-NEXT:    ret void
+//
+// AARCH64-LE-LABEL: define dso_local void @testBitfieldSingleValueBit(
+// AARCH64-LE-SAME: ptr noundef [[S:%.*]]) #[[ATTR0]] {
+// AARCH64-LE-NEXT:  [[ENTRY:.*:]]
+// AARCH64-LE-NEXT:    [[S_ADDR:%.*]] = alloca ptr, align 8
+// AARCH64-LE-NEXT:    store ptr [[S]], ptr [[S_ADDR]], align 8
+// AARCH64-LE-NEXT:    [[TMP0:%.*]] = load ptr, ptr [[S_ADDR]], align 8
+// AARCH64-LE-NEXT:    [[TMP1:%.*]] = getelementptr i8, ptr [[TMP0]], i32 0
+// AARCH64-LE-NEXT:    [[TMP2:%.*]] = load i8, ptr [[TMP1]], align 1
+// AARCH64-LE-NEXT:    [[TMP3:%.*]] = and i8 [[TMP2]], -4
+// AARCH64-LE-NEXT:    store i8 [[TMP3]], ptr [[TMP1]], align 1
+// AARCH64-LE-NEXT:    [[TMP4:%.*]] = getelementptr i8, ptr [[TMP0]], i32 0
+// AARCH64-LE-NEXT:    [[TMP5:%.*]] = load i8, ptr [[TMP4]], align 1
+// AARCH64-LE-NEXT:    [[TMP6:%.*]] = and i8 [[TMP5]], 7
+// AARCH64-LE-NEXT:    store i8 [[TMP6]], ptr [[TMP4]], align 1
+// AARCH64-LE-NEXT:    ret void
+//
+// AARCH64-BE-LABEL: define dso_local void @testBitfieldSingleValueBit(
+// AARCH64-BE-SAME: ptr noundef [[S:%.*]]) #[[ATTR0]] {
+// AARCH64-BE-NEXT:  [[ENTRY:.*:]]
+// AARCH64-BE-NEXT:    [[S_ADDR:%.*]] = alloca ptr, align 8
+// AARCH64-BE-NEXT:    store ptr [[S]], ptr [[S_ADDR]], align 8
+// AARCH64-BE-NEXT:    [[TMP0:%.*]] = load ptr, ptr [[S_ADDR]], align 8
+// AARCH64-BE-NEXT:    [[TMP1:%.*]] = getelementptr i8, ptr [[TMP0]], i32 0
+// AARCH64-BE-NEXT:    [[TMP2:%.*]] = load i8, ptr [[TMP1]], align 1
+// AARCH64-BE-NEXT:    [[TMP3:%.*]] = and i8 [[TMP2]], -32
+// AARCH64-BE-NEXT:    store i8 [[TMP3]], ptr [[TMP1]], align 1
+// AARCH64-BE-NEXT:    [[TMP4:%.*]] = getelementptr i8, ptr [[TMP0]], i32 0
+// AARCH64-BE-NEXT:    [[TMP5:%.*]] = load i8, ptr [[TMP4]], align 1
+// AARCH64-BE-NEXT:    [[TMP6:%.*]] = and i8 [[TMP5]], 63
+// AARCH64-BE-NEXT:    store i8 [[TMP6]], ptr [[TMP4]], align 1
+// AARCH64-BE-NEXT:    ret void
+//
+void testBitfieldSingleValueBit(struct BitfieldSingleValueBit *s) {
+  __builtin_clear_padding(s);
+}
+
+// Ordinary members share bitfield containers on Arm, and separate two runs of
+// bitfields. Clearing padding must preserve every ordinary member.
+struct BitfieldMixedMembers {
+  unsigned char prefix;
+  unsigned int first : 3;
+  unsigned char separator;
+  unsigned int second : 5;
+  unsigned short suffix;
+};
+
+// LINUX-LABEL: define dso_local void @testBitfieldMixedMembers(
+// LINUX-SAME: ptr noundef [[S:%.*]]) #[[ATTR0]] {
+// LINUX-NEXT:  [[ENTRY:.*:]]
+// LINUX-NEXT:    [[S_ADDR:%.*]] = alloca ptr, align 8
+// LINUX-NEXT:    store ptr [[S]], ptr [[S_ADDR]], align 8
+// LINUX-NEXT:    [[TMP0:%.*]] = load ptr, ptr [[S_ADDR]], align 8
+// LINUX-NEXT:    [[TMP1:%.*]] = getelementptr i8, ptr [[TMP0]], i32 1
+// LINUX-NEXT:    [[TMP2:%.*]] = load i8, ptr [[TMP1]], align 1
+// LINUX-NEXT:    [[TMP3:%.*]] = and i8 [[TMP2]], 7
+// LINUX-NEXT:    store i8 [[TMP3]], ptr [[TMP1]], align 1
+// LINUX-NEXT:    [[TMP4:%.*]] = getelementptr i8, ptr [[TMP0]], i32 3
+// LINUX-NEXT:    [[TMP5:%.*]] = load i8, ptr [[TMP4]], align 1
+// LINUX-NEXT:    [[TMP6:%.*]] = and i8 [[TMP5]], 31
+// LINUX-NEXT:    store i8 [[TMP6]], ptr [[TMP4]], align 1
+// LINUX-NEXT:    [[TMP7:%.*]] = getelementptr i8, ptr [[TMP0]], i32 6
+// LINUX-NEXT:    store i8 0, ptr [[TMP7]], align 2
+// LINUX-NEXT:    [[TMP8:%.*]] = getelementptr i8, ptr [[TMP0]], i32 7
+// LINUX-NEXT:    store i8 0, ptr [[TMP8]], align 1
+// LINUX-NEXT:    ret void
+//
+// WINDOWS-LABEL: define dso_local void @testBitfieldMixedMembers(
+// WINDOWS-SAME: ptr noundef [[S:%.*]]) #[[ATTR0]] {
+// WINDOWS-NEXT:  [[ENTRY:.*:]]
+// WINDOWS-NEXT:    [[S_ADDR:%.*]] = alloca ptr, align 8
+// WINDOWS-NEXT:    store ptr [[S]], ptr [[S_ADDR]], align 8
+// WINDOWS-NEXT:    [[TMP0:%.*]] = load ptr, ptr [[S_ADDR]], align 8
+// WINDOWS-NEXT:    [[TMP1:%.*]] = getelementptr i8, ptr [[TMP0]], i32 1
+// WINDOWS-NEXT:    store i8 0, ptr [[TMP1]], align 1
+// WINDOWS-NEXT:    [[TMP2:%.*]] = getelementptr i8, ptr [[TMP0]], i32 2
+// WINDOWS-NEXT:    store i8 0, ptr [[TMP2]], align 2
+// WINDOWS-NEXT:    [[TMP3:%.*]] = getelementptr i8, ptr [[TMP0]], i32 3
+// WINDOWS-NEXT:    store i8 0, ptr [[TMP3]], align 1
+// WINDOWS-NEXT:    [[TMP4:%.*]] = getelementptr i8, ptr [[TMP0]], i32 4
+// WINDOWS-NEXT:    [[TMP5:%.*]] = load i8, ptr [[TMP4]], align 4
+// WINDOWS-NEXT:    [[TMP6:%.*]] = and i8 [[TMP5]], 7
+// WINDOWS-NEXT:    store i8 [[TMP6]], ptr [[TMP4]], align 4
+// WINDOWS-NEXT:    [[TMP7:%.*]] = getelementptr i8, ptr [[TMP0]], i32 5
+// WINDOWS-NEXT:    store i8 0, ptr [[TMP7]], align 1
+// WINDOWS-NEXT:    [[TMP8:%.*]] = getelementptr i8, ptr [[TMP0]], i32 6
+// WINDOWS-NEXT:    store i8 0, ptr [[TMP8]], align 2
+// WINDOWS-NEXT:    [[TMP9:%.*]] = getelementptr i8, ptr [[TMP0]], i32 7
+// WINDOWS-NEXT:    store i8 0, ptr [[TMP9]], align 1
+// WINDOWS-NEXT:    [[TMP10:%.*]] = getelementptr i8, ptr [[TMP0]], i32 9
+// WINDOWS-NEXT:    store i8 0, ptr [[TMP10]], align 1
+// WINDOWS-NEXT:    [[TMP11:%.*]] = getelementptr i8, ptr [[TMP0]], i32 10
+// WINDOWS-NEXT:    store i8 0, ptr [[TMP11]], align 2
+// WINDOWS-NEXT:    [[TMP12:%.*]] = getelementptr i8, ptr [[TMP0]], i32 11
+// WINDOWS-NEXT:    store i8 0, ptr [[TMP12]], align 1
+// WINDOWS-NEXT:    [[TMP13:%.*]] = getelementptr i8, ptr [[TMP0]], i32 12
+// WINDOWS-NEXT:    [[TMP14:%.*]] = load i8, ptr [[TMP13]], align 4
+// WINDOWS-NEXT:    [[TMP15:%.*]] = and i8 [[TMP14]], 31
+// WINDOWS-NEXT:    store i8 [[TMP15]], ptr [[TMP13]], align 4
+// WINDOWS-NEXT:    [[TMP16:%.*]] = getelementptr i8, ptr [[TMP0]], i32 13
+// WINDOWS-NEXT:    store i8 0, ptr [[TMP16]], align 1
+// WINDOWS-NEXT:    [[TMP17:%.*]] = getelementptr i8, ptr [[TMP0]], i32 14
+// WINDOWS-NEXT:    store i8 0, ptr [[TMP17]], align 2
+// WINDOWS-NEXT:    [[TMP18:%.*]] = getelementptr i8, ptr [[TMP0]], i32 15
+// WINDOWS-NEXT:    store i8 0, ptr [[TMP18]], align 1
+// WINDOWS-NEXT:    [[TMP19:%.*]] = getelementptr i8, ptr [[TMP0]], i32 18
+// WINDOWS-NEXT:    store i8 0, ptr [[TMP19]], align 2
+// WINDOWS-NEXT:    [[TMP20:%.*]] = getelementptr i8, ptr [[TMP0]], i32 19
+// WINDOWS-NEXT:    store i8 0, ptr [[TMP20]], align 1
+// WINDOWS-NEXT:    ret void
+//
+// ARM-LE-LABEL: define dso_local void @testBitfieldMixedMembers(
+// ARM-LE-SAME: ptr noundef [[S:%.*]]) #[[ATTR0]] {
+// ARM-LE-NEXT:  [[ENTRY:.*:]]
+// ARM-LE-NEXT:    [[S_ADDR:%.*]] = alloca ptr, align 4
+// ARM-LE-NEXT:    store ptr [[S]], ptr [[S_ADDR]], align 4
+// ARM-LE-NEXT:    [[TMP0:%.*]] = load ptr, ptr [[S_ADDR]], align 4
+// ARM-LE-NEXT:    [[TMP1:%.*]] = getelementptr i8, ptr [[TMP0]], i32 1
+// ARM-LE-NEXT:    [[TMP2:%.*]] = load i8, ptr [[TMP1]], align 1
+// ARM-LE-NEXT:    [[TMP3:%.*]] = and i8 [[TMP2]], 7
+// ARM-LE-NEXT:    store i8 [[TMP3]], ptr [[TMP1]], align 1
+// ARM-LE-NEXT:    [[TMP4:%.*]] = getelementptr i8, ptr [[TMP0]], i32 3
+// ARM-LE-NEXT:    [[TMP5:%.*]] = load i8, ptr [[TMP4]], align 1
+// ARM-LE-NEXT:    [[TMP6:%.*]] = and i8 [[TMP5]], 31
+// ARM-LE-NEXT:    store i8 [[TMP6]], ptr [[TMP4]], align 1
+// ARM-LE-NEXT:    [[TMP7:%.*]] = getelementptr i8, ptr [[TMP0]], i32 6
+// ARM-LE-NEXT:    store i8 0, ptr [[TMP7]], align 2
+// ARM-LE-NEXT:    [[TMP8:%.*]] = getelementptr i8, ptr [[TMP0]], i32 7
+// ARM-LE-NEXT:    store i8 0, ptr [[TMP8]], align 1
+// ARM-LE-NEXT:    ret void
+//
+// ARM-BE-LABEL: define dso_local void @testBitfieldMixedMembers(
+// ARM-BE-SAME: ptr noundef [[S:%.*]]) #[[ATTR0]] {
+// ARM-BE-NEXT:  [[ENTRY:.*:]]
+// ARM-BE-NEXT:    [[S_ADDR:%.*]] = alloca ptr, align 4
+// ARM-BE-NEXT:    store ptr [[S]], ptr [[S_ADDR]], align 4
+// ARM-BE-NEXT:    [[TMP0:%.*]] = load ptr, ptr [[S_ADDR]], align 4
+// ARM-BE-NEXT:    [[TMP1:%.*]] = getelementptr i8, ptr [[TMP0]], i32 1
+// ARM-BE-NEXT:    [[TMP2:%.*]] = load i8, ptr [[TMP1]], align 1
+// ARM-BE-NEXT:    [[TMP3:%.*]] = and i8 [[TMP2]], -32
+// ARM-BE-NEXT:    store i8 [[TMP3]], ptr [[TMP1]], align 1
+// ARM-BE-NEXT:    [[TMP4:%.*]] = getelementptr i8, ptr [[TMP0]], i32 3
+// ARM-BE-NEXT:    [[TMP5:%.*]] = load i8, ptr [[TMP4]], align 1
+// ARM-BE-NEXT:    [[TMP6:%.*]] = and i8 [[TMP5]], -8
+// ARM-BE-NEXT:    store i8 [[TMP6]], ptr [[TMP4]], align 1
+// ARM-BE-NEXT:    [[TMP7:%.*]] = getelementptr i8, ptr [[TMP0]], i32 6
+// ARM-BE-NEXT:    store i8 0, ptr [[TMP7]], align 2
+// ARM-BE-NEXT:    [[TMP8:%.*]] = getelementptr i8, ptr [[TMP0]], i32 7
+// ARM-BE-NEXT:    store i8 0, ptr [[TMP8]], align 1
+// ARM-BE-NEXT:    ret void
+//
+// AARCH64-LE-LABEL: define dso_local void @testBitfieldMixedMembers(
+// AARCH64-LE-SAME: ptr noundef [[S:%.*]]) #[[ATTR0]] {
+// AARCH64-LE-NEXT:  [[ENTRY:.*:]]
+// AARCH64-LE-NEXT:    [[S_ADDR:%.*]] = alloca ptr, align 8
+// AARCH64-LE-NEXT:    store ptr [[S]], ptr [[S_ADDR]], align 8
+// AARCH64-LE-NEXT:    [[TMP0:%.*]] = load ptr, ptr [[S_ADDR]], align 8
+// AARCH64-LE-NEXT:    [[TMP1:%.*]] = getelementptr i8, ptr [[TMP0]], i32 1
+// AARCH64-LE-NEXT:    [[TMP2:%.*]] = load i8, ptr [[TMP1]], align 1
+// AARCH64-LE-NEXT:    [[TMP3:%.*]] = and i8 [[TMP2]], 7
+// AARCH64-LE-NEXT:    store i8 [[TMP3]], ptr [[TMP1]], align 1
+// AARCH64-LE-NEXT:    [[TMP4:%.*]] = getelementptr i8, ptr [[TMP0]], i32 3
+// AARCH64-LE-NEXT:    [[TMP5:%.*]] = load i8, ptr [[TMP4]], align 1
+// AARCH64-LE-NEXT:    [[TMP6:%.*]] = and i8 [[TMP5]], 31
+// AARCH64-LE-NEXT:    store i8 [[TMP6]], ptr [[TMP4]], align 1
+// AARCH64-LE-NEXT:    [[TMP7:%.*]] = getelementptr i8, ptr [[TMP0]], i32 6
+// AARCH64-LE-NEXT:    store i8 0, ptr [[TMP7]], align 2
+// AARCH64-LE-NEXT:    [[TMP8:%.*]] = getelementptr i8, ptr [[TMP0]], i32 7
+// AARCH64-LE-NEXT:    store i8 0, ptr [[TMP8]], align 1
+// AARCH64-LE-NEXT:    ret void
+//
+// AARCH64-BE-LABEL: define dso_local void @testBitfieldMixedMembers(
+// AARCH64-BE-SAME: ptr noundef [[S:%.*]]) #[[ATTR0]] {
+// AARCH64-BE-NEXT:  [[ENTRY:.*:]]
+// AARCH64-BE-NEXT:    [[S_ADDR:%.*]] = alloca ptr, align 8
+// AARCH64-BE-NEXT:    store ptr [[S]], ptr [[S_ADDR]], align 8
+// AARCH64-BE-NEXT:    [[TMP0:%.*]] = load ptr, ptr [[S_ADDR]], align 8
+// AARCH64-BE-NEXT:    [[TMP1:%.*]] = getelementptr i8, ptr [[TMP0]], i32 1
+// AARCH64-BE-NEXT:    [[TMP2:%.*]] = load i8, ptr [[TMP1]], align 1
+// AARCH64-BE-NEXT:    [[TMP3:%.*]] = and i8 [[TMP2]], -32
+// AARCH64-BE-NEXT:    store i8 [[TMP3]], ptr [[TMP1]], align 1
+// AARCH64-BE-NEXT:    [[TMP4:%.*]] = getelementptr i8, ptr [[TMP0]], i32 3
+// AARCH64-BE-NEXT:    [[TMP5:%.*]] = load i8, ptr [[TMP4]], align 1
+// AARCH64-BE-NEXT:    [[TMP6:%.*]] = and i8 [[TMP5]], -8
+// AARCH64-BE-NEXT:    store i8 [[TMP6]], ptr [[TMP4]], align 1
+// AARCH64-BE-NEXT:    [[TMP7:%.*]] = getelementptr i8, ptr [[TMP0]], i32 6
+// AARCH64-BE-NEXT:    store i8 0, ptr [[TMP7]], align 2
+// AARCH64-BE-NEXT:    [[TMP8:%.*]] = getelementptr i8, ptr [[TMP0]], i32 7
+// AARCH64-BE-NEXT:    store i8 0, ptr [[TMP8]], align 1
+// AARCH64-BE-NEXT:    ret void
+//
+void testBitfieldMixedMembers(struct BitfieldMixedMembers *s) {
+  __builtin_clear_padding(s);
+}
+
+// LINUX-LABEL: define dso_local void @testBitfieldMixedMembersVolatileObject(
+// LINUX-SAME: ptr noundef [[S:%.*]]) #[[ATTR0]] {
+// LINUX-NEXT:  [[ENTRY:.*:]]
+// LINUX-NEXT:    [[S_ADDR:%.*]] = alloca ptr, align 8
+// LINUX-NEXT:    store ptr [[S]], ptr [[S_ADDR]], align 8
+// LINUX-NEXT:    [[TMP0:%.*]] = load ptr, ptr [[S_ADDR]], align 8
+// LINUX-NEXT:    [[TMP1:%.*]] = getelementptr i8, ptr [[TMP0]], i32 1
+// LINUX-NEXT:    [[TMP2:%.*]] = load i8, ptr [[TMP1]], align 1
+// LINUX-NEXT:    [[TMP3:%.*]] = and i8 [[TMP2]], 7
+// LINUX-NEXT:    store i8 [[TMP3]], ptr [[TMP1]], align 1
+// LINUX-NEXT:    [[TMP4:%.*]] = getelementptr i8, ptr [[TMP0]], i32 3
+// LINUX-NEXT:    [[TMP5:%.*]] = load i8, ptr [[TMP4]], align 1
+// LINUX-NEXT:    [[TMP6:%.*]] = and i8 [[TMP5]], 31
+// LINUX-NEXT:    store i8 [[TMP6]], ptr [[TMP4]], align 1
+// LINUX-NEXT:    [[TMP7:%.*]] = getelementptr i8, ptr [[TMP0]], i32 6
+// LINUX-NEXT:    store i8 0, ptr [[TMP7]], align 2
+// LINUX-NEXT:    [[TMP8:%.*]] = getelementptr i8, ptr [[TMP0]], i32 7
+// LINUX-NEXT:    store i8 0, ptr [[TMP8]], align 1
+// LINUX-NEXT:    ret void
+//
+// WINDOWS-LABEL: define dso_local void @testBitfieldMixedMembersVolatileObject(
+// WINDOWS-SAME: ptr noundef [[S:%.*]]) #[[ATTR0]] {
+// WINDOWS-NEXT:  [[ENTRY:.*:]]
+// WINDOWS-NEXT:    [[S_ADDR:%.*]] = alloca ptr, align 8
+// WINDOWS-NEXT:    store ptr [[S]], ptr [[S_ADDR]], align 8
+// WINDOWS-NEXT:    [[TMP0:%.*]] = load ptr, ptr [[S_ADDR]], align 8
+// WINDOWS-NEXT:    [[TMP1:%.*]] = getelementptr i8, ptr [[TMP0]], i32 1
+// WINDOWS-NEXT:    store i8 0, ptr [[TMP1]], align 1
+// WINDOWS-NEXT:    [[TMP2:%.*]] = getelementptr i8, ptr [[TMP0]], i32 2
+// WINDOWS-NEXT:    store i8 0, ptr [[TMP2]], align 2
+// WINDOWS-NEXT:    [[TMP3:%.*]] = getelementptr i8, ptr [[TMP0]], i32 3
+// WINDOWS-NEXT:    store i8 0, ptr [[TMP3]], align 1
+// WINDOWS-NEXT:    [[TMP4:%.*]] = getelementptr i8, ptr [[TMP0]], i32 4
+// WINDOWS-NEXT:    [[TMP5:%.*]] = load i8, ptr [[TMP4]], align 4
+// WINDOWS-NEXT:    [[TMP6:%.*]] = and i8 [[TMP5]], 7
+// WINDOWS-NEXT:    store i8 [[TMP6]], ptr [[TMP4]], align 4
+// WINDOWS-NEXT:    [[TMP7:%.*]] = getelementptr i8, ptr [[TMP0]], i32 5
+// WINDOWS-NEXT:    store i8 0, ptr [[TMP7]], align 1
+// WINDOWS-NEXT:    [[TMP8:%.*]] = getelementptr i8, ptr [[TMP0]], i32 6
+// WINDOWS-NEXT:    store i8 0, ptr [[TMP8]], align 2
+// WINDOWS-NEXT:    [[TMP9:%.*]] = getelementptr i8, ptr [[TMP0]], i32 7
+// WINDOWS-NEXT:    store i8 0, ptr [[TMP9]], align 1
+// WINDOWS-NEXT:    [[TMP10:%.*]] = getelementptr i8, ptr [[TMP0]], i32 9
+// WINDOWS-NEXT:    store i8 0, ptr [[TMP10]], align 1
+// WINDOWS-NEXT:    [[TMP11:%.*]] = getelementptr i8, ptr [[TMP0]], i32 10
+// WINDOWS-NEXT:    store i8 0, ptr [[TMP11]], align 2
+// WINDOWS-NEXT:    [[TMP12:%.*]] = getelementptr i8, ptr [[TMP0]], i32 11
+// WINDOWS-NEXT:    store i8 0, ptr [[TMP12]], align 1
+// WINDOWS-NEXT:    [[TMP13:%.*]] = getelementptr i8, ptr [[TMP0]], i32 12
+// WINDOWS-NEXT:    [[TMP14:%.*]] = load i8, ptr [[TMP13]], align 4
+// WINDOWS-NEXT:    [[TMP15:%.*]] = and i8 [[TMP14]], 31
+// WINDOWS-NEXT:    store i8 [[TMP15]], ptr [[TMP13]], align 4
+// WINDOWS-NEXT:    [[TMP16:%.*]] = getelementptr i8, ptr [[TMP0]], i32 13
+// WINDOWS-NEXT:    store i8 0, ptr [[TMP16]], align 1
+// WINDOWS-NEXT:    [[TMP17:%.*]] = getelementptr i8, ptr [[TMP0]], i32 14
+// WINDOWS-NEXT:    store i8 0, ptr [[TMP17]], align 2
+// WINDOWS-NEXT:    [[TMP18:%.*]] = getelementptr i8, ptr [[TMP0]], i32 15
+// WINDOWS-NEXT:    store i8 0, ptr [[TMP18]], align 1
+// WINDOWS-NEXT:    [[TMP19:%.*]] = getelementptr i8, ptr [[TMP0]], i32 18
+// WINDOWS-NEXT:    store i8 0, ptr [[TMP19]], align 2
+// WINDOWS-NEXT:    [[TMP20:%.*]] = getelementptr i8, ptr [[TMP0]], i32 19
+// WINDOWS-NEXT:    store i8 0, ptr [[TMP20]], align 1
+// WINDOWS-NEXT:    ret void
+//
+// ARM-LE-LABEL: define dso_local void @testBitfieldMixedMembersVolatileObject(
+// ARM-LE-SAME: ptr noundef [[S:%.*]]) #[[ATTR0]] {
+// ARM-LE-NEXT:  [[ENTRY:.*:]]
+// ARM-LE-NEXT:    [[S_ADDR:%.*]] = alloca ptr, align 4
+// ARM-LE-NEXT:    store ptr [[S]], ptr [[S_ADDR]], align 4
+// ARM-LE-NEXT:    [[TMP0:%.*]] = load ptr, ptr [[S_ADDR]], align 4
+// ARM-LE-NEXT:    [[TMP1:%.*]] = getelementptr i8, ptr [[TMP0]], i32 1
+// ARM-LE-NEXT:    [[TMP2:%.*]] = load i8, ptr [[TMP1]], align 1
+// ARM-LE-NEXT:    [[TMP3:%.*]] = and i8 [[TMP2]], 7
+// ARM-LE-NEXT:    store i8 [[TMP3]], ptr [[TMP1]], align 1
+// ARM-LE-NEXT:    [[TMP4:%.*]] = getelementptr i8, ptr [[TMP0]], i32 3
+// ARM-LE-NEXT:    [[TMP5:%.*]] = load i8, ptr [[TMP4]], align 1
+// ARM-LE-NEXT:    [[TMP6:%.*]] = and i8 [[TMP5]], 31
+// ARM-LE-NEXT:    store i8 [[TMP6]], ptr [[TMP4]], align 1
+// ARM-LE-NEXT:    [[TMP7:%.*]] = getelementptr i8, ptr [[TMP0]], i32 6
+// ARM-LE-NEXT:    store i8 0, ptr [[TMP7]], align 2
+// ARM-LE-NEXT:    [[TMP8:%.*]] = getelementptr i8, ptr [[TMP0]], i32 7
+// ARM-LE-NEXT:    store i8 0, ptr [[TMP8]], align 1
+// ARM-LE-NEXT:    ret void
+//
+// ARM-BE-LABEL: define dso_local void @testBitfieldMixedMembersVolatileObject(
+// ARM-BE-SAME: ptr noundef [[S:%.*]]) #[[ATTR0]] {
+// ARM-BE-NEXT:  [[ENTRY:.*:]]
+// ARM-BE-NEXT:    [[S_ADDR:%.*]] = alloca ptr, align 4
+// ARM-BE-NEXT:    store ptr [[S]], ptr [[S_ADDR]], align 4
+// ARM-BE-NEXT:    [[TMP0:%.*]] = load ptr, ptr [[S_ADDR]], align 4
+// ARM-BE-NEXT:    [[TMP1:%.*]] = getelementptr i8, ptr [[TMP0]], i32 1
+// ARM-BE-NEXT:    [[TMP2:%.*]] = load i8, ptr [[TMP1]], align 1
+// ARM-BE-NEXT:    [[TMP3:%.*]] = and i8 [[TMP2]], -32
+// ARM-BE-NEXT:    store i8 [[TMP3]], ptr [[TMP1]], align 1
+// ARM-BE-NEXT:    [[TMP4:%.*]] = getelementptr i8, ptr [[TMP0]], i32 3
+// ARM-BE-NEXT:    [[TMP5:%.*]] = load i8, ptr [[TMP4]], align 1
+// ARM-BE-NEXT:    [[TMP6:%.*]] = and i8 [[TMP5]], -8
+// ARM-BE-NEXT:    store i8 [[TMP6]], ptr [[TMP4]], align 1
+// ARM-BE-NEXT:    [[TMP7:%.*]] = getelementptr i8, ptr [[TMP0]], i32 6
+// ARM-BE-NEXT:    store i8 0, ptr [[TMP7]], align 2
+// ARM-BE-NEXT:    [[TMP8:%.*]] = getelementptr i8, ptr [[TMP0]], i32 7
+// ARM-BE-NEXT:    store i8 0, ptr [[TMP8]], align 1
+// ARM-BE-NEXT:    ret void
+//
+// AARCH64-LE-LABEL: define dso_local void @testBitfieldMixedMembersVolatileObject(
+// AARCH64-LE-SAME: ptr noundef [[S:%.*]]) #[[ATTR0]] {
+// AARCH64-LE-NEXT:  [[ENTRY:.*:]]
+// AARCH64-LE-NEXT:    [[S_ADDR:%.*]] = alloca ptr, align 8
+// AARCH64-LE-NEXT:    store ptr [[S]], ptr [[S_ADDR]], align 8
+// AARCH64-LE-NEXT:    [[TMP0:%.*]] = load ptr, ptr [[S_ADDR]], align 8
+// AARCH64-LE-NEXT:    [[TMP1:%.*]] = getelementptr i8, ptr [[TMP0]], i32 1
+// AARCH64-LE-NEXT:    [[TMP2:%.*]] = load i8, ptr [[TMP1]], align 1
+// AARCH64-LE-NEXT:    [[TMP3:%.*]] = and i8 [[TMP2]], 7
+// AARCH64-LE-NEXT:    store i8 [[TMP3]], ptr [[TMP1]], align 1
+// AARCH64-LE-NEXT:    [[TMP4:%.*]] = getelementptr i8, ptr [[TMP0]], i32 3
+// AARCH64-LE-NEXT:    [[TMP5:%.*]] = load i8, ptr [[TMP4]], align 1
+// AARCH64-LE-NEXT:    [[TMP6:%.*]] = and i8 [[TMP5]], 31
+// AARCH64-LE-NEXT:    store i8 [[TMP6]], ptr [[TMP4]], align 1
+// AARCH64-LE-NEXT:    [[TMP7:%.*]] = getelementptr i8, ptr [[TMP0]], i32 6
+// AARCH64-LE-NEXT:    store i8 0, ptr [[TMP7]], align 2
+// AARCH64-LE-NEXT:    [[TMP8:%.*]] = getelementptr i8, ptr [[TMP0]], i32 7
+// AARCH64-LE-NEXT:    store i8 0, ptr [[TMP8]], align 1
+// AARCH64-LE-NEXT:    ret void
+//
+// AARCH64-BE-LABEL: define dso_local void @testBitfieldMixedMembersVolatileObject(
+// AARCH64-BE-SAME: ptr noundef [[S:%.*]]) #[[ATTR0]] {
+// AARCH64-BE-NEXT:  [[ENTRY:.*:]]
+// AARCH64-BE-NEXT:    [[S_ADDR:%.*]] = alloca ptr, align 8
+// AARCH64-BE-NEXT:    store ptr [[S]], ptr [[S_ADDR]], align 8
+// AARCH64-BE-NEXT:    [[TMP0:%.*]] = load ptr, ptr [[S_ADDR]], align 8
+// AARCH64-BE-NEXT:    [[TMP1:%.*]] = getelementptr i8, ptr [[TMP0]], i32 1
+// AARCH64-BE-NEXT:    [[TMP2:%.*]] = load i8, ptr [[TMP1]], align 1
+// AARCH64-BE-NEXT:    [[TMP3:%.*]] = and i8 [[TMP2]], -32
+// AARCH64-BE-NEXT:    store i8 [[TMP3]], ptr [[TMP1]], align 1
+// AARCH64-BE-NEXT:    [[TMP4:%.*]] = getelementptr i8, ptr [[TMP0]], i32 3
+// AARCH64-BE-NEXT:    [[TMP5:%.*]] = load i8, ptr [[TMP4]], align 1
+// AARCH64-BE-NEXT:    [[TMP6:%.*]] = and i8 [[TMP5]], -8
+// AARCH64-BE-NEXT:    store i8 [[TMP6]], ptr [[TMP4]], align 1
+// AARCH64-BE-NEXT:    [[TMP7:%.*]] = getelementptr i8, ptr [[TMP0]], i32 6
+// AARCH64-BE-NEXT:    store i8 0, ptr [[TMP7]], align 2
+// AARCH64-BE-NEXT:    [[TMP8:%.*]] = getelementptr i8, ptr [[TMP0]], i32 7
+// AARCH64-BE-NEXT:    store i8 0, ptr [[TMP8]], align 1
+// AARCH64-BE-NEXT:    ret void
+//
+void testBitfieldMixedMembersVolatileObject(
+    volatile struct BitfieldMixedMembers *s) {
+  __builtin_clear_padding(s);
+}
+
+struct BitfieldMixedVolatileMembers {
+  unsigned char prefix;
+  volatile unsigned int first : 3;
+  unsigned char separator;
+  volatile unsigned int second : 5;
+  unsigned short suffix;
+};
+
+// LINUX-LABEL: define dso_local void @testBitfieldMixedVolatileMembers(
+// LINUX-SAME: ptr noundef [[S:%.*]]) #[[ATTR0]] {
+// LINUX-NEXT:  [[ENTRY:.*:]]
+// LINUX-NEXT:    [[S_ADDR:%.*]] = alloca ptr, align 8
+// LINUX-NEXT:    store ptr [[S]], ptr [[S_ADDR]], align 8
+// LINUX-NEXT:    [[TMP0:%.*]] = load ptr, ptr [[S_ADDR]], align 8
+// LINUX-NEXT:    [[TMP1:%.*]] = getelementptr i8, ptr [[TMP0]], i32 1
+// LINUX-NEXT:    [[TMP2:%.*]] = load i8, ptr [[TMP1]], align 1
+// LINUX-NEXT:    [[TMP3:%.*]] = and i8 [[TMP2]], 7
+// LINUX-NEXT:    store i8 [[TMP3]], ptr [[TMP1]], align 1
+// LINUX-NEXT:    [[TMP4:%.*]] = getelementptr i8, ptr [[TMP0]], i32 3
+// LINUX-NEXT:    [[TMP5:%.*]] = load i8, ptr [[TMP4]], align 1
+// LINUX-NEXT:    [[TMP6:%.*]] = and i8 [[TMP5]], 31
+// LINUX-NEXT:    store i8 [[TMP6]], ptr [[TMP4]], align 1
+// LINUX-NEXT:    [[TMP7:%.*]] = getelementptr i8, ptr [[TMP0]], i32 6
+// LINUX-NEXT:    store i8 0, ptr [[TMP7]], align 2
+// LINUX-NEXT:    [[TMP8:%.*]] = getelementptr i8, ptr [[TMP0]], i32 7
+// LINUX-NEXT:    store i8 0, ptr [[TMP8]], align 1
+// LINUX-NEXT:    ret void
+//
+// WINDOWS-LABEL: define dso_local void @testBitfieldMixedVolatileMembers(
+// WINDOWS-SAME: ptr noundef [[S:%.*]]) #[[ATTR0]] {
+// WINDOWS-NEXT:  [[ENTRY:.*:]]
+// WINDOWS-NEXT:    [[S_ADDR:%.*]] = alloca ptr, align 8
+// WINDOWS-NEXT:    store ptr [[S]], ptr [[S_ADDR]], align 8
+// WINDOWS-NEXT:    [[TMP0:%.*]] = load ptr, ptr [[S_ADDR]], align 8
+// WINDOWS-NEXT:    [[TMP1:%.*]] = getelementptr i8, ptr [[TMP0]], i32 1
+// WINDOWS-NEXT:    store i8 0, ptr [[TMP1]], align 1
+// WINDOWS-NEXT:    [[TMP2:%.*]] = getelementptr i8, ptr [[TMP0]], i32 2
+// WINDOWS-NEXT:    store i8 0, ptr [[TMP2]], align 2
+// WINDOWS-NEXT:    [[TMP3:%.*]] = getelementptr i8, ptr [[TMP0]], i32 3
+// WINDOWS-NEXT:    store i8 0, ptr [[TMP3]], align 1
+// WINDOWS-NEXT:    [[TMP4:%.*]] = getelementptr i8, ptr [[TMP0]], i32 4
+// WINDOWS-NEXT:    [[TMP5:%.*]] = load i8, ptr [[TMP4]], align 4
+// WINDOWS-NEXT:    [[TMP6:%.*]] = and i8 [[TMP5]], 7
+// WINDOWS-NEXT:    store i8 [[TMP6]], ptr [[TMP4]], align 4
+// WINDOWS-NEXT:    [[TMP7:%.*]] = getelementptr i8, ptr [[TMP0]], i32 5
+// WINDOWS-NEXT:    store i8 0, ptr [[TMP7]], align 1
+// WINDOWS-NEXT:    [[TMP8:%.*]] = getelementptr i8, ptr [[TMP0]], i32 6
+// WINDOWS-NEXT:    store i8 0, ptr [[TMP8]], align 2
+// WINDOWS-NEXT:    [[TMP9:%.*]] = getelementptr i8, ptr [[TMP0]], i32 7
+// WINDOWS-NEXT:    store i8 0, ptr [[TMP9]], align 1
+// WINDOWS-NEXT:    [[TMP10:%.*]] = getelementptr i8, ptr [[TMP0]], i32 9
+// WINDOWS-NEXT:    store i8 0, ptr [[TMP10]], align 1
+// WINDOWS-NEXT:    [[TMP11:%.*]] = getelementptr i8, ptr [[TMP0]], i32 10
+// WINDOWS-NEXT:    store i8 0, ptr [[TMP11]], align 2
+// WINDOWS-NEXT:    [[TMP12:%.*]] = getelementptr i8, ptr [[TMP0]], i32 11
+// WINDOWS-NEXT:    store i8 0, ptr [[TMP12]], align 1
+// WINDOWS-NEXT:    [[TMP13:%.*]] = getelementptr i8, ptr [[TMP0]], i32 12
+// WINDOWS-NEXT:    [[TMP14:%.*]] = load i8, ptr [[TMP13]], align 4
+// WINDOWS-NEXT:    [[TMP15:%.*]] = and i8 [[TMP14]], 31
+// WINDOWS-NEXT:    store i8 [[TMP15]], ptr [[TMP13]], align 4
+// WINDOWS-NEXT:    [[TMP16:%.*]] = getelementptr i8, ptr [[TMP0]], i32 13
+// WINDOWS-NEXT:    store i8 0, ptr [[TMP16]], align 1
+// WINDOWS-NEXT:    [[TMP17:%.*]] = getelementptr i8, ptr [[TMP0]], i32 14
+// WINDOWS-NEXT:    store i8 0, ptr [[TMP17]], align 2
+// WINDOWS-NEXT:    [[TMP18:%.*]] = getelementptr i8, ptr [[TMP0]], i32 15
+// WINDOWS-NEXT:    store i8 0, ptr [[TMP18]], align 1
+// WINDOWS-NEXT:    [[TMP19:%.*]] = getelementptr i8, ptr [[TMP0]], i32 18
+// WINDOWS-NEXT:    store i8 0, ptr [[TMP19]], align 2
+// WINDOWS-NEXT:    [[TMP20:%.*]] = getelementptr i8, ptr [[TMP0]], i32 19
+// WINDOWS-NEXT:    store i8 0, ptr [[TMP20]], align 1
+// WINDOWS-NEXT:    ret void
+//
+// ARM-LE-LABEL: define dso_local void @testBitfieldMixedVolatileMembers(
+// ARM-LE-SAME: ptr noundef [[S:%.*]]) #[[ATTR0]] {
+// ARM-LE-NEXT:  [[ENTRY:.*:]]
+// ARM-LE-NEXT:    [[S_ADDR:%.*]] = alloca ptr, align 4
+// ARM-LE-NEXT:    store ptr [[S]], ptr [[S_ADDR]], align 4
+// ARM-LE-NEXT:    [[TMP0:%.*]] = load ptr, ptr [[S_ADDR]], align 4
+// ARM-LE-NEXT:    [[TMP1:%.*]] = getelementptr i8, ptr [[TMP0]], i32 1
+// ARM-LE-NEXT:    [[TMP2:%.*]] = load i8, ptr [[TMP1]], align 1
+// ARM-LE-NEXT:    [[TMP3:%.*]] = and i8 [[TMP2]], 7
+// ARM-LE-NEXT:    store i8 [[TMP3]], ptr [[TMP1]], align 1
+// ARM-LE-NEXT:    [[TMP4:%.*]] = getelementptr i8, ptr [[TMP0]], i32 3
+// ARM-LE-NEXT:    [[TMP5:%.*]] = load i8, ptr [[TMP4]], align 1
+// ARM-LE-NEXT:    [[TMP6:%.*]] = and i8 [[TMP5]], 31
+// ARM-LE-NEXT:    store i8 [[TMP6]], ptr [[TMP4]], align 1
+// ARM-LE-NEXT:    [[TMP7:%.*]] = getelementptr i8, ptr [[TMP0]], i32 6
+// ARM-LE-NEXT:    store i8 0, ptr [[TMP7]], align 2
+// ARM-LE-NEXT:    [[TMP8:%.*]] = getelementptr i8, ptr [[TMP0]], i32 7
+// ARM-LE-NEXT:    store i8 0, ptr [[TMP8]], align 1
+// ARM-LE-NEXT:    ret void
+//
+// ARM-BE-LABEL: define dso_local void @testBitfieldMixedVolatileMembers(
+// ARM-BE-SAME: ptr noundef [[S:%.*]]) #[[ATTR0]] {
+// ARM-BE-NEXT:  [[ENTRY:.*:]]
+// ARM-BE-NEXT:    [[S_ADDR:%.*]] = alloca ptr, align 4
+// ARM-BE-NEXT:    store ptr [[S]], ptr [[S_ADDR]], align 4
+// ARM-BE-NEXT:    [[TMP0:%.*]] = load ptr, ptr [[S_ADDR]], align 4
+// ARM-BE-NEXT:    [[TMP1:%.*]] = getelementptr i8, ptr [[TMP0]], i32 1
+// ARM-BE-NEXT:    [[TMP2:%.*]] = load i8, ptr [[TMP1]], align 1
+// ARM-BE-NEXT:    [[TMP3:%.*]] = and i8 [[TMP2]], -32
+// ARM-BE-NEXT:    store i8 [[TMP3]], ptr [[TMP1]], align 1
+// ARM-BE-NEXT:    [[TMP4:%.*]] = getelementptr i8, ptr [[TMP0]], i32 3
+// ARM-BE-NEXT:    [[TMP5:%.*]] = load i8, ptr [[TMP4]], align 1
+// ARM-BE-NEXT:    [[TMP6:%.*]] = and i8 [[TMP5]], -8
+// ARM-BE-NEXT:    store i8 [[TMP6]], ptr [[TMP4]], align 1
+// ARM-BE-NEXT:    [[TMP7:%.*]] = getelementptr i8, ptr [[TMP0]], i32 6
+// ARM-BE-NEXT:    store i8 0, ptr [[TMP7]], align 2
+// ARM-BE-NEXT:    [[TMP8:%.*]] = getelementptr i8, ptr [[TMP0]], i32 7
+// ARM-BE-NEXT:    store i8 0, ptr [[TMP8]], align 1
+// ARM-BE-NEXT:    ret void
+//
+// AARCH64-LE-LABEL: define dso_local void @testBitfieldMixedVolatileMembers(
+// AARCH64-LE-SAME: ptr noundef [[S:%.*]]) #[[ATTR0]] {
+// AARCH64-LE-NEXT:  [[ENTRY:.*:]]
+// AARCH64-LE-NEXT:    [[S_ADDR:%.*]] = alloca ptr, align 8
+// AARCH64-LE-NEXT:    store ptr [[S]], ptr [[S_ADDR]], align 8
+// AARCH64-LE-NEXT:    [[TMP0:%.*]] = load ptr, ptr [[S_ADDR]], align 8
+// AARCH64-LE-NEXT:    [[TMP1:%.*]] = getelementptr i8, ptr [[TMP0]], i32 1
+// AARCH64-LE-NEXT:    [[TMP2:%.*]] = load i8, ptr [[TMP1]], align 1
+// AARCH64-LE-NEXT:    [[TMP3:%.*]] = and i8 [[TMP2]], 7
+// AARCH64-LE-NEXT:    store i8 [[TMP3]], ptr [[TMP1]], align 1
+// AARCH64-LE-NEXT:    [[TMP4:%.*]] = getelementptr i8, ptr [[TMP0]], i32 3
+// AARCH64-LE-NEXT:    [[TMP5:%.*]] = load i8, ptr [[TMP4]], align 1
+// AARCH64-LE-NEXT:    [[TMP6:%.*]] = and i8 [[TMP5]], 31
+// AARCH64-LE-NEXT:    store i8 [[TMP6]], ptr [[TMP4]], align 1
+// AARCH64-LE-NEXT:    [[TMP7:%.*]] = getelementptr i8, ptr [[TMP0]], i32 6
+// AARCH64-LE-NEXT:    store i8 0, ptr [[TMP7]], align 2
+// AARCH64-LE-NEXT:    [[TMP8:%.*]] = getelementptr i8, ptr [[TMP0]], i32 7
+// AARCH64-LE-NEXT:    store i8 0, ptr [[TMP8]], align 1
+// AARCH64-LE-NEXT:    ret void
+//
+// AARCH64-BE-LABEL: define dso_local void @testBitfieldMixedVolatileMembers(
+// AARCH64-BE-SAME: ptr noundef [[S:%.*]]) #[[ATTR0]] {
+// AARCH64-BE-NEXT:  [[ENTRY:.*:]]
+// AARCH64-BE-NEXT:    [[S_ADDR:%.*]] = alloca ptr, align 8
+// AARCH64-BE-NEXT:    store ptr [[S]], ptr [[S_ADDR]], align 8
+// AARCH64-BE-NEXT:    [[TMP0:%.*]] = load ptr, ptr [[S_ADDR]], align 8
+// AARCH64-BE-NEXT:    [[TMP1:%.*]] = getelementptr i8, ptr [[TMP0]], i32 1
+// AARCH64-BE-NEXT:    [[TMP2:%.*]] = load i8, ptr [[TMP1]], align 1
+// AARCH64-BE-NEXT:    [[TMP3:%.*]] = and i8 [[TMP2]], -32
+// AARCH64-BE-NEXT:    store i8 [[TMP3]], ptr [[TMP1]], align 1
+// AARCH64-BE-NEXT:    [[TMP4:%.*]] = getelementptr i8, ptr [[TMP0]], i32 3
+// AARCH64-BE-NEXT:    [[TMP5:%.*]] = load i8, ptr [[TMP4]], align 1
+// AARCH64-BE-NEXT:    [[TMP6:%.*]] = and i8 [[TMP5]], -8
+// AARCH64-BE-NEXT:    store i8 [[TMP6]], ptr [[TMP4]], align 1
+// AARCH64-BE-NEXT:    [[TMP7:%.*]] = getelementptr i8, ptr [[TMP0]], i32 6
+// AARCH64-BE-NEXT:    store i8 0, ptr [[TMP7]], align 2
+// AARCH64-BE-NEXT:    [[TMP8:%.*]] = getelementptr i8, ptr [[TMP0]], i32 7
+// AARCH64-BE-NEXT:    store i8 0, ptr [[TMP8]], align 1
+// AARCH64-BE-NEXT:    ret void
+//
+void testBitfieldMixedVolatileMembers(struct BitfieldMixedVolatileMembers *s) {
+  __builtin_clear_padding(s);
+}
+
+// LINUX-LABEL: define dso_local void @testBitfieldMixedVolatileMembersAndObject(
+// LINUX-SAME: ptr noundef [[S:%.*]]) #[[ATTR0]] {
+// LINUX-NEXT:  [[ENTRY:.*:]]
+// LINUX-NEXT:    [[S_ADDR:%.*]] = alloca ptr, align 8
+// LINUX-NEXT:    store ptr [[S]], ptr [[S_ADDR]], align 8
+// LINUX-NEXT:    [[TMP0:%.*]] = load ptr, ptr [[S_ADDR]], align 8
+// LINUX-NEXT:    [[TMP1:%.*]] = getelementptr i8, ptr [[TMP0]], i32 1
+// LINUX-NEXT:    [[TMP2:%.*]] = load i8, ptr [[TMP1]], align 1
+// LINUX-NEXT:    [[TMP3:%.*]] = and i8 [[TMP2]], 7
+// LINUX-NEXT:    store i8 [[TMP3]], ptr [[TMP1]], align 1
+// LINUX-NEXT:    [[TMP4:%.*]] = getelementptr i8, ptr [[TMP0]], i32 3
+// LINUX-NEXT:    [[TMP5:%.*]] = load i8, ptr [[TMP4]], align 1
+// LINUX-NEXT:    [[TMP6:%.*]] = and i8 [[TMP5]], 31
+// LINUX-NEXT:    store i8 [[TMP6]], ptr [[TMP4]], align 1
+// LINUX-NEXT:    [[TMP7:%.*]] = getelementptr i8, ptr [[TMP0]], i32 6
+// LINUX-NEXT:    store i8 0, ptr [[TMP7]], align 2
+// LINUX-NEXT:    [[TMP8:%.*]] = getelementptr i8, ptr [[TMP0]], i32 7
+// LINUX-NEXT:    store i8 0, ptr [[TMP8]], align 1
+// LINUX-NEXT:    ret void
+//
+// WINDOWS-LABEL: define dso_local void @testBitfieldMixedVolatileMembersAndObject(
+// WINDOWS-SAME: ptr noundef [[S:%.*]]) #[[ATTR0]] {
+// WINDOWS-NEXT:  [[ENTRY:.*:]]
+// WINDOWS-NEXT:    [[S_ADDR:%.*]] = alloca ptr, align 8
+// WINDOWS-NEXT:    store ptr [[S]], ptr [[S_ADDR]], align 8
+// WINDOWS-NEXT:    [[TMP0:%.*]] = load ptr, ptr [[S_ADDR]], align 8
+// WINDOWS-NEXT:    [[TMP1:%.*]] = getelementptr i8, ptr [[TMP0]], i32 1
+// WINDOWS-NEXT:    store i8 0, ptr [[TMP1]], align 1
+// WINDOWS-NEXT:    [[TMP2:%.*]] = getelementptr i8, ptr [[TMP0]], i32 2
+// WINDOWS-NEXT:    store i8 0, ptr [[TMP2]], align 2
+// WINDOWS-NEXT:    [[TMP3:%.*]] = getelementptr i8, ptr [[TMP0]], i32 3
+// WINDOWS-NEXT:    store i8 0, ptr [[TMP3]], align 1
+// WINDOWS-NEXT:    [[TMP4:%.*]] = getelementptr i8, ptr [[TMP0]], i32 4
+// WINDOWS-NEXT:    [[TMP5:%.*]] = load i8, ptr [[TMP4]], align 4
+// WINDOWS-NEXT:    [[TMP6:%.*]] = and i8 [[TMP5]], 7
+// WINDOWS-NEXT:    store i8 [[TMP6]], ptr [[TMP4]], align 4
+// WINDOWS-NEXT:    [[TMP7:%.*]] = getelementptr i8, ptr [[TMP0]], i32 5
+// WINDOWS-NEXT:    store i8 0, ptr [[TMP7]], align 1
+// WINDOWS-NEXT:    [[TMP8:%.*]] = getelementptr i8, ptr [[TMP0]], i32 6
+// WINDOWS-NEXT:    store i8 0, ptr [[TMP8]], align 2
+// WINDOWS-NEXT:    [[TMP9:%.*]] = getelementptr i8, ptr [[TMP0]], i32 7
+// WINDOWS-NEXT:    store i8 0, ptr [[TMP9]], align 1
+// WINDOWS-NEXT:    [[TMP10:%.*]] = getelementptr i8, ptr [[TMP0]], i32 9
+// WINDOWS-NEXT:    store i8 0, ptr [[TMP10]], align 1
+// WINDOWS-NEXT:    [[TMP11:%.*]] = getelementptr i8, ptr [[TMP0]], i32 10
+// WINDOWS-NEXT:    store i8 0, ptr [[TMP11]], align 2
+// WINDOWS-NEXT:    [[TMP12:%.*]] = getelementptr i8, ptr [[TMP0]], i32 11
+// WINDOWS-NEXT:    store i8 0, ptr [[TMP12]], align 1
+// WINDOWS-NEXT:    [[TMP13:%.*]] = getelementptr i8, ptr [[TMP0]], i32 12
+// WINDOWS-NEXT:    [[TMP14:%.*]] = load i8, ptr [[TMP13]], align 4
+// WINDOWS-NEXT:    [[TMP15:%.*]] = and i8 [[TMP14]], 31
+// WINDOWS-NEXT:    store i8 [[TMP15]], ptr [[TMP13]], align 4
+// WINDOWS-NEXT:    [[TMP16:%.*]] = getelementptr i8, ptr [[TMP0]], i32 13
+// WINDOWS-NEXT:    store i8 0, ptr [[TMP16]], align 1
+// WINDOWS-NEXT:    [[TMP17:%.*]] = getelementptr i8, ptr [[TMP0]], i32 14
+// WINDOWS-NEXT:    store i8 0, ptr [[TMP17]], align 2
+// WINDOWS-NEXT:    [[TMP18:%.*]] = getelementptr i8, ptr [[TMP0]], i32 15
+// WINDOWS-NEXT:    store i8 0, ptr [[TMP18]], align 1
+// WINDOWS-NEXT:    [[TMP19:%.*]] = getelementptr i8, ptr [[TMP0]], i32 18
+// WINDOWS-NEXT:    store i8 0, ptr [[TMP19]], align 2
+// WINDOWS-NEXT:    [[TMP20:%.*]] = getelementptr i8, ptr [[TMP0]], i32 19
+// WINDOWS-NEXT:    store i8 0, ptr [[TMP20]], align 1
+// WINDOWS-NEXT:    ret void
+//
+// ARM-LE-LABEL: define dso_local void @testBitfieldMixedVolatileMembersAndObject(
+// ARM-LE-SAME: ptr noundef [[S:%.*]]) #[[ATTR0]] {
+// ARM-LE-NEXT:  [[ENTRY:.*:]]
+// ARM-LE-NEXT:    [[S_ADDR:%.*]] = alloca ptr, align 4
+// ARM-LE-NEXT:    store ptr [[S]], ptr [[S_ADDR]], align 4
+// ARM-LE-NEXT:    [[TMP0:%.*]] = load ptr, ptr [[S_ADDR]], align 4
+// ARM-LE-NEXT:    [[TMP1:%.*]] = getelementptr i8, ptr [[TMP0]], i32 1
+// ARM-LE-NEXT:    [[TMP2:%.*]] = load i8, ptr [[TMP1]], align 1
+// ARM-LE-NEXT:    [[TMP3:%.*]] = and i8 [[TMP2]], 7
+// ARM-LE-NEXT:    store i8 [[TMP3]], ptr [[TMP1]], align 1
+// ARM-LE-NEXT:    [[TMP4:%.*]] = getelementptr i8, ptr [[TMP0]], i32 3
+// ARM-LE-NEXT:    [[TMP5:%.*]] = load i8, ptr [[TMP4]], align 1
+// ARM-LE-NEXT:    [[TMP6:%.*]] = and i8 [[TMP5]], 31
+// ARM-LE-NEXT:    store i8 [[TMP6]], ptr [[TMP4]], align 1
+// ARM-LE-NEXT:    [[TMP7:%.*]] = getelementptr i8, ptr [[TMP0]], i32 6
+// ARM-LE-NEXT:    store i8 0, ptr [[TMP7]], align 2
+// ARM-LE-NEXT:    [[TMP8:%.*]] = getelementptr i8, ptr [[TMP0]], i32 7
+// ARM-LE-NEXT:    store i8 0, ptr [[TMP8]], align 1
+// ARM-LE-NEXT:    ret void
+//
+// ARM-BE-LABEL: define dso_local void @testBitfieldMixedVolatileMembersAndObject(
+// ARM-BE-SAME: ptr noundef [[S:%.*]]) #[[ATTR0]] {
+// ARM-BE-NEXT:  [[ENTRY:.*:]]
+// ARM-BE-NEXT:    [[S_ADDR:%.*]] = alloca ptr, align 4
+// ARM-BE-NEXT:    store ptr [[S]], ptr [[S_ADDR]], align 4
+// ARM-BE-NEXT:    [[TMP0:%.*]] = load ptr, ptr [[S_ADDR]], align 4
+// ARM-BE-NEXT:    [[TMP1:%.*]] = getelementptr i8, ptr [[TMP0]], i32 1
+// ARM-BE-NEXT:    [[TMP2:%.*]] = load i8, ptr [[TMP1]], align 1
+// ARM-BE-NEXT:    [[TMP3:%.*]] = and i8 [[TMP2]], -32
+// ARM-BE-NEXT:    store i8 [[TMP3]], ptr [[TMP1]], align 1
+// ARM-BE-NEXT:    [[TMP4:%.*]] = getelementptr i8, ptr [[TMP0]], i32 3
+// ARM-BE-NEXT:    [[TMP5:%.*]] = load i8, ptr [[TMP4]], align 1
+// ARM-BE-NEXT:    [[TMP6:%.*]] = and i8 [[TMP5]], -8
+// ARM-BE-NEXT:    store i8 [[TMP6]], ptr [[TMP4]], align 1
+// ARM-BE-NEXT:    [[TMP7:%.*]] = getelementptr i8, ptr [[TMP0]], i32 6
+// ARM-BE-NEXT:    store i8 0, ptr [[TMP7]], align 2
+// ARM-BE-NEXT:    [[TMP8:%.*]] = getelementptr i8, ptr [[TMP0]], i32 7
+// ARM-BE-NEXT:    store i8 0, ptr [[TMP8]], align 1
+// ARM-BE-NEXT:    ret void
+//
+// AARCH64-LE-LABEL: define dso_local void @testBitfieldMixedVolatileMembersAndObject(
+// AARCH64-LE-SAME: ptr noundef [[S:%.*]]) #[[ATTR0]] {
+// AARCH64-LE-NEXT:  [[ENTRY:.*:]]
+// AARCH64-LE-NEXT:    [[S_ADDR:%.*]] = alloca ptr, align 8
+// AARCH64-LE-NEXT:    store ptr [[S]], ptr [[S_ADDR]], align 8
+// AARCH64-LE-NEXT:    [[TMP0:%.*]] = load ptr, ptr [[S_ADDR]], align 8
+// AARCH64-LE-NEXT:    [[TMP1:%.*]] = getelementptr i8, ptr [[TMP0]], i32 1
+// AARCH64-LE-NEXT:    [[TMP2:%.*]] = load i8, ptr [[TMP1]], align 1
+// AARCH64-LE-NEXT:    [[TMP3:%.*]] = and i8 [[TMP2]], 7
+// AARCH64-LE-NEXT:    store i8 [[TMP3]], ptr [[TMP1]], align 1
+// AARCH64-LE-NEXT:    [[TMP4:%.*]] = getelementptr i8, ptr [[TMP0]], i32 3
+// AARCH64-LE-NEXT:    [[TMP5:%.*]] = load i8, ptr [[TMP4]], align 1
+// AARCH64-LE-NEXT:    [[TMP6:%.*]] = and i8 [[TMP5]], 31
+// AARCH64-LE-NEXT:    store i8 [[TMP6]], ptr [[TMP4]], align 1
+// AARCH64-LE-NEXT:    [[TMP7:%.*]] = getelementptr i8, ptr [[TMP0]], i32 6
+// AARCH64-LE-NEXT:    store i8 0, ptr [[TMP7]], align 2
+// AARCH64-LE-NEXT:    [[TMP8:%.*]] = getelementptr i8, ptr [[TMP0]], i32 7
+// AARCH64-LE-NEXT:    store i8 0, ptr [[TMP8]], align 1
+// AARCH64-LE-NEXT:    ret void
+//
+// AARCH64-BE-LABEL: define dso_local void @testBitfieldMixedVolatileMembersAndObject(
+// AARCH64-BE-SAME: ptr noundef [[S:%.*]]) #[[ATTR0]] {
+// AARCH64-BE-NEXT:  [[ENTRY:.*:]]
+// AARCH64-BE-NEXT:    [[S_ADDR:%.*]] = alloca ptr, align 8
+// AARCH64-BE-NEXT:    store ptr [[S]], ptr [[S_ADDR]], align 8
+// AARCH64-BE-NEXT:    [[TMP0:%.*]] = load ptr, ptr [[S_ADDR]], align 8
+// AARCH64-BE-NEXT:    [[TMP1:%.*]] = getelementptr i8, ptr [[TMP0]], i32 1
+// AARCH64-BE-NEXT:    [[TMP2:%.*]] = load i8, ptr [[TMP1]], align 1
+// AARCH64-BE-NEXT:    [[TMP3:%.*]] = and i8 [[TMP2]], -32
+// AARCH64-BE-NEXT:    store i8 [[TMP3]], ptr [[TMP1]], align 1
+// AARCH64-BE-NEXT:    [[TMP4:%.*]] = getelementptr i8, ptr [[TMP0]], i32 3
+// AARCH64-BE-NEXT:    [[TMP5:%.*]] = load i8, ptr [[TMP4]], align 1
+// AARCH64-BE-NEXT:    [[TMP6:%.*]] = and i8 [[TMP5]], -8
+// AARCH64-BE-NEXT:    store i8 [[TMP6]], ptr [[TMP4]], align 1
+// AARCH64-BE-NEXT:    [[TMP7:%.*]] = getelementptr i8, ptr [[TMP0]], i32 6
+// AARCH64-BE-NEXT:    store i8 0, ptr [[TMP7]], align 2
+// AARCH64-BE-NEXT:    [[TMP8:%.*]] = getelementptr i8, ptr [[TMP0]], i32 7
+// AARCH64-BE-NEXT:    store i8 0, ptr [[TMP8]], align 1
+// AARCH64-BE-NEXT:    ret void
+//
+void testBitfieldMixedVolatileMembersAndObject(
+    volatile struct BitfieldMixedVolatileMembers *s) {
+  __builtin_clear_padding(s);
+}
+
+// Both array elements have partial-byte padding, and neither starts at offset
+// zero in the enclosing object. Preserve the prefix and suffix too.
+struct BitfieldNestedArray {
+  unsigned char prefix[3];
+  struct BitfieldPartialEnds elements[2];
+  unsigned char suffix;
+};
+
+// LINUX-LABEL: define dso_local void @testBitfieldNestedArray(
+// LINUX-SAME: ptr noundef [[S:%.*]]) #[[ATTR0]] {
+// LINUX-NEXT:  [[ENTRY:.*:]]
+// LINUX-NEXT:    [[S_ADDR:%.*]] = alloca ptr, align 8
+// LINUX-NEXT:    store ptr [[S]], ptr [[S_ADDR]], align 8
+// LINUX-NEXT:    [[TMP0:%.*]] = load ptr, ptr [[S_ADDR]], align 8
+// LINUX-NEXT:    [[TMP1:%.*]] = getelementptr i8, ptr [[TMP0]], i32 3
+// LINUX-NEXT:    store i8 0, ptr [[TMP1]], align 1
+// LINUX-NEXT:    [[TMP2:%.*]] = getelementptr i8, ptr [[TMP0]], i32 4
+// LINUX-NEXT:    [[TMP3:%.*]] = load i8, ptr [[TMP2]], align 4
+// LINUX-NEXT:    [[TMP4:%.*]] = and i8 [[TMP3]], -8
+// LINUX-NEXT:    store i8 [[TMP4]], ptr [[TMP2]], align 4
+// LINUX-NEXT:    [[TMP5:%.*]] = getelementptr i8, ptr [[TMP0]], i32 6
+// LINUX-NEXT:    [[TMP6:%.*]] = load i8, ptr [[TMP5]], align 2
+// LINUX-NEXT:    [[TMP7:%.*]] = and i8 [[TMP6]], 31
+// LINUX-NEXT:    store i8 [[TMP7]], ptr [[TMP5]], align 2
+// LINUX-NEXT:    [[TMP8:%.*]] = getelementptr i8, ptr [[TMP0]], i32 7
+// LINUX-NEXT:    store i8 0, ptr [[TMP8]], align 1
+// LINUX-NEXT:    [[TMP9:%.*]] = getelementptr i8, ptr [[TMP0]], i32 8
+// LINUX-NEXT:    [[TMP10:%.*]] = load i8, ptr [[TMP9]], align 4
+// LINUX-NEXT:    [[TMP11:%.*]] = and i8 [[TMP10]], -8
+// LINUX-NEXT:    store i8 [[TMP11]], ptr [[TMP9]], align 4
+// LINUX-NEXT:    [[TMP12:%.*]] = getelementptr i8, ptr [[TMP0]], i32 10
+// LINUX-NEXT:    [[TMP13:%.*]] = load i8, ptr [[TMP12]], align 2
+// LINUX-NEXT:    [[TMP14:%.*]] = and i8 [[TMP13]], 31
+// LINUX-NEXT:    store i8 [[TMP14]], ptr [[TMP12]], align 2
+// LINUX-NEXT:    [[TMP15:%.*]] = getelementptr i8, ptr [[TMP0]], i32 11
+// LINUX-NEXT:    store i8 0, ptr [[TMP15]], align 1
+// LINUX-NEXT:    [[TMP16:%.*]] = getelementptr i8, ptr [[TMP0]], i32 13
+// LINUX-NEXT:    store i8 0, ptr [[TMP16]], align 1
+// LINUX-NEXT:    [[TMP17:%.*]] = getelementptr i8, ptr [[TMP0]], i32 14
+// LINUX-NEXT:    store i8 0, ptr [[TMP17]], align 2
+// LINUX-NEXT:    [[TMP18:%.*]] = getelementptr i8, ptr [[TMP0]], i32 15
+// LINUX-NEXT:    store i8 0, ptr [[TMP18]], align 1
+// LINUX-NEXT:    ret void
+//
+// WINDOWS-LABEL: define dso_local void @testBitfieldNestedArray(
+// WINDOWS-SAME: ptr noundef [[S:%.*]]) #[[ATTR0]] {
+// WINDOWS-NEXT:  [[ENTRY:.*:]]
+// WINDOWS-NEXT:    [[S_ADDR:%.*]] = alloca ptr, align 8
+// WINDOWS-NEXT:    store ptr [[S]], ptr [[S_ADDR]], align 8
+// WINDOWS-NEXT:    [[TMP0:%.*]] = load ptr, ptr [[S_ADDR]], align 8
+// WINDOWS-NEXT:    [[TMP1:%.*]] = getelementptr i8, ptr [[TMP0]], i32 3
+// WINDOWS-NEXT:    store i8 0, ptr [[TMP1]], align 1
+// WINDOWS-NEXT:    [[TMP2:%.*]] = getelementptr i8, ptr [[TMP0]], i32 4
+// WINDOWS-NEXT:    [[TMP3:%.*]] = load i8, ptr [[TMP2]], align 4
+// WINDOWS-NEXT:    [[TMP4:%.*]] = and i8 [[TMP3]], -8
+// WINDOWS-NEXT:    store i8 [[TMP4]], ptr [[TMP2]], align 4
+// WINDOWS-NEXT:    [[TMP5:%.*]] = getelementptr i8, ptr [[TMP0]], i32 6
+// WINDOWS-NEXT:    [[TMP6:%.*]] = load i8, ptr [[TMP5]], align 2
+// WINDOWS-NEXT:    [[TMP7:%.*]] = and i8 [[TMP6]], 31
+// WINDOWS-NEXT:    store i8 [[TMP7]], ptr [[TMP5]], align 2
+// WINDOWS-NEXT:    [[TMP8:%.*]] = getelementptr i8, ptr [[TMP0]], i32 7
+// WINDOWS-NEXT:    store i8 0, ptr [[TMP8]], align 1
+// WINDOWS-NEXT:    [[TMP9:%.*]] = getelementptr i8, ptr [[TMP0]], i32 8
+// WINDOWS-NEXT:    [[TMP10:%.*]] = load i8, ptr [[TMP9]], align 4
+// WINDOWS-NEXT:    [[TMP11:%.*]] = and i8 [[TMP10]], -8
+// WINDOWS-NEXT:    store i8 [[TMP11]], ptr [[TMP9]], align 4
+// WINDOWS-NEXT:    [[TMP12:%.*]] = getelementptr i8, ptr [[TMP0]], i32 10
+// WINDOWS-NEXT:    [[TMP13:%.*]] = load i8, ptr [[TMP12]], align 2
+// WINDOWS-NEXT:    [[TMP14:%.*]] = and i8 [[TMP13]], 31
+// WINDOWS-NEXT:    store i8 [[TMP14]], ptr [[TMP12]], align 2
+// WINDOWS-NEXT:    [[TMP15:%.*]] = getelementptr i8, ptr [[TMP0]], i32 11
+// WINDOWS-NEXT:    store i8 0, ptr [[TMP15]], align 1
+// WINDOWS-NEXT:    [[TMP16:%.*]] = getelementptr i8, ptr [[TMP0]], i32 13
+// WINDOWS-NEXT:    store i8 0, ptr [[TMP16]], align 1
+// WINDOWS-NEXT:    [[TMP17:%.*]] = getelementptr i8, ptr [[TMP0]], i32 14
+// WINDOWS-NEXT:    store i8 0, ptr [[TMP17]], align 2
+// WINDOWS-NEXT:    [[TMP18:%.*]] = getelementptr i8, ptr [[TMP0]], i32 15
+// WINDOWS-NEXT:    store i8 0, ptr [[TMP18]], align 1
+// WINDOWS-NEXT:    ret void
+//
+// ARM-LE-LABEL: define dso_local void @testBitfieldNestedArray(
+// ARM-LE-SAME: ptr noundef [[S:%.*]]) #[[ATTR0]] {
+// ARM-LE-NEXT:  [[ENTRY:.*:]]
+// ARM-LE-NEXT:    [[S_ADDR:%.*]] = alloca ptr, align 4
+// ARM-LE-NEXT:    store ptr [[S]], ptr [[S_ADDR]], align 4
+// ARM-LE-NEXT:    [[TMP0:%.*]] = load ptr, ptr [[S_ADDR]], align 4
+// ARM-LE-NEXT:    [[TMP1:%.*]] = getelementptr i8, ptr [[TMP0]], i32 3
+// ARM-LE-NEXT:    store i8 0, ptr [[TMP1]], align 1
+// ARM-LE-NEXT:    [[TMP2:%.*]] = getelementptr i8, ptr [[TMP0]], i32 4
+// ARM-LE-NEXT:    [[TMP3:%.*]] = load i8, ptr [[TMP2]], align 4
+// ARM-LE-NEXT:    [[TMP4:%.*]] = and i8 [[TMP3]], -8
+// ARM-LE-NEXT:    store i8 [[TMP4]], ptr [[TMP2]], align 4
+// ARM-LE-NEXT:    [[TMP5:%.*]] = getelementptr i8, ptr [[TMP0]], i32 6
+// ARM-LE-NEXT:    [[TMP6:%.*]] = load i8, ptr [[TMP5]], align 2
+// ARM-LE-NEXT:    [[TMP7:%.*]] = and i8 [[TMP6]], 31
+// ARM-LE-NEXT:    store i8 [[TMP7]], ptr [[TMP5]], align 2
+// ARM-LE-NEXT:    [[TMP8:%.*]] = getelementptr i8, ptr [[TMP0]], i32 7
+// ARM-LE-NEXT:    store i8 0, ptr [[TMP8]], align 1
+// ARM-LE-NEXT:    [[TMP9:%.*]] = getelementptr i8, ptr [[TMP0]], i32 8
+// ARM-LE-NEXT:    [[TMP10:%.*]] = load i8, ptr [[TMP9]], align 4
+// ARM-LE-NEXT:    [[TMP11:%.*]] = and i8 [[TMP10]], -8
+// ARM-LE-NEXT:    store i8 [[TMP11]], ptr [[TMP9]], align 4
+// ARM-LE-NEXT:    [[TMP12:%.*]] = getelementptr i8, ptr [[TMP0]], i32 10
+// ARM-LE-NEXT:    [[TMP13:%.*]] = load i8, ptr [[TMP12]], align 2
+// ARM-LE-NEXT:    [[TMP14:%.*]] = and i8 [[TMP13]], 31
+// ARM-LE-NEXT:    store i8 [[TMP14]], ptr [[TMP12]], align 2
+// ARM-LE-NEXT:    [[TMP15:%.*]] = getelementptr i8, ptr [[TMP0]], i32 11
+// ARM-LE-NEXT:    store i8 0, ptr [[TMP15]], align 1
+// ARM-LE-NEXT:    [[TMP16:%.*]] = getelementptr i8, ptr [[TMP0]], i32 13
+// ARM-LE-NEXT:    store i8 0, ptr [[TMP16]], align 1
+// ARM-LE-NEXT:    [[TMP17:%.*]] = getelementptr i8, ptr [[TMP0]], i32 14
+// ARM-LE-NEXT:    store i8 0, ptr [[TMP17]], align 2
+// ARM-LE-NEXT:    [[TMP18:%.*]] = getelementptr i8, ptr [[TMP0]], i32 15
+// ARM-LE-NEXT:    store i8 0, ptr [[TMP18]], align 1
+// ARM-LE-NEXT:    ret void
+//
+// ARM-BE-LABEL: define dso_local void @testBitfieldNestedArray(
+// ARM-BE-SAME: ptr noundef [[S:%.*]]) #[[ATTR0]] {
+// ARM-BE-NEXT:  [[ENTRY:.*:]]
+// ARM-BE-NEXT:    [[S_ADDR:%.*]] = alloca ptr, align 4
+// ARM-BE-NEXT:    store ptr [[S]], ptr [[S_ADDR]], align 4
+// ARM-BE-NEXT:    [[TMP0:%.*]] = load ptr, ptr [[S_ADDR]], align 4
+// ARM-BE-NEXT:    [[TMP1:%.*]] = getelementptr i8, ptr [[TMP0]], i32 3
+// ARM-BE-NEXT:    store i8 0, ptr [[TMP1]], align 1
+// ARM-BE-NEXT:    [[TMP2:%.*]] = getelementptr i8, ptr [[TMP0]], i32 4
+// ARM-BE-NEXT:    [[TMP3:%.*]] = load i8, ptr [[TMP2]], align 4
+// ARM-BE-NEXT:    [[TMP4:%.*]] = and i8 [[TMP3]], 31
+// ARM-BE-NEXT:    store i8 [[TMP4]], ptr [[TMP2]], align 4
+// ARM-BE-NEXT:    [[TMP5:%.*]] = getelementptr i8, ptr [[TMP0]], i32 6
+// ARM-BE-NEXT:    [[TMP6:%.*]] = load i8, ptr [[TMP5]], align 2
+// ARM-BE-NEXT:    [[TMP7:%.*]] = and i8 [[TMP6]], -8
+// ARM-BE-NEXT:    store i8 [[TMP7]], ptr [[TMP5]], align 2
+// ARM-BE-NEXT:    [[TMP8:%.*]] = getelementptr i8, ptr [[TMP0]], i32 7
+// ARM-BE-NEXT:    store i8 0, ptr [[TMP8]], align 1
+// ARM-BE-NEXT:    [[TMP9:%.*]] = getelementptr i8, ptr [[TMP0]], i32 8
+// ARM-BE-NEXT:    [[TMP10:%.*]] = load i8, ptr [[TMP9]], align 4
+// ARM-BE-NEXT:    [[TMP11:%.*]] = and i8 [[TMP10]], 31
+// ARM-BE-NEXT:    store i8 [[TMP11]], ptr [[TMP9]], align 4
+// ARM-BE-NEXT:    [[TMP12:%.*]] = getelementptr i8, ptr [[TMP0]], i32 10
+// ARM-BE-NEXT:    [[TMP13:%.*]] = load i8, ptr [[TMP12]], align 2
+// ARM-BE-NEXT:    [[TMP14:%.*]] = and i8 [[TMP13]], -8
+// ARM-BE-NEXT:    store i8 [[TMP14]], ptr [[TMP12]], align 2
+// ARM-BE-NEXT:    [[TMP15:%.*]] = getelementptr i8, ptr [[TMP0]], i32 11
+// ARM-BE-NEXT:    store i8 0, ptr [[TMP15]], align 1
+// ARM-BE-NEXT:    [[TMP16:%.*]] = getelementptr i8, ptr [[TMP0]], i32 13
+// ARM-BE-NEXT:    store i8 0, ptr [[TMP16]], align 1
+// ARM-BE-NEXT:    [[TMP17:%.*]] = getelementptr i8, ptr [[TMP0]], i32 14
+// ARM-BE-NEXT:    store i8 0, ptr [[TMP17]], align 2
+// ARM-BE-NEXT:    [[TMP18:%.*]] = getelementptr i8, ptr [[TMP0]], i32 15
+// ARM-BE-NEXT:    store i8 0, ptr [[TMP18]], align 1
+// ARM-BE-NEXT:    ret void
+//
+// AARCH64-LE-LABEL: define dso_local void @testBitfieldNestedArray(
+// AARCH64-LE-SAME: ptr noundef [[S:%.*]]) #[[ATTR0]] {
+// AARCH64-LE-NEXT:  [[ENTRY:.*:]]
+// AARCH64-LE-NEXT:    [[S_ADDR:%.*]] = alloca ptr, align 8
+// AARCH64-LE-NEXT:    store ptr [[S]], ptr [[S_ADDR]], align 8
+// AARCH64-LE-NEXT:    [[TMP0:%.*]] = load ptr, ptr [[S_ADDR]], align 8
+// AARCH64-LE-NEXT:    [[TMP1:%.*]] = getelementptr i8, ptr [[TMP0]], i32 3
+// AARCH64-LE-NEXT:    store i8 0, ptr [[TMP1]], align 1
+// AARCH64-LE-NEXT:    [[TMP2:%.*]] = getelementptr i8, ptr [[TMP0]], i32 4
+// AARCH64-LE-NEXT:    [[TMP3:%.*]] = load i8, ptr [[TMP2]], align 4
+// AARCH64-LE-NEXT:    [[TMP4:%.*]] = and i8 [[TMP3]], -8
+// AARCH64-LE-NEXT:    store i8 [[TMP4]], ptr [[TMP2]], align 4
+// AARCH64-LE-NEXT:    [[TMP5:%.*]] = getelementptr i8, ptr [[TMP0]], i32 6
+// AARCH64-LE-NEXT:    [[TMP6:%.*]] = load i8, ptr [[TMP5]], align 2
+// AARCH64-LE-NEXT:    [[TMP7:%.*]] = and i8 [[TMP6]], 31
+// AARCH64-LE-NEXT:    store i8 [[TMP7]], ptr [[TMP5]], align 2
+// AARCH64-LE-NEXT:    [[TMP8:%.*]] = getelementptr i8, ptr [[TMP0]], i32 7
+// AARCH64-LE-NEXT:    store i8 0, ptr [[TMP8]], align 1
+// AARCH64-LE-NEXT:    [[TMP9:%.*]] = getelementptr i8, ptr [[TMP0]], i32 8
+// AARCH64-LE-NEXT:    [[TMP10:%.*]] = load i8, ptr [[TMP9]], align 4
+// AARCH64-LE-NEXT:    [[TMP11:%.*]] = and i8 [[TMP10]], -8
+// AARCH64-LE-NEXT:    store i8 [[TMP11]], ptr [[TMP9]], align 4
+// AARCH64-LE-NEXT:    [[TMP12:%.*]] = getelementptr i8, ptr [[TMP0]], i32 10
+// AARCH64-LE-NEXT:    [[TMP13:%.*]] = load i8, ptr [[TMP12]], align 2
+// AARCH64-LE-NEXT:    [[TMP14:%.*]] = and i8 [[TMP13]], 31
+// AARCH64-LE-NEXT:    store i8 [[TMP14]], ptr [[TMP12]], align 2
+// AARCH64-LE-NEXT:    [[TMP15:%.*]] = getelementptr i8, ptr [[TMP0]], i32 11
+// AARCH64-LE-NEXT:    store i8 0, ptr [[TMP15]], align 1
+// AARCH64-LE-NEXT:    [[TMP16:%.*]] = getelementptr i8, ptr [[TMP0]], i32 13
+// AARCH64-LE-NEXT:    store i8 0, ptr [[TMP16]], align 1
+// AARCH64-LE-NEXT:    [[TMP17:%.*]] = getelementptr i8, ptr [[TMP0]], i32 14
+// AARCH64-LE-NEXT:    store i8 0, ptr [[TMP17]], align 2
+// AARCH64-LE-NEXT:    [[TMP18:%.*]] = getelementptr i8, ptr [[TMP0]], i32 15
+// AARCH64-LE-NEXT:    store i8 0, ptr [[TMP18]], align 1
+// AARCH64-LE-NEXT:    ret void
+//
+// AARCH64-BE-LABEL: define dso_local void @testBitfieldNestedArray(
+// AARCH64-BE-SAME: ptr noundef [[S:%.*]]) #[[ATTR0]] {
+// AARCH64-BE-NEXT:  [[ENTRY:.*:]]
+// AARCH64-BE-NEXT:    [[S_ADDR:%.*]] = alloca ptr, align 8
+// AARCH64-BE-NEXT:    store ptr [[S]], ptr [[S_ADDR]], align 8
+// AARCH64-BE-NEXT:    [[TMP0:%.*]] = load ptr, ptr [[S_ADDR]], align 8
+// AARCH64-BE-NEXT:    [[TMP1:%.*]] = getelementptr i8, ptr [[TMP0]], i32 3
+// AARCH64-BE-NEXT:    store i8 0, ptr [[TMP1]], align 1
+// AARCH64-BE-NEXT:    [[TMP2:%.*]] = getelementptr i8, ptr [[TMP0]], i32 4
+// AARCH64-BE-NEXT:    [[TMP3:%.*]] = load i8, ptr [[TMP2]], align 4
+// AARCH64-BE-NEXT:    [[TMP4:%.*]] = and i8 [[TMP3]], 31
+// AARCH64-BE-NEXT:    store i8 [[TMP4]], ptr [[TMP2]], align 4
+// AARCH64-BE-NEXT:    [[TMP5:%.*]] = getelementptr i8, ptr [[TMP0]], i32 6
+// AARCH64-BE-NEXT:    [[TMP6:%.*]] = load i8, ptr [[TMP5]], align 2
+// AARCH64-BE-NEXT:    [[TMP7:%.*]] = and i8 [[TMP6]], -8
+// AARCH64-BE-NEXT:    store i8 [[TMP7]], ptr [[TMP5]], align 2
+// AARCH64-BE-NEXT:    [[TMP8:%.*]] = getelementptr i8, ptr [[TMP0]], i32 7
+// AARCH64-BE-NEXT:    store i8 0, ptr [[TMP8]], align 1
+// AARCH64-BE-NEXT:    [[TMP9:%.*]] = getelementptr i8, ptr [[TMP0]], i32 8
+// AARCH64-BE-NEXT:    [[TMP10:%.*]] = load i8, ptr [[TMP9]], align 4
+// AARCH64-BE-NEXT:    [[TMP11:%.*]] = and i8 [[TMP10]], 31
+// AARCH64-BE-NEXT:    store i8 [[TMP11]], ptr [[TMP9]], align 4
+// AARCH64-BE-NEXT:    [[TMP12:%.*]] = getelementptr i8, ptr [[TMP0]], i32 10
+// AARCH64-BE-NEXT:    [[TMP13:%.*]] = load i8, ptr [[TMP12]], align 2
+// AARCH64-BE-NEXT:    [[TMP14:%.*]] = and i8 [[TMP13]], -8
+// AARCH64-BE-NEXT:    store i8 [[TMP14]], ptr [[TMP12]], align 2
+// AARCH64-BE-NEXT:    [[TMP15:%.*]] = getelementptr i8, ptr [[TMP0]], i32 11
+// AARCH64-BE-NEXT:    store i8 0, ptr [[TMP15]], align 1
+// AARCH64-BE-NEXT:    [[TMP16:%.*]] = getelementptr i8, ptr [[TMP0]], i32 13
+// AARCH64-BE-NEXT:    store i8 0, ptr [[TMP16]], align 1
+// AARCH64-BE-NEXT:    [[TMP17:%.*]] = getelementptr i8, ptr [[TMP0]], i32 14
+// AARCH64-BE-NEXT:    store i8 0, ptr [[TMP17]], align 2
+// AARCH64-BE-NEXT:    [[TMP18:%.*]] = getelementptr i8, ptr [[TMP0]], i32 15
+// AARCH64-BE-NEXT:    store i8 0, ptr [[TMP18]], align 1
+// AARCH64-BE-NEXT:    ret void
+//
+void testBitfieldNestedArray(struct BitfieldNestedArray *s) {
+  __builtin_clear_padding(s);
+}
+
+// The raw member occupies the first byte, including padding in the fields
+// member. Only padding shared by both union alternatives may be cleared.
+union BitfieldUnionOverlay {
+  struct BitfieldPartialEnds fields;
+  unsigned char raw;
+};
+
+// LINUX-LABEL: define dso_local void @testBitfieldUnionOverlay(
+// LINUX-SAME: ptr noundef [[S:%.*]]) #[[ATTR0]] {
+// LINUX-NEXT:  [[ENTRY:.*:]]
+// LINUX-NEXT:    [[S_ADDR:%.*]] = alloca ptr, align 8
+// LINUX-NEXT:    store ptr [[S]], ptr [[S_ADDR]], align 8
+// LINUX-NEXT:    [[TMP0:%.*]] = load ptr, ptr [[S_ADDR]], align 8
+// LINUX-NEXT:    [[TMP1:%.*]] = getelementptr i8, ptr [[TMP0]], i32 2
+// LINUX-NEXT:    [[TMP2:%.*]] = load i8, ptr [[TMP1]], align 2
+// LINUX-NEXT:    [[TMP3:%.*]] = and i8 [[TMP2]], 31
+// LINUX-NEXT:    store i8 [[TMP3]], ptr [[TMP1]], align 2
+// LINUX-NEXT:    [[TMP4:%.*]] = getelementptr i8, ptr [[TMP0]], i32 3
+// LINUX-NEXT:    store i8 0, ptr [[TMP4]], align 1
+// LINUX-NEXT:    ret void
+//
+// WINDOWS-LABEL: define dso_local void @testBitfieldUnionOverlay(
+// WINDOWS-SAME: ptr noundef [[S:%.*]]) #[[ATTR0]] {
+// WINDOWS-NEXT:  [[ENTRY:.*:]]
+// WINDOWS-NEXT:    [[S_ADDR:%.*]] = alloca ptr, align 8
+// WINDOWS-NEXT:    store ptr [[S]], ptr [[S_ADDR]], align 8
+// WINDOWS-NEXT:    [[TMP0:%.*]] = load ptr, ptr [[S_ADDR]], align 8
+// WINDOWS-NEXT:    [[TMP1:%.*]] = getelementptr i8, ptr [[TMP0]], i32 2
+// WINDOWS-NEXT:    [[TMP2:%.*]] = load i8, ptr [[TMP1]], align 2
+// WINDOWS-NEXT:    [[TMP3:%.*]] = and i8 [[TMP2]], 31
+// WINDOWS-NEXT:    store i8 [[TMP3]], ptr [[TMP1]], align 2
+// WINDOWS-NEXT:    [[TMP4:%.*]] = getelementptr i8, ptr [[TMP0]], i32 3
+// WINDOWS-NEXT:    store i8 0, ptr [[TMP4]], align 1
+// WINDOWS-NEXT:    ret void
+//
+// ARM-LE-LABEL: define dso_local void @testBitfieldUnionOverlay(
+// ARM-LE-SAME: ptr noundef [[S:%.*]]) #[[ATTR0]] {
+// ARM-LE-NEXT:  [[ENTRY:.*:]]
+// ARM-LE-NEXT:    [[S_ADDR:%.*]] = alloca ptr, align 4
+// ARM-LE-NEXT:    store ptr [[S]], ptr [[S_ADDR]], align 4
+// ARM-LE-NEXT:    [[TMP0:%.*]] = load ptr, ptr [[S_ADDR]], align 4
+// ARM-LE-NEXT:    [[TMP1:%.*]] = getelementptr i8, ptr [[TMP0]], i32 2
+// ARM-LE-NEXT:    [[TMP2:%.*]] = load i8, ptr [[TMP1]], align 2
+// ARM-LE-NEXT:    [[TMP3:%.*]] = and i8 [[TMP2]], 31
+// ARM-LE-NEXT:    store i8 [[TMP3]], ptr [[TMP1]], align 2
+// ARM-LE-NEXT:    [[TMP4:%.*]] = getelementptr i8, ptr [[TMP0]], i32 3
+// ARM-LE-NEXT:    store i8 0, ptr [[TMP4]], align 1
+// ARM-LE-NEXT:    ret void
+//
+// ARM-BE-LABEL: define dso_local void @testBitfieldUnionOverlay(
+// ARM-BE-SAME: ptr noundef [[S:%.*]]) #[[ATTR0]] {
+// ARM-BE-NEXT:  [[ENTRY:.*:]]
+// ARM-BE-NEXT:    [[S_ADDR:%.*]] = alloca ptr, align 4
+// ARM-BE-NEXT:    store ptr [[S]], ptr [[S_ADDR]], align 4
+// ARM-BE-NEXT:    [[TMP0:%.*]] = load ptr, ptr [[S_ADDR]], align 4
+// ARM-BE-NEXT:    [[TMP1:%.*]] = getelementptr i8, ptr [[TMP0]], i32 2
+// ARM-BE-NEXT:    [[TMP2:%.*]] = load i8, ptr [[TMP1]], align 2
+// ARM-BE-NEXT:    [[TMP3:%.*]] = and i8 [[TMP2]], -8
+// ARM-BE-NEXT:    store i8 [[TMP3]], ptr [[TMP1]], align 2
+// ARM-BE-NEXT:    [[TMP4:%.*]] = getelementptr i8, ptr [[TMP0]], i32 3
+// ARM-BE-NEXT:    store i8 0, ptr [[TMP4]], align 1
+// ARM-BE-NEXT:    ret void
+//
+// AARCH64-LE-LABEL: define dso_local void @testBitfieldUnionOverlay(
+// AARCH64-LE-SAME: ptr noundef [[S:%.*]]) #[[ATTR0]] {
+// AARCH64-LE-NEXT:  [[ENTRY:.*:]]
+// AARCH64-LE-NEXT:    [[S_ADDR:%.*]] = alloca ptr, align 8
+// AARCH64-LE-NEXT:    store ptr [[S]], ptr [[S_ADDR]], align 8
+// AARCH64-LE-NEXT:    [[TMP0:%.*]] = load ptr, ptr [[S_ADDR]], align 8
+// AARCH64-LE-NEXT:    [[TMP1:%.*]] = getelementptr i8, ptr [[TMP0]], i32 2
+// AARCH64-LE-NEXT:    [[TMP2:%.*]] = load i8, ptr [[TMP1]], align 2
+// AARCH64-LE-NEXT:    [[TMP3:%.*]] = and i8 [[TMP2]], 31
+// AARCH64-LE-NEXT:    store i8 [[TMP3]], ptr [[TMP1]], align 2
+// AARCH64-LE-NEXT:    [[TMP4:%.*]] = getelementptr i8, ptr [[TMP0]], i32 3
+// AARCH64-LE-NEXT:    store i8 0, ptr [[TMP4]], align 1
+// AARCH64-LE-NEXT:    ret void
+//
+// AARCH64-BE-LABEL: define dso_local void @testBitfieldUnionOverlay(
+// AARCH64-BE-SAME: ptr noundef [[S:%.*]]) #[[ATTR0]] {
+// AARCH64-BE-NEXT:  [[ENTRY:.*:]]
+// AARCH64-BE-NEXT:    [[S_ADDR:%.*]] = alloca ptr, align 8
+// AARCH64-BE-NEXT:    store ptr [[S]], ptr [[S_ADDR]], align 8
+// AARCH64-BE-NEXT:    [[TMP0:%.*]] = load ptr, ptr [[S_ADDR]], align 8
+// AARCH64-BE-NEXT:    [[TMP1:%.*]] = getelementptr i8, ptr [[TMP0]], i32 2
+// AARCH64-BE-NEXT:    [[TMP2:%.*]] = load i8, ptr [[TMP1]], align 2
+// AARCH64-BE-NEXT:    [[TMP3:%.*]] = and i8 [[TMP2]], -8
+// AARCH64-BE-NEXT:    store i8 [[TMP3]], ptr [[TMP1]], align 2
+// AARCH64-BE-NEXT:    [[TMP4:%.*]] = getelementptr i8, ptr [[TMP0]], i32 3
+// AARCH64-BE-NEXT:    store i8 0, ptr [[TMP4]], align 1
+// AARCH64-BE-NEXT:    ret void
+//
+void testBitfieldUnionOverlay(union BitfieldUnionOverlay *s) {
+  __builtin_clear_padding(s);
+}
+
+enum BitfieldState { StateZero, StateOne, StateTwo, StateThree };
+struct BitfieldBoolEnum {
+  _Bool flag : 1;
+  enum BitfieldState state : 2;
+  unsigned char payload : 3;
+};
+
+// LINUX-LABEL: define dso_local void @testBitfieldBoolEnum(
+// LINUX-SAME: ptr noundef [[S:%.*]]) #[[ATTR0]] {
+// LINUX-NEXT:  [[ENTRY:.*:]]
+// LINUX-NEXT:    [[S_ADDR:%.*]] = alloca ptr, align 8
+// LINUX-NEXT:    store ptr [[S]], ptr [[S_ADDR]], align 8
+// LINUX-NEXT:    [[TMP0:%.*]] = load ptr, ptr [[S_ADDR]], align 8
+// LINUX-NEXT:    [[TMP1:%.*]] = getelementptr i8, ptr [[TMP0]], i32 0
+// LINUX-NEXT:    [[TMP2:%.*]] = load i8, ptr [[TMP1]], align 4
+// LINUX-NEXT:    [[TMP3:%.*]] = and i8 [[TMP2]], 63
+// LINUX-NEXT:    store i8 [[TMP3]], ptr [[TMP1]], align 4
+// LINUX-NEXT:    [[TMP4:%.*]] = getelementptr i8, ptr [[TMP0]], i32 1
+// LINUX-NEXT:    store i8 0, ptr [[TMP4]], align 1
+// LINUX-NEXT:    [[TMP5:%.*]] = getelementptr i8, ptr [[TMP0]], i32 2
+// LINUX-NEXT:    store i8 0, ptr [[TMP5]], align 2
+// LINUX-NEXT:    [[TMP6:%.*]] = getelementptr i8, ptr [[TMP0]], i32 3
+// LINUX-NEXT:    store i8 0, ptr [[TMP6]], align 1
+// LINUX-NEXT:    ret void
+//
+// WINDOWS-LABEL: define dso_local void @testBitfieldBoolEnum(
+// WINDOWS-SAME: ptr noundef [[S:%.*]]) #[[ATTR0]] {
+// WINDOWS-NEXT:  [[ENTRY:.*:]]
+// WINDOWS-NEXT:    [[S_ADDR:%.*]] = alloca ptr, align 8
+// WINDOWS-NEXT:    store ptr [[S]], ptr [[S_ADDR]], align 8
+// WINDOWS-NEXT:    [[TMP0:%.*]] = load ptr, ptr [[S_ADDR]], align 8
+// WINDOWS-NEXT:    [[TMP1:%.*]] = getelementptr i8, ptr [[TMP0]], i32 0
+// WINDOWS-NEXT:    [[TMP2:%.*]] = load i8, ptr [[TMP1]], align 4
+// WINDOWS-NEXT:    [[TMP3:%.*]] = and i8 [[TMP2]], 1
+// WINDOWS-NEXT:    store i8 [[TMP3]], ptr [[TMP1]], align 4
+// WINDOWS-NEXT:    [[TMP4:%.*]] = getelementptr i8, ptr [[TMP0]], i32 1
+// WINDOWS-NEXT:    store i8 0, ptr [[TMP4]], align 1
+// WINDOWS-NEXT:    [[TMP5:%.*]] = getelementptr i8, ptr [[TMP0]], i32 2
+// WINDOWS-NEXT:    store i8 0, ptr [[TMP5]], align 2
+// WINDOWS-NEXT:    [[TMP6:%.*]] = getelementptr i8, ptr [[TMP0]], i32 3
+// WINDOWS-NEXT:    store i8 0, ptr [[TMP6]], align 1
+// WINDOWS-NEXT:    [[TMP7:%.*]] = getelementptr i8, ptr [[TMP0]], i32 4
+// WINDOWS-NEXT:    [[TMP8:%.*]] = load i8, ptr [[TMP7]], align 4
+// WINDOWS-NEXT:    [[TMP9:%.*]] = and i8 [[TMP8]], 3
+// WINDOWS-NEXT:    store i8 [[TMP9]], ptr [[TMP7]], align 4
+// WINDOWS-NEXT:    [[TMP10:%.*]] = getelementptr i8, ptr [[TMP0]], i32 5
+// WINDOWS-NEXT:    store i8 0, ptr [[TMP10]], align 1
+// WINDOWS-NEXT:    [[TMP11:%.*]] = getelementptr i8, ptr [[TMP0]], i32 6
+// WINDOWS-NEXT:    store i8 0, ptr [[TMP11]], align 2
+// WINDOWS-NEXT:    [[TMP12:%.*]] = getelementptr i8, ptr [[TMP0]], i32 7
+// WINDOWS-NEXT:    store i8 0, ptr [[TMP12]], align 1
+// WINDOWS-NEXT:    [[TMP13:%.*]] = getelementptr i8, ptr [[TMP0]], i32 8
+// WINDOWS-NEXT:    [[TMP14:%.*]] = load i8, ptr [[TMP13]], align 4
+// WINDOWS-NEXT:    [[TMP15:%.*]] = and i8 [[TMP14]], 7
+// WINDOWS-NEXT:    store i8 [[TMP15]], ptr [[TMP13]], align 4
+// WINDOWS-NEXT:    [[TMP16:%.*]] = getelementptr i8, ptr [[TMP0]], i32 9
+// WINDOWS-NEXT:    store i8 0, ptr [[TMP16]], align 1
+// WINDOWS-NEXT:    [[TMP17:%.*]] = getelementptr i8, ptr [[TMP0]], i32 10
+// WINDOWS-NEXT:    store i8 0, ptr [[TMP17]], align 2
+// WINDOWS-NEXT:    [[TMP18:%.*]] = getelementptr i8, ptr [[TMP0]], i32 11
+// WINDOWS-NEXT:    store i8 0, ptr [[TMP18]], align 1
+// WINDOWS-NEXT:    ret void
+//
+// ARM-LE-LABEL: define dso_local void @testBitfieldBoolEnum(
+// ARM-LE-SAME: ptr noundef [[S:%.*]]) #[[ATTR0]] {
+// ARM-LE-NEXT:  [[ENTRY:.*:]]
+// ARM-LE-NEXT:    [[S_ADDR:%.*]] = alloca ptr, align 4
+// ARM-LE-NEXT:    store ptr [[S]], ptr [[S_ADDR]], align 4
+// ARM-LE-NEXT:    [[TMP0:%.*]] = load ptr, ptr [[S_ADDR]], align 4
+// ARM-LE-NEXT:    [[TMP1:%.*]] = getelementptr i8, ptr [[TMP0]], i32 0
+// ARM-LE-NEXT:    [[TMP2:%.*]] = load i8, ptr [[TMP1]], align 4
+// ARM-LE-NEXT:    [[TMP3:%.*]] = and i8 [[TMP2]], 63
+// ARM-LE-NEXT:    store i8 [[TMP3]], ptr [[TMP1]], align 4
+// ARM-LE-NEXT:    [[TMP4:%.*]] = getelementptr i8, ptr [[TMP0]], i32 1
+// ARM-LE-NEXT:    store i8 0, ptr [[TMP4]], align 1
+// ARM-LE-NEXT:    [[TMP5:%.*]] = getelementptr i8, ptr [[TMP0]], i32 2
+// ARM-LE-NEXT:    store i8 0, ptr [[TMP5]], align 2
+// ARM-LE-NEXT:    [[TMP6:%.*]] = getelementptr i8, ptr [[TMP0]], i32 3
+// ARM-LE-NEXT:    store i8 0, ptr [[TMP6]], align 1
+// ARM-LE-NEXT:    ret void
+//
+// ARM-BE-LABEL: define dso_local void @testBitfieldBoolEnum(
+// ARM-BE-SAME: ptr noundef [[S:%.*]]) #[[ATTR0]] {
+// ARM-BE-NEXT:  [[ENTRY:.*:]]
+// ARM-BE-NEXT:    [[S_ADDR:%.*]] = alloca ptr, align 4
+// ARM-BE-NEXT:    store ptr [[S]], ptr [[S_ADDR]], align 4
+// ARM-BE-NEXT:    [[TMP0:%.*]] = load ptr, ptr [[S_ADDR]], align 4
+// ARM-BE-NEXT:    [[TMP1:%.*]] = getelementptr i8, ptr [[TMP0]], i32 0
+// ARM-BE-NEXT:    [[TMP2:%.*]] = load i8, ptr [[TMP1]], align 4
+// ARM-BE-NEXT:    [[TMP3:%.*]] = and i8 [[TMP2]], -4
+// ARM-BE-NEXT:    store i8 [[TMP3]], ptr [[TMP1]], align 4
+// ARM-BE-NEXT:    [[TMP4:%.*]] = getelementptr i8, ptr [[TMP0]], i32 1
+// ARM-BE-NEXT:    store i8 0, ptr [[TMP4]], align 1
+// ARM-BE-NEXT:    [[TMP5:%.*]] = getelementptr i8, ptr [[TMP0]], i32 2
+// ARM-BE-NEXT:    store i8 0, ptr [[TMP5]], align 2
+// ARM-BE-NEXT:    [[TMP6:%.*]] = getelementptr i8, ptr [[TMP0]], i32 3
+// ARM-BE-NEXT:    store i8 0, ptr [[TMP6]], align 1
+// ARM-BE-NEXT:    ret void
+//
+// AARCH64-LE-LABEL: define dso_local void @testBitfieldBoolEnum(
+// AARCH64-LE-SAME: ptr noundef [[S:%.*]]) #[[ATTR0]] {
+// AARCH64-LE-NEXT:  [[ENTRY:.*:]]
+// AARCH64-LE-NEXT:    [[S_ADDR:%.*]] = alloca ptr, align 8
+// AARCH64-LE-NEXT:    store ptr [[S]], ptr [[S_ADDR]], align 8
+// AARCH64-LE-NEXT:    [[TMP0:%.*]] = load ptr, ptr [[S_ADDR]], align 8
+// AARCH64-LE-NEXT:    [[TMP1:%.*]] = getelementptr i8, ptr [[TMP0]], i32 0
+// AARCH64-LE-NEXT:    [[TMP2:%.*]] = load i8, ptr [[TMP1]], align 4
+// AARCH64-LE-NEXT:    [[TMP3:%.*]] = and i8 [[TMP2]], 63
+// AARCH64-LE-NEXT:    store i8 [[TMP3]], ptr [[TMP1]], align 4
+// AARCH64-LE-NEXT:    [[TMP4:%.*]] = getelementptr i8, ptr [[TMP0]], i32 1
+// AARCH64-LE-NEXT:    store i8 0, ptr [[TMP4]], align 1
+// AARCH64-LE-NEXT:    [[TMP5:%.*]] = getelementptr i8, ptr [[TMP0]], i32 2
+// AARCH64-LE-NEXT:    store i8 0, ptr [[TMP5]], align 2
+// AARCH64-LE-NEXT:    [[TMP6:%.*]] = getelementptr i8, ptr [[TMP0]], i32 3
+// AARCH64-LE-NEXT:    store i8 0, ptr [[TMP6]], align 1
+// AARCH64-LE-NEXT:    ret void
+//
+// AARCH64-BE-LABEL: define dso_local void @testBitfieldBoolEnum(
+// AARCH64-BE-SAME: ptr noundef [[S:%.*]]) #[[ATTR0]] {
+// AARCH64-BE-NEXT:  [[ENTRY:.*:]]
+// AARCH64-BE-NEXT:    [[S_ADDR:%.*]] = alloca ptr, align 8
+// AARCH64-BE-NEXT:    store ptr [[S]], ptr [[S_ADDR]], align 8
+// AARCH64-BE-NEXT:    [[TMP0:%.*]] = load ptr, ptr [[S_ADDR]], align 8
+// AARCH64-BE-NEXT:    [[TMP1:%.*]] = getelementptr i8, ptr [[TMP0]], i32 0
+// AARCH64-BE-NEXT:    [[TMP2:%.*]] = load i8, ptr [[TMP1]], align 4
+// AARCH64-BE-NEXT:    [[TMP3:%.*]] = and i8 [[TMP2]], -4
+// AARCH64-BE-NEXT:    store i8 [[TMP3]], ptr [[TMP1]], align 4
+// AARCH64-BE-NEXT:    [[TMP4:%.*]] = getelementptr i8, ptr [[TMP0]], i32 1
+// AARCH64-BE-NEXT:    store i8 0, ptr [[TMP4]], align 1
+// AARCH64-BE-NEXT:    [[TMP5:%.*]] = getelementptr i8, ptr [[TMP0]], i32 2
+// AARCH64-BE-NEXT:    store i8 0, ptr [[TMP5]], align 2
+// AARCH64-BE-NEXT:    [[TMP6:%.*]] = getelementptr i8, ptr [[TMP0]], i32 3
+// AARCH64-BE-NEXT:    store i8 0, ptr [[TMP6]], align 1
+// AARCH64-BE-NEXT:    ret void
+//
+void testBitfieldBoolEnum(struct BitfieldBoolEnum *s) {
+  __builtin_clear_padding(s);
+}
+
+// Exercise bit-precise declared types and an occupied width greater than 32.
+struct BitfieldBitIntMembers {
+  unsigned _BitInt(9) small : 5;
+  signed _BitInt(17) medium : 13;
+  unsigned _BitInt(65) wide : 61;
+};
+
+// LINUX-LABEL: define dso_local void @testBitfieldBitIntMembers(
+// LINUX-SAME: ptr noundef [[S:%.*]]) #[[ATTR0]] {
+// LINUX-NEXT:  [[ENTRY:.*:]]
+// LINUX-NEXT:    [[S_ADDR:%.*]] = alloca ptr, align 8
+// LINUX-NEXT:    store ptr [[S]], ptr [[S_ADDR]], align 8
+// LINUX-NEXT:    [[TMP0:%.*]] = load ptr, ptr [[S_ADDR]], align 8
+// LINUX-NEXT:    [[TMP1:%.*]] = getelementptr i8, ptr [[TMP0]], i32 9
+// LINUX-NEXT:    [[TMP2:%.*]] = load i8, ptr [[TMP1]], align 1
+// LINUX-NEXT:    [[TMP3:%.*]] = and i8 [[TMP2]], 127
+// LINUX-NEXT:    store i8 [[TMP3]], ptr [[TMP1]], align 1
+// LINUX-NEXT:    [[TMP4:%.*]] = getelementptr i8, ptr [[TMP0]], i32 10
+// LINUX-NEXT:    store i8 0, ptr [[TMP4]], align 2
+// LINUX-NEXT:    [[TMP5:%.*]] = getelementptr i8, ptr [[TMP0]], i32 11
+// LINUX-NEXT:    store i8 0, ptr [[TMP5]], align 1
+// LINUX-NEXT:    [[TMP6:%.*]] = getelementptr i8, ptr [[TMP0]], i32 12
+// LINUX-NEXT:    store i8 0, ptr [[TMP6]], align 4
+// LINUX-NEXT:    [[TMP7:%.*]] = getelementptr i8, ptr [[TMP0]], i32 13
+// LINUX-NEXT:    store i8 0, ptr [[TMP7]], align 1
+// LINUX-NEXT:    [[TMP8:%.*]] = getelementptr i8, ptr [[TMP0]], i32 14
+// LINUX-NEXT:    store i8 0, ptr [[TMP8]], align 2
+// LINUX-NEXT:    [[TMP9:%.*]] = getelementptr i8, ptr [[TMP0]], i32 15
+// LINUX-NEXT:    store i8 0, ptr [[TMP9]], align 1
+// LINUX-NEXT:    ret void
+//
+// WINDOWS-LABEL: define dso_local void @testBitfieldBitIntMembers(
+// WINDOWS-SAME: ptr noundef [[S:%.*]]) #[[ATTR0]] {
+// WINDOWS-NEXT:  [[ENTRY:.*:]]
+// WINDOWS-NEXT:    [[S_ADDR:%.*]] = alloca ptr, align 8
+// WINDOWS-NEXT:    store ptr [[S]], ptr [[S_ADDR]], align 8
+// WINDOWS-NEXT:    [[TMP0:%.*]] = load ptr, ptr [[S_ADDR]], align 8
+// WINDOWS-NEXT:    [[TMP1:%.*]] = getelementptr i8, ptr [[TMP0]], i32 0
+// WINDOWS-NEXT:    [[TMP2:%.*]] = load i8, ptr [[TMP1]], align 8
+// WINDOWS-NEXT:    [[TMP3:%.*]] = and i8 [[TMP2]], 31
+// WINDOWS-NEXT:    store i8 [[TMP3]], ptr [[TMP1]], align 8
+// WINDOWS-NEXT:    [[TMP4:%.*]] = getelementptr i8, ptr [[TMP0]], i32 1
+// WINDOWS-NEXT:    store i8 0, ptr [[TMP4]], align 1
+// WINDOWS-NEXT:    [[TMP5:%.*]] = getelementptr i8, ptr [[TMP0]], i32 2
+// WINDOWS-NEXT:    store i8 0, ptr [[TMP5]], align 2
+// WINDOWS-NEXT:    [[TMP6:%.*]] = getelementptr i8, ptr [[TMP0]], i32 3
+// WINDOWS-NEXT:    store i8 0, ptr [[TMP6]], align 1
+// WINDOWS-NEXT:    [[TMP7:%.*]] = getelementptr i8, ptr [[TMP0]], i32 5
+// WINDOWS-NEXT:    [[TMP8:%.*]] = load i8, ptr [[TMP7]], align 1
+// WINDOWS-NEXT:    [[TMP9:%.*]] = and i8 [[TMP8]], 31
+// WINDOWS-NEXT:    store i8 [[TMP9]], ptr [[TMP7]], align 1
+// WINDOWS-NEXT:    [[TMP10:%.*]] = getelementptr i8, ptr [[TMP0]], i32 6
+// WINDOWS-NEXT:    store i8 0, ptr [[TMP10]], align 2
+// WINDOWS-NEXT:    [[TMP11:%.*]] = getelementptr i8, ptr [[TMP0]], i32 7
+// WINDOWS-NEXT:    store i8 0, ptr [[TMP11]], align 1
+// WINDOWS-NEXT:    [[TMP12:%.*]] = getelementptr i8, ptr [[TMP0]], i32 15
+// WINDOWS-NEXT:    [[TMP13:%.*]] = load i8, ptr [[TMP12]], align 1
+// WINDOWS-NEXT:    [[TMP14:%.*]] = and i8 [[TMP13]], 31
+// WINDOWS-NEXT:    store i8 [[TMP14]], ptr [[TMP12]], align 1
+// WINDOWS-NEXT:    [[TMP15:%.*]] = getelementptr i8, ptr [[TMP0]], i32 16
+// WINDOWS-NEXT:    store i8 0, ptr [[TMP15]], align 8
+// WINDOWS-NEXT:    [[TMP16:%.*]] = getelementptr i8, ptr [[TMP0]], i32 17
+// WINDOWS-NEXT:    store i8 0, ptr [[TMP16]], align 1
+// WINDOWS-NEXT:    [[TMP17:%.*]] = getelementptr i8, ptr [[TMP0]], i32 18
+// WINDOWS-NEXT:    store i8 0, ptr [[TMP17]], align 2
+// WINDOWS-NEXT:    [[TMP18:%.*]] = getelementptr i8, ptr [[TMP0]], i32 19
+// WINDOWS-NEXT:    store i8 0, ptr [[TMP18]], align 1
+// WINDOWS-NEXT:    [[TMP19:%.*]] = getelementptr i8, ptr [[TMP0]], i32 20
+// WINDOWS-NEXT:    store i8 0, ptr [[TMP19]], align 4
+// WINDOWS-NEXT:    [[TMP20:%.*]] = getelementptr i8, ptr [[TMP0]], i32 21
+// WINDOWS-NEXT:    store i8 0, ptr [[TMP20]], align 1
+// WINDOWS-NEXT:    [[TMP21:%.*]] = getelementptr i8, ptr [[TMP0]], i32 22
+// WINDOWS-NEXT:    store i8 0, ptr [[TMP21]], align 2
+// WINDOWS-NEXT:    [[TMP22:%.*]] = getelementptr i8, ptr [[TMP0]], i32 23
+// WINDOWS-NEXT:    store i8 0, ptr [[TMP22]], align 1
+// WINDOWS-NEXT:    ret void
+//
+// ARM-LE-LABEL: define dso_local void @testBitfieldBitIntMembers(
+// ARM-LE-SAME: ptr noundef [[S:%.*]]) #[[ATTR0]] {
+// ARM-LE-NEXT:  [[ENTRY:.*:]]
+// ARM-LE-NEXT:    [[S_ADDR:%.*]] = alloca ptr, align 4
+// ARM-LE-NEXT:    store ptr [[S]], ptr [[S_ADDR]], align 4
+// ARM-LE-NEXT:    [[TMP0:%.*]] = load ptr, ptr [[S_ADDR]], align 4
+// ARM-LE-NEXT:    [[TMP1:%.*]] = getelementptr i8, ptr [[TMP0]], i32 9
+// ARM-LE-NEXT:    [[TMP2:%.*]] = load i8, ptr [[TMP1]], align 1
+// ARM-LE-NEXT:    [[TMP3:%.*]] = and i8 [[TMP2]], 127
+// ARM-LE-NEXT:    store i8 [[TMP3]], ptr [[TMP1]], align 1
+// ARM-LE-NEXT:    [[TMP4:%.*]] = getelementptr i8, ptr [[TMP0]], i32 10
+// ARM-LE-NEXT:    store i8 0, ptr [[TMP4]], align 2
+// ARM-LE-NEXT:    [[TMP5:%.*]] = getelementptr i8, ptr [[TMP0]], i32 11
+// ARM-LE-NEXT:    store i8 0, ptr [[TMP5]], align 1
+// ARM-LE-NEXT:    [[TMP6:%.*]] = getelementptr i8, ptr [[TMP0]], i32 12
+// ARM-LE-NEXT:    store i8 0, ptr [[TMP6]], align 4
+// ARM-LE-NEXT:    [[TMP7:%.*]] = getelementptr i8, ptr [[TMP0]], i32 13
+// ARM-LE-NEXT:    store i8 0, ptr [[TMP7]], align 1
+// ARM-LE-NEXT:    [[TMP8:%.*]] = getelementptr i8, ptr [[TMP0]], i32 14
+// ARM-LE-NEXT:    store i8 0, ptr [[TMP8]], align 2
+// ARM-LE-NEXT:    [[TMP9:%.*]] = getelementptr i8, ptr [[TMP0]], i32 15
+// ARM-LE-NEXT:    store i8 0, ptr [[TMP9]], align 1
+// ARM-LE-NEXT:    ret void
+//
+// ARM-BE-LABEL: define dso_local void @testBitfieldBitIntMembers(
+// ARM-BE-SAME: ptr noundef [[S:%.*]]) #[[ATTR0]] {
+// ARM-BE-NEXT:  [[ENTRY:.*:]]
+// ARM-BE-NEXT:    [[S_ADDR:%.*]] = alloca ptr, align 4
+// ARM-BE-NEXT:    store ptr [[S]], ptr [[S_ADDR]], align 4
+// ARM-BE-NEXT:    [[TMP0:%.*]] = load ptr, ptr [[S_ADDR]], align 4
+// ARM-BE-NEXT:    [[TMP1:%.*]] = getelementptr i8, ptr [[TMP0]], i32 9
+// ARM-BE-NEXT:    [[TMP2:%.*]] = load i8, ptr [[TMP1]], align 1
+// ARM-BE-NEXT:    [[TMP3:%.*]] = and i8 [[TMP2]], -2
+// ARM-BE-NEXT:    store i8 [[TMP3]], ptr [[TMP1]], align 1
+// ARM-BE-NEXT:    [[TMP4:%.*]] = getelementptr i8, ptr [[TMP0]], i32 10
+// ARM-BE-NEXT:    store i8 0, ptr [[TMP4]], align 2
+// ARM-BE-NEXT:    [[TMP5:%.*]] = getelementptr i8, ptr [[TMP0]], i32 11
+// ARM-BE-NEXT:    store i8 0, ptr [[TMP5]], align 1
+// ARM-BE-NEXT:    [[TMP6:%.*]] = getelementptr i8, ptr [[TMP0]], i32 12
+// ARM-BE-NEXT:    store i8 0, ptr [[TMP6]], align 4
+// ARM-BE-NEXT:    [[TMP7:%.*]] = getelementptr i8, ptr [[TMP0]], i32 13
+// ARM-BE-NEXT:    store i8 0, ptr [[TMP7]], align 1
+// ARM-BE-NEXT:    [[TMP8:%.*]] = getelementptr i8, ptr [[TMP0]], i32 14
+// ARM-BE-NEXT:    store i8 0, ptr [[TMP8]], align 2
+// ARM-BE-NEXT:    [[TMP9:%.*]] = getelementptr i8, ptr [[TMP0]], i32 15
+// ARM-BE-NEXT:    store i8 0, ptr [[TMP9]], align 1
+// ARM-BE-NEXT:    ret void
+//
+// AARCH64-LE-LABEL: define dso_local void @testBitfieldBitIntMembers(
+// AARCH64-LE-SAME: ptr noundef [[S:%.*]]) #[[ATTR0]] {
+// AARCH64-LE-NEXT:  [[ENTRY:.*:]]
+// AARCH64-LE-NEXT:    [[S_ADDR:%.*]] = alloca ptr, align 8
+// AARCH64-LE-NEXT:    store ptr [[S]], ptr [[S_ADDR]], align 8
+// AARCH64-LE-NEXT:    [[TMP0:%.*]] = load ptr, ptr [[S_ADDR]], align 8
+// AARCH64-LE-NEXT:    [[TMP1:%.*]] = getelementptr i8, ptr [[TMP0]], i32 9
+// AARCH64-LE-NEXT:    [[TMP2:%.*]] = load i8, ptr [[TMP1]], align 1
+// AARCH64-LE-NEXT:    [[TMP3:%.*]] = and i8 [[TMP2]], 127
+// AARCH64-LE-NEXT:    store i8 [[TMP3]], ptr [[TMP1]], align 1
+// AARCH64-LE-NEXT:    [[TMP4:%.*]] = getelementptr i8, ptr [[TMP0]], i32 10
+// AARCH64-LE-NEXT:    store i8 0, ptr [[TMP4]], align 2
+// AARCH64-LE-NEXT:    [[TMP5:%.*]] = getelementptr i8, ptr [[TMP0]], i32 11
+// AARCH64-LE-NEXT:    store i8 0, ptr [[TMP5]], align 1
+// AARCH64-LE-NEXT:    [[TMP6:%.*]] = getelementptr i8, ptr [[TMP0]], i32 12
+// AARCH64-LE-NEXT:    store i8 0, ptr [[TMP6]], align 4
+// AARCH64-LE-NEXT:    [[TMP7:%.*]] = getelementptr i8, ptr [[TMP0]], i32 13
+// AARCH64-LE-NEXT:    store i8 0, ptr [[TMP7]], align 1
+// AARCH64-LE-NEXT:    [[TMP8:%.*]] = getelementptr i8, ptr [[TMP0]], i32 14
+// AARCH64-LE-NEXT:    store i8 0, ptr [[TMP8]], align 2
+// AARCH64-LE-NEXT:    [[TMP9:%.*]] = getelementptr i8, ptr [[TMP0]], i32 15
+// AARCH64-LE-NEXT:    store i8 0, ptr [[TMP9]], align 1
+// AARCH64-LE-NEXT:    ret void
+//
+// AARCH64-BE-LABEL: define dso_local void @testBitfieldBitIntMembers(
+// AARCH64-BE-SAME: ptr noundef [[S:%.*]]) #[[ATTR0]] {
+// AARCH64-BE-NEXT:  [[ENTRY:.*:]]
+// AARCH64-BE-NEXT:    [[S_ADDR:%.*]] = alloca ptr, align 8
+// AARCH64-BE-NEXT:    store ptr [[S]], ptr [[S_ADDR]], align 8
+// AARCH64-BE-NEXT:    [[TMP0:%.*]] = load ptr, ptr [[S_ADDR]], align 8
+// AARCH64-BE-NEXT:    [[TMP1:%.*]] = getelementptr i8, ptr [[TMP0]], i32 9
+// AARCH64-BE-NEXT:    [[TMP2:%.*]] = load i8, ptr [[TMP1]], align 1
+// AARCH64-BE-NEXT:    [[TMP3:%.*]] = and i8 [[TMP2]], -2
+// AARCH64-BE-NEXT:    store i8 [[TMP3]], ptr [[TMP1]], align 1
+// AARCH64-BE-NEXT:    [[TMP4:%.*]] = getelementptr i8, ptr [[TMP0]], i32 10
+// AARCH64-BE-NEXT:    store i8 0, ptr [[TMP4]], align 2
+// AARCH64-BE-NEXT:    [[TMP5:%.*]] = getelementptr i8, ptr [[TMP0]], i32 11
+// AARCH64-BE-NEXT:    store i8 0, ptr [[TMP5]], align 1
+// AARCH64-BE-NEXT:    [[TMP6:%.*]] = getelementptr i8, ptr [[TMP0]], i32 12
+// AARCH64-BE-NEXT:    store i8 0, ptr [[TMP6]], align 4
+// AARCH64-BE-NEXT:    [[TMP7:%.*]] = getelementptr i8, ptr [[TMP0]], i32 13
+// AARCH64-BE-NEXT:    store i8 0, ptr [[TMP7]], align 1
+// AARCH64-BE-NEXT:    [[TMP8:%.*]] = getelementptr i8, ptr [[TMP0]], i32 14
+// AARCH64-BE-NEXT:    store i8 0, ptr [[TMP8]], align 2
+// AARCH64-BE-NEXT:    [[TMP9:%.*]] = getelementptr i8, ptr [[TMP0]], i32 15
+// AARCH64-BE-NEXT:    store i8 0, ptr [[TMP9]], align 1
+// AARCH64-BE-NEXT:    ret void
+//
+void testBitfieldBitIntMembers(struct BitfieldBitIntMembers *s) {
+  __builtin_clear_padding(s);
+}
+
+// Packing permits the wide field to straddle its declared type's usual
+// storage-unit boundary on Arm. Ordinary members surround the bitfield run.
+struct __attribute__((packed)) BitfieldPackedMembers {
+  unsigned char prefix;
+  unsigned int first : 5;
+  unsigned long long wide : 60;
+  unsigned char suffix;
+};
+
+// LINUX-LABEL: define dso_local void @testBitfieldPackedMembers(
+// LINUX-SAME: ptr noundef [[S:%.*]]) #[[ATTR0]] {
+// LINUX-NEXT:  [[ENTRY:.*:]]
+// LINUX-NEXT:    [[S_ADDR:%.*]] = alloca ptr, align 8
+// LINUX-NEXT:    store ptr [[S]], ptr [[S_ADDR]], align 8
+// LINUX-NEXT:    [[TMP0:%.*]] = load ptr, ptr [[S_ADDR]], align 8
+// LINUX-NEXT:    [[TMP1:%.*]] = getelementptr i8, ptr [[TMP0]], i32 9
+// LINUX-NEXT:    [[TMP2:%.*]] = load i8, ptr [[TMP1]], align 1
+// LINUX-NEXT:    [[TMP3:%.*]] = and i8 [[TMP2]], 1
+// LINUX-NEXT:    store i8 [[TMP3]], ptr [[TMP1]], align 1
+// LINUX-NEXT:    ret void
+//
+// WINDOWS-LABEL: define dso_local void @testBitfieldPackedMembers(
+// WINDOWS-SAME: ptr noundef [[S:%.*]]) #[[ATTR0]] {
+// WINDOWS-NEXT:  [[ENTRY:.*:]]
+// WINDOWS-NEXT:    [[S_ADDR:%.*]] = alloca ptr, align 8
+// WINDOWS-NEXT:    store ptr [[S]], ptr [[S_ADDR]], align 8
+// WINDOWS-NEXT:    [[TMP0:%.*]] = load ptr, ptr [[S_ADDR]], align 8
+// WINDOWS-NEXT:    [[TMP1:%.*]] = getelementptr i8, ptr [[TMP0]], i32 1
+// WINDOWS-NEXT:    [[TMP2:%.*]] = load i8, ptr [[TMP1]], align 1
+// WINDOWS-NEXT:    [[TMP3:%.*]] = and i8 [[TMP2]], 31
+// WINDOWS-NEXT:    store i8 [[TMP3]], ptr [[TMP1]], align 1
+// WINDOWS-NEXT:    [[TMP4:%.*]] = getelementptr i8, ptr [[TMP0]], i32 2
+// WINDOWS-NEXT:    store i8 0, ptr [[TMP4]], align 1
+// WINDOWS-NEXT:    [[TMP5:%.*]] = getelementptr i8, ptr [[TMP0]], i32 3
+// WINDOWS-NEXT:    store i8 0, ptr [[TMP5]], align 1
+// WINDOWS-NEXT:    [[TMP6:%.*]] = getelementptr i8, ptr [[TMP0]], i32 4
+// WINDOWS-NEXT:    store i8 0, ptr [[TMP6]], align 1
+// WINDOWS-NEXT:    [[TMP7:%.*]] = getelementptr i8, ptr [[TMP0]], i32 12
+// WINDOWS-NEXT:    [[TMP8:%.*]] = load i8, ptr [[TMP7]], align 1
+// WINDOWS-NEXT:    [[TMP9:%.*]] = and i8 [[TMP8]], 15
+// WINDOWS-NEXT:    store i8 [[TMP9]], ptr [[TMP7]], align 1
+// WINDOWS-NEXT:    ret void
+//
+// ARM-LE-LABEL: define dso_local void @testBitfieldPackedMembers(
+// ARM-LE-SAME: ptr noundef [[S:%.*]]) #[[ATTR0]] {
+// ARM-LE-NEXT:  [[ENTRY:.*:]]
+// ARM-LE-NEXT:    [[S_ADDR:%.*]] = alloca ptr, align 4
+// ARM-LE-NEXT:    store ptr [[S]], ptr [[S_ADDR]], align 4
+// ARM-LE-NEXT:    [[TMP0:%.*]] = load ptr, ptr [[S_ADDR]], align 4
+// ARM-LE-NEXT:    [[TMP1:%.*]] = getelementptr i8, ptr [[TMP0]], i32 9
+// ARM-LE-NEXT:    [[TMP2:%.*]] = load i8, ptr [[TMP1]], align 1
+// ARM-LE-NEXT:    [[TMP3:%.*]] = and i8 [[TMP2]], 1
+// ARM-LE-NEXT:    store i8 [[TMP3]], ptr [[TMP1]], align 1
+// ARM-LE-NEXT:    ret void
+//
+// ARM-BE-LABEL: define dso_local void @testBitfieldPackedMembers(
+// ARM-BE-SAME: ptr noundef [[S:%.*]]) #[[ATTR0]] {
+// ARM-BE-NEXT:  [[ENTRY:.*:]]
+// ARM-BE-NEXT:    [[S_ADDR:%.*]] = alloca ptr, align 4
+// ARM-BE-NEXT:    store ptr [[S]], ptr [[S_ADDR]], align 4
+// ARM-BE-NEXT:    [[TMP0:%.*]] = load ptr, ptr [[S_ADDR]], align 4
+// ARM-BE-NEXT:    [[TMP1:%.*]] = getelementptr i8, ptr [[TMP0]], i32 9
+// ARM-BE-NEXT:    [[TMP2:%.*]] = load i8, ptr [[TMP1]], align 1
+// ARM-BE-NEXT:    [[TMP3:%.*]] = and i8 [[TMP2]], -128
+// ARM-BE-NEXT:    store i8 [[TMP3]], ptr [[TMP1]], align 1
+// ARM-BE-NEXT:    ret void
+//
+// AARCH64-LE-LABEL: define dso_local void @testBitfieldPackedMembers(
+// AARCH64-LE-SAME: ptr noundef [[S:%.*]]) #[[ATTR0]] {
+// AARCH64-LE-NEXT:  [[ENTRY:.*:]]
+// AARCH64-LE-NEXT:    [[S_ADDR:%.*]] = alloca ptr, align 8
+// AARCH64-LE-NEXT:    store ptr [[S]], ptr [[S_ADDR]], align 8
+// AARCH64-LE-NEXT:    [[TMP0:%.*]] = load ptr, ptr [[S_ADDR]], align 8
+// AARCH64-LE-NEXT:    [[TMP1:%.*]] = getelementptr i8, ptr [[TMP0]], i32 9
+// AARCH64-LE-NEXT:    [[TMP2:%.*]] = load i8, ptr [[TMP1]], align 1
+// AARCH64-LE-NEXT:    [[TMP3:%.*]] = and i8 [[TMP2]], 1
+// AARCH64-LE-NEXT:    store i8 [[TMP3]], ptr [[TMP1]], align 1
+// AARCH64-LE-NEXT:    ret void
+//
+// AARCH64-BE-LABEL: define dso_local void @testBitfieldPackedMembers(
+// AARCH64-BE-SAME: ptr noundef [[S:%.*]]) #[[ATTR0]] {
+// AARCH64-BE-NEXT:  [[ENTRY:.*:]]
+// AARCH64-BE-NEXT:    [[S_ADDR:%.*]] = alloca ptr, align 8
+// AARCH64-BE-NEXT:    store ptr [[S]], ptr [[S_ADDR]], align 8
+// AARCH64-BE-NEXT:    [[TMP0:%.*]] = load ptr, ptr [[S_ADDR]], align 8
+// AARCH64-BE-NEXT:    [[TMP1:%.*]] = getelementptr i8, ptr [[TMP0]], i32 9
+// AARCH64-BE-NEXT:    [[TMP2:%.*]] = load i8, ptr [[TMP1]], align 1
+// AARCH64-BE-NEXT:    [[TMP3:%.*]] = and i8 [[TMP2]], -128
+// AARCH64-BE-NEXT:    store i8 [[TMP3]], ptr [[TMP1]], align 1
+// AARCH64-BE-NEXT:    ret void
+//
+void testBitfieldPackedMembers(struct BitfieldPackedMembers *s) {
+  __builtin_clear_padding(s);
+}
+
+// There are no occupied intervals: clear the entire object representation.
+struct BitfieldUnnamedOnly {
+  unsigned int : 3;
+  unsigned int : 0;
+  unsigned char : 2;
+};
+
+// LINUX-LABEL: define dso_local void @testBitfieldUnnamedOnly(
+// LINUX-SAME: ptr noundef [[S:%.*]]) #[[ATTR0]] {
+// LINUX-NEXT:  [[ENTRY:.*:]]
+// LINUX-NEXT:    [[S_ADDR:%.*]] = alloca ptr, align 8
+// LINUX-NEXT:    store ptr [[S]], ptr [[S_ADDR]], align 8
+// LINUX-NEXT:    [[TMP0:%.*]] = load ptr, ptr [[S_ADDR]], align 8
+// LINUX-NEXT:    [[TMP1:%.*]] = getelementptr i8, ptr [[TMP0]], i32 0
+// LINUX-NEXT:    store i8 0, ptr [[TMP1]], align 1
+// LINUX-NEXT:    [[TMP2:%.*]] = getelementptr i8, ptr [[TMP0]], i32 1
+// LINUX-NEXT:    store i8 0, ptr [[TMP2]], align 1
+// LINUX-NEXT:    [[TMP3:%.*]] = getelementptr i8, ptr [[TMP0]], i32 2
+// LINUX-NEXT:    store i8 0, ptr [[TMP3]], align 1
+// LINUX-NEXT:    [[TMP4:%.*]] = getelementptr i8, ptr [[TMP0]], i32 3
+// LINUX-NEXT:    store i8 0, ptr [[TMP4]], align 1
+// LINUX-NEXT:    [[TMP5:%.*]] = getelementptr i8, ptr [[TMP0]], i32 4
+// LINUX-NEXT:    store i8 0, ptr [[TMP5]], align 1
+// LINUX-NEXT:    ret void
+//
+// WINDOWS-LABEL: define dso_local void @testBitfieldUnnamedOnly(
+// WINDOWS-SAME: ptr noundef [[S:%.*]]) #[[ATTR0]] {
+// WINDOWS-NEXT:  [[ENTRY:.*:]]
+// WINDOWS-NEXT:    [[S_ADDR:%.*]] = alloca ptr, align 8
+// WINDOWS-NEXT:    store ptr [[S]], ptr [[S_ADDR]], align 8
+// WINDOWS-NEXT:    [[TMP0:%.*]] = load ptr, ptr [[S_ADDR]], align 8
+// WINDOWS-NEXT:    [[TMP1:%.*]] = getelementptr i8, ptr [[TMP0]], i32 0
+// WINDOWS-NEXT:    store i8 0, ptr [[TMP1]], align 4
+// WINDOWS-NEXT:    [[TMP2:%.*]] = getelementptr i8, ptr [[TMP0]], i32 1
+// WINDOWS-NEXT:    store i8 0, ptr [[TMP2]], align 1
+// WINDOWS-NEXT:    [[TMP3:%.*]] = getelementptr i8, ptr [[TMP0]], i32 2
+// WINDOWS-NEXT:    store i8 0, ptr [[TMP3]], align 2
+// WINDOWS-NEXT:    [[TMP4:%.*]] = getelementptr i8, ptr [[TMP0]], i32 3
+// WINDOWS-NEXT:    store i8 0, ptr [[TMP4]], align 1
+// WINDOWS-NEXT:    [[TMP5:%.*]] = getelementptr i8, ptr [[TMP0]], i32 4
+// WINDOWS-NEXT:    store i8 0, ptr [[TMP5]], align 4
+// WINDOWS-NEXT:    [[TMP6:%.*]] = getelementptr i8, ptr [[TMP0]], i32 5
+// WINDOWS-NEXT:    store i8 0, ptr [[TMP6]], align 1
+// WINDOWS-NEXT:    [[TMP7:%.*]] = getelementptr i8, ptr [[TMP0]], i32 6
+// WINDOWS-NEXT:    store i8 0, ptr [[TMP7]], align 2
+// WINDOWS-NEXT:    [[TMP8:%.*]] = getelementptr i8, ptr [[TMP0]], i32 7
+// WINDOWS-NEXT:    store i8 0, ptr [[TMP8]], align 1
+// WINDOWS-NEXT:    ret void
+//
+// ARM-LE-LABEL: define dso_local void @testBitfieldUnnamedOnly(
+// ARM-LE-SAME: ptr noundef [[S:%.*]]) #[[ATTR0]] {
+// ARM-LE-NEXT:  [[ENTRY:.*:]]
+// ARM-LE-NEXT:    [[S_ADDR:%.*]] = alloca ptr, align 4
+// ARM-LE-NEXT:    store ptr [[S]], ptr [[S_ADDR]], align 4
+// ARM-LE-NEXT:    [[TMP0:%.*]] = load ptr, ptr [[S_ADDR]], align 4
+// ARM-LE-NEXT:    [[TMP1:%.*]] = getelementptr i8, ptr [[TMP0]], i32 0
+// ARM-LE-NEXT:    store i8 0, ptr [[TMP1]], align 4
+// ARM-LE-NEXT:    [[TMP2:%.*]] = getelementptr i8, ptr [[TMP0]], i32 1
+// ARM-LE-NEXT:    store i8 0, ptr [[TMP2]], align 1
+// ARM-LE-NEXT:    [[TMP3:%.*]] = getelementptr i8, ptr [[TMP0]], i32 2
+// ARM-LE-NEXT:    store i8 0, ptr [[TMP3]], align 2
+// ARM-LE-NEXT:    [[TMP4:%.*]] = getelementptr i8, ptr [[TMP0]], i32 3
+// ARM-LE-NEXT:    store i8 0, ptr [[TMP4]], align 1
+// ARM-LE-NEXT:    [[TMP5:%.*]] = getelementptr i8, ptr [[TMP0]], i32 4
+// ARM-LE-NEXT:    store i8 0, ptr [[TMP5]], align 4
+// ARM-LE-NEXT:    [[TMP6:%.*]] = getelementptr i8, ptr [[TMP0]], i32 5
+// ARM-LE-NEXT:    store i8 0, ptr [[TMP6]], align 1
+// ARM-LE-NEXT:    [[TMP7:%.*]] = getelementptr i8, ptr [[TMP0]], i32 6
+// ARM-LE-NEXT:    store i8 0, ptr [[TMP7]], align 2
+// ARM-LE-NEXT:    [[TMP8:%.*]] = getelementptr i8, ptr [[TMP0]], i32 7
+// ARM-LE-NEXT:    store i8 0, ptr [[TMP8]], align 1
+// ARM-LE-NEXT:    ret void
+//
+// ARM-BE-LABEL: define dso_local void @testBitfieldUnnamedOnly(
+// ARM-BE-SAME: ptr noundef [[S:%.*]]) #[[ATTR0]] {
+// ARM-BE-NEXT:  [[ENTRY:.*:]]
+// ARM-BE-NEXT:    [[S_ADDR:%.*]] = alloca ptr, align 4
+// ARM-BE-NEXT:    store ptr [[S]], ptr [[S_ADDR]], align 4
+// ARM-BE-NEXT:    [[TMP0:%.*]] = load ptr, ptr [[S_ADDR]], align 4
+// ARM-BE-NEXT:    [[TMP1:%.*]] = getelementptr i8, ptr [[TMP0]], i32 0
+// ARM-BE-NEXT:    store i8 0, ptr [[TMP1]], align 4
+// ARM-BE-NEXT:    [[TMP2:%.*]] = getelementptr i8, ptr [[TMP0]], i32 1
+// ARM-BE-NEXT:    store i8 0, ptr [[TMP2]], align 1
+// ARM-BE-NEXT:    [[TMP3:%.*]] = getelementptr i8, ptr [[TMP0]], i32 2
+// ARM-BE-NEXT:    store i8 0, ptr [[TMP3]], align 2
+// ARM-BE-NEXT:    [[TMP4:%.*]] = getelementptr i8, ptr [[TMP0]], i32 3
+// ARM-BE-NEXT:    store i8 0, ptr [[TMP4]], align 1
+// ARM-BE-NEXT:    [[TMP5:%.*]] = getelementptr i8, ptr [[TMP0]], i32 4
+// ARM-BE-NEXT:    store i8 0, ptr [[TMP5]], align 4
+// ARM-BE-NEXT:    [[TMP6:%.*]] = getelementptr i8, ptr [[TMP0]], i32 5
+// ARM-BE-NEXT:    store i8 0, ptr [[TMP6]], align 1
+// ARM-BE-NEXT:    [[TMP7:%.*]] = getelementptr i8, ptr [[TMP0]], i32 6
+// ARM-BE-NEXT:    store i8 0, ptr [[TMP7]], align 2
+// ARM-BE-NEXT:    [[TMP8:%.*]] = getelementptr i8, ptr [[TMP0]], i32 7
+// ARM-BE-NEXT:    store i8 0, ptr [[TMP8]], align 1
+// ARM-BE-NEXT:    ret void
+//
+// AARCH64-LE-LABEL: define dso_local void @testBitfieldUnnamedOnly(
+// AARCH64-LE-SAME: ptr noundef [[S:%.*]]) #[[ATTR0]] {
+// AARCH64-LE-NEXT:  [[ENTRY:.*:]]
+// AARCH64-LE-NEXT:    [[S_ADDR:%.*]] = alloca ptr, align 8
+// AARCH64-LE-NEXT:    store ptr [[S]], ptr [[S_ADDR]], align 8
+// AARCH64-LE-NEXT:    [[TMP0:%.*]] = load ptr, ptr [[S_ADDR]], align 8
+// AARCH64-LE-NEXT:    [[TMP1:%.*]] = getelementptr i8, ptr [[TMP0]], i32 0
+// AARCH64-LE-NEXT:    store i8 0, ptr [[TMP1]], align 4
+// AARCH64-LE-NEXT:    [[TMP2:%.*]] = getelementptr i8, ptr [[TMP0]], i32 1
+// AARCH64-LE-NEXT:    store i8 0, ptr [[TMP2]], align 1
+// AARCH64-LE-NEXT:    [[TMP3:%.*]] = getelementptr i8, ptr [[TMP0]], i32 2
+// AARCH64-LE-NEXT:    store i8 0, ptr [[TMP3]], align 2
+// AARCH64-LE-NEXT:    [[TMP4:%.*]] = getelementptr i8, ptr [[TMP0]], i32 3
+// AARCH64-LE-NEXT:    store i8 0, ptr [[TMP4]], align 1
+// AARCH64-LE-NEXT:    [[TMP5:%.*]] = getelementptr i8, ptr [[TMP0]], i32 4
+// AARCH64-LE-NEXT:    store i8 0, ptr [[TMP5]], align 4
+// AARCH64-LE-NEXT:    [[TMP6:%.*]] = getelementptr i8, ptr [[TMP0]], i32 5
+// AARCH64-LE-NEXT:    store i8 0, ptr [[TMP6]], align 1
+// AARCH64-LE-NEXT:    [[TMP7:%.*]] = getelementptr i8, ptr [[TMP0]], i32 6
+// AARCH64-LE-NEXT:    store i8 0, ptr [[TMP7]], align 2
+// AARCH64-LE-NEXT:    [[TMP8:%.*]] = getelementptr i8, ptr [[TMP0]], i32 7
+// AARCH64-LE-NEXT:    store i8 0, ptr [[TMP8]], align 1
+// AARCH64-LE-NEXT:    ret void
+//
+// AARCH64-BE-LABEL: define dso_local void @testBitfieldUnnamedOnly(
+// AARCH64-BE-SAME: ptr noundef [[S:%.*]]) #[[ATTR0]] {
+// AARCH64-BE-NEXT:  [[ENTRY:.*:]]
+// AARCH64-BE-NEXT:    [[S_ADDR:%.*]] = alloca ptr, align 8
+// AARCH64-BE-NEXT:    store ptr [[S]], ptr [[S_ADDR]], align 8
+// AARCH64-BE-NEXT:    [[TMP0:%.*]] = load ptr, ptr [[S_ADDR]], align 8
+// AARCH64-BE-NEXT:    [[TMP1:%.*]] = getelementptr i8, ptr [[TMP0]], i32 0
+// AARCH64-BE-NEXT:    store i8 0, ptr [[TMP1]], align 4
+// AARCH64-BE-NEXT:    [[TMP2:%.*]] = getelementptr i8, ptr [[TMP0]], i32 1
+// AARCH64-BE-NEXT:    store i8 0, ptr [[TMP2]], align 1
+// AARCH64-BE-NEXT:    [[TMP3:%.*]] = getelementptr i8, ptr [[TMP0]], i32 2
+// AARCH64-BE-NEXT:    store i8 0, ptr [[TMP3]], align 2
+// AARCH64-BE-NEXT:    [[TMP4:%.*]] = getelementptr i8, ptr [[TMP0]], i32 3
+// AARCH64-BE-NEXT:    store i8 0, ptr [[TMP4]], align 1
+// AARCH64-BE-NEXT:    [[TMP5:%.*]] = getelementptr i8, ptr [[TMP0]], i32 4
+// AARCH64-BE-NEXT:    store i8 0, ptr [[TMP5]], align 4
+// AARCH64-BE-NEXT:    [[TMP6:%.*]] = getelementptr i8, ptr [[TMP0]], i32 5
+// AARCH64-BE-NEXT:    store i8 0, ptr [[TMP6]], align 1
+// AARCH64-BE-NEXT:    [[TMP7:%.*]] = getelementptr i8, ptr [[TMP0]], i32 6
+// AARCH64-BE-NEXT:    store i8 0, ptr [[TMP7]], align 2
+// AARCH64-BE-NEXT:    [[TMP8:%.*]] = getelementptr i8, ptr [[TMP0]], i32 7
+// AARCH64-BE-NEXT:    store i8 0, ptr [[TMP8]], align 1
+// AARCH64-BE-NEXT:    ret void
+//
+void testBitfieldUnnamedOnly(struct BitfieldUnnamedOnly *s) {
+  __builtin_clear_padding(s);
+}
diff --git a/clang/test/CodeGenCXX/builtin-clear-padding-codegen.cpp b/clang/test/CodeGenCXX/builtin-clear-padding-codegen.cpp
index 2de4645fa62a8..858b43168a67d 100644
--- a/clang/test/CodeGenCXX/builtin-clear-padding-codegen.cpp
+++ b/clang/test/CodeGenCXX/builtin-clear-padding-codegen.cpp
@@ -1,15 +1,16 @@
 // NOTE: Assertions have been autogenerated by utils/update_cc_test_checks.py UTC_ARGS: --version 5
+// Test functions use C linkage so the updater recognizes their Windows names.
 // RUN: %clang_cc1 -std=c++20 -triple=x86_64-linux-gnu -emit-llvm -o - %s | FileCheck %s --check-prefix=LINUX
 // RUN: %clang_cc1 -std=c++20 -triple=x86_64-pc-windows-msvc -emit-llvm -o - %s | FileCheck %s --check-prefix=WINDOWS
 // RUN: %clang_cc1 -std=c++20 -triple=armv7a-none-eabi -emit-llvm -o - %s | FileCheck %s --check-prefix=ARM-LE
 // RUN: %clang_cc1 -std=c++20 -triple=armebv7a-none-eabi -emit-llvm -o - %s | FileCheck %s --check-prefix=ARM-BE
-// RUN: %clang_cc1 -std=c++20 -triple=aarch64-none-eabi -emit-llvm -o - %s | FileCheck %s --check-prefix=ARM-LE
-// RUN: %clang_cc1 -std=c++20 -triple=aarch64_be-none-eabi -emit-llvm -o - %s | FileCheck %s --check-prefix=ARM-BE
+// RUN: %clang_cc1 -std=c++20 -triple=aarch64-none-eabi -emit-llvm -o - %s | FileCheck %s --check-prefix=AARCH64-LE
+// RUN: %clang_cc1 -std=c++20 -triple=aarch64_be-none-eabi -emit-llvm -o - %s | FileCheck %s --check-prefix=AARCH64-BE
 
 
 struct Empty {};
 
-// LINUX-LABEL: define dso_local void @_Z9testEmptyP5Empty(
+// LINUX-LABEL: define dso_local void @testEmpty(
 // LINUX-SAME: ptr noundef [[E:%.*]]) #[[ATTR0:[0-9]+]] {
 // LINUX-NEXT:  [[ENTRY:.*:]]
 // LINUX-NEXT:    [[E_ADDR:%.*]] = alloca ptr, align 8
@@ -19,22 +20,64 @@ struct Empty {};
 // LINUX-NEXT:    store i8 0, ptr [[TMP1]], align 1
 // LINUX-NEXT:    ret void
 //
-// WINDOWS-LABEL: define dso_local void @"?testEmpty@@YAXPEAUEmpty@@@Z"(ptr noundef %e) #0 {
-// WINDOWS-NEXT: entry:
-// WINDOWS-NEXT:   %e.addr = alloca ptr, align 8
-// WINDOWS-NEXT:   store ptr %e, ptr %e.addr, align 8
-// WINDOWS-NEXT:   %0 = load ptr, ptr %e.addr, align 8
-// WINDOWS-NEXT:   %1 = getelementptr i8, ptr %0, i32 0
-// WINDOWS-NEXT:   store i8 0, ptr %1, align 1
-// WINDOWS-NEXT:   ret void
+// WINDOWS-LABEL: define dso_local void @testEmpty(
+// WINDOWS-SAME: ptr noundef [[E:%.*]]) #[[ATTR0:[0-9]+]] {
+// WINDOWS-NEXT:  [[ENTRY:.*:]]
+// WINDOWS-NEXT:    [[E_ADDR:%.*]] = alloca ptr, align 8
+// WINDOWS-NEXT:    store ptr [[E]], ptr [[E_ADDR]], align 8
+// WINDOWS-NEXT:    [[TMP0:%.*]] = load ptr, ptr [[E_ADDR]], align 8
+// WINDOWS-NEXT:    [[TMP1:%.*]] = getelementptr i8, ptr [[TMP0]], i32 0
+// WINDOWS-NEXT:    store i8 0, ptr [[TMP1]], align 1
+// WINDOWS-NEXT:    ret void
 //
-void testEmpty(Empty *e) {
+// ARM-LE-LABEL: define dso_local void @testEmpty(
+// ARM-LE-SAME: ptr noundef [[E:%.*]]) #[[ATTR0:[0-9]+]] {
+// ARM-LE-NEXT:  [[ENTRY:.*:]]
+// ARM-LE-NEXT:    [[E_ADDR:%.*]] = alloca ptr, align 4
+// ARM-LE-NEXT:    store ptr [[E]], ptr [[E_ADDR]], align 4
+// ARM-LE-NEXT:    [[TMP0:%.*]] = load ptr, ptr [[E_ADDR]], align 4
+// ARM-LE-NEXT:    [[TMP1:%.*]] = getelementptr i8, ptr [[TMP0]], i32 0
+// ARM-LE-NEXT:    store i8 0, ptr [[TMP1]], align 1
+// ARM-LE-NEXT:    ret void
+//
+// ARM-BE-LABEL: define dso_local void @testEmpty(
+// ARM-BE-SAME: ptr noundef [[E:%.*]]) #[[ATTR0:[0-9]+]] {
+// ARM-BE-NEXT:  [[ENTRY:.*:]]
+// ARM-BE-NEXT:    [[E_ADDR:%.*]] = alloca ptr, align 4
+// ARM-BE-NEXT:    store ptr [[E]], ptr [[E_ADDR]], align 4
+// ARM-BE-NEXT:    [[TMP0:%.*]] = load ptr, ptr [[E_ADDR]], align 4
+// ARM-BE-NEXT:    [[TMP1:%.*]] = getelementptr i8, ptr [[TMP0]], i32 0
+// ARM-BE-NEXT:    store i8 0, ptr [[TMP1]], align 1
+// ARM-BE-NEXT:    ret void
+//
+// AARCH64-LE-LABEL: define dso_local void @testEmpty(
+// AARCH64-LE-SAME: ptr noundef [[E:%.*]]) #[[ATTR0:[0-9]+]] {
+// AARCH64-LE-NEXT:  [[ENTRY:.*:]]
+// AARCH64-LE-NEXT:    [[E_ADDR:%.*]] = alloca ptr, align 8
+// AARCH64-LE-NEXT:    store ptr [[E]], ptr [[E_ADDR]], align 8
+// AARCH64-LE-NEXT:    [[TMP0:%.*]] = load ptr, ptr [[E_ADDR]], align 8
+// AARCH64-LE-NEXT:    [[TMP1:%.*]] = getelementptr i8, ptr [[TMP0]], i32 0
+// AARCH64-LE-NEXT:    store i8 0, ptr [[TMP1]], align 1
+// AARCH64-LE-NEXT:    ret void
+//
+// AARCH64-BE-LABEL: define dso_local void @testEmpty(
+// AARCH64-BE-SAME: ptr noundef [[E:%.*]]) #[[ATTR0:[0-9]+]] {
+// AARCH64-BE-NEXT:  [[ENTRY:.*:]]
+// AARCH64-BE-NEXT:    [[E_ADDR:%.*]] = alloca ptr, align 8
+// AARCH64-BE-NEXT:    store ptr [[E]], ptr [[E_ADDR]], align 8
+// AARCH64-BE-NEXT:    [[TMP0:%.*]] = load ptr, ptr [[E_ADDR]], align 8
+// AARCH64-BE-NEXT:    [[TMP1:%.*]] = getelementptr i8, ptr [[TMP0]], i32 0
+// AARCH64-BE-NEXT:    store i8 0, ptr [[TMP1]], align 1
+// AARCH64-BE-NEXT:    ret void
+//
+extern "C" void testEmpty(Empty *e) {
   // This should clear the one byte that Empty occupies.
   __builtin_clear_padding(e);
 }
 
 
-// LINUX-LABEL: define dso_local void @_Z22testPrimitiveNoPaddingPi(
+
+// LINUX-LABEL: define dso_local void @testPrimitiveNoPadding(
 // LINUX-SAME: ptr noundef [[I:%.*]]) #[[ATTR0]] {
 // LINUX-NEXT:  [[ENTRY:.*:]]
 // LINUX-NEXT:    [[I_ADDR:%.*]] = alloca ptr, align 8
@@ -42,20 +85,54 @@ void testEmpty(Empty *e) {
 // LINUX-NEXT:    [[TMP0:%.*]] = load ptr, ptr [[I_ADDR]], align 8
 // LINUX-NEXT:    ret void
 //
-// WINDOWS-LABEL: define dso_local void @"?testPrimitiveNoPadding@@YAXPEAH at Z"(ptr noundef %i) #0 {
-// WINDOWS-NEXT: entry:
-// WINDOWS-NEXT:   %i.addr = alloca ptr, align 8
-// WINDOWS-NEXT:   store ptr %i, ptr %i.addr, align 8
-// WINDOWS-NEXT:   %0 = load ptr, ptr %i.addr, align 8
-// WINDOWS-NEXT:   ret void
+// WINDOWS-LABEL: define dso_local void @testPrimitiveNoPadding(
+// WINDOWS-SAME: ptr noundef [[I:%.*]]) #[[ATTR0]] {
+// WINDOWS-NEXT:  [[ENTRY:.*:]]
+// WINDOWS-NEXT:    [[I_ADDR:%.*]] = alloca ptr, align 8
+// WINDOWS-NEXT:    store ptr [[I]], ptr [[I_ADDR]], align 8
+// WINDOWS-NEXT:    [[TMP0:%.*]] = load ptr, ptr [[I_ADDR]], align 8
+// WINDOWS-NEXT:    ret void
+//
+// ARM-LE-LABEL: define dso_local void @testPrimitiveNoPadding(
+// ARM-LE-SAME: ptr noundef [[I:%.*]]) #[[ATTR0]] {
+// ARM-LE-NEXT:  [[ENTRY:.*:]]
+// ARM-LE-NEXT:    [[I_ADDR:%.*]] = alloca ptr, align 4
+// ARM-LE-NEXT:    store ptr [[I]], ptr [[I_ADDR]], align 4
+// ARM-LE-NEXT:    [[TMP0:%.*]] = load ptr, ptr [[I_ADDR]], align 4
+// ARM-LE-NEXT:    ret void
+//
+// ARM-BE-LABEL: define dso_local void @testPrimitiveNoPadding(
+// ARM-BE-SAME: ptr noundef [[I:%.*]]) #[[ATTR0]] {
+// ARM-BE-NEXT:  [[ENTRY:.*:]]
+// ARM-BE-NEXT:    [[I_ADDR:%.*]] = alloca ptr, align 4
+// ARM-BE-NEXT:    store ptr [[I]], ptr [[I_ADDR]], align 4
+// ARM-BE-NEXT:    [[TMP0:%.*]] = load ptr, ptr [[I_ADDR]], align 4
+// ARM-BE-NEXT:    ret void
+//
+// AARCH64-LE-LABEL: define dso_local void @testPrimitiveNoPadding(
+// AARCH64-LE-SAME: ptr noundef [[I:%.*]]) #[[ATTR0]] {
+// AARCH64-LE-NEXT:  [[ENTRY:.*:]]
+// AARCH64-LE-NEXT:    [[I_ADDR:%.*]] = alloca ptr, align 8
+// AARCH64-LE-NEXT:    store ptr [[I]], ptr [[I_ADDR]], align 8
+// AARCH64-LE-NEXT:    [[TMP0:%.*]] = load ptr, ptr [[I_ADDR]], align 8
+// AARCH64-LE-NEXT:    ret void
 //
-void testPrimitiveNoPadding(int *i) {
+// AARCH64-BE-LABEL: define dso_local void @testPrimitiveNoPadding(
+// AARCH64-BE-SAME: ptr noundef [[I:%.*]]) #[[ATTR0]] {
+// AARCH64-BE-NEXT:  [[ENTRY:.*:]]
+// AARCH64-BE-NEXT:    [[I_ADDR:%.*]] = alloca ptr, align 8
+// AARCH64-BE-NEXT:    store ptr [[I]], ptr [[I_ADDR]], align 8
+// AARCH64-BE-NEXT:    [[TMP0:%.*]] = load ptr, ptr [[I_ADDR]], align 8
+// AARCH64-BE-NEXT:    ret void
+//
+extern "C" void testPrimitiveNoPadding(int *i) {
   // This should not clear any padding, since int has no padding.
   __builtin_clear_padding(i);
 }
 
 
-// LINUX-LABEL: define dso_local void @_Z23testPrimitiveLongDoublePe(
+
+// LINUX-LABEL: define dso_local void @testPrimitiveLongDouble(
 // LINUX-SAME: ptr noundef [[LD:%.*]]) #[[ATTR0]] {
 // LINUX-NEXT:  [[ENTRY:.*:]]
 // LINUX-NEXT:    [[LD_ADDR:%.*]] = alloca ptr, align 8
@@ -75,19 +152,53 @@ void testPrimitiveNoPadding(int *i) {
 // LINUX-NEXT:    store i8 0, ptr [[TMP6]], align 1
 // LINUX-NEXT:    ret void
 //
-// WINDOWS-LABEL: define dso_local void @"?testPrimitiveLongDouble@@YAXPEAO at Z"(ptr noundef %ld) #0 {
-// WINDOWS-NEXT: entry:
-// WINDOWS-NEXT:   %ld.addr = alloca ptr, align 8
-// WINDOWS-NEXT:   store ptr %ld, ptr %ld.addr, align 8
-// WINDOWS-NEXT:   %0 = load ptr, ptr %ld.addr, align 8
-// WINDOWS-NEXT:   ret void
+// WINDOWS-LABEL: define dso_local void @testPrimitiveLongDouble(
+// WINDOWS-SAME: ptr noundef [[LD:%.*]]) #[[ATTR0]] {
+// WINDOWS-NEXT:  [[ENTRY:.*:]]
+// WINDOWS-NEXT:    [[LD_ADDR:%.*]] = alloca ptr, align 8
+// WINDOWS-NEXT:    store ptr [[LD]], ptr [[LD_ADDR]], align 8
+// WINDOWS-NEXT:    [[TMP0:%.*]] = load ptr, ptr [[LD_ADDR]], align 8
+// WINDOWS-NEXT:    ret void
+//
+// ARM-LE-LABEL: define dso_local void @testPrimitiveLongDouble(
+// ARM-LE-SAME: ptr noundef [[LD:%.*]]) #[[ATTR0]] {
+// ARM-LE-NEXT:  [[ENTRY:.*:]]
+// ARM-LE-NEXT:    [[LD_ADDR:%.*]] = alloca ptr, align 4
+// ARM-LE-NEXT:    store ptr [[LD]], ptr [[LD_ADDR]], align 4
+// ARM-LE-NEXT:    [[TMP0:%.*]] = load ptr, ptr [[LD_ADDR]], align 4
+// ARM-LE-NEXT:    ret void
+//
+// ARM-BE-LABEL: define dso_local void @testPrimitiveLongDouble(
+// ARM-BE-SAME: ptr noundef [[LD:%.*]]) #[[ATTR0]] {
+// ARM-BE-NEXT:  [[ENTRY:.*:]]
+// ARM-BE-NEXT:    [[LD_ADDR:%.*]] = alloca ptr, align 4
+// ARM-BE-NEXT:    store ptr [[LD]], ptr [[LD_ADDR]], align 4
+// ARM-BE-NEXT:    [[TMP0:%.*]] = load ptr, ptr [[LD_ADDR]], align 4
+// ARM-BE-NEXT:    ret void
 //
-void testPrimitiveLongDouble(long double *ld) {
+// AARCH64-LE-LABEL: define dso_local void @testPrimitiveLongDouble(
+// AARCH64-LE-SAME: ptr noundef [[LD:%.*]]) #[[ATTR0]] {
+// AARCH64-LE-NEXT:  [[ENTRY:.*:]]
+// AARCH64-LE-NEXT:    [[LD_ADDR:%.*]] = alloca ptr, align 8
+// AARCH64-LE-NEXT:    store ptr [[LD]], ptr [[LD_ADDR]], align 8
+// AARCH64-LE-NEXT:    [[TMP0:%.*]] = load ptr, ptr [[LD_ADDR]], align 8
+// AARCH64-LE-NEXT:    ret void
+//
+// AARCH64-BE-LABEL: define dso_local void @testPrimitiveLongDouble(
+// AARCH64-BE-SAME: ptr noundef [[LD:%.*]]) #[[ATTR0]] {
+// AARCH64-BE-NEXT:  [[ENTRY:.*:]]
+// AARCH64-BE-NEXT:    [[LD_ADDR:%.*]] = alloca ptr, align 8
+// AARCH64-BE-NEXT:    store ptr [[LD]], ptr [[LD_ADDR]], align 8
+// AARCH64-BE-NEXT:    [[TMP0:%.*]] = load ptr, ptr [[LD_ADDR]], align 8
+// AARCH64-BE-NEXT:    ret void
+//
+extern "C" void testPrimitiveLongDouble(long double *ld) {
   // padding [10, 15] on x86
   __builtin_clear_padding(ld);
 }
 
-// LINUX-LABEL: define dso_local void @_Z10testBitIntPDB97_(
+
+// LINUX-LABEL: define dso_local void @testBitInt(
 // LINUX-SAME: ptr noundef [[BI:%.*]]) #[[ATTR0]] {
 // LINUX-NEXT:  [[ENTRY:.*:]]
 // LINUX-NEXT:    [[BI_ADDR:%.*]] = alloca ptr, align 8
@@ -105,107 +216,281 @@ void testPrimitiveLongDouble(long double *ld) {
 // LINUX-NEXT:    store i8 0, ptr [[TMP6]], align 1
 // LINUX-NEXT:    ret void
 //
-// WINDOWS-LABEL: define dso_local void @"?testBitInt@@YAXPEAU?$_BitInt@$0GB@@__clang@@@Z"(ptr noundef %bi) #0 {
-// WINDOWS-NEXT: entry:
-// WINDOWS-NEXT:   %bi.addr = alloca ptr, align 8
-// WINDOWS-NEXT:   store ptr %bi, ptr %bi.addr, align 8
-// WINDOWS-NEXT:   %0 = load ptr, ptr %bi.addr, align 8
-// WINDOWS-NEXT:   %1 = getelementptr i8, ptr %0, i32 12
-// WINDOWS-NEXT:   %2 = load i8, ptr %1, align 4
-// WINDOWS-NEXT:   %3 = and i8 %2, 1
-// WINDOWS-NEXT:   store i8 %3, ptr %1, align 4
-// WINDOWS-NEXT:   %4 = getelementptr i8, ptr %0, i32 13
-// WINDOWS-NEXT:   store i8 0, ptr %4, align 1
-// WINDOWS-NEXT:   %5 = getelementptr i8, ptr %0, i32 14
-// WINDOWS-NEXT:   store i8 0, ptr %5, align 2
-// WINDOWS-NEXT:   %6 = getelementptr i8, ptr %0, i32 15
-// WINDOWS-NEXT:   store i8 0, ptr %6, align 1
-// WINDOWS-NEXT:   ret void
-//
-// ARM-LE-LABEL: define dso_local void @_Z10testBitIntPDB97_(
-// ARM-LE:      [[OBJ:%.*]] = load ptr, ptr {{.*}}, align {{[0-9]+}}
-// ARM-LE-NEXT: [[BYTE12:%.*]] = getelementptr i8, ptr [[OBJ]], i32 12
-// ARM-LE-NEXT: [[VALUE:%.*]] = load i8, ptr [[BYTE12]], align 4
-// ARM-LE-NEXT: [[MASKED:%.*]] = and i8 [[VALUE]], 1
-// ARM-LE-NEXT: store i8 [[MASKED]], ptr [[BYTE12]], align 4
-// ARM-LE-NEXT: [[BYTE13:%.*]] = getelementptr i8, ptr [[OBJ]], i32 13
-// ARM-LE-NEXT: store i8 0, ptr [[BYTE13]], align 1
-// ARM-LE-NEXT: [[BYTE14:%.*]] = getelementptr i8, ptr [[OBJ]], i32 14
-// ARM-LE-NEXT: store i8 0, ptr [[BYTE14]], align 2
-// ARM-LE-NEXT: [[BYTE15:%.*]] = getelementptr i8, ptr [[OBJ]], i32 15
-// ARM-LE-NEXT: store i8 0, ptr [[BYTE15]], align 1
-// ARM-LE-NEXT: ret void
-//
-// ARM-BE-LABEL: define dso_local void @_Z10testBitIntPDB97_(
-// ARM-BE:      [[OBJ:%.*]] = load ptr, ptr {{.*}}, align {{[0-9]+}}
-// ARM-BE-NEXT: [[BYTE0:%.*]] = getelementptr i8, ptr [[OBJ]], i32 0
-// ARM-BE-NEXT: store i8 0, ptr [[BYTE0]], align {{[0-9]+}}
-// ARM-BE-NEXT: [[BYTE1:%.*]] = getelementptr i8, ptr [[OBJ]], i32 1
-// ARM-BE-NEXT: store i8 0, ptr [[BYTE1]], align 1
-// ARM-BE-NEXT: [[BYTE2:%.*]] = getelementptr i8, ptr [[OBJ]], i32 2
-// ARM-BE-NEXT: store i8 0, ptr [[BYTE2]], align 2
-// ARM-BE-NEXT: [[BYTE3:%.*]] = getelementptr i8, ptr [[OBJ]], i32 3
-// ARM-BE-NEXT: [[VALUE:%.*]] = load i8, ptr [[BYTE3]], align 1
-// ARM-BE-NEXT: [[MASKED:%.*]] = and i8 [[VALUE]], 1
-// ARM-BE-NEXT: store i8 [[MASKED]], ptr [[BYTE3]], align 1
-// ARM-BE-NEXT: ret void
-//
-void testBitInt(_BitInt(97) *bi) {
+// WINDOWS-LABEL: define dso_local void @testBitInt(
+// WINDOWS-SAME: ptr noundef [[BI:%.*]]) #[[ATTR0]] {
+// WINDOWS-NEXT:  [[ENTRY:.*:]]
+// WINDOWS-NEXT:    [[BI_ADDR:%.*]] = alloca ptr, align 8
+// WINDOWS-NEXT:    store ptr [[BI]], ptr [[BI_ADDR]], align 8
+// WINDOWS-NEXT:    [[TMP0:%.*]] = load ptr, ptr [[BI_ADDR]], align 8
+// WINDOWS-NEXT:    [[TMP1:%.*]] = getelementptr i8, ptr [[TMP0]], i32 12
+// WINDOWS-NEXT:    [[TMP2:%.*]] = load i8, ptr [[TMP1]], align 4
+// WINDOWS-NEXT:    [[TMP3:%.*]] = and i8 [[TMP2]], 1
+// WINDOWS-NEXT:    store i8 [[TMP3]], ptr [[TMP1]], align 4
+// WINDOWS-NEXT:    [[TMP4:%.*]] = getelementptr i8, ptr [[TMP0]], i32 13
+// WINDOWS-NEXT:    store i8 0, ptr [[TMP4]], align 1
+// WINDOWS-NEXT:    [[TMP5:%.*]] = getelementptr i8, ptr [[TMP0]], i32 14
+// WINDOWS-NEXT:    store i8 0, ptr [[TMP5]], align 2
+// WINDOWS-NEXT:    [[TMP6:%.*]] = getelementptr i8, ptr [[TMP0]], i32 15
+// WINDOWS-NEXT:    store i8 0, ptr [[TMP6]], align 1
+// WINDOWS-NEXT:    ret void
+//
+// ARM-LE-LABEL: define dso_local void @testBitInt(
+// ARM-LE-SAME: ptr noundef [[BI:%.*]]) #[[ATTR0]] {
+// ARM-LE-NEXT:  [[ENTRY:.*:]]
+// ARM-LE-NEXT:    [[BI_ADDR:%.*]] = alloca ptr, align 4
+// ARM-LE-NEXT:    store ptr [[BI]], ptr [[BI_ADDR]], align 4
+// ARM-LE-NEXT:    [[TMP0:%.*]] = load ptr, ptr [[BI_ADDR]], align 4
+// ARM-LE-NEXT:    [[TMP1:%.*]] = getelementptr i8, ptr [[TMP0]], i32 12
+// ARM-LE-NEXT:    [[TMP2:%.*]] = load i8, ptr [[TMP1]], align 4
+// ARM-LE-NEXT:    [[TMP3:%.*]] = and i8 [[TMP2]], 1
+// ARM-LE-NEXT:    store i8 [[TMP3]], ptr [[TMP1]], align 4
+// ARM-LE-NEXT:    [[TMP4:%.*]] = getelementptr i8, ptr [[TMP0]], i32 13
+// ARM-LE-NEXT:    store i8 0, ptr [[TMP4]], align 1
+// ARM-LE-NEXT:    [[TMP5:%.*]] = getelementptr i8, ptr [[TMP0]], i32 14
+// ARM-LE-NEXT:    store i8 0, ptr [[TMP5]], align 2
+// ARM-LE-NEXT:    [[TMP6:%.*]] = getelementptr i8, ptr [[TMP0]], i32 15
+// ARM-LE-NEXT:    store i8 0, ptr [[TMP6]], align 1
+// ARM-LE-NEXT:    ret void
+//
+// ARM-BE-LABEL: define dso_local void @testBitInt(
+// ARM-BE-SAME: ptr noundef [[BI:%.*]]) #[[ATTR0]] {
+// ARM-BE-NEXT:  [[ENTRY:.*:]]
+// ARM-BE-NEXT:    [[BI_ADDR:%.*]] = alloca ptr, align 4
+// ARM-BE-NEXT:    store ptr [[BI]], ptr [[BI_ADDR]], align 4
+// ARM-BE-NEXT:    [[TMP0:%.*]] = load ptr, ptr [[BI_ADDR]], align 4
+// ARM-BE-NEXT:    [[TMP1:%.*]] = getelementptr i8, ptr [[TMP0]], i32 0
+// ARM-BE-NEXT:    store i8 0, ptr [[TMP1]], align 8
+// ARM-BE-NEXT:    [[TMP2:%.*]] = getelementptr i8, ptr [[TMP0]], i32 1
+// ARM-BE-NEXT:    store i8 0, ptr [[TMP2]], align 1
+// ARM-BE-NEXT:    [[TMP3:%.*]] = getelementptr i8, ptr [[TMP0]], i32 2
+// ARM-BE-NEXT:    store i8 0, ptr [[TMP3]], align 2
+// ARM-BE-NEXT:    [[TMP4:%.*]] = getelementptr i8, ptr [[TMP0]], i32 3
+// ARM-BE-NEXT:    [[TMP5:%.*]] = load i8, ptr [[TMP4]], align 1
+// ARM-BE-NEXT:    [[TMP6:%.*]] = and i8 [[TMP5]], 1
+// ARM-BE-NEXT:    store i8 [[TMP6]], ptr [[TMP4]], align 1
+// ARM-BE-NEXT:    ret void
+//
+// AARCH64-LE-LABEL: define dso_local void @testBitInt(
+// AARCH64-LE-SAME: ptr noundef [[BI:%.*]]) #[[ATTR0]] {
+// AARCH64-LE-NEXT:  [[ENTRY:.*:]]
+// AARCH64-LE-NEXT:    [[BI_ADDR:%.*]] = alloca ptr, align 8
+// AARCH64-LE-NEXT:    store ptr [[BI]], ptr [[BI_ADDR]], align 8
+// AARCH64-LE-NEXT:    [[TMP0:%.*]] = load ptr, ptr [[BI_ADDR]], align 8
+// AARCH64-LE-NEXT:    [[TMP1:%.*]] = getelementptr i8, ptr [[TMP0]], i32 12
+// AARCH64-LE-NEXT:    [[TMP2:%.*]] = load i8, ptr [[TMP1]], align 4
+// AARCH64-LE-NEXT:    [[TMP3:%.*]] = and i8 [[TMP2]], 1
+// AARCH64-LE-NEXT:    store i8 [[TMP3]], ptr [[TMP1]], align 4
+// AARCH64-LE-NEXT:    [[TMP4:%.*]] = getelementptr i8, ptr [[TMP0]], i32 13
+// AARCH64-LE-NEXT:    store i8 0, ptr [[TMP4]], align 1
+// AARCH64-LE-NEXT:    [[TMP5:%.*]] = getelementptr i8, ptr [[TMP0]], i32 14
+// AARCH64-LE-NEXT:    store i8 0, ptr [[TMP5]], align 2
+// AARCH64-LE-NEXT:    [[TMP6:%.*]] = getelementptr i8, ptr [[TMP0]], i32 15
+// AARCH64-LE-NEXT:    store i8 0, ptr [[TMP6]], align 1
+// AARCH64-LE-NEXT:    ret void
+//
+// AARCH64-BE-LABEL: define dso_local void @testBitInt(
+// AARCH64-BE-SAME: ptr noundef [[BI:%.*]]) #[[ATTR0]] {
+// AARCH64-BE-NEXT:  [[ENTRY:.*:]]
+// AARCH64-BE-NEXT:    [[BI_ADDR:%.*]] = alloca ptr, align 8
+// AARCH64-BE-NEXT:    store ptr [[BI]], ptr [[BI_ADDR]], align 8
+// AARCH64-BE-NEXT:    [[TMP0:%.*]] = load ptr, ptr [[BI_ADDR]], align 8
+// AARCH64-BE-NEXT:    [[TMP1:%.*]] = getelementptr i8, ptr [[TMP0]], i32 0
+// AARCH64-BE-NEXT:    store i8 0, ptr [[TMP1]], align 16
+// AARCH64-BE-NEXT:    [[TMP2:%.*]] = getelementptr i8, ptr [[TMP0]], i32 1
+// AARCH64-BE-NEXT:    store i8 0, ptr [[TMP2]], align 1
+// AARCH64-BE-NEXT:    [[TMP3:%.*]] = getelementptr i8, ptr [[TMP0]], i32 2
+// AARCH64-BE-NEXT:    store i8 0, ptr [[TMP3]], align 2
+// AARCH64-BE-NEXT:    [[TMP4:%.*]] = getelementptr i8, ptr [[TMP0]], i32 3
+// AARCH64-BE-NEXT:    [[TMP5:%.*]] = load i8, ptr [[TMP4]], align 1
+// AARCH64-BE-NEXT:    [[TMP6:%.*]] = and i8 [[TMP5]], 1
+// AARCH64-BE-NEXT:    store i8 [[TMP6]], ptr [[TMP4]], align 1
+// AARCH64-BE-NEXT:    ret void
+//
+extern "C" void testBitInt(_BitInt(97) *bi) {
   // Storage is widened to 128 bits; clear bits [97, 128).
   __builtin_clear_padding(bi);
 }
 
-// ARM-LE-LABEL: define dso_local void @_Z21testBitIntByteAlignedPDB24_(
-// ARM-LE:      [[OBJ:%.*]] = load ptr, ptr {{.*}}, align {{[0-9]+}}
-// ARM-LE-NEXT: [[BYTE3:%.*]] = getelementptr i8, ptr [[OBJ]], i32 3
-// ARM-LE-NEXT: store i8 0, ptr [[BYTE3]], align 1
-// ARM-LE-NEXT: ret void
+
+// LINUX-LABEL: define dso_local void @testBitIntByteAligned(
+// LINUX-SAME: ptr noundef [[BI:%.*]]) #[[ATTR0]] {
+// LINUX-NEXT:  [[ENTRY:.*:]]
+// LINUX-NEXT:    [[BI_ADDR:%.*]] = alloca ptr, align 8
+// LINUX-NEXT:    store ptr [[BI]], ptr [[BI_ADDR]], align 8
+// LINUX-NEXT:    [[TMP0:%.*]] = load ptr, ptr [[BI_ADDR]], align 8
+// LINUX-NEXT:    [[TMP1:%.*]] = getelementptr i8, ptr [[TMP0]], i32 3
+// LINUX-NEXT:    store i8 0, ptr [[TMP1]], align 1
+// LINUX-NEXT:    ret void
+//
+// WINDOWS-LABEL: define dso_local void @testBitIntByteAligned(
+// WINDOWS-SAME: ptr noundef [[BI:%.*]]) #[[ATTR0]] {
+// WINDOWS-NEXT:  [[ENTRY:.*:]]
+// WINDOWS-NEXT:    [[BI_ADDR:%.*]] = alloca ptr, align 8
+// WINDOWS-NEXT:    store ptr [[BI]], ptr [[BI_ADDR]], align 8
+// WINDOWS-NEXT:    [[TMP0:%.*]] = load ptr, ptr [[BI_ADDR]], align 8
+// WINDOWS-NEXT:    [[TMP1:%.*]] = getelementptr i8, ptr [[TMP0]], i32 3
+// WINDOWS-NEXT:    store i8 0, ptr [[TMP1]], align 1
+// WINDOWS-NEXT:    ret void
+//
+// ARM-LE-LABEL: define dso_local void @testBitIntByteAligned(
+// ARM-LE-SAME: ptr noundef [[BI:%.*]]) #[[ATTR0]] {
+// ARM-LE-NEXT:  [[ENTRY:.*:]]
+// ARM-LE-NEXT:    [[BI_ADDR:%.*]] = alloca ptr, align 4
+// ARM-LE-NEXT:    store ptr [[BI]], ptr [[BI_ADDR]], align 4
+// ARM-LE-NEXT:    [[TMP0:%.*]] = load ptr, ptr [[BI_ADDR]], align 4
+// ARM-LE-NEXT:    [[TMP1:%.*]] = getelementptr i8, ptr [[TMP0]], i32 3
+// ARM-LE-NEXT:    store i8 0, ptr [[TMP1]], align 1
+// ARM-LE-NEXT:    ret void
+//
+// ARM-BE-LABEL: define dso_local void @testBitIntByteAligned(
+// ARM-BE-SAME: ptr noundef [[BI:%.*]]) #[[ATTR0]] {
+// ARM-BE-NEXT:  [[ENTRY:.*:]]
+// ARM-BE-NEXT:    [[BI_ADDR:%.*]] = alloca ptr, align 4
+// ARM-BE-NEXT:    store ptr [[BI]], ptr [[BI_ADDR]], align 4
+// ARM-BE-NEXT:    [[TMP0:%.*]] = load ptr, ptr [[BI_ADDR]], align 4
+// ARM-BE-NEXT:    [[TMP1:%.*]] = getelementptr i8, ptr [[TMP0]], i32 0
+// ARM-BE-NEXT:    store i8 0, ptr [[TMP1]], align 4
+// ARM-BE-NEXT:    ret void
+//
+// AARCH64-LE-LABEL: define dso_local void @testBitIntByteAligned(
+// AARCH64-LE-SAME: ptr noundef [[BI:%.*]]) #[[ATTR0]] {
+// AARCH64-LE-NEXT:  [[ENTRY:.*:]]
+// AARCH64-LE-NEXT:    [[BI_ADDR:%.*]] = alloca ptr, align 8
+// AARCH64-LE-NEXT:    store ptr [[BI]], ptr [[BI_ADDR]], align 8
+// AARCH64-LE-NEXT:    [[TMP0:%.*]] = load ptr, ptr [[BI_ADDR]], align 8
+// AARCH64-LE-NEXT:    [[TMP1:%.*]] = getelementptr i8, ptr [[TMP0]], i32 3
+// AARCH64-LE-NEXT:    store i8 0, ptr [[TMP1]], align 1
+// AARCH64-LE-NEXT:    ret void
 //
-// ARM-BE-LABEL: define dso_local void @_Z21testBitIntByteAlignedPDB24_(
-// ARM-BE:      [[OBJ:%.*]] = load ptr, ptr {{.*}}, align {{[0-9]+}}
-// ARM-BE-NEXT: [[BYTE0:%.*]] = getelementptr i8, ptr [[OBJ]], i32 0
-// ARM-BE-NEXT: store i8 0, ptr [[BYTE0]], align 4
-// ARM-BE-NEXT: ret void
+// AARCH64-BE-LABEL: define dso_local void @testBitIntByteAligned(
+// AARCH64-BE-SAME: ptr noundef [[BI:%.*]]) #[[ATTR0]] {
+// AARCH64-BE-NEXT:  [[ENTRY:.*:]]
+// AARCH64-BE-NEXT:    [[BI_ADDR:%.*]] = alloca ptr, align 8
+// AARCH64-BE-NEXT:    store ptr [[BI]], ptr [[BI_ADDR]], align 8
+// AARCH64-BE-NEXT:    [[TMP0:%.*]] = load ptr, ptr [[BI_ADDR]], align 8
+// AARCH64-BE-NEXT:    [[TMP1:%.*]] = getelementptr i8, ptr [[TMP0]], i32 0
+// AARCH64-BE-NEXT:    store i8 0, ptr [[TMP1]], align 4
+// AARCH64-BE-NEXT:    ret void
 //
-void testBitIntByteAligned(_BitInt(24) *bi) {
+extern "C" void testBitIntByteAligned(_BitInt(24) *bi) {
   // Storage is widened to 32 bits; clear the leading or trailing byte.
   __builtin_clear_padding(bi);
 }
 
-// ARM-LE-LABEL: define dso_local void @_Z18testVolatileBitIntPVDB97_(
-// ARM-LE:      [[OBJ:%.*]] = load ptr, ptr {{.*}}, align {{[0-9]+}}
-// ARM-LE-NEXT: [[BYTE12:%.*]] = getelementptr i8, ptr [[OBJ]], i32 12
-// ARM-LE-NEXT: [[VALUE:%.*]] = load i8, ptr [[BYTE12]], align 4
-// ARM-LE-NEXT: [[MASKED:%.*]] = and i8 [[VALUE]], 1
-// ARM-LE-NEXT: store i8 [[MASKED]], ptr [[BYTE12]], align 4
-// ARM-LE-NEXT: [[BYTE13:%.*]] = getelementptr i8, ptr [[OBJ]], i32 13
-// ARM-LE-NEXT: store i8 0, ptr [[BYTE13]], align 1
-// ARM-LE-NEXT: [[BYTE14:%.*]] = getelementptr i8, ptr [[OBJ]], i32 14
-// ARM-LE-NEXT: store i8 0, ptr [[BYTE14]], align 2
-// ARM-LE-NEXT: [[BYTE15:%.*]] = getelementptr i8, ptr [[OBJ]], i32 15
-// ARM-LE-NEXT: store i8 0, ptr [[BYTE15]], align 1
-// ARM-LE-NEXT: ret void
-//
-// ARM-BE-LABEL: define dso_local void @_Z18testVolatileBitIntPVDB97_(
-// ARM-BE:      [[OBJ:%.*]] = load ptr, ptr {{.*}}, align {{[0-9]+}}
-// ARM-BE-NEXT: [[BYTE0:%.*]] = getelementptr i8, ptr [[OBJ]], i32 0
-// ARM-BE-NEXT: store i8 0, ptr [[BYTE0]], align {{[0-9]+}}
-// ARM-BE-NEXT: [[BYTE1:%.*]] = getelementptr i8, ptr [[OBJ]], i32 1
-// ARM-BE-NEXT: store i8 0, ptr [[BYTE1]], align 1
-// ARM-BE-NEXT: [[BYTE2:%.*]] = getelementptr i8, ptr [[OBJ]], i32 2
-// ARM-BE-NEXT: store i8 0, ptr [[BYTE2]], align 2
-// ARM-BE-NEXT: [[BYTE3:%.*]] = getelementptr i8, ptr [[OBJ]], i32 3
-// ARM-BE-NEXT: [[VALUE:%.*]] = load i8, ptr [[BYTE3]], align 1
-// ARM-BE-NEXT: [[MASKED:%.*]] = and i8 [[VALUE]], 1
-// ARM-BE-NEXT: store i8 [[MASKED]], ptr [[BYTE3]], align 1
-// ARM-BE-NEXT: ret void
-//
-void testVolatileBitInt(volatile _BitInt(97) *bi) {
+// LINUX-LABEL: define dso_local void @testVolatileBitInt(
+// LINUX-SAME: ptr noundef [[BI:%.*]]) #[[ATTR0]] {
+// LINUX-NEXT:  [[ENTRY:.*:]]
+// LINUX-NEXT:    [[BI_ADDR:%.*]] = alloca ptr, align 8
+// LINUX-NEXT:    store ptr [[BI]], ptr [[BI_ADDR]], align 8
+// LINUX-NEXT:    [[TMP0:%.*]] = load ptr, ptr [[BI_ADDR]], align 8
+// LINUX-NEXT:    [[TMP1:%.*]] = getelementptr i8, ptr [[TMP0]], i32 12
+// LINUX-NEXT:    [[TMP2:%.*]] = load i8, ptr [[TMP1]], align 4
+// LINUX-NEXT:    [[TMP3:%.*]] = and i8 [[TMP2]], 1
+// LINUX-NEXT:    store i8 [[TMP3]], ptr [[TMP1]], align 4
+// LINUX-NEXT:    [[TMP4:%.*]] = getelementptr i8, ptr [[TMP0]], i32 13
+// LINUX-NEXT:    store i8 0, ptr [[TMP4]], align 1
+// LINUX-NEXT:    [[TMP5:%.*]] = getelementptr i8, ptr [[TMP0]], i32 14
+// LINUX-NEXT:    store i8 0, ptr [[TMP5]], align 2
+// LINUX-NEXT:    [[TMP6:%.*]] = getelementptr i8, ptr [[TMP0]], i32 15
+// LINUX-NEXT:    store i8 0, ptr [[TMP6]], align 1
+// LINUX-NEXT:    ret void
+//
+// WINDOWS-LABEL: define dso_local void @testVolatileBitInt(
+// WINDOWS-SAME: ptr noundef [[BI:%.*]]) #[[ATTR0]] {
+// WINDOWS-NEXT:  [[ENTRY:.*:]]
+// WINDOWS-NEXT:    [[BI_ADDR:%.*]] = alloca ptr, align 8
+// WINDOWS-NEXT:    store ptr [[BI]], ptr [[BI_ADDR]], align 8
+// WINDOWS-NEXT:    [[TMP0:%.*]] = load ptr, ptr [[BI_ADDR]], align 8
+// WINDOWS-NEXT:    [[TMP1:%.*]] = getelementptr i8, ptr [[TMP0]], i32 12
+// WINDOWS-NEXT:    [[TMP2:%.*]] = load i8, ptr [[TMP1]], align 4
+// WINDOWS-NEXT:    [[TMP3:%.*]] = and i8 [[TMP2]], 1
+// WINDOWS-NEXT:    store i8 [[TMP3]], ptr [[TMP1]], align 4
+// WINDOWS-NEXT:    [[TMP4:%.*]] = getelementptr i8, ptr [[TMP0]], i32 13
+// WINDOWS-NEXT:    store i8 0, ptr [[TMP4]], align 1
+// WINDOWS-NEXT:    [[TMP5:%.*]] = getelementptr i8, ptr [[TMP0]], i32 14
+// WINDOWS-NEXT:    store i8 0, ptr [[TMP5]], align 2
+// WINDOWS-NEXT:    [[TMP6:%.*]] = getelementptr i8, ptr [[TMP0]], i32 15
+// WINDOWS-NEXT:    store i8 0, ptr [[TMP6]], align 1
+// WINDOWS-NEXT:    ret void
+//
+// ARM-LE-LABEL: define dso_local void @testVolatileBitInt(
+// ARM-LE-SAME: ptr noundef [[BI:%.*]]) #[[ATTR0]] {
+// ARM-LE-NEXT:  [[ENTRY:.*:]]
+// ARM-LE-NEXT:    [[BI_ADDR:%.*]] = alloca ptr, align 4
+// ARM-LE-NEXT:    store ptr [[BI]], ptr [[BI_ADDR]], align 4
+// ARM-LE-NEXT:    [[TMP0:%.*]] = load ptr, ptr [[BI_ADDR]], align 4
+// ARM-LE-NEXT:    [[TMP1:%.*]] = getelementptr i8, ptr [[TMP0]], i32 12
+// ARM-LE-NEXT:    [[TMP2:%.*]] = load i8, ptr [[TMP1]], align 4
+// ARM-LE-NEXT:    [[TMP3:%.*]] = and i8 [[TMP2]], 1
+// ARM-LE-NEXT:    store i8 [[TMP3]], ptr [[TMP1]], align 4
+// ARM-LE-NEXT:    [[TMP4:%.*]] = getelementptr i8, ptr [[TMP0]], i32 13
+// ARM-LE-NEXT:    store i8 0, ptr [[TMP4]], align 1
+// ARM-LE-NEXT:    [[TMP5:%.*]] = getelementptr i8, ptr [[TMP0]], i32 14
+// ARM-LE-NEXT:    store i8 0, ptr [[TMP5]], align 2
+// ARM-LE-NEXT:    [[TMP6:%.*]] = getelementptr i8, ptr [[TMP0]], i32 15
+// ARM-LE-NEXT:    store i8 0, ptr [[TMP6]], align 1
+// ARM-LE-NEXT:    ret void
+//
+// ARM-BE-LABEL: define dso_local void @testVolatileBitInt(
+// ARM-BE-SAME: ptr noundef [[BI:%.*]]) #[[ATTR0]] {
+// ARM-BE-NEXT:  [[ENTRY:.*:]]
+// ARM-BE-NEXT:    [[BI_ADDR:%.*]] = alloca ptr, align 4
+// ARM-BE-NEXT:    store ptr [[BI]], ptr [[BI_ADDR]], align 4
+// ARM-BE-NEXT:    [[TMP0:%.*]] = load ptr, ptr [[BI_ADDR]], align 4
+// ARM-BE-NEXT:    [[TMP1:%.*]] = getelementptr i8, ptr [[TMP0]], i32 0
+// ARM-BE-NEXT:    store i8 0, ptr [[TMP1]], align 8
+// ARM-BE-NEXT:    [[TMP2:%.*]] = getelementptr i8, ptr [[TMP0]], i32 1
+// ARM-BE-NEXT:    store i8 0, ptr [[TMP2]], align 1
+// ARM-BE-NEXT:    [[TMP3:%.*]] = getelementptr i8, ptr [[TMP0]], i32 2
+// ARM-BE-NEXT:    store i8 0, ptr [[TMP3]], align 2
+// ARM-BE-NEXT:    [[TMP4:%.*]] = getelementptr i8, ptr [[TMP0]], i32 3
+// ARM-BE-NEXT:    [[TMP5:%.*]] = load i8, ptr [[TMP4]], align 1
+// ARM-BE-NEXT:    [[TMP6:%.*]] = and i8 [[TMP5]], 1
+// ARM-BE-NEXT:    store i8 [[TMP6]], ptr [[TMP4]], align 1
+// ARM-BE-NEXT:    ret void
+//
+// AARCH64-LE-LABEL: define dso_local void @testVolatileBitInt(
+// AARCH64-LE-SAME: ptr noundef [[BI:%.*]]) #[[ATTR0]] {
+// AARCH64-LE-NEXT:  [[ENTRY:.*:]]
+// AARCH64-LE-NEXT:    [[BI_ADDR:%.*]] = alloca ptr, align 8
+// AARCH64-LE-NEXT:    store ptr [[BI]], ptr [[BI_ADDR]], align 8
+// AARCH64-LE-NEXT:    [[TMP0:%.*]] = load ptr, ptr [[BI_ADDR]], align 8
+// AARCH64-LE-NEXT:    [[TMP1:%.*]] = getelementptr i8, ptr [[TMP0]], i32 12
+// AARCH64-LE-NEXT:    [[TMP2:%.*]] = load i8, ptr [[TMP1]], align 4
+// AARCH64-LE-NEXT:    [[TMP3:%.*]] = and i8 [[TMP2]], 1
+// AARCH64-LE-NEXT:    store i8 [[TMP3]], ptr [[TMP1]], align 4
+// AARCH64-LE-NEXT:    [[TMP4:%.*]] = getelementptr i8, ptr [[TMP0]], i32 13
+// AARCH64-LE-NEXT:    store i8 0, ptr [[TMP4]], align 1
+// AARCH64-LE-NEXT:    [[TMP5:%.*]] = getelementptr i8, ptr [[TMP0]], i32 14
+// AARCH64-LE-NEXT:    store i8 0, ptr [[TMP5]], align 2
+// AARCH64-LE-NEXT:    [[TMP6:%.*]] = getelementptr i8, ptr [[TMP0]], i32 15
+// AARCH64-LE-NEXT:    store i8 0, ptr [[TMP6]], align 1
+// AARCH64-LE-NEXT:    ret void
+//
+// AARCH64-BE-LABEL: define dso_local void @testVolatileBitInt(
+// AARCH64-BE-SAME: ptr noundef [[BI:%.*]]) #[[ATTR0]] {
+// AARCH64-BE-NEXT:  [[ENTRY:.*:]]
+// AARCH64-BE-NEXT:    [[BI_ADDR:%.*]] = alloca ptr, align 8
+// AARCH64-BE-NEXT:    store ptr [[BI]], ptr [[BI_ADDR]], align 8
+// AARCH64-BE-NEXT:    [[TMP0:%.*]] = load ptr, ptr [[BI_ADDR]], align 8
+// AARCH64-BE-NEXT:    [[TMP1:%.*]] = getelementptr i8, ptr [[TMP0]], i32 0
+// AARCH64-BE-NEXT:    store i8 0, ptr [[TMP1]], align 16
+// AARCH64-BE-NEXT:    [[TMP2:%.*]] = getelementptr i8, ptr [[TMP0]], i32 1
+// AARCH64-BE-NEXT:    store i8 0, ptr [[TMP2]], align 1
+// AARCH64-BE-NEXT:    [[TMP3:%.*]] = getelementptr i8, ptr [[TMP0]], i32 2
+// AARCH64-BE-NEXT:    store i8 0, ptr [[TMP3]], align 2
+// AARCH64-BE-NEXT:    [[TMP4:%.*]] = getelementptr i8, ptr [[TMP0]], i32 3
+// AARCH64-BE-NEXT:    [[TMP5:%.*]] = load i8, ptr [[TMP4]], align 1
+// AARCH64-BE-NEXT:    [[TMP6:%.*]] = and i8 [[TMP5]], 1
+// AARCH64-BE-NEXT:    store i8 [[TMP6]], ptr [[TMP4]], align 1
+// AARCH64-BE-NEXT:    ret void
+//
+extern "C" void testVolatileBitInt(volatile _BitInt(97) *bi) {
   __builtin_clear_padding(bi);
 }
 
 
-// LINUX-LABEL: define dso_local void @_Z30testPrimitiveComplexLongDoublePCe(
+// LINUX-LABEL: define dso_local void @testPrimitiveComplexLongDouble(
 // LINUX-SAME: ptr noundef [[C:%.*]]) #[[ATTR0]] {
 // LINUX-NEXT:  [[ENTRY:.*:]]
 // LINUX-NEXT:    [[C_ADDR:%.*]] = alloca ptr, align 8
@@ -237,24 +522,58 @@ void testVolatileBitInt(volatile _BitInt(97) *bi) {
 // LINUX-NEXT:    store i8 0, ptr [[TMP12]], align 1
 // LINUX-NEXT:    ret void
 //
-// WINDOWS-LABEL: define dso_local void @"?testPrimitiveComplexLongDouble@@YAXPEAU?$_Complex at O@__clang@@@Z"(ptr noundef %c) #0 {
-// WINDOWS-NEXT: entry:
-// WINDOWS-NEXT:   %c.addr = alloca ptr, align 8
-// WINDOWS-NEXT:   store ptr %c, ptr %c.addr, align 8
-// WINDOWS-NEXT:   %0 = load ptr, ptr %c.addr, align 8
-// WINDOWS-NEXT:   ret void
+// WINDOWS-LABEL: define dso_local void @testPrimitiveComplexLongDouble(
+// WINDOWS-SAME: ptr noundef [[C:%.*]]) #[[ATTR0]] {
+// WINDOWS-NEXT:  [[ENTRY:.*:]]
+// WINDOWS-NEXT:    [[C_ADDR:%.*]] = alloca ptr, align 8
+// WINDOWS-NEXT:    store ptr [[C]], ptr [[C_ADDR]], align 8
+// WINDOWS-NEXT:    [[TMP0:%.*]] = load ptr, ptr [[C_ADDR]], align 8
+// WINDOWS-NEXT:    ret void
+//
+// ARM-LE-LABEL: define dso_local void @testPrimitiveComplexLongDouble(
+// ARM-LE-SAME: ptr noundef [[C:%.*]]) #[[ATTR0]] {
+// ARM-LE-NEXT:  [[ENTRY:.*:]]
+// ARM-LE-NEXT:    [[C_ADDR:%.*]] = alloca ptr, align 4
+// ARM-LE-NEXT:    store ptr [[C]], ptr [[C_ADDR]], align 4
+// ARM-LE-NEXT:    [[TMP0:%.*]] = load ptr, ptr [[C_ADDR]], align 4
+// ARM-LE-NEXT:    ret void
+//
+// ARM-BE-LABEL: define dso_local void @testPrimitiveComplexLongDouble(
+// ARM-BE-SAME: ptr noundef [[C:%.*]]) #[[ATTR0]] {
+// ARM-BE-NEXT:  [[ENTRY:.*:]]
+// ARM-BE-NEXT:    [[C_ADDR:%.*]] = alloca ptr, align 4
+// ARM-BE-NEXT:    store ptr [[C]], ptr [[C_ADDR]], align 4
+// ARM-BE-NEXT:    [[TMP0:%.*]] = load ptr, ptr [[C_ADDR]], align 4
+// ARM-BE-NEXT:    ret void
 //
-void testPrimitiveComplexLongDouble(_Complex long double *c) {
+// AARCH64-LE-LABEL: define dso_local void @testPrimitiveComplexLongDouble(
+// AARCH64-LE-SAME: ptr noundef [[C:%.*]]) #[[ATTR0]] {
+// AARCH64-LE-NEXT:  [[ENTRY:.*:]]
+// AARCH64-LE-NEXT:    [[C_ADDR:%.*]] = alloca ptr, align 8
+// AARCH64-LE-NEXT:    store ptr [[C]], ptr [[C_ADDR]], align 8
+// AARCH64-LE-NEXT:    [[TMP0:%.*]] = load ptr, ptr [[C_ADDR]], align 8
+// AARCH64-LE-NEXT:    ret void
+//
+// AARCH64-BE-LABEL: define dso_local void @testPrimitiveComplexLongDouble(
+// AARCH64-BE-SAME: ptr noundef [[C:%.*]]) #[[ATTR0]] {
+// AARCH64-BE-NEXT:  [[ENTRY:.*:]]
+// AARCH64-BE-NEXT:    [[C_ADDR:%.*]] = alloca ptr, align 8
+// AARCH64-BE-NEXT:    store ptr [[C]], ptr [[C_ADDR]], align 8
+// AARCH64-BE-NEXT:    [[TMP0:%.*]] = load ptr, ptr [[C_ADDR]], align 8
+// AARCH64-BE-NEXT:    ret void
+//
+extern "C" void testPrimitiveComplexLongDouble(_Complex long double *c) {
   // padding [10, 15] and [26, 31] on x86
   __builtin_clear_padding(c);
 }
 
+
 union U1 {
   int i;
   char c;
 };
 
-// LINUX-LABEL: define dso_local void @_Z24testUnionDifferentLengthP2U1(
+// LINUX-LABEL: define dso_local void @testUnionDifferentLength(
 // LINUX-SAME: ptr noundef [[U:%.*]]) #[[ATTR0]] {
 // LINUX-NEXT:  [[ENTRY:.*:]]
 // LINUX-NEXT:    [[U_ADDR:%.*]] = alloca ptr, align 8
@@ -262,19 +581,53 @@ union U1 {
 // LINUX-NEXT:    [[TMP0:%.*]] = load ptr, ptr [[U_ADDR]], align 8
 // LINUX-NEXT:    ret void
 //
-// WINDOWS-LABEL: define dso_local void @"?testUnionDifferentLength@@YAXPEATU1@@@Z"(ptr noundef %u) #0 {
-// WINDOWS-NEXT: entry:
-// WINDOWS-NEXT:   %u.addr = alloca ptr, align 8
-// WINDOWS-NEXT:   store ptr %u, ptr %u.addr, align 8
-// WINDOWS-NEXT:   %0 = load ptr, ptr %u.addr, align 8
-// WINDOWS-NEXT:   ret void
+// WINDOWS-LABEL: define dso_local void @testUnionDifferentLength(
+// WINDOWS-SAME: ptr noundef [[U:%.*]]) #[[ATTR0]] {
+// WINDOWS-NEXT:  [[ENTRY:.*:]]
+// WINDOWS-NEXT:    [[U_ADDR:%.*]] = alloca ptr, align 8
+// WINDOWS-NEXT:    store ptr [[U]], ptr [[U_ADDR]], align 8
+// WINDOWS-NEXT:    [[TMP0:%.*]] = load ptr, ptr [[U_ADDR]], align 8
+// WINDOWS-NEXT:    ret void
+//
+// ARM-LE-LABEL: define dso_local void @testUnionDifferentLength(
+// ARM-LE-SAME: ptr noundef [[U:%.*]]) #[[ATTR0]] {
+// ARM-LE-NEXT:  [[ENTRY:.*:]]
+// ARM-LE-NEXT:    [[U_ADDR:%.*]] = alloca ptr, align 4
+// ARM-LE-NEXT:    store ptr [[U]], ptr [[U_ADDR]], align 4
+// ARM-LE-NEXT:    [[TMP0:%.*]] = load ptr, ptr [[U_ADDR]], align 4
+// ARM-LE-NEXT:    ret void
 //
-void testUnionDifferentLength(U1 *u) {
+// ARM-BE-LABEL: define dso_local void @testUnionDifferentLength(
+// ARM-BE-SAME: ptr noundef [[U:%.*]]) #[[ATTR0]] {
+// ARM-BE-NEXT:  [[ENTRY:.*:]]
+// ARM-BE-NEXT:    [[U_ADDR:%.*]] = alloca ptr, align 4
+// ARM-BE-NEXT:    store ptr [[U]], ptr [[U_ADDR]], align 4
+// ARM-BE-NEXT:    [[TMP0:%.*]] = load ptr, ptr [[U_ADDR]], align 4
+// ARM-BE-NEXT:    ret void
+//
+// AARCH64-LE-LABEL: define dso_local void @testUnionDifferentLength(
+// AARCH64-LE-SAME: ptr noundef [[U:%.*]]) #[[ATTR0]] {
+// AARCH64-LE-NEXT:  [[ENTRY:.*:]]
+// AARCH64-LE-NEXT:    [[U_ADDR:%.*]] = alloca ptr, align 8
+// AARCH64-LE-NEXT:    store ptr [[U]], ptr [[U_ADDR]], align 8
+// AARCH64-LE-NEXT:    [[TMP0:%.*]] = load ptr, ptr [[U_ADDR]], align 8
+// AARCH64-LE-NEXT:    ret void
+//
+// AARCH64-BE-LABEL: define dso_local void @testUnionDifferentLength(
+// AARCH64-BE-SAME: ptr noundef [[U:%.*]]) #[[ATTR0]] {
+// AARCH64-BE-NEXT:  [[ENTRY:.*:]]
+// AARCH64-BE-NEXT:    [[U_ADDR:%.*]] = alloca ptr, align 8
+// AARCH64-BE-NEXT:    store ptr [[U]], ptr [[U_ADDR]], align 8
+// AARCH64-BE-NEXT:    [[TMP0:%.*]] = load ptr, ptr [[U_ADDR]], align 8
+// AARCH64-BE-NEXT:    ret void
+//
+extern "C" void testUnionDifferentLength(U1 *u) {
   // This should not clear the object representation bits of the non-active member.
   __builtin_clear_padding(u);
 }
 
 
+
 struct S {
   alignas(8) char c1;
 };
@@ -284,7 +637,7 @@ union U2 {
   char c2;
 };
 
-// LINUX-LABEL: define dso_local void @_Z35testUnionTailPaddingOfLongestMemberP2U2(
+// LINUX-LABEL: define dso_local void @testUnionTailPaddingOfLongestMember(
 // LINUX-SAME: ptr noundef [[U:%.*]]) #[[ATTR0]] {
 // LINUX-NEXT:  [[ENTRY:.*:]]
 // LINUX-NEXT:    [[U_ADDR:%.*]] = alloca ptr, align 8
@@ -306,34 +659,124 @@ union U2 {
 // LINUX-NEXT:    store i8 0, ptr [[TMP7]], align 1
 // LINUX-NEXT:    ret void
 //
-// WINDOWS-LABEL: define dso_local void @"?testUnionTailPaddingOfLongestMember@@YAXPEATU2@@@Z"(ptr noundef %u) #0 {
-// WINDOWS-NEXT: entry:
-// WINDOWS-NEXT:   %u.addr = alloca ptr, align 8
-// WINDOWS-NEXT:   store ptr %u, ptr %u.addr, align 8
-// WINDOWS-NEXT:   %0 = load ptr, ptr %u.addr, align 8
-// WINDOWS-NEXT:   %1 = getelementptr i8, ptr %0, i32 1
-// WINDOWS-NEXT:   store i8 0, ptr %1, align 1
-// WINDOWS-NEXT:   %2 = getelementptr i8, ptr %0, i32 2
-// WINDOWS-NEXT:   store i8 0, ptr %2, align 2
-// WINDOWS-NEXT:   %3 = getelementptr i8, ptr %0, i32 3
-// WINDOWS-NEXT:   store i8 0, ptr %3, align 1
-// WINDOWS-NEXT:   %4 = getelementptr i8, ptr %0, i32 4
-// WINDOWS-NEXT:   store i8 0, ptr %4, align 4
-// WINDOWS-NEXT:   %5 = getelementptr i8, ptr %0, i32 5
-// WINDOWS-NEXT:   store i8 0, ptr %5, align 1
-// WINDOWS-NEXT:   %6 = getelementptr i8, ptr %0, i32 6
-// WINDOWS-NEXT:   store i8 0, ptr %6, align 2
-// WINDOWS-NEXT:   %7 = getelementptr i8, ptr %0, i32 7
-// WINDOWS-NEXT:   store i8 0, ptr %7, align 1
-// WINDOWS-NEXT:   ret void
-//
-void testUnionTailPaddingOfLongestMember(U2 *u) {
+// WINDOWS-LABEL: define dso_local void @testUnionTailPaddingOfLongestMember(
+// WINDOWS-SAME: ptr noundef [[U:%.*]]) #[[ATTR0]] {
+// WINDOWS-NEXT:  [[ENTRY:.*:]]
+// WINDOWS-NEXT:    [[U_ADDR:%.*]] = alloca ptr, align 8
+// WINDOWS-NEXT:    store ptr [[U]], ptr [[U_ADDR]], align 8
+// WINDOWS-NEXT:    [[TMP0:%.*]] = load ptr, ptr [[U_ADDR]], align 8
+// WINDOWS-NEXT:    [[TMP1:%.*]] = getelementptr i8, ptr [[TMP0]], i32 1
+// WINDOWS-NEXT:    store i8 0, ptr [[TMP1]], align 1
+// WINDOWS-NEXT:    [[TMP2:%.*]] = getelementptr i8, ptr [[TMP0]], i32 2
+// WINDOWS-NEXT:    store i8 0, ptr [[TMP2]], align 2
+// WINDOWS-NEXT:    [[TMP3:%.*]] = getelementptr i8, ptr [[TMP0]], i32 3
+// WINDOWS-NEXT:    store i8 0, ptr [[TMP3]], align 1
+// WINDOWS-NEXT:    [[TMP4:%.*]] = getelementptr i8, ptr [[TMP0]], i32 4
+// WINDOWS-NEXT:    store i8 0, ptr [[TMP4]], align 4
+// WINDOWS-NEXT:    [[TMP5:%.*]] = getelementptr i8, ptr [[TMP0]], i32 5
+// WINDOWS-NEXT:    store i8 0, ptr [[TMP5]], align 1
+// WINDOWS-NEXT:    [[TMP6:%.*]] = getelementptr i8, ptr [[TMP0]], i32 6
+// WINDOWS-NEXT:    store i8 0, ptr [[TMP6]], align 2
+// WINDOWS-NEXT:    [[TMP7:%.*]] = getelementptr i8, ptr [[TMP0]], i32 7
+// WINDOWS-NEXT:    store i8 0, ptr [[TMP7]], align 1
+// WINDOWS-NEXT:    ret void
+//
+// ARM-LE-LABEL: define dso_local void @testUnionTailPaddingOfLongestMember(
+// ARM-LE-SAME: ptr noundef [[U:%.*]]) #[[ATTR0]] {
+// ARM-LE-NEXT:  [[ENTRY:.*:]]
+// ARM-LE-NEXT:    [[U_ADDR:%.*]] = alloca ptr, align 4
+// ARM-LE-NEXT:    store ptr [[U]], ptr [[U_ADDR]], align 4
+// ARM-LE-NEXT:    [[TMP0:%.*]] = load ptr, ptr [[U_ADDR]], align 4
+// ARM-LE-NEXT:    [[TMP1:%.*]] = getelementptr i8, ptr [[TMP0]], i32 1
+// ARM-LE-NEXT:    store i8 0, ptr [[TMP1]], align 1
+// ARM-LE-NEXT:    [[TMP2:%.*]] = getelementptr i8, ptr [[TMP0]], i32 2
+// ARM-LE-NEXT:    store i8 0, ptr [[TMP2]], align 2
+// ARM-LE-NEXT:    [[TMP3:%.*]] = getelementptr i8, ptr [[TMP0]], i32 3
+// ARM-LE-NEXT:    store i8 0, ptr [[TMP3]], align 1
+// ARM-LE-NEXT:    [[TMP4:%.*]] = getelementptr i8, ptr [[TMP0]], i32 4
+// ARM-LE-NEXT:    store i8 0, ptr [[TMP4]], align 4
+// ARM-LE-NEXT:    [[TMP5:%.*]] = getelementptr i8, ptr [[TMP0]], i32 5
+// ARM-LE-NEXT:    store i8 0, ptr [[TMP5]], align 1
+// ARM-LE-NEXT:    [[TMP6:%.*]] = getelementptr i8, ptr [[TMP0]], i32 6
+// ARM-LE-NEXT:    store i8 0, ptr [[TMP6]], align 2
+// ARM-LE-NEXT:    [[TMP7:%.*]] = getelementptr i8, ptr [[TMP0]], i32 7
+// ARM-LE-NEXT:    store i8 0, ptr [[TMP7]], align 1
+// ARM-LE-NEXT:    ret void
+//
+// ARM-BE-LABEL: define dso_local void @testUnionTailPaddingOfLongestMember(
+// ARM-BE-SAME: ptr noundef [[U:%.*]]) #[[ATTR0]] {
+// ARM-BE-NEXT:  [[ENTRY:.*:]]
+// ARM-BE-NEXT:    [[U_ADDR:%.*]] = alloca ptr, align 4
+// ARM-BE-NEXT:    store ptr [[U]], ptr [[U_ADDR]], align 4
+// ARM-BE-NEXT:    [[TMP0:%.*]] = load ptr, ptr [[U_ADDR]], align 4
+// ARM-BE-NEXT:    [[TMP1:%.*]] = getelementptr i8, ptr [[TMP0]], i32 1
+// ARM-BE-NEXT:    store i8 0, ptr [[TMP1]], align 1
+// ARM-BE-NEXT:    [[TMP2:%.*]] = getelementptr i8, ptr [[TMP0]], i32 2
+// ARM-BE-NEXT:    store i8 0, ptr [[TMP2]], align 2
+// ARM-BE-NEXT:    [[TMP3:%.*]] = getelementptr i8, ptr [[TMP0]], i32 3
+// ARM-BE-NEXT:    store i8 0, ptr [[TMP3]], align 1
+// ARM-BE-NEXT:    [[TMP4:%.*]] = getelementptr i8, ptr [[TMP0]], i32 4
+// ARM-BE-NEXT:    store i8 0, ptr [[TMP4]], align 4
+// ARM-BE-NEXT:    [[TMP5:%.*]] = getelementptr i8, ptr [[TMP0]], i32 5
+// ARM-BE-NEXT:    store i8 0, ptr [[TMP5]], align 1
+// ARM-BE-NEXT:    [[TMP6:%.*]] = getelementptr i8, ptr [[TMP0]], i32 6
+// ARM-BE-NEXT:    store i8 0, ptr [[TMP6]], align 2
+// ARM-BE-NEXT:    [[TMP7:%.*]] = getelementptr i8, ptr [[TMP0]], i32 7
+// ARM-BE-NEXT:    store i8 0, ptr [[TMP7]], align 1
+// ARM-BE-NEXT:    ret void
+//
+// AARCH64-LE-LABEL: define dso_local void @testUnionTailPaddingOfLongestMember(
+// AARCH64-LE-SAME: ptr noundef [[U:%.*]]) #[[ATTR0]] {
+// AARCH64-LE-NEXT:  [[ENTRY:.*:]]
+// AARCH64-LE-NEXT:    [[U_ADDR:%.*]] = alloca ptr, align 8
+// AARCH64-LE-NEXT:    store ptr [[U]], ptr [[U_ADDR]], align 8
+// AARCH64-LE-NEXT:    [[TMP0:%.*]] = load ptr, ptr [[U_ADDR]], align 8
+// AARCH64-LE-NEXT:    [[TMP1:%.*]] = getelementptr i8, ptr [[TMP0]], i32 1
+// AARCH64-LE-NEXT:    store i8 0, ptr [[TMP1]], align 1
+// AARCH64-LE-NEXT:    [[TMP2:%.*]] = getelementptr i8, ptr [[TMP0]], i32 2
+// AARCH64-LE-NEXT:    store i8 0, ptr [[TMP2]], align 2
+// AARCH64-LE-NEXT:    [[TMP3:%.*]] = getelementptr i8, ptr [[TMP0]], i32 3
+// AARCH64-LE-NEXT:    store i8 0, ptr [[TMP3]], align 1
+// AARCH64-LE-NEXT:    [[TMP4:%.*]] = getelementptr i8, ptr [[TMP0]], i32 4
+// AARCH64-LE-NEXT:    store i8 0, ptr [[TMP4]], align 4
+// AARCH64-LE-NEXT:    [[TMP5:%.*]] = getelementptr i8, ptr [[TMP0]], i32 5
+// AARCH64-LE-NEXT:    store i8 0, ptr [[TMP5]], align 1
+// AARCH64-LE-NEXT:    [[TMP6:%.*]] = getelementptr i8, ptr [[TMP0]], i32 6
+// AARCH64-LE-NEXT:    store i8 0, ptr [[TMP6]], align 2
+// AARCH64-LE-NEXT:    [[TMP7:%.*]] = getelementptr i8, ptr [[TMP0]], i32 7
+// AARCH64-LE-NEXT:    store i8 0, ptr [[TMP7]], align 1
+// AARCH64-LE-NEXT:    ret void
+//
+// AARCH64-BE-LABEL: define dso_local void @testUnionTailPaddingOfLongestMember(
+// AARCH64-BE-SAME: ptr noundef [[U:%.*]]) #[[ATTR0]] {
+// AARCH64-BE-NEXT:  [[ENTRY:.*:]]
+// AARCH64-BE-NEXT:    [[U_ADDR:%.*]] = alloca ptr, align 8
+// AARCH64-BE-NEXT:    store ptr [[U]], ptr [[U_ADDR]], align 8
+// AARCH64-BE-NEXT:    [[TMP0:%.*]] = load ptr, ptr [[U_ADDR]], align 8
+// AARCH64-BE-NEXT:    [[TMP1:%.*]] = getelementptr i8, ptr [[TMP0]], i32 1
+// AARCH64-BE-NEXT:    store i8 0, ptr [[TMP1]], align 1
+// AARCH64-BE-NEXT:    [[TMP2:%.*]] = getelementptr i8, ptr [[TMP0]], i32 2
+// AARCH64-BE-NEXT:    store i8 0, ptr [[TMP2]], align 2
+// AARCH64-BE-NEXT:    [[TMP3:%.*]] = getelementptr i8, ptr [[TMP0]], i32 3
+// AARCH64-BE-NEXT:    store i8 0, ptr [[TMP3]], align 1
+// AARCH64-BE-NEXT:    [[TMP4:%.*]] = getelementptr i8, ptr [[TMP0]], i32 4
+// AARCH64-BE-NEXT:    store i8 0, ptr [[TMP4]], align 4
+// AARCH64-BE-NEXT:    [[TMP5:%.*]] = getelementptr i8, ptr [[TMP0]], i32 5
+// AARCH64-BE-NEXT:    store i8 0, ptr [[TMP5]], align 1
+// AARCH64-BE-NEXT:    [[TMP6:%.*]] = getelementptr i8, ptr [[TMP0]], i32 6
+// AARCH64-BE-NEXT:    store i8 0, ptr [[TMP6]], align 2
+// AARCH64-BE-NEXT:    [[TMP7:%.*]] = getelementptr i8, ptr [[TMP0]], i32 7
+// AARCH64-BE-NEXT:    store i8 0, ptr [[TMP7]], align 1
+// AARCH64-BE-NEXT:    ret void
+//
+extern "C" void testUnionTailPaddingOfLongestMember(U2 *u) {
   // This should clear the tail padding of the longest member.
   // [1 - 7]
   __builtin_clear_padding(u);
 }
 
 
+
 struct alignas(4) Foo {
   char a;
   alignas(2) char b;
@@ -355,7 +798,7 @@ struct alignas(4) Baz : Foo {
 // %struct.Foo = type { i8, i8, i8, i8 }
 // %struct.Bar = type { i8, i8, i8, i8 }
 
-// LINUX-LABEL: define dso_local void @_Z33testStructPaddingInBetweenMembersP3Baz(
+// LINUX-LABEL: define dso_local void @testStructPaddingInBetweenMembers(
 // LINUX-SAME: ptr noundef [[BAZ:%.*]]) #[[ATTR0]] {
 // LINUX-NEXT:  [[ENTRY:.*:]]
 // LINUX-NEXT:    [[BAZ_ADDR:%.*]] = alloca ptr, align 8
@@ -377,33 +820,123 @@ struct alignas(4) Baz : Foo {
 // LINUX-NEXT:    store i8 0, ptr [[TMP7]], align 1
 // LINUX-NEXT:    ret void
 //
-// WINDOWS-LABEL: define dso_local void @"?testStructPaddingInBetweenMembers@@YAXPEAUBaz@@@Z"(ptr noundef %baz) #0 {
-// WINDOWS-NEXT: entry:
-// WINDOWS-NEXT:   %baz.addr = alloca ptr, align 8
-// WINDOWS-NEXT:   store ptr %baz, ptr %baz.addr, align 8
-// WINDOWS-NEXT:   %0 = load ptr, ptr %baz.addr, align 8
-// WINDOWS-NEXT:   %1 = getelementptr i8, ptr %0, i32 1
-// WINDOWS-NEXT:   store i8 0, ptr %1, align 1
-// WINDOWS-NEXT:   %2 = getelementptr i8, ptr %0, i32 3
-// WINDOWS-NEXT:   store i8 0, ptr %2, align 1
-// WINDOWS-NEXT:   %3 = getelementptr i8, ptr %0, i32 5
-// WINDOWS-NEXT:   store i8 0, ptr %3, align 1
-// WINDOWS-NEXT:   %4 = getelementptr i8, ptr %0, i32 6
-// WINDOWS-NEXT:   store i8 0, ptr %4, align 2
-// WINDOWS-NEXT:   %5 = getelementptr i8, ptr %0, i32 7
-// WINDOWS-NEXT:   store i8 0, ptr %5, align 1
-// WINDOWS-NEXT:   %6 = getelementptr i8, ptr %0, i32 9
-// WINDOWS-NEXT:   store i8 0, ptr %6, align 1
-// WINDOWS-NEXT:   %7 = getelementptr i8, ptr %0, i32 11
-// WINDOWS-NEXT:   store i8 0, ptr %7, align 1
-// WINDOWS-NEXT:   ret void
-//
-void testStructPaddingInBetweenMembers(Baz *baz) {
+// WINDOWS-LABEL: define dso_local void @testStructPaddingInBetweenMembers(
+// WINDOWS-SAME: ptr noundef [[BAZ:%.*]]) #[[ATTR0]] {
+// WINDOWS-NEXT:  [[ENTRY:.*:]]
+// WINDOWS-NEXT:    [[BAZ_ADDR:%.*]] = alloca ptr, align 8
+// WINDOWS-NEXT:    store ptr [[BAZ]], ptr [[BAZ_ADDR]], align 8
+// WINDOWS-NEXT:    [[TMP0:%.*]] = load ptr, ptr [[BAZ_ADDR]], align 8
+// WINDOWS-NEXT:    [[TMP1:%.*]] = getelementptr i8, ptr [[TMP0]], i32 1
+// WINDOWS-NEXT:    store i8 0, ptr [[TMP1]], align 1
+// WINDOWS-NEXT:    [[TMP2:%.*]] = getelementptr i8, ptr [[TMP0]], i32 3
+// WINDOWS-NEXT:    store i8 0, ptr [[TMP2]], align 1
+// WINDOWS-NEXT:    [[TMP3:%.*]] = getelementptr i8, ptr [[TMP0]], i32 5
+// WINDOWS-NEXT:    store i8 0, ptr [[TMP3]], align 1
+// WINDOWS-NEXT:    [[TMP4:%.*]] = getelementptr i8, ptr [[TMP0]], i32 6
+// WINDOWS-NEXT:    store i8 0, ptr [[TMP4]], align 2
+// WINDOWS-NEXT:    [[TMP5:%.*]] = getelementptr i8, ptr [[TMP0]], i32 7
+// WINDOWS-NEXT:    store i8 0, ptr [[TMP5]], align 1
+// WINDOWS-NEXT:    [[TMP6:%.*]] = getelementptr i8, ptr [[TMP0]], i32 9
+// WINDOWS-NEXT:    store i8 0, ptr [[TMP6]], align 1
+// WINDOWS-NEXT:    [[TMP7:%.*]] = getelementptr i8, ptr [[TMP0]], i32 11
+// WINDOWS-NEXT:    store i8 0, ptr [[TMP7]], align 1
+// WINDOWS-NEXT:    ret void
+//
+// ARM-LE-LABEL: define dso_local void @testStructPaddingInBetweenMembers(
+// ARM-LE-SAME: ptr noundef [[BAZ:%.*]]) #[[ATTR0]] {
+// ARM-LE-NEXT:  [[ENTRY:.*:]]
+// ARM-LE-NEXT:    [[BAZ_ADDR:%.*]] = alloca ptr, align 4
+// ARM-LE-NEXT:    store ptr [[BAZ]], ptr [[BAZ_ADDR]], align 4
+// ARM-LE-NEXT:    [[TMP0:%.*]] = load ptr, ptr [[BAZ_ADDR]], align 4
+// ARM-LE-NEXT:    [[TMP1:%.*]] = getelementptr i8, ptr [[TMP0]], i32 1
+// ARM-LE-NEXT:    store i8 0, ptr [[TMP1]], align 1
+// ARM-LE-NEXT:    [[TMP2:%.*]] = getelementptr i8, ptr [[TMP0]], i32 3
+// ARM-LE-NEXT:    store i8 0, ptr [[TMP2]], align 1
+// ARM-LE-NEXT:    [[TMP3:%.*]] = getelementptr i8, ptr [[TMP0]], i32 5
+// ARM-LE-NEXT:    store i8 0, ptr [[TMP3]], align 1
+// ARM-LE-NEXT:    [[TMP4:%.*]] = getelementptr i8, ptr [[TMP0]], i32 6
+// ARM-LE-NEXT:    store i8 0, ptr [[TMP4]], align 2
+// ARM-LE-NEXT:    [[TMP5:%.*]] = getelementptr i8, ptr [[TMP0]], i32 7
+// ARM-LE-NEXT:    store i8 0, ptr [[TMP5]], align 1
+// ARM-LE-NEXT:    [[TMP6:%.*]] = getelementptr i8, ptr [[TMP0]], i32 9
+// ARM-LE-NEXT:    store i8 0, ptr [[TMP6]], align 1
+// ARM-LE-NEXT:    [[TMP7:%.*]] = getelementptr i8, ptr [[TMP0]], i32 11
+// ARM-LE-NEXT:    store i8 0, ptr [[TMP7]], align 1
+// ARM-LE-NEXT:    ret void
+//
+// ARM-BE-LABEL: define dso_local void @testStructPaddingInBetweenMembers(
+// ARM-BE-SAME: ptr noundef [[BAZ:%.*]]) #[[ATTR0]] {
+// ARM-BE-NEXT:  [[ENTRY:.*:]]
+// ARM-BE-NEXT:    [[BAZ_ADDR:%.*]] = alloca ptr, align 4
+// ARM-BE-NEXT:    store ptr [[BAZ]], ptr [[BAZ_ADDR]], align 4
+// ARM-BE-NEXT:    [[TMP0:%.*]] = load ptr, ptr [[BAZ_ADDR]], align 4
+// ARM-BE-NEXT:    [[TMP1:%.*]] = getelementptr i8, ptr [[TMP0]], i32 1
+// ARM-BE-NEXT:    store i8 0, ptr [[TMP1]], align 1
+// ARM-BE-NEXT:    [[TMP2:%.*]] = getelementptr i8, ptr [[TMP0]], i32 3
+// ARM-BE-NEXT:    store i8 0, ptr [[TMP2]], align 1
+// ARM-BE-NEXT:    [[TMP3:%.*]] = getelementptr i8, ptr [[TMP0]], i32 5
+// ARM-BE-NEXT:    store i8 0, ptr [[TMP3]], align 1
+// ARM-BE-NEXT:    [[TMP4:%.*]] = getelementptr i8, ptr [[TMP0]], i32 6
+// ARM-BE-NEXT:    store i8 0, ptr [[TMP4]], align 2
+// ARM-BE-NEXT:    [[TMP5:%.*]] = getelementptr i8, ptr [[TMP0]], i32 7
+// ARM-BE-NEXT:    store i8 0, ptr [[TMP5]], align 1
+// ARM-BE-NEXT:    [[TMP6:%.*]] = getelementptr i8, ptr [[TMP0]], i32 9
+// ARM-BE-NEXT:    store i8 0, ptr [[TMP6]], align 1
+// ARM-BE-NEXT:    [[TMP7:%.*]] = getelementptr i8, ptr [[TMP0]], i32 11
+// ARM-BE-NEXT:    store i8 0, ptr [[TMP7]], align 1
+// ARM-BE-NEXT:    ret void
+//
+// AARCH64-LE-LABEL: define dso_local void @testStructPaddingInBetweenMembers(
+// AARCH64-LE-SAME: ptr noundef [[BAZ:%.*]]) #[[ATTR0]] {
+// AARCH64-LE-NEXT:  [[ENTRY:.*:]]
+// AARCH64-LE-NEXT:    [[BAZ_ADDR:%.*]] = alloca ptr, align 8
+// AARCH64-LE-NEXT:    store ptr [[BAZ]], ptr [[BAZ_ADDR]], align 8
+// AARCH64-LE-NEXT:    [[TMP0:%.*]] = load ptr, ptr [[BAZ_ADDR]], align 8
+// AARCH64-LE-NEXT:    [[TMP1:%.*]] = getelementptr i8, ptr [[TMP0]], i32 1
+// AARCH64-LE-NEXT:    store i8 0, ptr [[TMP1]], align 1
+// AARCH64-LE-NEXT:    [[TMP2:%.*]] = getelementptr i8, ptr [[TMP0]], i32 3
+// AARCH64-LE-NEXT:    store i8 0, ptr [[TMP2]], align 1
+// AARCH64-LE-NEXT:    [[TMP3:%.*]] = getelementptr i8, ptr [[TMP0]], i32 5
+// AARCH64-LE-NEXT:    store i8 0, ptr [[TMP3]], align 1
+// AARCH64-LE-NEXT:    [[TMP4:%.*]] = getelementptr i8, ptr [[TMP0]], i32 6
+// AARCH64-LE-NEXT:    store i8 0, ptr [[TMP4]], align 2
+// AARCH64-LE-NEXT:    [[TMP5:%.*]] = getelementptr i8, ptr [[TMP0]], i32 7
+// AARCH64-LE-NEXT:    store i8 0, ptr [[TMP5]], align 1
+// AARCH64-LE-NEXT:    [[TMP6:%.*]] = getelementptr i8, ptr [[TMP0]], i32 9
+// AARCH64-LE-NEXT:    store i8 0, ptr [[TMP6]], align 1
+// AARCH64-LE-NEXT:    [[TMP7:%.*]] = getelementptr i8, ptr [[TMP0]], i32 11
+// AARCH64-LE-NEXT:    store i8 0, ptr [[TMP7]], align 1
+// AARCH64-LE-NEXT:    ret void
+//
+// AARCH64-BE-LABEL: define dso_local void @testStructPaddingInBetweenMembers(
+// AARCH64-BE-SAME: ptr noundef [[BAZ:%.*]]) #[[ATTR0]] {
+// AARCH64-BE-NEXT:  [[ENTRY:.*:]]
+// AARCH64-BE-NEXT:    [[BAZ_ADDR:%.*]] = alloca ptr, align 8
+// AARCH64-BE-NEXT:    store ptr [[BAZ]], ptr [[BAZ_ADDR]], align 8
+// AARCH64-BE-NEXT:    [[TMP0:%.*]] = load ptr, ptr [[BAZ_ADDR]], align 8
+// AARCH64-BE-NEXT:    [[TMP1:%.*]] = getelementptr i8, ptr [[TMP0]], i32 1
+// AARCH64-BE-NEXT:    store i8 0, ptr [[TMP1]], align 1
+// AARCH64-BE-NEXT:    [[TMP2:%.*]] = getelementptr i8, ptr [[TMP0]], i32 3
+// AARCH64-BE-NEXT:    store i8 0, ptr [[TMP2]], align 1
+// AARCH64-BE-NEXT:    [[TMP3:%.*]] = getelementptr i8, ptr [[TMP0]], i32 5
+// AARCH64-BE-NEXT:    store i8 0, ptr [[TMP3]], align 1
+// AARCH64-BE-NEXT:    [[TMP4:%.*]] = getelementptr i8, ptr [[TMP0]], i32 6
+// AARCH64-BE-NEXT:    store i8 0, ptr [[TMP4]], align 2
+// AARCH64-BE-NEXT:    [[TMP5:%.*]] = getelementptr i8, ptr [[TMP0]], i32 7
+// AARCH64-BE-NEXT:    store i8 0, ptr [[TMP5]], align 1
+// AARCH64-BE-NEXT:    [[TMP6:%.*]] = getelementptr i8, ptr [[TMP0]], i32 9
+// AARCH64-BE-NEXT:    store i8 0, ptr [[TMP6]], align 1
+// AARCH64-BE-NEXT:    [[TMP7:%.*]] = getelementptr i8, ptr [[TMP0]], i32 11
+// AARCH64-BE-NEXT:    store i8 0, ptr [[TMP7]], align 1
+// AARCH64-BE-NEXT:    ret void
+//
+extern "C" void testStructPaddingInBetweenMembers(Baz *baz) {
   // this should clear all the padding in between various members
   __builtin_clear_padding(baz);
 }
 
-// LINUX-LABEL: define dso_local void @_Z18testStructVolatilePV3Baz(
+
+// LINUX-LABEL: define dso_local void @testStructVolatile(
 // LINUX-SAME: ptr noundef [[BAZ:%.*]]) #[[ATTR0]] {
 // LINUX-NEXT:  [[ENTRY:.*:]]
 // LINUX-NEXT:    [[BAZ_ADDR:%.*]] = alloca ptr, align 8
@@ -425,32 +958,122 @@ void testStructPaddingInBetweenMembers(Baz *baz) {
 // LINUX-NEXT:    store i8 0, ptr [[TMP7]], align 1
 // LINUX-NEXT:    ret void
 //
-// WINDOWS-LABEL: define dso_local void @"?testStructVolatile@@YAXPECUBaz@@@Z"(ptr noundef %baz) #0 {
-// WINDOWS-NEXT: entry:
-// WINDOWS-NEXT:   %baz.addr = alloca ptr, align 8
-// WINDOWS-NEXT:   store ptr %baz, ptr %baz.addr, align 8
-// WINDOWS-NEXT:   %0 = load ptr, ptr %baz.addr, align 8
-// WINDOWS-NEXT:   %1 = getelementptr i8, ptr %0, i32 1
-// WINDOWS-NEXT:   store i8 0, ptr %1, align 1
-// WINDOWS-NEXT:   %2 = getelementptr i8, ptr %0, i32 3
-// WINDOWS-NEXT:   store i8 0, ptr %2, align 1
-// WINDOWS-NEXT:   %3 = getelementptr i8, ptr %0, i32 5
-// WINDOWS-NEXT:   store i8 0, ptr %3, align 1
-// WINDOWS-NEXT:   %4 = getelementptr i8, ptr %0, i32 6
-// WINDOWS-NEXT:   store i8 0, ptr %4, align 2
-// WINDOWS-NEXT:   %5 = getelementptr i8, ptr %0, i32 7
-// WINDOWS-NEXT:   store i8 0, ptr %5, align 1
-// WINDOWS-NEXT:   %6 = getelementptr i8, ptr %0, i32 9
-// WINDOWS-NEXT:   store i8 0, ptr %6, align 1
-// WINDOWS-NEXT:   %7 = getelementptr i8, ptr %0, i32 11
-// WINDOWS-NEXT:   store i8 0, ptr %7, align 1
-// WINDOWS-NEXT:   ret void
-//
-void testStructVolatile(volatile Baz *baz) {
+// WINDOWS-LABEL: define dso_local void @testStructVolatile(
+// WINDOWS-SAME: ptr noundef [[BAZ:%.*]]) #[[ATTR0]] {
+// WINDOWS-NEXT:  [[ENTRY:.*:]]
+// WINDOWS-NEXT:    [[BAZ_ADDR:%.*]] = alloca ptr, align 8
+// WINDOWS-NEXT:    store ptr [[BAZ]], ptr [[BAZ_ADDR]], align 8
+// WINDOWS-NEXT:    [[TMP0:%.*]] = load ptr, ptr [[BAZ_ADDR]], align 8
+// WINDOWS-NEXT:    [[TMP1:%.*]] = getelementptr i8, ptr [[TMP0]], i32 1
+// WINDOWS-NEXT:    store i8 0, ptr [[TMP1]], align 1
+// WINDOWS-NEXT:    [[TMP2:%.*]] = getelementptr i8, ptr [[TMP0]], i32 3
+// WINDOWS-NEXT:    store i8 0, ptr [[TMP2]], align 1
+// WINDOWS-NEXT:    [[TMP3:%.*]] = getelementptr i8, ptr [[TMP0]], i32 5
+// WINDOWS-NEXT:    store i8 0, ptr [[TMP3]], align 1
+// WINDOWS-NEXT:    [[TMP4:%.*]] = getelementptr i8, ptr [[TMP0]], i32 6
+// WINDOWS-NEXT:    store i8 0, ptr [[TMP4]], align 2
+// WINDOWS-NEXT:    [[TMP5:%.*]] = getelementptr i8, ptr [[TMP0]], i32 7
+// WINDOWS-NEXT:    store i8 0, ptr [[TMP5]], align 1
+// WINDOWS-NEXT:    [[TMP6:%.*]] = getelementptr i8, ptr [[TMP0]], i32 9
+// WINDOWS-NEXT:    store i8 0, ptr [[TMP6]], align 1
+// WINDOWS-NEXT:    [[TMP7:%.*]] = getelementptr i8, ptr [[TMP0]], i32 11
+// WINDOWS-NEXT:    store i8 0, ptr [[TMP7]], align 1
+// WINDOWS-NEXT:    ret void
+//
+// ARM-LE-LABEL: define dso_local void @testStructVolatile(
+// ARM-LE-SAME: ptr noundef [[BAZ:%.*]]) #[[ATTR0]] {
+// ARM-LE-NEXT:  [[ENTRY:.*:]]
+// ARM-LE-NEXT:    [[BAZ_ADDR:%.*]] = alloca ptr, align 4
+// ARM-LE-NEXT:    store ptr [[BAZ]], ptr [[BAZ_ADDR]], align 4
+// ARM-LE-NEXT:    [[TMP0:%.*]] = load ptr, ptr [[BAZ_ADDR]], align 4
+// ARM-LE-NEXT:    [[TMP1:%.*]] = getelementptr i8, ptr [[TMP0]], i32 1
+// ARM-LE-NEXT:    store i8 0, ptr [[TMP1]], align 1
+// ARM-LE-NEXT:    [[TMP2:%.*]] = getelementptr i8, ptr [[TMP0]], i32 3
+// ARM-LE-NEXT:    store i8 0, ptr [[TMP2]], align 1
+// ARM-LE-NEXT:    [[TMP3:%.*]] = getelementptr i8, ptr [[TMP0]], i32 5
+// ARM-LE-NEXT:    store i8 0, ptr [[TMP3]], align 1
+// ARM-LE-NEXT:    [[TMP4:%.*]] = getelementptr i8, ptr [[TMP0]], i32 6
+// ARM-LE-NEXT:    store i8 0, ptr [[TMP4]], align 2
+// ARM-LE-NEXT:    [[TMP5:%.*]] = getelementptr i8, ptr [[TMP0]], i32 7
+// ARM-LE-NEXT:    store i8 0, ptr [[TMP5]], align 1
+// ARM-LE-NEXT:    [[TMP6:%.*]] = getelementptr i8, ptr [[TMP0]], i32 9
+// ARM-LE-NEXT:    store i8 0, ptr [[TMP6]], align 1
+// ARM-LE-NEXT:    [[TMP7:%.*]] = getelementptr i8, ptr [[TMP0]], i32 11
+// ARM-LE-NEXT:    store i8 0, ptr [[TMP7]], align 1
+// ARM-LE-NEXT:    ret void
+//
+// ARM-BE-LABEL: define dso_local void @testStructVolatile(
+// ARM-BE-SAME: ptr noundef [[BAZ:%.*]]) #[[ATTR0]] {
+// ARM-BE-NEXT:  [[ENTRY:.*:]]
+// ARM-BE-NEXT:    [[BAZ_ADDR:%.*]] = alloca ptr, align 4
+// ARM-BE-NEXT:    store ptr [[BAZ]], ptr [[BAZ_ADDR]], align 4
+// ARM-BE-NEXT:    [[TMP0:%.*]] = load ptr, ptr [[BAZ_ADDR]], align 4
+// ARM-BE-NEXT:    [[TMP1:%.*]] = getelementptr i8, ptr [[TMP0]], i32 1
+// ARM-BE-NEXT:    store i8 0, ptr [[TMP1]], align 1
+// ARM-BE-NEXT:    [[TMP2:%.*]] = getelementptr i8, ptr [[TMP0]], i32 3
+// ARM-BE-NEXT:    store i8 0, ptr [[TMP2]], align 1
+// ARM-BE-NEXT:    [[TMP3:%.*]] = getelementptr i8, ptr [[TMP0]], i32 5
+// ARM-BE-NEXT:    store i8 0, ptr [[TMP3]], align 1
+// ARM-BE-NEXT:    [[TMP4:%.*]] = getelementptr i8, ptr [[TMP0]], i32 6
+// ARM-BE-NEXT:    store i8 0, ptr [[TMP4]], align 2
+// ARM-BE-NEXT:    [[TMP5:%.*]] = getelementptr i8, ptr [[TMP0]], i32 7
+// ARM-BE-NEXT:    store i8 0, ptr [[TMP5]], align 1
+// ARM-BE-NEXT:    [[TMP6:%.*]] = getelementptr i8, ptr [[TMP0]], i32 9
+// ARM-BE-NEXT:    store i8 0, ptr [[TMP6]], align 1
+// ARM-BE-NEXT:    [[TMP7:%.*]] = getelementptr i8, ptr [[TMP0]], i32 11
+// ARM-BE-NEXT:    store i8 0, ptr [[TMP7]], align 1
+// ARM-BE-NEXT:    ret void
+//
+// AARCH64-LE-LABEL: define dso_local void @testStructVolatile(
+// AARCH64-LE-SAME: ptr noundef [[BAZ:%.*]]) #[[ATTR0]] {
+// AARCH64-LE-NEXT:  [[ENTRY:.*:]]
+// AARCH64-LE-NEXT:    [[BAZ_ADDR:%.*]] = alloca ptr, align 8
+// AARCH64-LE-NEXT:    store ptr [[BAZ]], ptr [[BAZ_ADDR]], align 8
+// AARCH64-LE-NEXT:    [[TMP0:%.*]] = load ptr, ptr [[BAZ_ADDR]], align 8
+// AARCH64-LE-NEXT:    [[TMP1:%.*]] = getelementptr i8, ptr [[TMP0]], i32 1
+// AARCH64-LE-NEXT:    store i8 0, ptr [[TMP1]], align 1
+// AARCH64-LE-NEXT:    [[TMP2:%.*]] = getelementptr i8, ptr [[TMP0]], i32 3
+// AARCH64-LE-NEXT:    store i8 0, ptr [[TMP2]], align 1
+// AARCH64-LE-NEXT:    [[TMP3:%.*]] = getelementptr i8, ptr [[TMP0]], i32 5
+// AARCH64-LE-NEXT:    store i8 0, ptr [[TMP3]], align 1
+// AARCH64-LE-NEXT:    [[TMP4:%.*]] = getelementptr i8, ptr [[TMP0]], i32 6
+// AARCH64-LE-NEXT:    store i8 0, ptr [[TMP4]], align 2
+// AARCH64-LE-NEXT:    [[TMP5:%.*]] = getelementptr i8, ptr [[TMP0]], i32 7
+// AARCH64-LE-NEXT:    store i8 0, ptr [[TMP5]], align 1
+// AARCH64-LE-NEXT:    [[TMP6:%.*]] = getelementptr i8, ptr [[TMP0]], i32 9
+// AARCH64-LE-NEXT:    store i8 0, ptr [[TMP6]], align 1
+// AARCH64-LE-NEXT:    [[TMP7:%.*]] = getelementptr i8, ptr [[TMP0]], i32 11
+// AARCH64-LE-NEXT:    store i8 0, ptr [[TMP7]], align 1
+// AARCH64-LE-NEXT:    ret void
+//
+// AARCH64-BE-LABEL: define dso_local void @testStructVolatile(
+// AARCH64-BE-SAME: ptr noundef [[BAZ:%.*]]) #[[ATTR0]] {
+// AARCH64-BE-NEXT:  [[ENTRY:.*:]]
+// AARCH64-BE-NEXT:    [[BAZ_ADDR:%.*]] = alloca ptr, align 8
+// AARCH64-BE-NEXT:    store ptr [[BAZ]], ptr [[BAZ_ADDR]], align 8
+// AARCH64-BE-NEXT:    [[TMP0:%.*]] = load ptr, ptr [[BAZ_ADDR]], align 8
+// AARCH64-BE-NEXT:    [[TMP1:%.*]] = getelementptr i8, ptr [[TMP0]], i32 1
+// AARCH64-BE-NEXT:    store i8 0, ptr [[TMP1]], align 1
+// AARCH64-BE-NEXT:    [[TMP2:%.*]] = getelementptr i8, ptr [[TMP0]], i32 3
+// AARCH64-BE-NEXT:    store i8 0, ptr [[TMP2]], align 1
+// AARCH64-BE-NEXT:    [[TMP3:%.*]] = getelementptr i8, ptr [[TMP0]], i32 5
+// AARCH64-BE-NEXT:    store i8 0, ptr [[TMP3]], align 1
+// AARCH64-BE-NEXT:    [[TMP4:%.*]] = getelementptr i8, ptr [[TMP0]], i32 6
+// AARCH64-BE-NEXT:    store i8 0, ptr [[TMP4]], align 2
+// AARCH64-BE-NEXT:    [[TMP5:%.*]] = getelementptr i8, ptr [[TMP0]], i32 7
+// AARCH64-BE-NEXT:    store i8 0, ptr [[TMP5]], align 1
+// AARCH64-BE-NEXT:    [[TMP6:%.*]] = getelementptr i8, ptr [[TMP0]], i32 9
+// AARCH64-BE-NEXT:    store i8 0, ptr [[TMP6]], align 1
+// AARCH64-BE-NEXT:    [[TMP7:%.*]] = getelementptr i8, ptr [[TMP0]], i32 11
+// AARCH64-BE-NEXT:    store i8 0, ptr [[TMP7]], align 1
+// AARCH64-BE-NEXT:    ret void
+//
+extern "C" void testStructVolatile(volatile Baz *baz) {
   // this should clear all the padding in between various members
   __builtin_clear_padding(baz);
 }
 
+
 class S1 {
   int x;
   char c;
@@ -461,7 +1084,7 @@ class S2 {
   bool b;
 };
 
-// LINUX-LABEL: define dso_local void @_Z19testNoUniqueAddressP2S2(
+// LINUX-LABEL: define dso_local void @testNoUniqueAddress(
 // LINUX-SAME: ptr noundef [[S:%.*]]) #[[ATTR0]] {
 // LINUX-NEXT:  [[ENTRY:.*:]]
 // LINUX-NEXT:    [[S_ADDR:%.*]] = alloca ptr, align 8
@@ -473,36 +1096,86 @@ class S2 {
 // LINUX-NEXT:    store i8 0, ptr [[TMP2]], align 1
 // LINUX-NEXT:    ret void
 //
-// WINDOWS-LABEL: define dso_local void @"?testNoUniqueAddress@@YAXPEAVS2@@@Z"(ptr noundef %s) #0 {
-// WINDOWS-NEXT: entry:
-// WINDOWS-NEXT:   %s.addr = alloca ptr, align 8
-// WINDOWS-NEXT:   store ptr %s, ptr %s.addr, align 8
-// WINDOWS-NEXT:   %0 = load ptr, ptr %s.addr, align 8
-// WINDOWS-NEXT:   %1 = getelementptr i8, ptr %0, i32 5
-// WINDOWS-NEXT:   store i8 0, ptr %1, align 1
-// WINDOWS-NEXT:   %2 = getelementptr i8, ptr %0, i32 6
-// WINDOWS-NEXT:   store i8 0, ptr %2, align 2
-// WINDOWS-NEXT:   %3 = getelementptr i8, ptr %0, i32 7
-// WINDOWS-NEXT:   store i8 0, ptr %3, align 1
-// WINDOWS-NEXT:   %4 = getelementptr i8, ptr %0, i32 9
-// WINDOWS-NEXT:   store i8 0, ptr %4, align 1
-// WINDOWS-NEXT:   %5 = getelementptr i8, ptr %0, i32 10
-// WINDOWS-NEXT:   store i8 0, ptr %5, align 2
-// WINDOWS-NEXT:   %6 = getelementptr i8, ptr %0, i32 11
-// WINDOWS-NEXT:   store i8 0, ptr %6, align 1
-// WINDOWS-NEXT:   ret void
-//
-void testNoUniqueAddress(S2 *s) {
+// WINDOWS-LABEL: define dso_local void @testNoUniqueAddress(
+// WINDOWS-SAME: ptr noundef [[S:%.*]]) #[[ATTR0]] {
+// WINDOWS-NEXT:  [[ENTRY:.*:]]
+// WINDOWS-NEXT:    [[S_ADDR:%.*]] = alloca ptr, align 8
+// WINDOWS-NEXT:    store ptr [[S]], ptr [[S_ADDR]], align 8
+// WINDOWS-NEXT:    [[TMP0:%.*]] = load ptr, ptr [[S_ADDR]], align 8
+// WINDOWS-NEXT:    [[TMP1:%.*]] = getelementptr i8, ptr [[TMP0]], i32 5
+// WINDOWS-NEXT:    store i8 0, ptr [[TMP1]], align 1
+// WINDOWS-NEXT:    [[TMP2:%.*]] = getelementptr i8, ptr [[TMP0]], i32 6
+// WINDOWS-NEXT:    store i8 0, ptr [[TMP2]], align 2
+// WINDOWS-NEXT:    [[TMP3:%.*]] = getelementptr i8, ptr [[TMP0]], i32 7
+// WINDOWS-NEXT:    store i8 0, ptr [[TMP3]], align 1
+// WINDOWS-NEXT:    [[TMP4:%.*]] = getelementptr i8, ptr [[TMP0]], i32 9
+// WINDOWS-NEXT:    store i8 0, ptr [[TMP4]], align 1
+// WINDOWS-NEXT:    [[TMP5:%.*]] = getelementptr i8, ptr [[TMP0]], i32 10
+// WINDOWS-NEXT:    store i8 0, ptr [[TMP5]], align 2
+// WINDOWS-NEXT:    [[TMP6:%.*]] = getelementptr i8, ptr [[TMP0]], i32 11
+// WINDOWS-NEXT:    store i8 0, ptr [[TMP6]], align 1
+// WINDOWS-NEXT:    ret void
+//
+// ARM-LE-LABEL: define dso_local void @testNoUniqueAddress(
+// ARM-LE-SAME: ptr noundef [[S:%.*]]) #[[ATTR0]] {
+// ARM-LE-NEXT:  [[ENTRY:.*:]]
+// ARM-LE-NEXT:    [[S_ADDR:%.*]] = alloca ptr, align 4
+// ARM-LE-NEXT:    store ptr [[S]], ptr [[S_ADDR]], align 4
+// ARM-LE-NEXT:    [[TMP0:%.*]] = load ptr, ptr [[S_ADDR]], align 4
+// ARM-LE-NEXT:    [[TMP1:%.*]] = getelementptr i8, ptr [[TMP0]], i32 6
+// ARM-LE-NEXT:    store i8 0, ptr [[TMP1]], align 2
+// ARM-LE-NEXT:    [[TMP2:%.*]] = getelementptr i8, ptr [[TMP0]], i32 7
+// ARM-LE-NEXT:    store i8 0, ptr [[TMP2]], align 1
+// ARM-LE-NEXT:    ret void
+//
+// ARM-BE-LABEL: define dso_local void @testNoUniqueAddress(
+// ARM-BE-SAME: ptr noundef [[S:%.*]]) #[[ATTR0]] {
+// ARM-BE-NEXT:  [[ENTRY:.*:]]
+// ARM-BE-NEXT:    [[S_ADDR:%.*]] = alloca ptr, align 4
+// ARM-BE-NEXT:    store ptr [[S]], ptr [[S_ADDR]], align 4
+// ARM-BE-NEXT:    [[TMP0:%.*]] = load ptr, ptr [[S_ADDR]], align 4
+// ARM-BE-NEXT:    [[TMP1:%.*]] = getelementptr i8, ptr [[TMP0]], i32 6
+// ARM-BE-NEXT:    store i8 0, ptr [[TMP1]], align 2
+// ARM-BE-NEXT:    [[TMP2:%.*]] = getelementptr i8, ptr [[TMP0]], i32 7
+// ARM-BE-NEXT:    store i8 0, ptr [[TMP2]], align 1
+// ARM-BE-NEXT:    ret void
+//
+// AARCH64-LE-LABEL: define dso_local void @testNoUniqueAddress(
+// AARCH64-LE-SAME: ptr noundef [[S:%.*]]) #[[ATTR0]] {
+// AARCH64-LE-NEXT:  [[ENTRY:.*:]]
+// AARCH64-LE-NEXT:    [[S_ADDR:%.*]] = alloca ptr, align 8
+// AARCH64-LE-NEXT:    store ptr [[S]], ptr [[S_ADDR]], align 8
+// AARCH64-LE-NEXT:    [[TMP0:%.*]] = load ptr, ptr [[S_ADDR]], align 8
+// AARCH64-LE-NEXT:    [[TMP1:%.*]] = getelementptr i8, ptr [[TMP0]], i32 6
+// AARCH64-LE-NEXT:    store i8 0, ptr [[TMP1]], align 2
+// AARCH64-LE-NEXT:    [[TMP2:%.*]] = getelementptr i8, ptr [[TMP0]], i32 7
+// AARCH64-LE-NEXT:    store i8 0, ptr [[TMP2]], align 1
+// AARCH64-LE-NEXT:    ret void
+//
+// AARCH64-BE-LABEL: define dso_local void @testNoUniqueAddress(
+// AARCH64-BE-SAME: ptr noundef [[S:%.*]]) #[[ATTR0]] {
+// AARCH64-BE-NEXT:  [[ENTRY:.*:]]
+// AARCH64-BE-NEXT:    [[S_ADDR:%.*]] = alloca ptr, align 8
+// AARCH64-BE-NEXT:    store ptr [[S]], ptr [[S_ADDR]], align 8
+// AARCH64-BE-NEXT:    [[TMP0:%.*]] = load ptr, ptr [[S_ADDR]], align 8
+// AARCH64-BE-NEXT:    [[TMP1:%.*]] = getelementptr i8, ptr [[TMP0]], i32 6
+// AARCH64-BE-NEXT:    store i8 0, ptr [[TMP1]], align 2
+// AARCH64-BE-NEXT:    [[TMP2:%.*]] = getelementptr i8, ptr [[TMP0]], i32 7
+// AARCH64-BE-NEXT:    store i8 0, ptr [[TMP2]], align 1
+// AARCH64-BE-NEXT:    ret void
+//
+extern "C" void testNoUniqueAddress(S2 *s) {
   // "x [0-3]",  "c" , "b", PAD [6-7]
   __builtin_clear_padding(s);
 }
 
+
 struct S3 {
   long double l;
   bool b;
 };
 
-// LINUX-LABEL: define dso_local void @_Z24testStructWithLongDoubleP2S3(
+// LINUX-LABEL: define dso_local void @testStructWithLongDouble(
 // LINUX-SAME: ptr noundef [[S:%.*]]) #[[ATTR0]] {
 // LINUX-NEXT:  [[ENTRY:.*:]]
 // LINUX-NEXT:    [[S_ADDR:%.*]] = alloca ptr, align 8
@@ -552,32 +1225,154 @@ struct S3 {
 // LINUX-NEXT:    store i8 0, ptr [[TMP21]], align 1
 // LINUX-NEXT:    ret void
 //
-// WINDOWS-LABEL: define dso_local void @"?testStructWithLongDouble@@YAXPEAUS3@@@Z"(ptr noundef %s) #0 {
-// WINDOWS-NEXT: entry:
-// WINDOWS-NEXT:   %s.addr = alloca ptr, align 8
-// WINDOWS-NEXT:   store ptr %s, ptr %s.addr, align 8
-// WINDOWS-NEXT:   %0 = load ptr, ptr %s.addr, align 8
-// WINDOWS-NEXT:   %1 = getelementptr i8, ptr %0, i32 9
-// WINDOWS-NEXT:   store i8 0, ptr %1, align 1
-// WINDOWS-NEXT:   %2 = getelementptr i8, ptr %0, i32 10
-// WINDOWS-NEXT:   store i8 0, ptr %2, align 2
-// WINDOWS-NEXT:   %3 = getelementptr i8, ptr %0, i32 11
-// WINDOWS-NEXT:   store i8 0, ptr %3, align 1
-// WINDOWS-NEXT:   %4 = getelementptr i8, ptr %0, i32 12
-// WINDOWS-NEXT:   store i8 0, ptr %4, align 4
-// WINDOWS-NEXT:   %5 = getelementptr i8, ptr %0, i32 13
-// WINDOWS-NEXT:   store i8 0, ptr %5, align 1
-// WINDOWS-NEXT:   %6 = getelementptr i8, ptr %0, i32 14
-// WINDOWS-NEXT:   store i8 0, ptr %6, align 2
-// WINDOWS-NEXT:   %7 = getelementptr i8, ptr %0, i32 15
-// WINDOWS-NEXT:   store i8 0, ptr %7, align 1
-// WINDOWS-NEXT:   ret void
-//
-void testStructWithLongDouble(S3 *s) {
+// WINDOWS-LABEL: define dso_local void @testStructWithLongDouble(
+// WINDOWS-SAME: ptr noundef [[S:%.*]]) #[[ATTR0]] {
+// WINDOWS-NEXT:  [[ENTRY:.*:]]
+// WINDOWS-NEXT:    [[S_ADDR:%.*]] = alloca ptr, align 8
+// WINDOWS-NEXT:    store ptr [[S]], ptr [[S_ADDR]], align 8
+// WINDOWS-NEXT:    [[TMP0:%.*]] = load ptr, ptr [[S_ADDR]], align 8
+// WINDOWS-NEXT:    [[TMP1:%.*]] = getelementptr i8, ptr [[TMP0]], i32 9
+// WINDOWS-NEXT:    store i8 0, ptr [[TMP1]], align 1
+// WINDOWS-NEXT:    [[TMP2:%.*]] = getelementptr i8, ptr [[TMP0]], i32 10
+// WINDOWS-NEXT:    store i8 0, ptr [[TMP2]], align 2
+// WINDOWS-NEXT:    [[TMP3:%.*]] = getelementptr i8, ptr [[TMP0]], i32 11
+// WINDOWS-NEXT:    store i8 0, ptr [[TMP3]], align 1
+// WINDOWS-NEXT:    [[TMP4:%.*]] = getelementptr i8, ptr [[TMP0]], i32 12
+// WINDOWS-NEXT:    store i8 0, ptr [[TMP4]], align 4
+// WINDOWS-NEXT:    [[TMP5:%.*]] = getelementptr i8, ptr [[TMP0]], i32 13
+// WINDOWS-NEXT:    store i8 0, ptr [[TMP5]], align 1
+// WINDOWS-NEXT:    [[TMP6:%.*]] = getelementptr i8, ptr [[TMP0]], i32 14
+// WINDOWS-NEXT:    store i8 0, ptr [[TMP6]], align 2
+// WINDOWS-NEXT:    [[TMP7:%.*]] = getelementptr i8, ptr [[TMP0]], i32 15
+// WINDOWS-NEXT:    store i8 0, ptr [[TMP7]], align 1
+// WINDOWS-NEXT:    ret void
+//
+// ARM-LE-LABEL: define dso_local void @testStructWithLongDouble(
+// ARM-LE-SAME: ptr noundef [[S:%.*]]) #[[ATTR0]] {
+// ARM-LE-NEXT:  [[ENTRY:.*:]]
+// ARM-LE-NEXT:    [[S_ADDR:%.*]] = alloca ptr, align 4
+// ARM-LE-NEXT:    store ptr [[S]], ptr [[S_ADDR]], align 4
+// ARM-LE-NEXT:    [[TMP0:%.*]] = load ptr, ptr [[S_ADDR]], align 4
+// ARM-LE-NEXT:    [[TMP1:%.*]] = getelementptr i8, ptr [[TMP0]], i32 9
+// ARM-LE-NEXT:    store i8 0, ptr [[TMP1]], align 1
+// ARM-LE-NEXT:    [[TMP2:%.*]] = getelementptr i8, ptr [[TMP0]], i32 10
+// ARM-LE-NEXT:    store i8 0, ptr [[TMP2]], align 2
+// ARM-LE-NEXT:    [[TMP3:%.*]] = getelementptr i8, ptr [[TMP0]], i32 11
+// ARM-LE-NEXT:    store i8 0, ptr [[TMP3]], align 1
+// ARM-LE-NEXT:    [[TMP4:%.*]] = getelementptr i8, ptr [[TMP0]], i32 12
+// ARM-LE-NEXT:    store i8 0, ptr [[TMP4]], align 4
+// ARM-LE-NEXT:    [[TMP5:%.*]] = getelementptr i8, ptr [[TMP0]], i32 13
+// ARM-LE-NEXT:    store i8 0, ptr [[TMP5]], align 1
+// ARM-LE-NEXT:    [[TMP6:%.*]] = getelementptr i8, ptr [[TMP0]], i32 14
+// ARM-LE-NEXT:    store i8 0, ptr [[TMP6]], align 2
+// ARM-LE-NEXT:    [[TMP7:%.*]] = getelementptr i8, ptr [[TMP0]], i32 15
+// ARM-LE-NEXT:    store i8 0, ptr [[TMP7]], align 1
+// ARM-LE-NEXT:    ret void
+//
+// ARM-BE-LABEL: define dso_local void @testStructWithLongDouble(
+// ARM-BE-SAME: ptr noundef [[S:%.*]]) #[[ATTR0]] {
+// ARM-BE-NEXT:  [[ENTRY:.*:]]
+// ARM-BE-NEXT:    [[S_ADDR:%.*]] = alloca ptr, align 4
+// ARM-BE-NEXT:    store ptr [[S]], ptr [[S_ADDR]], align 4
+// ARM-BE-NEXT:    [[TMP0:%.*]] = load ptr, ptr [[S_ADDR]], align 4
+// ARM-BE-NEXT:    [[TMP1:%.*]] = getelementptr i8, ptr [[TMP0]], i32 9
+// ARM-BE-NEXT:    store i8 0, ptr [[TMP1]], align 1
+// ARM-BE-NEXT:    [[TMP2:%.*]] = getelementptr i8, ptr [[TMP0]], i32 10
+// ARM-BE-NEXT:    store i8 0, ptr [[TMP2]], align 2
+// ARM-BE-NEXT:    [[TMP3:%.*]] = getelementptr i8, ptr [[TMP0]], i32 11
+// ARM-BE-NEXT:    store i8 0, ptr [[TMP3]], align 1
+// ARM-BE-NEXT:    [[TMP4:%.*]] = getelementptr i8, ptr [[TMP0]], i32 12
+// ARM-BE-NEXT:    store i8 0, ptr [[TMP4]], align 4
+// ARM-BE-NEXT:    [[TMP5:%.*]] = getelementptr i8, ptr [[TMP0]], i32 13
+// ARM-BE-NEXT:    store i8 0, ptr [[TMP5]], align 1
+// ARM-BE-NEXT:    [[TMP6:%.*]] = getelementptr i8, ptr [[TMP0]], i32 14
+// ARM-BE-NEXT:    store i8 0, ptr [[TMP6]], align 2
+// ARM-BE-NEXT:    [[TMP7:%.*]] = getelementptr i8, ptr [[TMP0]], i32 15
+// ARM-BE-NEXT:    store i8 0, ptr [[TMP7]], align 1
+// ARM-BE-NEXT:    ret void
+//
+// AARCH64-LE-LABEL: define dso_local void @testStructWithLongDouble(
+// AARCH64-LE-SAME: ptr noundef [[S:%.*]]) #[[ATTR0]] {
+// AARCH64-LE-NEXT:  [[ENTRY:.*:]]
+// AARCH64-LE-NEXT:    [[S_ADDR:%.*]] = alloca ptr, align 8
+// AARCH64-LE-NEXT:    store ptr [[S]], ptr [[S_ADDR]], align 8
+// AARCH64-LE-NEXT:    [[TMP0:%.*]] = load ptr, ptr [[S_ADDR]], align 8
+// AARCH64-LE-NEXT:    [[TMP1:%.*]] = getelementptr i8, ptr [[TMP0]], i32 17
+// AARCH64-LE-NEXT:    store i8 0, ptr [[TMP1]], align 1
+// AARCH64-LE-NEXT:    [[TMP2:%.*]] = getelementptr i8, ptr [[TMP0]], i32 18
+// AARCH64-LE-NEXT:    store i8 0, ptr [[TMP2]], align 2
+// AARCH64-LE-NEXT:    [[TMP3:%.*]] = getelementptr i8, ptr [[TMP0]], i32 19
+// AARCH64-LE-NEXT:    store i8 0, ptr [[TMP3]], align 1
+// AARCH64-LE-NEXT:    [[TMP4:%.*]] = getelementptr i8, ptr [[TMP0]], i32 20
+// AARCH64-LE-NEXT:    store i8 0, ptr [[TMP4]], align 4
+// AARCH64-LE-NEXT:    [[TMP5:%.*]] = getelementptr i8, ptr [[TMP0]], i32 21
+// AARCH64-LE-NEXT:    store i8 0, ptr [[TMP5]], align 1
+// AARCH64-LE-NEXT:    [[TMP6:%.*]] = getelementptr i8, ptr [[TMP0]], i32 22
+// AARCH64-LE-NEXT:    store i8 0, ptr [[TMP6]], align 2
+// AARCH64-LE-NEXT:    [[TMP7:%.*]] = getelementptr i8, ptr [[TMP0]], i32 23
+// AARCH64-LE-NEXT:    store i8 0, ptr [[TMP7]], align 1
+// AARCH64-LE-NEXT:    [[TMP8:%.*]] = getelementptr i8, ptr [[TMP0]], i32 24
+// AARCH64-LE-NEXT:    store i8 0, ptr [[TMP8]], align 8
+// AARCH64-LE-NEXT:    [[TMP9:%.*]] = getelementptr i8, ptr [[TMP0]], i32 25
+// AARCH64-LE-NEXT:    store i8 0, ptr [[TMP9]], align 1
+// AARCH64-LE-NEXT:    [[TMP10:%.*]] = getelementptr i8, ptr [[TMP0]], i32 26
+// AARCH64-LE-NEXT:    store i8 0, ptr [[TMP10]], align 2
+// AARCH64-LE-NEXT:    [[TMP11:%.*]] = getelementptr i8, ptr [[TMP0]], i32 27
+// AARCH64-LE-NEXT:    store i8 0, ptr [[TMP11]], align 1
+// AARCH64-LE-NEXT:    [[TMP12:%.*]] = getelementptr i8, ptr [[TMP0]], i32 28
+// AARCH64-LE-NEXT:    store i8 0, ptr [[TMP12]], align 4
+// AARCH64-LE-NEXT:    [[TMP13:%.*]] = getelementptr i8, ptr [[TMP0]], i32 29
+// AARCH64-LE-NEXT:    store i8 0, ptr [[TMP13]], align 1
+// AARCH64-LE-NEXT:    [[TMP14:%.*]] = getelementptr i8, ptr [[TMP0]], i32 30
+// AARCH64-LE-NEXT:    store i8 0, ptr [[TMP14]], align 2
+// AARCH64-LE-NEXT:    [[TMP15:%.*]] = getelementptr i8, ptr [[TMP0]], i32 31
+// AARCH64-LE-NEXT:    store i8 0, ptr [[TMP15]], align 1
+// AARCH64-LE-NEXT:    ret void
+//
+// AARCH64-BE-LABEL: define dso_local void @testStructWithLongDouble(
+// AARCH64-BE-SAME: ptr noundef [[S:%.*]]) #[[ATTR0]] {
+// AARCH64-BE-NEXT:  [[ENTRY:.*:]]
+// AARCH64-BE-NEXT:    [[S_ADDR:%.*]] = alloca ptr, align 8
+// AARCH64-BE-NEXT:    store ptr [[S]], ptr [[S_ADDR]], align 8
+// AARCH64-BE-NEXT:    [[TMP0:%.*]] = load ptr, ptr [[S_ADDR]], align 8
+// AARCH64-BE-NEXT:    [[TMP1:%.*]] = getelementptr i8, ptr [[TMP0]], i32 17
+// AARCH64-BE-NEXT:    store i8 0, ptr [[TMP1]], align 1
+// AARCH64-BE-NEXT:    [[TMP2:%.*]] = getelementptr i8, ptr [[TMP0]], i32 18
+// AARCH64-BE-NEXT:    store i8 0, ptr [[TMP2]], align 2
+// AARCH64-BE-NEXT:    [[TMP3:%.*]] = getelementptr i8, ptr [[TMP0]], i32 19
+// AARCH64-BE-NEXT:    store i8 0, ptr [[TMP3]], align 1
+// AARCH64-BE-NEXT:    [[TMP4:%.*]] = getelementptr i8, ptr [[TMP0]], i32 20
+// AARCH64-BE-NEXT:    store i8 0, ptr [[TMP4]], align 4
+// AARCH64-BE-NEXT:    [[TMP5:%.*]] = getelementptr i8, ptr [[TMP0]], i32 21
+// AARCH64-BE-NEXT:    store i8 0, ptr [[TMP5]], align 1
+// AARCH64-BE-NEXT:    [[TMP6:%.*]] = getelementptr i8, ptr [[TMP0]], i32 22
+// AARCH64-BE-NEXT:    store i8 0, ptr [[TMP6]], align 2
+// AARCH64-BE-NEXT:    [[TMP7:%.*]] = getelementptr i8, ptr [[TMP0]], i32 23
+// AARCH64-BE-NEXT:    store i8 0, ptr [[TMP7]], align 1
+// AARCH64-BE-NEXT:    [[TMP8:%.*]] = getelementptr i8, ptr [[TMP0]], i32 24
+// AARCH64-BE-NEXT:    store i8 0, ptr [[TMP8]], align 8
+// AARCH64-BE-NEXT:    [[TMP9:%.*]] = getelementptr i8, ptr [[TMP0]], i32 25
+// AARCH64-BE-NEXT:    store i8 0, ptr [[TMP9]], align 1
+// AARCH64-BE-NEXT:    [[TMP10:%.*]] = getelementptr i8, ptr [[TMP0]], i32 26
+// AARCH64-BE-NEXT:    store i8 0, ptr [[TMP10]], align 2
+// AARCH64-BE-NEXT:    [[TMP11:%.*]] = getelementptr i8, ptr [[TMP0]], i32 27
+// AARCH64-BE-NEXT:    store i8 0, ptr [[TMP11]], align 1
+// AARCH64-BE-NEXT:    [[TMP12:%.*]] = getelementptr i8, ptr [[TMP0]], i32 28
+// AARCH64-BE-NEXT:    store i8 0, ptr [[TMP12]], align 4
+// AARCH64-BE-NEXT:    [[TMP13:%.*]] = getelementptr i8, ptr [[TMP0]], i32 29
+// AARCH64-BE-NEXT:    store i8 0, ptr [[TMP13]], align 1
+// AARCH64-BE-NEXT:    [[TMP14:%.*]] = getelementptr i8, ptr [[TMP0]], i32 30
+// AARCH64-BE-NEXT:    store i8 0, ptr [[TMP14]], align 2
+// AARCH64-BE-NEXT:    [[TMP15:%.*]] = getelementptr i8, ptr [[TMP0]], i32 31
+// AARCH64-BE-NEXT:    store i8 0, ptr [[TMP15]], align 1
+// AARCH64-BE-NEXT:    ret void
+//
+extern "C" void testStructWithLongDouble(S3 *s) {
   // "long double data[0-9]", PAD [10-15], "b", PAD [17-31]
   __builtin_clear_padding(s);
 }
 
+
 struct B {
   int i;
 };
@@ -585,7 +1380,7 @@ struct S4 : Empty, B {
   bool b;
 };
 
-// LINUX-LABEL: define dso_local void @_Z23testStructWithEmptyBaseP2S4(
+// LINUX-LABEL: define dso_local void @testStructWithEmptyBase(
 // LINUX-SAME: ptr noundef [[S:%.*]]) #[[ATTR0]] {
 // LINUX-NEXT:  [[ENTRY:.*:]]
 // LINUX-NEXT:    [[S_ADDR:%.*]] = alloca ptr, align 8
@@ -599,24 +1394,82 @@ struct S4 : Empty, B {
 // LINUX-NEXT:    store i8 0, ptr [[TMP3]], align 1
 // LINUX-NEXT:    ret void
 //
-// WINDOWS-LABEL: define dso_local void @"?testStructWithEmptyBase@@YAXPEAUS4@@@Z"(ptr noundef %s) #0 {
-// WINDOWS-NEXT: entry:
-// WINDOWS-NEXT:   %s.addr = alloca ptr, align 8
-// WINDOWS-NEXT:   store ptr %s, ptr %s.addr, align 8
-// WINDOWS-NEXT:   %0 = load ptr, ptr %s.addr, align 8
-// WINDOWS-NEXT:   %1 = getelementptr i8, ptr %0, i32 5
-// WINDOWS-NEXT:   store i8 0, ptr %1, align 1
-// WINDOWS-NEXT:   %2 = getelementptr i8, ptr %0, i32 6
-// WINDOWS-NEXT:   store i8 0, ptr %2, align 2
-// WINDOWS-NEXT:   %3 = getelementptr i8, ptr %0, i32 7
-// WINDOWS-NEXT:   store i8 0, ptr %3, align 1
-// WINDOWS-NEXT:   ret void
-//
-void testStructWithEmptyBase(S4 *s) {
+// WINDOWS-LABEL: define dso_local void @testStructWithEmptyBase(
+// WINDOWS-SAME: ptr noundef [[S:%.*]]) #[[ATTR0]] {
+// WINDOWS-NEXT:  [[ENTRY:.*:]]
+// WINDOWS-NEXT:    [[S_ADDR:%.*]] = alloca ptr, align 8
+// WINDOWS-NEXT:    store ptr [[S]], ptr [[S_ADDR]], align 8
+// WINDOWS-NEXT:    [[TMP0:%.*]] = load ptr, ptr [[S_ADDR]], align 8
+// WINDOWS-NEXT:    [[TMP1:%.*]] = getelementptr i8, ptr [[TMP0]], i32 5
+// WINDOWS-NEXT:    store i8 0, ptr [[TMP1]], align 1
+// WINDOWS-NEXT:    [[TMP2:%.*]] = getelementptr i8, ptr [[TMP0]], i32 6
+// WINDOWS-NEXT:    store i8 0, ptr [[TMP2]], align 2
+// WINDOWS-NEXT:    [[TMP3:%.*]] = getelementptr i8, ptr [[TMP0]], i32 7
+// WINDOWS-NEXT:    store i8 0, ptr [[TMP3]], align 1
+// WINDOWS-NEXT:    ret void
+//
+// ARM-LE-LABEL: define dso_local void @testStructWithEmptyBase(
+// ARM-LE-SAME: ptr noundef [[S:%.*]]) #[[ATTR0]] {
+// ARM-LE-NEXT:  [[ENTRY:.*:]]
+// ARM-LE-NEXT:    [[S_ADDR:%.*]] = alloca ptr, align 4
+// ARM-LE-NEXT:    store ptr [[S]], ptr [[S_ADDR]], align 4
+// ARM-LE-NEXT:    [[TMP0:%.*]] = load ptr, ptr [[S_ADDR]], align 4
+// ARM-LE-NEXT:    [[TMP1:%.*]] = getelementptr i8, ptr [[TMP0]], i32 5
+// ARM-LE-NEXT:    store i8 0, ptr [[TMP1]], align 1
+// ARM-LE-NEXT:    [[TMP2:%.*]] = getelementptr i8, ptr [[TMP0]], i32 6
+// ARM-LE-NEXT:    store i8 0, ptr [[TMP2]], align 2
+// ARM-LE-NEXT:    [[TMP3:%.*]] = getelementptr i8, ptr [[TMP0]], i32 7
+// ARM-LE-NEXT:    store i8 0, ptr [[TMP3]], align 1
+// ARM-LE-NEXT:    ret void
+//
+// ARM-BE-LABEL: define dso_local void @testStructWithEmptyBase(
+// ARM-BE-SAME: ptr noundef [[S:%.*]]) #[[ATTR0]] {
+// ARM-BE-NEXT:  [[ENTRY:.*:]]
+// ARM-BE-NEXT:    [[S_ADDR:%.*]] = alloca ptr, align 4
+// ARM-BE-NEXT:    store ptr [[S]], ptr [[S_ADDR]], align 4
+// ARM-BE-NEXT:    [[TMP0:%.*]] = load ptr, ptr [[S_ADDR]], align 4
+// ARM-BE-NEXT:    [[TMP1:%.*]] = getelementptr i8, ptr [[TMP0]], i32 5
+// ARM-BE-NEXT:    store i8 0, ptr [[TMP1]], align 1
+// ARM-BE-NEXT:    [[TMP2:%.*]] = getelementptr i8, ptr [[TMP0]], i32 6
+// ARM-BE-NEXT:    store i8 0, ptr [[TMP2]], align 2
+// ARM-BE-NEXT:    [[TMP3:%.*]] = getelementptr i8, ptr [[TMP0]], i32 7
+// ARM-BE-NEXT:    store i8 0, ptr [[TMP3]], align 1
+// ARM-BE-NEXT:    ret void
+//
+// AARCH64-LE-LABEL: define dso_local void @testStructWithEmptyBase(
+// AARCH64-LE-SAME: ptr noundef [[S:%.*]]) #[[ATTR0]] {
+// AARCH64-LE-NEXT:  [[ENTRY:.*:]]
+// AARCH64-LE-NEXT:    [[S_ADDR:%.*]] = alloca ptr, align 8
+// AARCH64-LE-NEXT:    store ptr [[S]], ptr [[S_ADDR]], align 8
+// AARCH64-LE-NEXT:    [[TMP0:%.*]] = load ptr, ptr [[S_ADDR]], align 8
+// AARCH64-LE-NEXT:    [[TMP1:%.*]] = getelementptr i8, ptr [[TMP0]], i32 5
+// AARCH64-LE-NEXT:    store i8 0, ptr [[TMP1]], align 1
+// AARCH64-LE-NEXT:    [[TMP2:%.*]] = getelementptr i8, ptr [[TMP0]], i32 6
+// AARCH64-LE-NEXT:    store i8 0, ptr [[TMP2]], align 2
+// AARCH64-LE-NEXT:    [[TMP3:%.*]] = getelementptr i8, ptr [[TMP0]], i32 7
+// AARCH64-LE-NEXT:    store i8 0, ptr [[TMP3]], align 1
+// AARCH64-LE-NEXT:    ret void
+//
+// AARCH64-BE-LABEL: define dso_local void @testStructWithEmptyBase(
+// AARCH64-BE-SAME: ptr noundef [[S:%.*]]) #[[ATTR0]] {
+// AARCH64-BE-NEXT:  [[ENTRY:.*:]]
+// AARCH64-BE-NEXT:    [[S_ADDR:%.*]] = alloca ptr, align 8
+// AARCH64-BE-NEXT:    store ptr [[S]], ptr [[S_ADDR]], align 8
+// AARCH64-BE-NEXT:    [[TMP0:%.*]] = load ptr, ptr [[S_ADDR]], align 8
+// AARCH64-BE-NEXT:    [[TMP1:%.*]] = getelementptr i8, ptr [[TMP0]], i32 5
+// AARCH64-BE-NEXT:    store i8 0, ptr [[TMP1]], align 1
+// AARCH64-BE-NEXT:    [[TMP2:%.*]] = getelementptr i8, ptr [[TMP0]], i32 6
+// AARCH64-BE-NEXT:    store i8 0, ptr [[TMP2]], align 2
+// AARCH64-BE-NEXT:    [[TMP3:%.*]] = getelementptr i8, ptr [[TMP0]], i32 7
+// AARCH64-BE-NEXT:    store i8 0, ptr [[TMP3]], align 1
+// AARCH64-BE-NEXT:    ret void
+//
+extern "C" void testStructWithEmptyBase(S4 *s) {
   // "i" [0-3], "b" [4], PAD [5-7]
   __builtin_clear_padding(s);
 }
 
+
 struct B1 {
   char c1;
 };
@@ -626,7 +1479,7 @@ struct B2 {
 struct S5 : B1, B2 {
 };
 
-// LINUX-LABEL: define dso_local void @_Z23testPaddingBetweenBasesP2S5(
+// LINUX-LABEL: define dso_local void @testPaddingBetweenBases(
 // LINUX-SAME: ptr noundef [[S:%.*]]) #[[ATTR0]] {
 // LINUX-NEXT:  [[ENTRY:.*:]]
 // LINUX-NEXT:    [[S_ADDR:%.*]] = alloca ptr, align 8
@@ -646,30 +1499,112 @@ struct S5 : B1, B2 {
 // LINUX-NEXT:    store i8 0, ptr [[TMP6]], align 1
 // LINUX-NEXT:    ret void
 //
-// WINDOWS-LABEL: define dso_local void @"?testPaddingBetweenBases@@YAXPEAUS5@@@Z"(ptr noundef %s) #0 {
-// WINDOWS-NEXT: entry:
-// WINDOWS-NEXT:   %s.addr = alloca ptr, align 8
-// WINDOWS-NEXT:   store ptr %s, ptr %s.addr, align 8
-// WINDOWS-NEXT:   %0 = load ptr, ptr %s.addr, align 8
-// WINDOWS-NEXT:   %1 = getelementptr i8, ptr %0, i32 1
-// WINDOWS-NEXT:   store i8 0, ptr %1, align 1
-// WINDOWS-NEXT:   %2 = getelementptr i8, ptr %0, i32 2
-// WINDOWS-NEXT:   store i8 0, ptr %2, align 2
-// WINDOWS-NEXT:   %3 = getelementptr i8, ptr %0, i32 3
-// WINDOWS-NEXT:   store i8 0, ptr %3, align 1
-// WINDOWS-NEXT:   %4 = getelementptr i8, ptr %0, i32 5
-// WINDOWS-NEXT:   store i8 0, ptr %4, align 1
-// WINDOWS-NEXT:   %5 = getelementptr i8, ptr %0, i32 6
-// WINDOWS-NEXT:   store i8 0, ptr %5, align 2
-// WINDOWS-NEXT:   %6 = getelementptr i8, ptr %0, i32 7
-// WINDOWS-NEXT:   store i8 0, ptr %6, align 1
-// WINDOWS-NEXT:   ret void
-//
-void testPaddingBetweenBases(S5 *s) {
+// WINDOWS-LABEL: define dso_local void @testPaddingBetweenBases(
+// WINDOWS-SAME: ptr noundef [[S:%.*]]) #[[ATTR0]] {
+// WINDOWS-NEXT:  [[ENTRY:.*:]]
+// WINDOWS-NEXT:    [[S_ADDR:%.*]] = alloca ptr, align 8
+// WINDOWS-NEXT:    store ptr [[S]], ptr [[S_ADDR]], align 8
+// WINDOWS-NEXT:    [[TMP0:%.*]] = load ptr, ptr [[S_ADDR]], align 8
+// WINDOWS-NEXT:    [[TMP1:%.*]] = getelementptr i8, ptr [[TMP0]], i32 1
+// WINDOWS-NEXT:    store i8 0, ptr [[TMP1]], align 1
+// WINDOWS-NEXT:    [[TMP2:%.*]] = getelementptr i8, ptr [[TMP0]], i32 2
+// WINDOWS-NEXT:    store i8 0, ptr [[TMP2]], align 2
+// WINDOWS-NEXT:    [[TMP3:%.*]] = getelementptr i8, ptr [[TMP0]], i32 3
+// WINDOWS-NEXT:    store i8 0, ptr [[TMP3]], align 1
+// WINDOWS-NEXT:    [[TMP4:%.*]] = getelementptr i8, ptr [[TMP0]], i32 5
+// WINDOWS-NEXT:    store i8 0, ptr [[TMP4]], align 1
+// WINDOWS-NEXT:    [[TMP5:%.*]] = getelementptr i8, ptr [[TMP0]], i32 6
+// WINDOWS-NEXT:    store i8 0, ptr [[TMP5]], align 2
+// WINDOWS-NEXT:    [[TMP6:%.*]] = getelementptr i8, ptr [[TMP0]], i32 7
+// WINDOWS-NEXT:    store i8 0, ptr [[TMP6]], align 1
+// WINDOWS-NEXT:    ret void
+//
+// ARM-LE-LABEL: define dso_local void @testPaddingBetweenBases(
+// ARM-LE-SAME: ptr noundef [[S:%.*]]) #[[ATTR0]] {
+// ARM-LE-NEXT:  [[ENTRY:.*:]]
+// ARM-LE-NEXT:    [[S_ADDR:%.*]] = alloca ptr, align 4
+// ARM-LE-NEXT:    store ptr [[S]], ptr [[S_ADDR]], align 4
+// ARM-LE-NEXT:    [[TMP0:%.*]] = load ptr, ptr [[S_ADDR]], align 4
+// ARM-LE-NEXT:    [[TMP1:%.*]] = getelementptr i8, ptr [[TMP0]], i32 1
+// ARM-LE-NEXT:    store i8 0, ptr [[TMP1]], align 1
+// ARM-LE-NEXT:    [[TMP2:%.*]] = getelementptr i8, ptr [[TMP0]], i32 2
+// ARM-LE-NEXT:    store i8 0, ptr [[TMP2]], align 2
+// ARM-LE-NEXT:    [[TMP3:%.*]] = getelementptr i8, ptr [[TMP0]], i32 3
+// ARM-LE-NEXT:    store i8 0, ptr [[TMP3]], align 1
+// ARM-LE-NEXT:    [[TMP4:%.*]] = getelementptr i8, ptr [[TMP0]], i32 5
+// ARM-LE-NEXT:    store i8 0, ptr [[TMP4]], align 1
+// ARM-LE-NEXT:    [[TMP5:%.*]] = getelementptr i8, ptr [[TMP0]], i32 6
+// ARM-LE-NEXT:    store i8 0, ptr [[TMP5]], align 2
+// ARM-LE-NEXT:    [[TMP6:%.*]] = getelementptr i8, ptr [[TMP0]], i32 7
+// ARM-LE-NEXT:    store i8 0, ptr [[TMP6]], align 1
+// ARM-LE-NEXT:    ret void
+//
+// ARM-BE-LABEL: define dso_local void @testPaddingBetweenBases(
+// ARM-BE-SAME: ptr noundef [[S:%.*]]) #[[ATTR0]] {
+// ARM-BE-NEXT:  [[ENTRY:.*:]]
+// ARM-BE-NEXT:    [[S_ADDR:%.*]] = alloca ptr, align 4
+// ARM-BE-NEXT:    store ptr [[S]], ptr [[S_ADDR]], align 4
+// ARM-BE-NEXT:    [[TMP0:%.*]] = load ptr, ptr [[S_ADDR]], align 4
+// ARM-BE-NEXT:    [[TMP1:%.*]] = getelementptr i8, ptr [[TMP0]], i32 1
+// ARM-BE-NEXT:    store i8 0, ptr [[TMP1]], align 1
+// ARM-BE-NEXT:    [[TMP2:%.*]] = getelementptr i8, ptr [[TMP0]], i32 2
+// ARM-BE-NEXT:    store i8 0, ptr [[TMP2]], align 2
+// ARM-BE-NEXT:    [[TMP3:%.*]] = getelementptr i8, ptr [[TMP0]], i32 3
+// ARM-BE-NEXT:    store i8 0, ptr [[TMP3]], align 1
+// ARM-BE-NEXT:    [[TMP4:%.*]] = getelementptr i8, ptr [[TMP0]], i32 5
+// ARM-BE-NEXT:    store i8 0, ptr [[TMP4]], align 1
+// ARM-BE-NEXT:    [[TMP5:%.*]] = getelementptr i8, ptr [[TMP0]], i32 6
+// ARM-BE-NEXT:    store i8 0, ptr [[TMP5]], align 2
+// ARM-BE-NEXT:    [[TMP6:%.*]] = getelementptr i8, ptr [[TMP0]], i32 7
+// ARM-BE-NEXT:    store i8 0, ptr [[TMP6]], align 1
+// ARM-BE-NEXT:    ret void
+//
+// AARCH64-LE-LABEL: define dso_local void @testPaddingBetweenBases(
+// AARCH64-LE-SAME: ptr noundef [[S:%.*]]) #[[ATTR0]] {
+// AARCH64-LE-NEXT:  [[ENTRY:.*:]]
+// AARCH64-LE-NEXT:    [[S_ADDR:%.*]] = alloca ptr, align 8
+// AARCH64-LE-NEXT:    store ptr [[S]], ptr [[S_ADDR]], align 8
+// AARCH64-LE-NEXT:    [[TMP0:%.*]] = load ptr, ptr [[S_ADDR]], align 8
+// AARCH64-LE-NEXT:    [[TMP1:%.*]] = getelementptr i8, ptr [[TMP0]], i32 1
+// AARCH64-LE-NEXT:    store i8 0, ptr [[TMP1]], align 1
+// AARCH64-LE-NEXT:    [[TMP2:%.*]] = getelementptr i8, ptr [[TMP0]], i32 2
+// AARCH64-LE-NEXT:    store i8 0, ptr [[TMP2]], align 2
+// AARCH64-LE-NEXT:    [[TMP3:%.*]] = getelementptr i8, ptr [[TMP0]], i32 3
+// AARCH64-LE-NEXT:    store i8 0, ptr [[TMP3]], align 1
+// AARCH64-LE-NEXT:    [[TMP4:%.*]] = getelementptr i8, ptr [[TMP0]], i32 5
+// AARCH64-LE-NEXT:    store i8 0, ptr [[TMP4]], align 1
+// AARCH64-LE-NEXT:    [[TMP5:%.*]] = getelementptr i8, ptr [[TMP0]], i32 6
+// AARCH64-LE-NEXT:    store i8 0, ptr [[TMP5]], align 2
+// AARCH64-LE-NEXT:    [[TMP6:%.*]] = getelementptr i8, ptr [[TMP0]], i32 7
+// AARCH64-LE-NEXT:    store i8 0, ptr [[TMP6]], align 1
+// AARCH64-LE-NEXT:    ret void
+//
+// AARCH64-BE-LABEL: define dso_local void @testPaddingBetweenBases(
+// AARCH64-BE-SAME: ptr noundef [[S:%.*]]) #[[ATTR0]] {
+// AARCH64-BE-NEXT:  [[ENTRY:.*:]]
+// AARCH64-BE-NEXT:    [[S_ADDR:%.*]] = alloca ptr, align 8
+// AARCH64-BE-NEXT:    store ptr [[S]], ptr [[S_ADDR]], align 8
+// AARCH64-BE-NEXT:    [[TMP0:%.*]] = load ptr, ptr [[S_ADDR]], align 8
+// AARCH64-BE-NEXT:    [[TMP1:%.*]] = getelementptr i8, ptr [[TMP0]], i32 1
+// AARCH64-BE-NEXT:    store i8 0, ptr [[TMP1]], align 1
+// AARCH64-BE-NEXT:    [[TMP2:%.*]] = getelementptr i8, ptr [[TMP0]], i32 2
+// AARCH64-BE-NEXT:    store i8 0, ptr [[TMP2]], align 2
+// AARCH64-BE-NEXT:    [[TMP3:%.*]] = getelementptr i8, ptr [[TMP0]], i32 3
+// AARCH64-BE-NEXT:    store i8 0, ptr [[TMP3]], align 1
+// AARCH64-BE-NEXT:    [[TMP4:%.*]] = getelementptr i8, ptr [[TMP0]], i32 5
+// AARCH64-BE-NEXT:    store i8 0, ptr [[TMP4]], align 1
+// AARCH64-BE-NEXT:    [[TMP5:%.*]] = getelementptr i8, ptr [[TMP0]], i32 6
+// AARCH64-BE-NEXT:    store i8 0, ptr [[TMP5]], align 2
+// AARCH64-BE-NEXT:    [[TMP6:%.*]] = getelementptr i8, ptr [[TMP0]], i32 7
+// AARCH64-BE-NEXT:    store i8 0, ptr [[TMP6]], align 1
+// AARCH64-BE-NEXT:    ret void
+//
+extern "C" void testPaddingBetweenBases(S5 *s) {
   // "c1" [0], PAD [1-3] , "c2" [4], PAD [5-7]
   __builtin_clear_padding(s);
 }
 
+
 struct B3 {
   char c1;
 };
@@ -682,7 +1617,7 @@ struct S6 : B3, B4 {
   alignas(4) char c3;
 };
 
-// LINUX-LABEL: define dso_local void @_Z24testPaddingAfterLastBaseP2S6(
+// LINUX-LABEL: define dso_local void @testPaddingAfterLastBase(
 // LINUX-SAME: ptr noundef [[S:%.*]]) #[[ATTR0]] {
 // LINUX-NEXT:  [[ENTRY:.*:]]
 // LINUX-NEXT:    [[S_ADDR:%.*]] = alloca ptr, align 8
@@ -700,32 +1635,108 @@ struct S6 : B3, B4 {
 // LINUX-NEXT:    store i8 0, ptr [[TMP5]], align 1
 // LINUX-NEXT:    ret void
 //
-// WINDOWS-LABEL: define dso_local void @"?testPaddingAfterLastBase@@YAXPEAUS6@@@Z"(ptr noundef %s) #0 {
-// WINDOWS-NEXT: entry:
-// WINDOWS-NEXT:   %s.addr = alloca ptr, align 8
-// WINDOWS-NEXT:   store ptr %s, ptr %s.addr, align 8
-// WINDOWS-NEXT:   %0 = load ptr, ptr %s.addr, align 8
-// WINDOWS-NEXT:   %1 = getelementptr i8, ptr %0, i32 2
-// WINDOWS-NEXT:   store i8 0, ptr %1, align 2
-// WINDOWS-NEXT:   %2 = getelementptr i8, ptr %0, i32 3
-// WINDOWS-NEXT:   store i8 0, ptr %2, align 1
-// WINDOWS-NEXT:   %3 = getelementptr i8, ptr %0, i32 5
-// WINDOWS-NEXT:   store i8 0, ptr %3, align 1
-// WINDOWS-NEXT:   %4 = getelementptr i8, ptr %0, i32 6
-// WINDOWS-NEXT:   store i8 0, ptr %4, align 2
-// WINDOWS-NEXT:   %5 = getelementptr i8, ptr %0, i32 7
-// WINDOWS-NEXT:   store i8 0, ptr %5, align 1
-// WINDOWS-NEXT:   ret void
-//
-void testPaddingAfterLastBase(S6 *s) {
+// WINDOWS-LABEL: define dso_local void @testPaddingAfterLastBase(
+// WINDOWS-SAME: ptr noundef [[S:%.*]]) #[[ATTR0]] {
+// WINDOWS-NEXT:  [[ENTRY:.*:]]
+// WINDOWS-NEXT:    [[S_ADDR:%.*]] = alloca ptr, align 8
+// WINDOWS-NEXT:    store ptr [[S]], ptr [[S_ADDR]], align 8
+// WINDOWS-NEXT:    [[TMP0:%.*]] = load ptr, ptr [[S_ADDR]], align 8
+// WINDOWS-NEXT:    [[TMP1:%.*]] = getelementptr i8, ptr [[TMP0]], i32 2
+// WINDOWS-NEXT:    store i8 0, ptr [[TMP1]], align 2
+// WINDOWS-NEXT:    [[TMP2:%.*]] = getelementptr i8, ptr [[TMP0]], i32 3
+// WINDOWS-NEXT:    store i8 0, ptr [[TMP2]], align 1
+// WINDOWS-NEXT:    [[TMP3:%.*]] = getelementptr i8, ptr [[TMP0]], i32 5
+// WINDOWS-NEXT:    store i8 0, ptr [[TMP3]], align 1
+// WINDOWS-NEXT:    [[TMP4:%.*]] = getelementptr i8, ptr [[TMP0]], i32 6
+// WINDOWS-NEXT:    store i8 0, ptr [[TMP4]], align 2
+// WINDOWS-NEXT:    [[TMP5:%.*]] = getelementptr i8, ptr [[TMP0]], i32 7
+// WINDOWS-NEXT:    store i8 0, ptr [[TMP5]], align 1
+// WINDOWS-NEXT:    ret void
+//
+// ARM-LE-LABEL: define dso_local void @testPaddingAfterLastBase(
+// ARM-LE-SAME: ptr noundef [[S:%.*]]) #[[ATTR0]] {
+// ARM-LE-NEXT:  [[ENTRY:.*:]]
+// ARM-LE-NEXT:    [[S_ADDR:%.*]] = alloca ptr, align 4
+// ARM-LE-NEXT:    store ptr [[S]], ptr [[S_ADDR]], align 4
+// ARM-LE-NEXT:    [[TMP0:%.*]] = load ptr, ptr [[S_ADDR]], align 4
+// ARM-LE-NEXT:    [[TMP1:%.*]] = getelementptr i8, ptr [[TMP0]], i32 2
+// ARM-LE-NEXT:    store i8 0, ptr [[TMP1]], align 2
+// ARM-LE-NEXT:    [[TMP2:%.*]] = getelementptr i8, ptr [[TMP0]], i32 3
+// ARM-LE-NEXT:    store i8 0, ptr [[TMP2]], align 1
+// ARM-LE-NEXT:    [[TMP3:%.*]] = getelementptr i8, ptr [[TMP0]], i32 5
+// ARM-LE-NEXT:    store i8 0, ptr [[TMP3]], align 1
+// ARM-LE-NEXT:    [[TMP4:%.*]] = getelementptr i8, ptr [[TMP0]], i32 6
+// ARM-LE-NEXT:    store i8 0, ptr [[TMP4]], align 2
+// ARM-LE-NEXT:    [[TMP5:%.*]] = getelementptr i8, ptr [[TMP0]], i32 7
+// ARM-LE-NEXT:    store i8 0, ptr [[TMP5]], align 1
+// ARM-LE-NEXT:    ret void
+//
+// ARM-BE-LABEL: define dso_local void @testPaddingAfterLastBase(
+// ARM-BE-SAME: ptr noundef [[S:%.*]]) #[[ATTR0]] {
+// ARM-BE-NEXT:  [[ENTRY:.*:]]
+// ARM-BE-NEXT:    [[S_ADDR:%.*]] = alloca ptr, align 4
+// ARM-BE-NEXT:    store ptr [[S]], ptr [[S_ADDR]], align 4
+// ARM-BE-NEXT:    [[TMP0:%.*]] = load ptr, ptr [[S_ADDR]], align 4
+// ARM-BE-NEXT:    [[TMP1:%.*]] = getelementptr i8, ptr [[TMP0]], i32 2
+// ARM-BE-NEXT:    store i8 0, ptr [[TMP1]], align 2
+// ARM-BE-NEXT:    [[TMP2:%.*]] = getelementptr i8, ptr [[TMP0]], i32 3
+// ARM-BE-NEXT:    store i8 0, ptr [[TMP2]], align 1
+// ARM-BE-NEXT:    [[TMP3:%.*]] = getelementptr i8, ptr [[TMP0]], i32 5
+// ARM-BE-NEXT:    store i8 0, ptr [[TMP3]], align 1
+// ARM-BE-NEXT:    [[TMP4:%.*]] = getelementptr i8, ptr [[TMP0]], i32 6
+// ARM-BE-NEXT:    store i8 0, ptr [[TMP4]], align 2
+// ARM-BE-NEXT:    [[TMP5:%.*]] = getelementptr i8, ptr [[TMP0]], i32 7
+// ARM-BE-NEXT:    store i8 0, ptr [[TMP5]], align 1
+// ARM-BE-NEXT:    ret void
+//
+// AARCH64-LE-LABEL: define dso_local void @testPaddingAfterLastBase(
+// AARCH64-LE-SAME: ptr noundef [[S:%.*]]) #[[ATTR0]] {
+// AARCH64-LE-NEXT:  [[ENTRY:.*:]]
+// AARCH64-LE-NEXT:    [[S_ADDR:%.*]] = alloca ptr, align 8
+// AARCH64-LE-NEXT:    store ptr [[S]], ptr [[S_ADDR]], align 8
+// AARCH64-LE-NEXT:    [[TMP0:%.*]] = load ptr, ptr [[S_ADDR]], align 8
+// AARCH64-LE-NEXT:    [[TMP1:%.*]] = getelementptr i8, ptr [[TMP0]], i32 2
+// AARCH64-LE-NEXT:    store i8 0, ptr [[TMP1]], align 2
+// AARCH64-LE-NEXT:    [[TMP2:%.*]] = getelementptr i8, ptr [[TMP0]], i32 3
+// AARCH64-LE-NEXT:    store i8 0, ptr [[TMP2]], align 1
+// AARCH64-LE-NEXT:    [[TMP3:%.*]] = getelementptr i8, ptr [[TMP0]], i32 5
+// AARCH64-LE-NEXT:    store i8 0, ptr [[TMP3]], align 1
+// AARCH64-LE-NEXT:    [[TMP4:%.*]] = getelementptr i8, ptr [[TMP0]], i32 6
+// AARCH64-LE-NEXT:    store i8 0, ptr [[TMP4]], align 2
+// AARCH64-LE-NEXT:    [[TMP5:%.*]] = getelementptr i8, ptr [[TMP0]], i32 7
+// AARCH64-LE-NEXT:    store i8 0, ptr [[TMP5]], align 1
+// AARCH64-LE-NEXT:    ret void
+//
+// AARCH64-BE-LABEL: define dso_local void @testPaddingAfterLastBase(
+// AARCH64-BE-SAME: ptr noundef [[S:%.*]]) #[[ATTR0]] {
+// AARCH64-BE-NEXT:  [[ENTRY:.*:]]
+// AARCH64-BE-NEXT:    [[S_ADDR:%.*]] = alloca ptr, align 8
+// AARCH64-BE-NEXT:    store ptr [[S]], ptr [[S_ADDR]], align 8
+// AARCH64-BE-NEXT:    [[TMP0:%.*]] = load ptr, ptr [[S_ADDR]], align 8
+// AARCH64-BE-NEXT:    [[TMP1:%.*]] = getelementptr i8, ptr [[TMP0]], i32 2
+// AARCH64-BE-NEXT:    store i8 0, ptr [[TMP1]], align 2
+// AARCH64-BE-NEXT:    [[TMP2:%.*]] = getelementptr i8, ptr [[TMP0]], i32 3
+// AARCH64-BE-NEXT:    store i8 0, ptr [[TMP2]], align 1
+// AARCH64-BE-NEXT:    [[TMP3:%.*]] = getelementptr i8, ptr [[TMP0]], i32 5
+// AARCH64-BE-NEXT:    store i8 0, ptr [[TMP3]], align 1
+// AARCH64-BE-NEXT:    [[TMP4:%.*]] = getelementptr i8, ptr [[TMP0]], i32 6
+// AARCH64-BE-NEXT:    store i8 0, ptr [[TMP4]], align 2
+// AARCH64-BE-NEXT:    [[TMP5:%.*]] = getelementptr i8, ptr [[TMP0]], i32 7
+// AARCH64-BE-NEXT:    store i8 0, ptr [[TMP5]], align 1
+// AARCH64-BE-NEXT:    ret void
+//
+extern "C" void testPaddingAfterLastBase(S6 *s) {
   // "c1"[0], "c2"[1], PAD [2-3], "c3" [4], PAD [5-7]
   __builtin_clear_padding(s);
 }
 
 
+
 struct VirtualBase {
   unsigned int x;
+//
   virtual int call() { return x; };
+//
   virtual ~VirtualBase() = default;
 };
 
@@ -734,11 +1745,12 @@ struct NonVirtualBase {
 };
 
 struct S7 : VirtualBase, NonVirtualBase {
+//
   virtual int call() override { return 5; }
   bool z;
 };
 
-// LINUX-LABEL: define dso_local void @_Z10testVtable2S7(
+// LINUX-LABEL: define dso_local void @testVtable(
 // LINUX-SAME: ptr nofreeobj noundef align 8 dereferenceable(16) [[S:%.*]]) #[[ATTR0]] {
 // LINUX-NEXT:  [[ENTRY:.*:]]
 // LINUX-NEXT:    [[S_INDIRECT_ADDR:%.*]] = alloca ptr, align 8
@@ -749,52 +1761,101 @@ struct S7 : VirtualBase, NonVirtualBase {
 // LINUX-NEXT:    store i8 0, ptr [[TMP1]], align 1
 // LINUX-NEXT:    ret void
 //
-// WINDOWS-LABEL: define dso_local void @"?testVtable@@YAXUS7@@@Z"(ptr nofreeobj noundef align 8 dead_on_return dereferenceable(24) %s) #0 {
-// WINDOWS-NEXT: entry:
-// WINDOWS-NEXT:   %s.indirect_addr = alloca ptr, align 8
-// WINDOWS-NEXT:   store ptr %s, ptr %s.indirect_addr, align 8
-// WINDOWS-NEXT:   %0 = getelementptr i8, ptr %s, i32 12
-// WINDOWS-NEXT:   store i8 0, ptr %0, align 4
-// WINDOWS-NEXT:   %1 = getelementptr i8, ptr %s, i32 13
-// WINDOWS-NEXT:   store i8 0, ptr %1, align 1
-// WINDOWS-NEXT:   %2 = getelementptr i8, ptr %s, i32 14
-// WINDOWS-NEXT:   store i8 0, ptr %2, align 2
-// WINDOWS-NEXT:   %3 = getelementptr i8, ptr %s, i32 15
-// WINDOWS-NEXT:   store i8 0, ptr %3, align 1
-// WINDOWS-NEXT:   %4 = getelementptr i8, ptr %s, i32 18
-// WINDOWS-NEXT:   store i8 0, ptr %4, align 2
-// WINDOWS-NEXT:   %5 = getelementptr i8, ptr %s, i32 19
-// WINDOWS-NEXT:   store i8 0, ptr %5, align 1
-// WINDOWS-NEXT:   %6 = getelementptr i8, ptr %s, i32 20
-// WINDOWS-NEXT:   store i8 0, ptr %6, align 4
-// WINDOWS-NEXT:   %7 = getelementptr i8, ptr %s, i32 21
-// WINDOWS-NEXT:   store i8 0, ptr %7, align 1
-// WINDOWS-NEXT:   %8 = getelementptr i8, ptr %s, i32 22
-// WINDOWS-NEXT:   store i8 0, ptr %8, align 2
-// WINDOWS-NEXT:   %9 = getelementptr i8, ptr %s, i32 23
-// WINDOWS-NEXT:   store i8 0, ptr %9, align 1
-// WINDOWS-NEXT:   call void @"??1S7@@UEAA at XZ"(ptr noundef nonnull align 8 dead_on_return(24) dereferenceable(24) %s) #1
-// WINDOWS-NEXT:   ret void
-//
-void testVtable(S7 s) {
+// WINDOWS-LABEL: define dso_local void @testVtable(
+// WINDOWS-SAME: ptr nofreeobj noundef align 8 dead_on_return dereferenceable(24) [[S:%.*]]) #[[ATTR0]] {
+// WINDOWS-NEXT:  [[ENTRY:.*:]]
+// WINDOWS-NEXT:    [[S_INDIRECT_ADDR:%.*]] = alloca ptr, align 8
+// WINDOWS-NEXT:    store ptr [[S]], ptr [[S_INDIRECT_ADDR]], align 8
+// WINDOWS-NEXT:    [[TMP0:%.*]] = getelementptr i8, ptr [[S]], i32 12
+// WINDOWS-NEXT:    store i8 0, ptr [[TMP0]], align 4
+// WINDOWS-NEXT:    [[TMP1:%.*]] = getelementptr i8, ptr [[S]], i32 13
+// WINDOWS-NEXT:    store i8 0, ptr [[TMP1]], align 1
+// WINDOWS-NEXT:    [[TMP2:%.*]] = getelementptr i8, ptr [[S]], i32 14
+// WINDOWS-NEXT:    store i8 0, ptr [[TMP2]], align 2
+// WINDOWS-NEXT:    [[TMP3:%.*]] = getelementptr i8, ptr [[S]], i32 15
+// WINDOWS-NEXT:    store i8 0, ptr [[TMP3]], align 1
+// WINDOWS-NEXT:    [[TMP4:%.*]] = getelementptr i8, ptr [[S]], i32 18
+// WINDOWS-NEXT:    store i8 0, ptr [[TMP4]], align 2
+// WINDOWS-NEXT:    [[TMP5:%.*]] = getelementptr i8, ptr [[S]], i32 19
+// WINDOWS-NEXT:    store i8 0, ptr [[TMP5]], align 1
+// WINDOWS-NEXT:    [[TMP6:%.*]] = getelementptr i8, ptr [[S]], i32 20
+// WINDOWS-NEXT:    store i8 0, ptr [[TMP6]], align 4
+// WINDOWS-NEXT:    [[TMP7:%.*]] = getelementptr i8, ptr [[S]], i32 21
+// WINDOWS-NEXT:    store i8 0, ptr [[TMP7]], align 1
+// WINDOWS-NEXT:    [[TMP8:%.*]] = getelementptr i8, ptr [[S]], i32 22
+// WINDOWS-NEXT:    store i8 0, ptr [[TMP8]], align 2
+// WINDOWS-NEXT:    [[TMP9:%.*]] = getelementptr i8, ptr [[S]], i32 23
+// WINDOWS-NEXT:    store i8 0, ptr [[TMP9]], align 1
+// WINDOWS-NEXT:    call void @"??1S7@@UEAA at XZ"(ptr noundef nonnull align 8 dead_on_return(24) dereferenceable(24) [[S]]) #[[ATTR1:[0-9]+]]
+// WINDOWS-NEXT:    ret void
+//
+// ARM-LE-LABEL: define dso_local void @testVtable(
+// ARM-LE-SAME: ptr nofreeobj noundef align 4 dereferenceable(12) [[S:%.*]]) #[[ATTR0]] {
+// ARM-LE-NEXT:  [[ENTRY:.*:]]
+// ARM-LE-NEXT:    [[S_INDIRECT_ADDR:%.*]] = alloca ptr, align 4
+// ARM-LE-NEXT:    store ptr [[S]], ptr [[S_INDIRECT_ADDR]], align 4
+// ARM-LE-NEXT:    [[TMP0:%.*]] = getelementptr i8, ptr [[S]], i32 10
+// ARM-LE-NEXT:    store i8 0, ptr [[TMP0]], align 2
+// ARM-LE-NEXT:    [[TMP1:%.*]] = getelementptr i8, ptr [[S]], i32 11
+// ARM-LE-NEXT:    store i8 0, ptr [[TMP1]], align 1
+// ARM-LE-NEXT:    ret void
+//
+// ARM-BE-LABEL: define dso_local void @testVtable(
+// ARM-BE-SAME: ptr nofreeobj noundef align 4 dereferenceable(12) [[S:%.*]]) #[[ATTR0]] {
+// ARM-BE-NEXT:  [[ENTRY:.*:]]
+// ARM-BE-NEXT:    [[S_INDIRECT_ADDR:%.*]] = alloca ptr, align 4
+// ARM-BE-NEXT:    store ptr [[S]], ptr [[S_INDIRECT_ADDR]], align 4
+// ARM-BE-NEXT:    [[TMP0:%.*]] = getelementptr i8, ptr [[S]], i32 10
+// ARM-BE-NEXT:    store i8 0, ptr [[TMP0]], align 2
+// ARM-BE-NEXT:    [[TMP1:%.*]] = getelementptr i8, ptr [[S]], i32 11
+// ARM-BE-NEXT:    store i8 0, ptr [[TMP1]], align 1
+// ARM-BE-NEXT:    ret void
+//
+// AARCH64-LE-LABEL: define dso_local void @testVtable(
+// AARCH64-LE-SAME: ptr nofreeobj noundef align 8 dereferenceable(16) [[S:%.*]]) #[[ATTR0]] {
+// AARCH64-LE-NEXT:  [[ENTRY:.*:]]
+// AARCH64-LE-NEXT:    [[S_INDIRECT_ADDR:%.*]] = alloca ptr, align 8
+// AARCH64-LE-NEXT:    store ptr [[S]], ptr [[S_INDIRECT_ADDR]], align 8
+// AARCH64-LE-NEXT:    [[TMP0:%.*]] = getelementptr i8, ptr [[S]], i32 14
+// AARCH64-LE-NEXT:    store i8 0, ptr [[TMP0]], align 2
+// AARCH64-LE-NEXT:    [[TMP1:%.*]] = getelementptr i8, ptr [[S]], i32 15
+// AARCH64-LE-NEXT:    store i8 0, ptr [[TMP1]], align 1
+// AARCH64-LE-NEXT:    ret void
+//
+// AARCH64-BE-LABEL: define dso_local void @testVtable(
+// AARCH64-BE-SAME: ptr nofreeobj noundef align 8 dereferenceable(16) [[S:%.*]]) #[[ATTR0]] {
+// AARCH64-BE-NEXT:  [[ENTRY:.*:]]
+// AARCH64-BE-NEXT:    [[S_INDIRECT_ADDR:%.*]] = alloca ptr, align 8
+// AARCH64-BE-NEXT:    store ptr [[S]], ptr [[S_INDIRECT_ADDR]], align 8
+// AARCH64-BE-NEXT:    [[TMP0:%.*]] = getelementptr i8, ptr [[S]], i32 14
+// AARCH64-BE-NEXT:    store i8 0, ptr [[TMP0]], align 2
+// AARCH64-BE-NEXT:    [[TMP1:%.*]] = getelementptr i8, ptr [[S]], i32 15
+// AARCH64-BE-NEXT:    store i8 0, ptr [[TMP1]], align 1
+// AARCH64-BE-NEXT:    ret void
+//
+extern "C" void testVtable(S7 s) {
   // "vtable ptr" [0-7], "x" [8-11], "y" [12], "z" [13], PAD [14-15]
   __builtin_clear_padding(&s);
 }
 
+
 struct VirtualBase1 {
   unsigned int x1;
+//
   virtual int call1() { return x1; };
   virtual ~VirtualBase1() = default;
 };
 
 struct VirtualBase2 {
   unsigned int x2;
+//
   virtual int call2() { return x2; };
   virtual ~VirtualBase2() = default;
 };
 
 struct VirtualBase3 {
   unsigned int x3;
+//
   virtual int call3() { return x3; };
   virtual ~VirtualBase3() = default;
 };
@@ -803,11 +1864,12 @@ struct NonVirtualBase1 {
   char y;
 };
 struct S8 : VirtualBase1, VirtualBase2, NonVirtualBase1, VirtualBase3 {
+//
   virtual int call1() override { return 5; }
   bool z;
 };
 
-// LINUX-LABEL: define dso_local void @_Z23testMultipleBasesVtable2S8(
+// LINUX-LABEL: define dso_local void @testMultipleBasesVtable(
 // LINUX-SAME: ptr nofreeobj noundef align 8 dereferenceable(48) [[S:%.*]]) #[[ATTR0]] {
 // LINUX-NEXT:  [[ENTRY:.*:]]
 // LINUX-NEXT:    [[S_INDIRECT_ADDR:%.*]] = alloca ptr, align 8
@@ -834,70 +1896,167 @@ struct S8 : VirtualBase1, VirtualBase2, NonVirtualBase1, VirtualBase3 {
 // LINUX-NEXT:    store i8 0, ptr [[TMP9]], align 1
 // LINUX-NEXT:    ret void
 //
-// WINDOWS-LABEL: define dso_local void @"?testMultipleBasesVtable@@YAXUS8@@@Z"(ptr nofreeobj noundef align 8 dead_on_return dereferenceable(56) %s) #0 {
-// WINDOWS-NEXT: entry:
-// WINDOWS-NEXT:   %s.indirect_addr = alloca ptr, align 8
-// WINDOWS-NEXT:   store ptr %s, ptr %s.indirect_addr, align 8
-// WINDOWS-NEXT:   %0 = getelementptr i8, ptr %s, i32 12
-// WINDOWS-NEXT:   store i8 0, ptr %0, align 4
-// WINDOWS-NEXT:   %1 = getelementptr i8, ptr %s, i32 13
-// WINDOWS-NEXT:   store i8 0, ptr %1, align 1
-// WINDOWS-NEXT:   %2 = getelementptr i8, ptr %s, i32 14
-// WINDOWS-NEXT:   store i8 0, ptr %2, align 2
-// WINDOWS-NEXT:   %3 = getelementptr i8, ptr %s, i32 15
-// WINDOWS-NEXT:   store i8 0, ptr %3, align 1
-// WINDOWS-NEXT:   %4 = getelementptr i8, ptr %s, i32 28
-// WINDOWS-NEXT:   store i8 0, ptr %4, align 4
-// WINDOWS-NEXT:   %5 = getelementptr i8, ptr %s, i32 29
-// WINDOWS-NEXT:   store i8 0, ptr %5, align 1
-// WINDOWS-NEXT:   %6 = getelementptr i8, ptr %s, i32 30
-// WINDOWS-NEXT:   store i8 0, ptr %6, align 2
-// WINDOWS-NEXT:   %7 = getelementptr i8, ptr %s, i32 31
-// WINDOWS-NEXT:   store i8 0, ptr %7, align 1
-// WINDOWS-NEXT:   %8 = getelementptr i8, ptr %s, i32 44
-// WINDOWS-NEXT:   store i8 0, ptr %8, align 4
-// WINDOWS-NEXT:   %9 = getelementptr i8, ptr %s, i32 45
-// WINDOWS-NEXT:   store i8 0, ptr %9, align 1
-// WINDOWS-NEXT:   %10 = getelementptr i8, ptr %s, i32 46
-// WINDOWS-NEXT:   store i8 0, ptr %10, align 2
-// WINDOWS-NEXT:   %11 = getelementptr i8, ptr %s, i32 47
-// WINDOWS-NEXT:   store i8 0, ptr %11, align 1
-// WINDOWS-NEXT:   %12 = getelementptr i8, ptr %s, i32 50
-// WINDOWS-NEXT:   store i8 0, ptr %12, align 2
-// WINDOWS-NEXT:   %13 = getelementptr i8, ptr %s, i32 51
-// WINDOWS-NEXT:   store i8 0, ptr %13, align 1
-// WINDOWS-NEXT:   %14 = getelementptr i8, ptr %s, i32 52
-// WINDOWS-NEXT:   store i8 0, ptr %14, align 4
-// WINDOWS-NEXT:   %15 = getelementptr i8, ptr %s, i32 53
-// WINDOWS-NEXT:   store i8 0, ptr %15, align 1
-// WINDOWS-NEXT:   %16 = getelementptr i8, ptr %s, i32 54
-// WINDOWS-NEXT:   store i8 0, ptr %16, align 2
-// WINDOWS-NEXT:   %17 = getelementptr i8, ptr %s, i32 55
-// WINDOWS-NEXT:   store i8 0, ptr %17, align 1
-// WINDOWS-NEXT:   call void @"??1S8@@UEAA at XZ"(ptr noundef nonnull align 8 dead_on_return(56) dereferenceable(56) %s) #1
-// WINDOWS-NEXT:   ret void
-//
-void testMultipleBasesVtable(S8 s) {
+// WINDOWS-LABEL: define dso_local void @testMultipleBasesVtable(
+// WINDOWS-SAME: ptr nofreeobj noundef align 8 dead_on_return dereferenceable(56) [[S:%.*]]) #[[ATTR0]] {
+// WINDOWS-NEXT:  [[ENTRY:.*:]]
+// WINDOWS-NEXT:    [[S_INDIRECT_ADDR:%.*]] = alloca ptr, align 8
+// WINDOWS-NEXT:    store ptr [[S]], ptr [[S_INDIRECT_ADDR]], align 8
+// WINDOWS-NEXT:    [[TMP0:%.*]] = getelementptr i8, ptr [[S]], i32 12
+// WINDOWS-NEXT:    store i8 0, ptr [[TMP0]], align 4
+// WINDOWS-NEXT:    [[TMP1:%.*]] = getelementptr i8, ptr [[S]], i32 13
+// WINDOWS-NEXT:    store i8 0, ptr [[TMP1]], align 1
+// WINDOWS-NEXT:    [[TMP2:%.*]] = getelementptr i8, ptr [[S]], i32 14
+// WINDOWS-NEXT:    store i8 0, ptr [[TMP2]], align 2
+// WINDOWS-NEXT:    [[TMP3:%.*]] = getelementptr i8, ptr [[S]], i32 15
+// WINDOWS-NEXT:    store i8 0, ptr [[TMP3]], align 1
+// WINDOWS-NEXT:    [[TMP4:%.*]] = getelementptr i8, ptr [[S]], i32 28
+// WINDOWS-NEXT:    store i8 0, ptr [[TMP4]], align 4
+// WINDOWS-NEXT:    [[TMP5:%.*]] = getelementptr i8, ptr [[S]], i32 29
+// WINDOWS-NEXT:    store i8 0, ptr [[TMP5]], align 1
+// WINDOWS-NEXT:    [[TMP6:%.*]] = getelementptr i8, ptr [[S]], i32 30
+// WINDOWS-NEXT:    store i8 0, ptr [[TMP6]], align 2
+// WINDOWS-NEXT:    [[TMP7:%.*]] = getelementptr i8, ptr [[S]], i32 31
+// WINDOWS-NEXT:    store i8 0, ptr [[TMP7]], align 1
+// WINDOWS-NEXT:    [[TMP8:%.*]] = getelementptr i8, ptr [[S]], i32 44
+// WINDOWS-NEXT:    store i8 0, ptr [[TMP8]], align 4
+// WINDOWS-NEXT:    [[TMP9:%.*]] = getelementptr i8, ptr [[S]], i32 45
+// WINDOWS-NEXT:    store i8 0, ptr [[TMP9]], align 1
+// WINDOWS-NEXT:    [[TMP10:%.*]] = getelementptr i8, ptr [[S]], i32 46
+// WINDOWS-NEXT:    store i8 0, ptr [[TMP10]], align 2
+// WINDOWS-NEXT:    [[TMP11:%.*]] = getelementptr i8, ptr [[S]], i32 47
+// WINDOWS-NEXT:    store i8 0, ptr [[TMP11]], align 1
+// WINDOWS-NEXT:    [[TMP12:%.*]] = getelementptr i8, ptr [[S]], i32 50
+// WINDOWS-NEXT:    store i8 0, ptr [[TMP12]], align 2
+// WINDOWS-NEXT:    [[TMP13:%.*]] = getelementptr i8, ptr [[S]], i32 51
+// WINDOWS-NEXT:    store i8 0, ptr [[TMP13]], align 1
+// WINDOWS-NEXT:    [[TMP14:%.*]] = getelementptr i8, ptr [[S]], i32 52
+// WINDOWS-NEXT:    store i8 0, ptr [[TMP14]], align 4
+// WINDOWS-NEXT:    [[TMP15:%.*]] = getelementptr i8, ptr [[S]], i32 53
+// WINDOWS-NEXT:    store i8 0, ptr [[TMP15]], align 1
+// WINDOWS-NEXT:    [[TMP16:%.*]] = getelementptr i8, ptr [[S]], i32 54
+// WINDOWS-NEXT:    store i8 0, ptr [[TMP16]], align 2
+// WINDOWS-NEXT:    [[TMP17:%.*]] = getelementptr i8, ptr [[S]], i32 55
+// WINDOWS-NEXT:    store i8 0, ptr [[TMP17]], align 1
+// WINDOWS-NEXT:    call void @"??1S8@@UEAA at XZ"(ptr noundef nonnull align 8 dead_on_return(56) dereferenceable(56) [[S]]) #[[ATTR1]]
+// WINDOWS-NEXT:    ret void
+//
+// ARM-LE-LABEL: define dso_local void @testMultipleBasesVtable(
+// ARM-LE-SAME: ptr nofreeobj noundef align 4 dereferenceable(32) [[S:%.*]]) #[[ATTR0]] {
+// ARM-LE-NEXT:  [[ENTRY:.*:]]
+// ARM-LE-NEXT:    [[S_INDIRECT_ADDR:%.*]] = alloca ptr, align 4
+// ARM-LE-NEXT:    store ptr [[S]], ptr [[S_INDIRECT_ADDR]], align 4
+// ARM-LE-NEXT:    [[TMP0:%.*]] = getelementptr i8, ptr [[S]], i32 17
+// ARM-LE-NEXT:    store i8 0, ptr [[TMP0]], align 1
+// ARM-LE-NEXT:    [[TMP1:%.*]] = getelementptr i8, ptr [[S]], i32 18
+// ARM-LE-NEXT:    store i8 0, ptr [[TMP1]], align 2
+// ARM-LE-NEXT:    [[TMP2:%.*]] = getelementptr i8, ptr [[S]], i32 19
+// ARM-LE-NEXT:    store i8 0, ptr [[TMP2]], align 1
+// ARM-LE-NEXT:    [[TMP3:%.*]] = getelementptr i8, ptr [[S]], i32 29
+// ARM-LE-NEXT:    store i8 0, ptr [[TMP3]], align 1
+// ARM-LE-NEXT:    [[TMP4:%.*]] = getelementptr i8, ptr [[S]], i32 30
+// ARM-LE-NEXT:    store i8 0, ptr [[TMP4]], align 2
+// ARM-LE-NEXT:    [[TMP5:%.*]] = getelementptr i8, ptr [[S]], i32 31
+// ARM-LE-NEXT:    store i8 0, ptr [[TMP5]], align 1
+// ARM-LE-NEXT:    ret void
+//
+// ARM-BE-LABEL: define dso_local void @testMultipleBasesVtable(
+// ARM-BE-SAME: ptr nofreeobj noundef align 4 dereferenceable(32) [[S:%.*]]) #[[ATTR0]] {
+// ARM-BE-NEXT:  [[ENTRY:.*:]]
+// ARM-BE-NEXT:    [[S_INDIRECT_ADDR:%.*]] = alloca ptr, align 4
+// ARM-BE-NEXT:    store ptr [[S]], ptr [[S_INDIRECT_ADDR]], align 4
+// ARM-BE-NEXT:    [[TMP0:%.*]] = getelementptr i8, ptr [[S]], i32 17
+// ARM-BE-NEXT:    store i8 0, ptr [[TMP0]], align 1
+// ARM-BE-NEXT:    [[TMP1:%.*]] = getelementptr i8, ptr [[S]], i32 18
+// ARM-BE-NEXT:    store i8 0, ptr [[TMP1]], align 2
+// ARM-BE-NEXT:    [[TMP2:%.*]] = getelementptr i8, ptr [[S]], i32 19
+// ARM-BE-NEXT:    store i8 0, ptr [[TMP2]], align 1
+// ARM-BE-NEXT:    [[TMP3:%.*]] = getelementptr i8, ptr [[S]], i32 29
+// ARM-BE-NEXT:    store i8 0, ptr [[TMP3]], align 1
+// ARM-BE-NEXT:    [[TMP4:%.*]] = getelementptr i8, ptr [[S]], i32 30
+// ARM-BE-NEXT:    store i8 0, ptr [[TMP4]], align 2
+// ARM-BE-NEXT:    [[TMP5:%.*]] = getelementptr i8, ptr [[S]], i32 31
+// ARM-BE-NEXT:    store i8 0, ptr [[TMP5]], align 1
+// ARM-BE-NEXT:    ret void
+//
+// AARCH64-LE-LABEL: define dso_local void @testMultipleBasesVtable(
+// AARCH64-LE-SAME: ptr nofreeobj noundef align 8 dereferenceable(48) [[S:%.*]]) #[[ATTR0]] {
+// AARCH64-LE-NEXT:  [[ENTRY:.*:]]
+// AARCH64-LE-NEXT:    [[S_INDIRECT_ADDR:%.*]] = alloca ptr, align 8
+// AARCH64-LE-NEXT:    store ptr [[S]], ptr [[S_INDIRECT_ADDR]], align 8
+// AARCH64-LE-NEXT:    [[TMP0:%.*]] = getelementptr i8, ptr [[S]], i32 12
+// AARCH64-LE-NEXT:    store i8 0, ptr [[TMP0]], align 4
+// AARCH64-LE-NEXT:    [[TMP1:%.*]] = getelementptr i8, ptr [[S]], i32 13
+// AARCH64-LE-NEXT:    store i8 0, ptr [[TMP1]], align 1
+// AARCH64-LE-NEXT:    [[TMP2:%.*]] = getelementptr i8, ptr [[S]], i32 14
+// AARCH64-LE-NEXT:    store i8 0, ptr [[TMP2]], align 2
+// AARCH64-LE-NEXT:    [[TMP3:%.*]] = getelementptr i8, ptr [[S]], i32 15
+// AARCH64-LE-NEXT:    store i8 0, ptr [[TMP3]], align 1
+// AARCH64-LE-NEXT:    [[TMP4:%.*]] = getelementptr i8, ptr [[S]], i32 29
+// AARCH64-LE-NEXT:    store i8 0, ptr [[TMP4]], align 1
+// AARCH64-LE-NEXT:    [[TMP5:%.*]] = getelementptr i8, ptr [[S]], i32 30
+// AARCH64-LE-NEXT:    store i8 0, ptr [[TMP5]], align 2
+// AARCH64-LE-NEXT:    [[TMP6:%.*]] = getelementptr i8, ptr [[S]], i32 31
+// AARCH64-LE-NEXT:    store i8 0, ptr [[TMP6]], align 1
+// AARCH64-LE-NEXT:    [[TMP7:%.*]] = getelementptr i8, ptr [[S]], i32 45
+// AARCH64-LE-NEXT:    store i8 0, ptr [[TMP7]], align 1
+// AARCH64-LE-NEXT:    [[TMP8:%.*]] = getelementptr i8, ptr [[S]], i32 46
+// AARCH64-LE-NEXT:    store i8 0, ptr [[TMP8]], align 2
+// AARCH64-LE-NEXT:    [[TMP9:%.*]] = getelementptr i8, ptr [[S]], i32 47
+// AARCH64-LE-NEXT:    store i8 0, ptr [[TMP9]], align 1
+// AARCH64-LE-NEXT:    ret void
+//
+// AARCH64-BE-LABEL: define dso_local void @testMultipleBasesVtable(
+// AARCH64-BE-SAME: ptr nofreeobj noundef align 8 dereferenceable(48) [[S:%.*]]) #[[ATTR0]] {
+// AARCH64-BE-NEXT:  [[ENTRY:.*:]]
+// AARCH64-BE-NEXT:    [[S_INDIRECT_ADDR:%.*]] = alloca ptr, align 8
+// AARCH64-BE-NEXT:    store ptr [[S]], ptr [[S_INDIRECT_ADDR]], align 8
+// AARCH64-BE-NEXT:    [[TMP0:%.*]] = getelementptr i8, ptr [[S]], i32 12
+// AARCH64-BE-NEXT:    store i8 0, ptr [[TMP0]], align 4
+// AARCH64-BE-NEXT:    [[TMP1:%.*]] = getelementptr i8, ptr [[S]], i32 13
+// AARCH64-BE-NEXT:    store i8 0, ptr [[TMP1]], align 1
+// AARCH64-BE-NEXT:    [[TMP2:%.*]] = getelementptr i8, ptr [[S]], i32 14
+// AARCH64-BE-NEXT:    store i8 0, ptr [[TMP2]], align 2
+// AARCH64-BE-NEXT:    [[TMP3:%.*]] = getelementptr i8, ptr [[S]], i32 15
+// AARCH64-BE-NEXT:    store i8 0, ptr [[TMP3]], align 1
+// AARCH64-BE-NEXT:    [[TMP4:%.*]] = getelementptr i8, ptr [[S]], i32 29
+// AARCH64-BE-NEXT:    store i8 0, ptr [[TMP4]], align 1
+// AARCH64-BE-NEXT:    [[TMP5:%.*]] = getelementptr i8, ptr [[S]], i32 30
+// AARCH64-BE-NEXT:    store i8 0, ptr [[TMP5]], align 2
+// AARCH64-BE-NEXT:    [[TMP6:%.*]] = getelementptr i8, ptr [[S]], i32 31
+// AARCH64-BE-NEXT:    store i8 0, ptr [[TMP6]], align 1
+// AARCH64-BE-NEXT:    [[TMP7:%.*]] = getelementptr i8, ptr [[S]], i32 45
+// AARCH64-BE-NEXT:    store i8 0, ptr [[TMP7]], align 1
+// AARCH64-BE-NEXT:    [[TMP8:%.*]] = getelementptr i8, ptr [[S]], i32 46
+// AARCH64-BE-NEXT:    store i8 0, ptr [[TMP8]], align 2
+// AARCH64-BE-NEXT:    [[TMP9:%.*]] = getelementptr i8, ptr [[S]], i32 47
+// AARCH64-BE-NEXT:    store i8 0, ptr [[TMP9]], align 1
+// AARCH64-BE-NEXT:    ret void
+//
+extern "C" void testMultipleBasesVtable(S8 s) {
   // "vtable ptr" [0-7], "x1" [8-11], PAD "[12-15]",
   // "vtable ptr" [16-23], "x2" [24-27], "y" [28], PAD "[29-31]",
   // "vtable_ptr" [32-39], "x3" [40-43], "z" [44], PAD [45-47]
   __builtin_clear_padding(&s);
 }
 
+
 struct VirtualChain1 {
   unsigned int x1;
+//
   virtual int call1() { return x1; };
   virtual ~VirtualChain1() = default;
 };
 
 struct VirtualChain2 : VirtualChain1 {
   unsigned int x2;
+//
   virtual int call2() { return x2; };
   virtual ~VirtualChain2() = default;
 };
 
 struct VirtualChain3 : VirtualChain2 {
   unsigned int x3;
+//
   virtual int call3() { return x3; };
   virtual ~VirtualChain3() = default;
 };
@@ -910,7 +2069,7 @@ struct S9 : NonVirtualBase2, VirtualChain3 {
   bool z;
 };
 
-// LINUX-LABEL: define dso_local void @_Z16testVirtualChain2S9(
+// LINUX-LABEL: define dso_local void @testVirtualChain(
 // LINUX-SAME: ptr nofreeobj noundef align 8 dereferenceable(24) [[S:%.*]]) #[[ATTR0]] {
 // LINUX-NEXT:  [[ENTRY:.*:]]
 // LINUX-NEXT:    [[S_INDIRECT_ADDR:%.*]] = alloca ptr, align 8
@@ -921,50 +2080,95 @@ struct S9 : NonVirtualBase2, VirtualChain3 {
 // LINUX-NEXT:    store i8 0, ptr [[TMP1]], align 1
 // LINUX-NEXT:    ret void
 //
-// WINDOWS-LABEL: define dso_local void @"?testVirtualChain@@YAXUS9@@@Z"(ptr nofreeobj noundef align 8 dead_on_return dereferenceable(40) %s) #0 {
-// WINDOWS-NEXT: entry:
-// WINDOWS-NEXT:   %s.indirect_addr = alloca ptr, align 8
-// WINDOWS-NEXT:   store ptr %s, ptr %s.indirect_addr, align 8
-// WINDOWS-NEXT:   %0 = getelementptr i8, ptr %s, i32 12
-// WINDOWS-NEXT:   store i8 0, ptr %0, align 4
-// WINDOWS-NEXT:   %1 = getelementptr i8, ptr %s, i32 13
-// WINDOWS-NEXT:   store i8 0, ptr %1, align 1
-// WINDOWS-NEXT:   %2 = getelementptr i8, ptr %s, i32 14
-// WINDOWS-NEXT:   store i8 0, ptr %2, align 2
-// WINDOWS-NEXT:   %3 = getelementptr i8, ptr %s, i32 15
-// WINDOWS-NEXT:   store i8 0, ptr %3, align 1
-// WINDOWS-NEXT:   %4 = getelementptr i8, ptr %s, i32 20
-// WINDOWS-NEXT:   store i8 0, ptr %4, align 4
-// WINDOWS-NEXT:   %5 = getelementptr i8, ptr %s, i32 21
-// WINDOWS-NEXT:   store i8 0, ptr %5, align 1
-// WINDOWS-NEXT:   %6 = getelementptr i8, ptr %s, i32 22
-// WINDOWS-NEXT:   store i8 0, ptr %6, align 2
-// WINDOWS-NEXT:   %7 = getelementptr i8, ptr %s, i32 23
-// WINDOWS-NEXT:   store i8 0, ptr %7, align 1
-// WINDOWS-NEXT:   %8 = getelementptr i8, ptr %s, i32 28
-// WINDOWS-NEXT:   store i8 0, ptr %8, align 4
-// WINDOWS-NEXT:   %9 = getelementptr i8, ptr %s, i32 29
-// WINDOWS-NEXT:   store i8 0, ptr %9, align 1
-// WINDOWS-NEXT:   %10 = getelementptr i8, ptr %s, i32 30
-// WINDOWS-NEXT:   store i8 0, ptr %10, align 2
-// WINDOWS-NEXT:   %11 = getelementptr i8, ptr %s, i32 31
-// WINDOWS-NEXT:   store i8 0, ptr %11, align 1
-// WINDOWS-NEXT:   %12 = getelementptr i8, ptr %s, i32 34
-// WINDOWS-NEXT:   store i8 0, ptr %12, align 2
-// WINDOWS-NEXT:   %13 = getelementptr i8, ptr %s, i32 35
-// WINDOWS-NEXT:   store i8 0, ptr %13, align 1
-// WINDOWS-NEXT:   %14 = getelementptr i8, ptr %s, i32 36
-// WINDOWS-NEXT:   store i8 0, ptr %14, align 4
-// WINDOWS-NEXT:   %15 = getelementptr i8, ptr %s, i32 37
-// WINDOWS-NEXT:   store i8 0, ptr %15, align 1
-// WINDOWS-NEXT:   %16 = getelementptr i8, ptr %s, i32 38
-// WINDOWS-NEXT:   store i8 0, ptr %16, align 2
-// WINDOWS-NEXT:   %17 = getelementptr i8, ptr %s, i32 39
-// WINDOWS-NEXT:   store i8 0, ptr %17, align 1
-// WINDOWS-NEXT:   call void @"??1S9@@UEAA at XZ"(ptr noundef nonnull align 8 dead_on_return(40) dereferenceable(40) %s) #1
-// WINDOWS-NEXT:   ret void
-//
-void testVirtualChain(S9 s) {
+// WINDOWS-LABEL: define dso_local void @testVirtualChain(
+// WINDOWS-SAME: ptr nofreeobj noundef align 8 dead_on_return dereferenceable(40) [[S:%.*]]) #[[ATTR0]] {
+// WINDOWS-NEXT:  [[ENTRY:.*:]]
+// WINDOWS-NEXT:    [[S_INDIRECT_ADDR:%.*]] = alloca ptr, align 8
+// WINDOWS-NEXT:    store ptr [[S]], ptr [[S_INDIRECT_ADDR]], align 8
+// WINDOWS-NEXT:    [[TMP0:%.*]] = getelementptr i8, ptr [[S]], i32 12
+// WINDOWS-NEXT:    store i8 0, ptr [[TMP0]], align 4
+// WINDOWS-NEXT:    [[TMP1:%.*]] = getelementptr i8, ptr [[S]], i32 13
+// WINDOWS-NEXT:    store i8 0, ptr [[TMP1]], align 1
+// WINDOWS-NEXT:    [[TMP2:%.*]] = getelementptr i8, ptr [[S]], i32 14
+// WINDOWS-NEXT:    store i8 0, ptr [[TMP2]], align 2
+// WINDOWS-NEXT:    [[TMP3:%.*]] = getelementptr i8, ptr [[S]], i32 15
+// WINDOWS-NEXT:    store i8 0, ptr [[TMP3]], align 1
+// WINDOWS-NEXT:    [[TMP4:%.*]] = getelementptr i8, ptr [[S]], i32 20
+// WINDOWS-NEXT:    store i8 0, ptr [[TMP4]], align 4
+// WINDOWS-NEXT:    [[TMP5:%.*]] = getelementptr i8, ptr [[S]], i32 21
+// WINDOWS-NEXT:    store i8 0, ptr [[TMP5]], align 1
+// WINDOWS-NEXT:    [[TMP6:%.*]] = getelementptr i8, ptr [[S]], i32 22
+// WINDOWS-NEXT:    store i8 0, ptr [[TMP6]], align 2
+// WINDOWS-NEXT:    [[TMP7:%.*]] = getelementptr i8, ptr [[S]], i32 23
+// WINDOWS-NEXT:    store i8 0, ptr [[TMP7]], align 1
+// WINDOWS-NEXT:    [[TMP8:%.*]] = getelementptr i8, ptr [[S]], i32 28
+// WINDOWS-NEXT:    store i8 0, ptr [[TMP8]], align 4
+// WINDOWS-NEXT:    [[TMP9:%.*]] = getelementptr i8, ptr [[S]], i32 29
+// WINDOWS-NEXT:    store i8 0, ptr [[TMP9]], align 1
+// WINDOWS-NEXT:    [[TMP10:%.*]] = getelementptr i8, ptr [[S]], i32 30
+// WINDOWS-NEXT:    store i8 0, ptr [[TMP10]], align 2
+// WINDOWS-NEXT:    [[TMP11:%.*]] = getelementptr i8, ptr [[S]], i32 31
+// WINDOWS-NEXT:    store i8 0, ptr [[TMP11]], align 1
+// WINDOWS-NEXT:    [[TMP12:%.*]] = getelementptr i8, ptr [[S]], i32 34
+// WINDOWS-NEXT:    store i8 0, ptr [[TMP12]], align 2
+// WINDOWS-NEXT:    [[TMP13:%.*]] = getelementptr i8, ptr [[S]], i32 35
+// WINDOWS-NEXT:    store i8 0, ptr [[TMP13]], align 1
+// WINDOWS-NEXT:    [[TMP14:%.*]] = getelementptr i8, ptr [[S]], i32 36
+// WINDOWS-NEXT:    store i8 0, ptr [[TMP14]], align 4
+// WINDOWS-NEXT:    [[TMP15:%.*]] = getelementptr i8, ptr [[S]], i32 37
+// WINDOWS-NEXT:    store i8 0, ptr [[TMP15]], align 1
+// WINDOWS-NEXT:    [[TMP16:%.*]] = getelementptr i8, ptr [[S]], i32 38
+// WINDOWS-NEXT:    store i8 0, ptr [[TMP16]], align 2
+// WINDOWS-NEXT:    [[TMP17:%.*]] = getelementptr i8, ptr [[S]], i32 39
+// WINDOWS-NEXT:    store i8 0, ptr [[TMP17]], align 1
+// WINDOWS-NEXT:    call void @"??1S9@@UEAA at XZ"(ptr noundef nonnull align 8 dead_on_return(40) dereferenceable(40) [[S]]) #[[ATTR1]]
+// WINDOWS-NEXT:    ret void
+//
+// ARM-LE-LABEL: define dso_local void @testVirtualChain(
+// ARM-LE-SAME: ptr nofreeobj noundef align 4 dereferenceable(20) [[S:%.*]]) #[[ATTR0]] {
+// ARM-LE-NEXT:  [[ENTRY:.*:]]
+// ARM-LE-NEXT:    [[S_INDIRECT_ADDR:%.*]] = alloca ptr, align 4
+// ARM-LE-NEXT:    store ptr [[S]], ptr [[S_INDIRECT_ADDR]], align 4
+// ARM-LE-NEXT:    [[TMP0:%.*]] = getelementptr i8, ptr [[S]], i32 18
+// ARM-LE-NEXT:    store i8 0, ptr [[TMP0]], align 2
+// ARM-LE-NEXT:    [[TMP1:%.*]] = getelementptr i8, ptr [[S]], i32 19
+// ARM-LE-NEXT:    store i8 0, ptr [[TMP1]], align 1
+// ARM-LE-NEXT:    ret void
+//
+// ARM-BE-LABEL: define dso_local void @testVirtualChain(
+// ARM-BE-SAME: ptr nofreeobj noundef align 4 dereferenceable(20) [[S:%.*]]) #[[ATTR0]] {
+// ARM-BE-NEXT:  [[ENTRY:.*:]]
+// ARM-BE-NEXT:    [[S_INDIRECT_ADDR:%.*]] = alloca ptr, align 4
+// ARM-BE-NEXT:    store ptr [[S]], ptr [[S_INDIRECT_ADDR]], align 4
+// ARM-BE-NEXT:    [[TMP0:%.*]] = getelementptr i8, ptr [[S]], i32 18
+// ARM-BE-NEXT:    store i8 0, ptr [[TMP0]], align 2
+// ARM-BE-NEXT:    [[TMP1:%.*]] = getelementptr i8, ptr [[S]], i32 19
+// ARM-BE-NEXT:    store i8 0, ptr [[TMP1]], align 1
+// ARM-BE-NEXT:    ret void
+//
+// AARCH64-LE-LABEL: define dso_local void @testVirtualChain(
+// AARCH64-LE-SAME: ptr nofreeobj noundef align 8 dereferenceable(24) [[S:%.*]]) #[[ATTR0]] {
+// AARCH64-LE-NEXT:  [[ENTRY:.*:]]
+// AARCH64-LE-NEXT:    [[S_INDIRECT_ADDR:%.*]] = alloca ptr, align 8
+// AARCH64-LE-NEXT:    store ptr [[S]], ptr [[S_INDIRECT_ADDR]], align 8
+// AARCH64-LE-NEXT:    [[TMP0:%.*]] = getelementptr i8, ptr [[S]], i32 22
+// AARCH64-LE-NEXT:    store i8 0, ptr [[TMP0]], align 2
+// AARCH64-LE-NEXT:    [[TMP1:%.*]] = getelementptr i8, ptr [[S]], i32 23
+// AARCH64-LE-NEXT:    store i8 0, ptr [[TMP1]], align 1
+// AARCH64-LE-NEXT:    ret void
+//
+// AARCH64-BE-LABEL: define dso_local void @testVirtualChain(
+// AARCH64-BE-SAME: ptr nofreeobj noundef align 8 dereferenceable(24) [[S:%.*]]) #[[ATTR0]] {
+// AARCH64-BE-NEXT:  [[ENTRY:.*:]]
+// AARCH64-BE-NEXT:    [[S_INDIRECT_ADDR:%.*]] = alloca ptr, align 8
+// AARCH64-BE-NEXT:    store ptr [[S]], ptr [[S_INDIRECT_ADDR]], align 8
+// AARCH64-BE-NEXT:    [[TMP0:%.*]] = getelementptr i8, ptr [[S]], i32 22
+// AARCH64-BE-NEXT:    store i8 0, ptr [[TMP0]], align 2
+// AARCH64-BE-NEXT:    [[TMP1:%.*]] = getelementptr i8, ptr [[S]], i32 23
+// AARCH64-BE-NEXT:    store i8 0, ptr [[TMP1]], align 1
+// AARCH64-BE-NEXT:    ret void
+//
+extern "C" void testVirtualChain(S9 s) {
   // This should clear the padding after the bool z.
   // base reordered
   // "vtable ptr" [0-7],  "x1" [8-11], x2 [12-15], x3 [16-19],
@@ -973,6 +2177,7 @@ void testVirtualChain(S9 s) {
 }
 
 
+
 struct Base {
   int x;
 };
@@ -990,7 +2195,7 @@ struct S10 : D1, D2 {
   bool s;
 };
 
-// LINUX-LABEL: define dso_local void @_Z22testVirtualInheritance3S10(
+// LINUX-LABEL: define dso_local void @testVirtualInheritance(
 // LINUX-SAME: ptr nofreeobj noundef align 8 dead_on_return dereferenceable(40) [[S:%.*]]) #[[ATTR0]] {
 // LINUX-NEXT:  [[ENTRY:.*:]]
 // LINUX-NEXT:    [[S_INDIRECT_ADDR:%.*]] = alloca ptr, align 8
@@ -1015,47 +2220,132 @@ struct S10 : D1, D2 {
 // LINUX-NEXT:    store i8 0, ptr [[TMP8]], align 1
 // LINUX-NEXT:    ret void
 //
-// WINDOWS-LABEL: define dso_local void @"?testVirtualInheritance@@YAXUS10@@@Z"(ptr nofreeobj noundef align 8 dead_on_return dereferenceable(48) %s) #0 {
-// WINDOWS-NEXT: entry:
-// WINDOWS-NEXT:   %s.indirect_addr = alloca ptr, align 8
-// WINDOWS-NEXT:   store ptr %s, ptr %s.indirect_addr, align 8
-// WINDOWS-NEXT:   %0 = getelementptr i8, ptr %s, i32 13
-// WINDOWS-NEXT:   store i8 0, ptr %0, align 1
-// WINDOWS-NEXT:   %1 = getelementptr i8, ptr %s, i32 14
-// WINDOWS-NEXT:   store i8 0, ptr %1, align 2
-// WINDOWS-NEXT:   %2 = getelementptr i8, ptr %s, i32 15
-// WINDOWS-NEXT:   store i8 0, ptr %2, align 1
-// WINDOWS-NEXT:   %3 = getelementptr i8, ptr %s, i32 29
-// WINDOWS-NEXT:   store i8 0, ptr %3, align 1
-// WINDOWS-NEXT:   %4 = getelementptr i8, ptr %s, i32 30
-// WINDOWS-NEXT:   store i8 0, ptr %4, align 2
-// WINDOWS-NEXT:   %5 = getelementptr i8, ptr %s, i32 31
-// WINDOWS-NEXT:   store i8 0, ptr %5, align 1
-// WINDOWS-NEXT:   %6 = getelementptr i8, ptr %s, i32 33
-// WINDOWS-NEXT:   store i8 0, ptr %6, align 1
-// WINDOWS-NEXT:   %7 = getelementptr i8, ptr %s, i32 34
-// WINDOWS-NEXT:   store i8 0, ptr %7, align 2
-// WINDOWS-NEXT:   %8 = getelementptr i8, ptr %s, i32 35
-// WINDOWS-NEXT:   store i8 0, ptr %8, align 1
-// WINDOWS-NEXT:   %9 = getelementptr i8, ptr %s, i32 36
-// WINDOWS-NEXT:   store i8 0, ptr %9, align 4
-// WINDOWS-NEXT:   %10 = getelementptr i8, ptr %s, i32 37
-// WINDOWS-NEXT:   store i8 0, ptr %10, align 1
-// WINDOWS-NEXT:   %11 = getelementptr i8, ptr %s, i32 38
-// WINDOWS-NEXT:   store i8 0, ptr %11, align 2
-// WINDOWS-NEXT:   %12 = getelementptr i8, ptr %s, i32 39
-// WINDOWS-NEXT:   store i8 0, ptr %12, align 1
-// WINDOWS-NEXT:   %13 = getelementptr i8, ptr %s, i32 44
-// WINDOWS-NEXT:   store i8 0, ptr %13, align 4
-// WINDOWS-NEXT:   %14 = getelementptr i8, ptr %s, i32 45
-// WINDOWS-NEXT:   store i8 0, ptr %14, align 1
-// WINDOWS-NEXT:   %15 = getelementptr i8, ptr %s, i32 46
-// WINDOWS-NEXT:   store i8 0, ptr %15, align 2
-// WINDOWS-NEXT:   %16 = getelementptr i8, ptr %s, i32 47
-// WINDOWS-NEXT:   store i8 0, ptr %16, align 1
-// WINDOWS-NEXT:   ret void
-//
-void testVirtualInheritance(S10 s) {
+// WINDOWS-LABEL: define dso_local void @testVirtualInheritance(
+// WINDOWS-SAME: ptr nofreeobj noundef align 8 dead_on_return dereferenceable(48) [[S:%.*]]) #[[ATTR0]] {
+// WINDOWS-NEXT:  [[ENTRY:.*:]]
+// WINDOWS-NEXT:    [[S_INDIRECT_ADDR:%.*]] = alloca ptr, align 8
+// WINDOWS-NEXT:    store ptr [[S]], ptr [[S_INDIRECT_ADDR]], align 8
+// WINDOWS-NEXT:    [[TMP0:%.*]] = getelementptr i8, ptr [[S]], i32 13
+// WINDOWS-NEXT:    store i8 0, ptr [[TMP0]], align 1
+// WINDOWS-NEXT:    [[TMP1:%.*]] = getelementptr i8, ptr [[S]], i32 14
+// WINDOWS-NEXT:    store i8 0, ptr [[TMP1]], align 2
+// WINDOWS-NEXT:    [[TMP2:%.*]] = getelementptr i8, ptr [[S]], i32 15
+// WINDOWS-NEXT:    store i8 0, ptr [[TMP2]], align 1
+// WINDOWS-NEXT:    [[TMP3:%.*]] = getelementptr i8, ptr [[S]], i32 29
+// WINDOWS-NEXT:    store i8 0, ptr [[TMP3]], align 1
+// WINDOWS-NEXT:    [[TMP4:%.*]] = getelementptr i8, ptr [[S]], i32 30
+// WINDOWS-NEXT:    store i8 0, ptr [[TMP4]], align 2
+// WINDOWS-NEXT:    [[TMP5:%.*]] = getelementptr i8, ptr [[S]], i32 31
+// WINDOWS-NEXT:    store i8 0, ptr [[TMP5]], align 1
+// WINDOWS-NEXT:    [[TMP6:%.*]] = getelementptr i8, ptr [[S]], i32 33
+// WINDOWS-NEXT:    store i8 0, ptr [[TMP6]], align 1
+// WINDOWS-NEXT:    [[TMP7:%.*]] = getelementptr i8, ptr [[S]], i32 34
+// WINDOWS-NEXT:    store i8 0, ptr [[TMP7]], align 2
+// WINDOWS-NEXT:    [[TMP8:%.*]] = getelementptr i8, ptr [[S]], i32 35
+// WINDOWS-NEXT:    store i8 0, ptr [[TMP8]], align 1
+// WINDOWS-NEXT:    [[TMP9:%.*]] = getelementptr i8, ptr [[S]], i32 36
+// WINDOWS-NEXT:    store i8 0, ptr [[TMP9]], align 4
+// WINDOWS-NEXT:    [[TMP10:%.*]] = getelementptr i8, ptr [[S]], i32 37
+// WINDOWS-NEXT:    store i8 0, ptr [[TMP10]], align 1
+// WINDOWS-NEXT:    [[TMP11:%.*]] = getelementptr i8, ptr [[S]], i32 38
+// WINDOWS-NEXT:    store i8 0, ptr [[TMP11]], align 2
+// WINDOWS-NEXT:    [[TMP12:%.*]] = getelementptr i8, ptr [[S]], i32 39
+// WINDOWS-NEXT:    store i8 0, ptr [[TMP12]], align 1
+// WINDOWS-NEXT:    [[TMP13:%.*]] = getelementptr i8, ptr [[S]], i32 44
+// WINDOWS-NEXT:    store i8 0, ptr [[TMP13]], align 4
+// WINDOWS-NEXT:    [[TMP14:%.*]] = getelementptr i8, ptr [[S]], i32 45
+// WINDOWS-NEXT:    store i8 0, ptr [[TMP14]], align 1
+// WINDOWS-NEXT:    [[TMP15:%.*]] = getelementptr i8, ptr [[S]], i32 46
+// WINDOWS-NEXT:    store i8 0, ptr [[TMP15]], align 2
+// WINDOWS-NEXT:    [[TMP16:%.*]] = getelementptr i8, ptr [[S]], i32 47
+// WINDOWS-NEXT:    store i8 0, ptr [[TMP16]], align 1
+// WINDOWS-NEXT:    ret void
+//
+// ARM-LE-LABEL: define dso_local void @testVirtualInheritance(
+// ARM-LE-SAME: ptr nofreeobj noundef align 4 dead_on_return dereferenceable(28) [[S:%.*]]) #[[ATTR0]] {
+// ARM-LE-NEXT:  [[ENTRY:.*:]]
+// ARM-LE-NEXT:    [[S_INDIRECT_ADDR:%.*]] = alloca ptr, align 4
+// ARM-LE-NEXT:    store ptr [[S]], ptr [[S_INDIRECT_ADDR]], align 4
+// ARM-LE-NEXT:    [[TMP0:%.*]] = getelementptr i8, ptr [[S]], i32 9
+// ARM-LE-NEXT:    store i8 0, ptr [[TMP0]], align 1
+// ARM-LE-NEXT:    [[TMP1:%.*]] = getelementptr i8, ptr [[S]], i32 10
+// ARM-LE-NEXT:    store i8 0, ptr [[TMP1]], align 2
+// ARM-LE-NEXT:    [[TMP2:%.*]] = getelementptr i8, ptr [[S]], i32 11
+// ARM-LE-NEXT:    store i8 0, ptr [[TMP2]], align 1
+// ARM-LE-NEXT:    [[TMP3:%.*]] = getelementptr i8, ptr [[S]], i32 22
+// ARM-LE-NEXT:    store i8 0, ptr [[TMP3]], align 2
+// ARM-LE-NEXT:    [[TMP4:%.*]] = getelementptr i8, ptr [[S]], i32 23
+// ARM-LE-NEXT:    store i8 0, ptr [[TMP4]], align 1
+// ARM-LE-NEXT:    ret void
+//
+// ARM-BE-LABEL: define dso_local void @testVirtualInheritance(
+// ARM-BE-SAME: ptr nofreeobj noundef align 4 dead_on_return dereferenceable(28) [[S:%.*]]) #[[ATTR0]] {
+// ARM-BE-NEXT:  [[ENTRY:.*:]]
+// ARM-BE-NEXT:    [[S_INDIRECT_ADDR:%.*]] = alloca ptr, align 4
+// ARM-BE-NEXT:    store ptr [[S]], ptr [[S_INDIRECT_ADDR]], align 4
+// ARM-BE-NEXT:    [[TMP0:%.*]] = getelementptr i8, ptr [[S]], i32 9
+// ARM-BE-NEXT:    store i8 0, ptr [[TMP0]], align 1
+// ARM-BE-NEXT:    [[TMP1:%.*]] = getelementptr i8, ptr [[S]], i32 10
+// ARM-BE-NEXT:    store i8 0, ptr [[TMP1]], align 2
+// ARM-BE-NEXT:    [[TMP2:%.*]] = getelementptr i8, ptr [[S]], i32 11
+// ARM-BE-NEXT:    store i8 0, ptr [[TMP2]], align 1
+// ARM-BE-NEXT:    [[TMP3:%.*]] = getelementptr i8, ptr [[S]], i32 22
+// ARM-BE-NEXT:    store i8 0, ptr [[TMP3]], align 2
+// ARM-BE-NEXT:    [[TMP4:%.*]] = getelementptr i8, ptr [[S]], i32 23
+// ARM-BE-NEXT:    store i8 0, ptr [[TMP4]], align 1
+// ARM-BE-NEXT:    ret void
+//
+// AARCH64-LE-LABEL: define dso_local void @testVirtualInheritance(
+// AARCH64-LE-SAME: ptr nofreeobj noundef align 8 dead_on_return dereferenceable(40) [[S:%.*]]) #[[ATTR0]] {
+// AARCH64-LE-NEXT:  [[ENTRY:.*:]]
+// AARCH64-LE-NEXT:    [[S_INDIRECT_ADDR:%.*]] = alloca ptr, align 8
+// AARCH64-LE-NEXT:    store ptr [[S]], ptr [[S_INDIRECT_ADDR]], align 8
+// AARCH64-LE-NEXT:    [[TMP0:%.*]] = getelementptr i8, ptr [[S]], i32 13
+// AARCH64-LE-NEXT:    store i8 0, ptr [[TMP0]], align 1
+// AARCH64-LE-NEXT:    [[TMP1:%.*]] = getelementptr i8, ptr [[S]], i32 14
+// AARCH64-LE-NEXT:    store i8 0, ptr [[TMP1]], align 2
+// AARCH64-LE-NEXT:    [[TMP2:%.*]] = getelementptr i8, ptr [[S]], i32 15
+// AARCH64-LE-NEXT:    store i8 0, ptr [[TMP2]], align 1
+// AARCH64-LE-NEXT:    [[TMP3:%.*]] = getelementptr i8, ptr [[S]], i32 30
+// AARCH64-LE-NEXT:    store i8 0, ptr [[TMP3]], align 2
+// AARCH64-LE-NEXT:    [[TMP4:%.*]] = getelementptr i8, ptr [[S]], i32 31
+// AARCH64-LE-NEXT:    store i8 0, ptr [[TMP4]], align 1
+// AARCH64-LE-NEXT:    [[TMP5:%.*]] = getelementptr i8, ptr [[S]], i32 36
+// AARCH64-LE-NEXT:    store i8 0, ptr [[TMP5]], align 4
+// AARCH64-LE-NEXT:    [[TMP6:%.*]] = getelementptr i8, ptr [[S]], i32 37
+// AARCH64-LE-NEXT:    store i8 0, ptr [[TMP6]], align 1
+// AARCH64-LE-NEXT:    [[TMP7:%.*]] = getelementptr i8, ptr [[S]], i32 38
+// AARCH64-LE-NEXT:    store i8 0, ptr [[TMP7]], align 2
+// AARCH64-LE-NEXT:    [[TMP8:%.*]] = getelementptr i8, ptr [[S]], i32 39
+// AARCH64-LE-NEXT:    store i8 0, ptr [[TMP8]], align 1
+// AARCH64-LE-NEXT:    ret void
+//
+// AARCH64-BE-LABEL: define dso_local void @testVirtualInheritance(
+// AARCH64-BE-SAME: ptr nofreeobj noundef align 8 dead_on_return dereferenceable(40) [[S:%.*]]) #[[ATTR0]] {
+// AARCH64-BE-NEXT:  [[ENTRY:.*:]]
+// AARCH64-BE-NEXT:    [[S_INDIRECT_ADDR:%.*]] = alloca ptr, align 8
+// AARCH64-BE-NEXT:    store ptr [[S]], ptr [[S_INDIRECT_ADDR]], align 8
+// AARCH64-BE-NEXT:    [[TMP0:%.*]] = getelementptr i8, ptr [[S]], i32 13
+// AARCH64-BE-NEXT:    store i8 0, ptr [[TMP0]], align 1
+// AARCH64-BE-NEXT:    [[TMP1:%.*]] = getelementptr i8, ptr [[S]], i32 14
+// AARCH64-BE-NEXT:    store i8 0, ptr [[TMP1]], align 2
+// AARCH64-BE-NEXT:    [[TMP2:%.*]] = getelementptr i8, ptr [[S]], i32 15
+// AARCH64-BE-NEXT:    store i8 0, ptr [[TMP2]], align 1
+// AARCH64-BE-NEXT:    [[TMP3:%.*]] = getelementptr i8, ptr [[S]], i32 30
+// AARCH64-BE-NEXT:    store i8 0, ptr [[TMP3]], align 2
+// AARCH64-BE-NEXT:    [[TMP4:%.*]] = getelementptr i8, ptr [[S]], i32 31
+// AARCH64-BE-NEXT:    store i8 0, ptr [[TMP4]], align 1
+// AARCH64-BE-NEXT:    [[TMP5:%.*]] = getelementptr i8, ptr [[S]], i32 36
+// AARCH64-BE-NEXT:    store i8 0, ptr [[TMP5]], align 4
+// AARCH64-BE-NEXT:    [[TMP6:%.*]] = getelementptr i8, ptr [[S]], i32 37
+// AARCH64-BE-NEXT:    store i8 0, ptr [[TMP6]], align 1
+// AARCH64-BE-NEXT:    [[TMP7:%.*]] = getelementptr i8, ptr [[S]], i32 38
+// AARCH64-BE-NEXT:    store i8 0, ptr [[TMP7]], align 2
+// AARCH64-BE-NEXT:    [[TMP8:%.*]] = getelementptr i8, ptr [[S]], i32 39
+// AARCH64-BE-NEXT:    store i8 0, ptr [[TMP8]], align 1
+// AARCH64-BE-NEXT:    ret void
+//
+extern "C" void testVirtualInheritance(S10 s) {
   // note derived member placed before the virtual base
   // "vtable ptr" [0-7],  "d1" [8-11], "b1" [12], PAD [13-15],
   // "vtable ptr" [16-23],  "d2" [24-27], "b2" [28], s [29],  PAD [30-31],
@@ -1063,13 +2353,14 @@ void testVirtualInheritance(S10 s) {
   __builtin_clear_padding(&s);
 }
 
+
 struct S10_2 : virtual Base {
   int d1;
   bool b;
   virtual ~S10_2() = default;
 };
 
-// LINUX-LABEL: define dso_local void @_Z40testVirtualInheritanceAndVirtualFunction5S10_2(
+// LINUX-LABEL: define dso_local void @testVirtualInheritanceAndVirtualFunction(
 // LINUX-SAME: ptr nofreeobj noundef align 8 dereferenceable(24) [[S:%.*]]) #[[ATTR0]] {
 // LINUX-NEXT:  [[ENTRY:.*:]]
 // LINUX-NEXT:    [[S_INDIRECT_ADDR:%.*]] = alloca ptr, align 8
@@ -1090,31 +2381,101 @@ struct S10_2 : virtual Base {
 // LINUX-NEXT:    store i8 0, ptr [[TMP6]], align 1
 // LINUX-NEXT:    ret void
 //
-// WINDOWS-LABEL: define dso_local void @"?testVirtualInheritanceAndVirtualFunction@@YAXUS10_2@@@Z"(ptr nofreeobj noundef align 8 dead_on_return dereferenceable(32) %s) #0 {
-// WINDOWS-NEXT: entry:
-// WINDOWS-NEXT:   %s.indirect_addr = alloca ptr, align 8
-// WINDOWS-NEXT:   store ptr %s, ptr %s.indirect_addr, align 8
-// WINDOWS-NEXT:   %0 = getelementptr i8, ptr %s, i32 21
-// WINDOWS-NEXT:   store i8 0, ptr %0, align 1
-// WINDOWS-NEXT:   %1 = getelementptr i8, ptr %s, i32 22
-// WINDOWS-NEXT:   store i8 0, ptr %1, align 2
-// WINDOWS-NEXT:   %2 = getelementptr i8, ptr %s, i32 23
-// WINDOWS-NEXT:   store i8 0, ptr %2, align 1
-// WINDOWS-NEXT:   %3 = getelementptr i8, ptr %s, i32 28
-// WINDOWS-NEXT:   store i8 0, ptr %3, align 4
-// WINDOWS-NEXT:   %4 = getelementptr i8, ptr %s, i32 29
-// WINDOWS-NEXT:   store i8 0, ptr %4, align 1
-// WINDOWS-NEXT:   %5 = getelementptr i8, ptr %s, i32 30
-// WINDOWS-NEXT:   store i8 0, ptr %5, align 2
-// WINDOWS-NEXT:   %6 = getelementptr i8, ptr %s, i32 31
-// WINDOWS-NEXT:   store i8 0, ptr %6, align 1
-// WINDOWS-NEXT:   call void @"??_DS10_2@@QEAAXXZ"(ptr noundef nonnull align 8 dereferenceable(24) %s) #1
-// WINDOWS-NEXT:   ret void
-//
-void testVirtualInheritanceAndVirtualFunction(S10_2 s) {
+// WINDOWS-LABEL: define dso_local void @testVirtualInheritanceAndVirtualFunction(
+// WINDOWS-SAME: ptr nofreeobj noundef align 8 dead_on_return dereferenceable(32) [[S:%.*]]) #[[ATTR0]] {
+// WINDOWS-NEXT:  [[ENTRY:.*:]]
+// WINDOWS-NEXT:    [[S_INDIRECT_ADDR:%.*]] = alloca ptr, align 8
+// WINDOWS-NEXT:    store ptr [[S]], ptr [[S_INDIRECT_ADDR]], align 8
+// WINDOWS-NEXT:    [[TMP0:%.*]] = getelementptr i8, ptr [[S]], i32 21
+// WINDOWS-NEXT:    store i8 0, ptr [[TMP0]], align 1
+// WINDOWS-NEXT:    [[TMP1:%.*]] = getelementptr i8, ptr [[S]], i32 22
+// WINDOWS-NEXT:    store i8 0, ptr [[TMP1]], align 2
+// WINDOWS-NEXT:    [[TMP2:%.*]] = getelementptr i8, ptr [[S]], i32 23
+// WINDOWS-NEXT:    store i8 0, ptr [[TMP2]], align 1
+// WINDOWS-NEXT:    [[TMP3:%.*]] = getelementptr i8, ptr [[S]], i32 28
+// WINDOWS-NEXT:    store i8 0, ptr [[TMP3]], align 4
+// WINDOWS-NEXT:    [[TMP4:%.*]] = getelementptr i8, ptr [[S]], i32 29
+// WINDOWS-NEXT:    store i8 0, ptr [[TMP4]], align 1
+// WINDOWS-NEXT:    [[TMP5:%.*]] = getelementptr i8, ptr [[S]], i32 30
+// WINDOWS-NEXT:    store i8 0, ptr [[TMP5]], align 2
+// WINDOWS-NEXT:    [[TMP6:%.*]] = getelementptr i8, ptr [[S]], i32 31
+// WINDOWS-NEXT:    store i8 0, ptr [[TMP6]], align 1
+// WINDOWS-NEXT:    call void @"??_DS10_2@@QEAAXXZ"(ptr noundef nonnull align 8 dereferenceable(24) [[S]]) #[[ATTR1]]
+// WINDOWS-NEXT:    ret void
+//
+// ARM-LE-LABEL: define dso_local void @testVirtualInheritanceAndVirtualFunction(
+// ARM-LE-SAME: ptr nofreeobj noundef align 4 dereferenceable(16) [[S:%.*]]) #[[ATTR0]] {
+// ARM-LE-NEXT:  [[ENTRY:.*:]]
+// ARM-LE-NEXT:    [[S_INDIRECT_ADDR:%.*]] = alloca ptr, align 4
+// ARM-LE-NEXT:    store ptr [[S]], ptr [[S_INDIRECT_ADDR]], align 4
+// ARM-LE-NEXT:    [[TMP0:%.*]] = getelementptr i8, ptr [[S]], i32 9
+// ARM-LE-NEXT:    store i8 0, ptr [[TMP0]], align 1
+// ARM-LE-NEXT:    [[TMP1:%.*]] = getelementptr i8, ptr [[S]], i32 10
+// ARM-LE-NEXT:    store i8 0, ptr [[TMP1]], align 2
+// ARM-LE-NEXT:    [[TMP2:%.*]] = getelementptr i8, ptr [[S]], i32 11
+// ARM-LE-NEXT:    store i8 0, ptr [[TMP2]], align 1
+// ARM-LE-NEXT:    ret void
+//
+// ARM-BE-LABEL: define dso_local void @testVirtualInheritanceAndVirtualFunction(
+// ARM-BE-SAME: ptr nofreeobj noundef align 4 dereferenceable(16) [[S:%.*]]) #[[ATTR0]] {
+// ARM-BE-NEXT:  [[ENTRY:.*:]]
+// ARM-BE-NEXT:    [[S_INDIRECT_ADDR:%.*]] = alloca ptr, align 4
+// ARM-BE-NEXT:    store ptr [[S]], ptr [[S_INDIRECT_ADDR]], align 4
+// ARM-BE-NEXT:    [[TMP0:%.*]] = getelementptr i8, ptr [[S]], i32 9
+// ARM-BE-NEXT:    store i8 0, ptr [[TMP0]], align 1
+// ARM-BE-NEXT:    [[TMP1:%.*]] = getelementptr i8, ptr [[S]], i32 10
+// ARM-BE-NEXT:    store i8 0, ptr [[TMP1]], align 2
+// ARM-BE-NEXT:    [[TMP2:%.*]] = getelementptr i8, ptr [[S]], i32 11
+// ARM-BE-NEXT:    store i8 0, ptr [[TMP2]], align 1
+// ARM-BE-NEXT:    ret void
+//
+// AARCH64-LE-LABEL: define dso_local void @testVirtualInheritanceAndVirtualFunction(
+// AARCH64-LE-SAME: ptr nofreeobj noundef align 8 dereferenceable(24) [[S:%.*]]) #[[ATTR0]] {
+// AARCH64-LE-NEXT:  [[ENTRY:.*:]]
+// AARCH64-LE-NEXT:    [[S_INDIRECT_ADDR:%.*]] = alloca ptr, align 8
+// AARCH64-LE-NEXT:    store ptr [[S]], ptr [[S_INDIRECT_ADDR]], align 8
+// AARCH64-LE-NEXT:    [[TMP0:%.*]] = getelementptr i8, ptr [[S]], i32 13
+// AARCH64-LE-NEXT:    store i8 0, ptr [[TMP0]], align 1
+// AARCH64-LE-NEXT:    [[TMP1:%.*]] = getelementptr i8, ptr [[S]], i32 14
+// AARCH64-LE-NEXT:    store i8 0, ptr [[TMP1]], align 2
+// AARCH64-LE-NEXT:    [[TMP2:%.*]] = getelementptr i8, ptr [[S]], i32 15
+// AARCH64-LE-NEXT:    store i8 0, ptr [[TMP2]], align 1
+// AARCH64-LE-NEXT:    [[TMP3:%.*]] = getelementptr i8, ptr [[S]], i32 20
+// AARCH64-LE-NEXT:    store i8 0, ptr [[TMP3]], align 4
+// AARCH64-LE-NEXT:    [[TMP4:%.*]] = getelementptr i8, ptr [[S]], i32 21
+// AARCH64-LE-NEXT:    store i8 0, ptr [[TMP4]], align 1
+// AARCH64-LE-NEXT:    [[TMP5:%.*]] = getelementptr i8, ptr [[S]], i32 22
+// AARCH64-LE-NEXT:    store i8 0, ptr [[TMP5]], align 2
+// AARCH64-LE-NEXT:    [[TMP6:%.*]] = getelementptr i8, ptr [[S]], i32 23
+// AARCH64-LE-NEXT:    store i8 0, ptr [[TMP6]], align 1
+// AARCH64-LE-NEXT:    ret void
+//
+// AARCH64-BE-LABEL: define dso_local void @testVirtualInheritanceAndVirtualFunction(
+// AARCH64-BE-SAME: ptr nofreeobj noundef align 8 dereferenceable(24) [[S:%.*]]) #[[ATTR0]] {
+// AARCH64-BE-NEXT:  [[ENTRY:.*:]]
+// AARCH64-BE-NEXT:    [[S_INDIRECT_ADDR:%.*]] = alloca ptr, align 8
+// AARCH64-BE-NEXT:    store ptr [[S]], ptr [[S_INDIRECT_ADDR]], align 8
+// AARCH64-BE-NEXT:    [[TMP0:%.*]] = getelementptr i8, ptr [[S]], i32 13
+// AARCH64-BE-NEXT:    store i8 0, ptr [[TMP0]], align 1
+// AARCH64-BE-NEXT:    [[TMP1:%.*]] = getelementptr i8, ptr [[S]], i32 14
+// AARCH64-BE-NEXT:    store i8 0, ptr [[TMP1]], align 2
+// AARCH64-BE-NEXT:    [[TMP2:%.*]] = getelementptr i8, ptr [[S]], i32 15
+// AARCH64-BE-NEXT:    store i8 0, ptr [[TMP2]], align 1
+// AARCH64-BE-NEXT:    [[TMP3:%.*]] = getelementptr i8, ptr [[S]], i32 20
+// AARCH64-BE-NEXT:    store i8 0, ptr [[TMP3]], align 4
+// AARCH64-BE-NEXT:    [[TMP4:%.*]] = getelementptr i8, ptr [[S]], i32 21
+// AARCH64-BE-NEXT:    store i8 0, ptr [[TMP4]], align 1
+// AARCH64-BE-NEXT:    [[TMP5:%.*]] = getelementptr i8, ptr [[S]], i32 22
+// AARCH64-BE-NEXT:    store i8 0, ptr [[TMP5]], align 2
+// AARCH64-BE-NEXT:    [[TMP6:%.*]] = getelementptr i8, ptr [[S]], i32 23
+// AARCH64-BE-NEXT:    store i8 0, ptr [[TMP6]], align 1
+// AARCH64-BE-NEXT:    ret void
+//
+extern "C" void testVirtualInheritanceAndVirtualFunction(S10_2 s) {
   __builtin_clear_padding(&s);
 }
 
+
 struct S11 {
   // will usually occupy 2 bytes:
   unsigned char b1 : 3; // 1st 3 bits (in 1st byte) are b1
@@ -1123,7 +2484,7 @@ struct S11 {
   unsigned char b4 : 2; // 2 bits for b4 - next (and final) bits in the 2nd byte
 };
 
-// LINUX-LABEL: define dso_local void @_Z13testBitFieldsP3S11(
+// LINUX-LABEL: define dso_local void @testBitFields(
 // LINUX-SAME: ptr noundef [[S:%.*]]) #[[ATTR0]] {
 // LINUX-NEXT:  [[ENTRY:.*:]]
 // LINUX-NEXT:    [[S_ADDR:%.*]] = alloca ptr, align 8
@@ -1135,96 +2496,1053 @@ struct S11 {
 // LINUX-NEXT:    store i8 [[TMP3]], ptr [[TMP1]], align 1
 // LINUX-NEXT:    ret void
 //
-// WINDOWS-LABEL: define dso_local void @"?testBitFields@@YAXPEAUS11@@@Z"(ptr noundef %s) #0 {
-// WINDOWS-NEXT: entry:
-// WINDOWS-NEXT:   %s.addr = alloca ptr, align 8
-// WINDOWS-NEXT:   store ptr %s, ptr %s.addr, align 8
-// WINDOWS-NEXT:   %0 = load ptr, ptr %s.addr, align 8
-// WINDOWS-NEXT:   %1 = getelementptr i8, ptr %0, i32 0
-// WINDOWS-NEXT:   %2 = load i8, ptr %1, align 1
-// WINDOWS-NEXT:   %3 = and i8 %2, 31
-// WINDOWS-NEXT:   store i8 %3, ptr %1, align 1
-// WINDOWS-NEXT:   ret void
-//
-void testBitFields(S11 *s) {
+// WINDOWS-LABEL: define dso_local void @testBitFields(
+// WINDOWS-SAME: ptr noundef [[S:%.*]]) #[[ATTR0]] {
+// WINDOWS-NEXT:  [[ENTRY:.*:]]
+// WINDOWS-NEXT:    [[S_ADDR:%.*]] = alloca ptr, align 8
+// WINDOWS-NEXT:    store ptr [[S]], ptr [[S_ADDR]], align 8
+// WINDOWS-NEXT:    [[TMP0:%.*]] = load ptr, ptr [[S_ADDR]], align 8
+// WINDOWS-NEXT:    [[TMP1:%.*]] = getelementptr i8, ptr [[TMP0]], i32 0
+// WINDOWS-NEXT:    [[TMP2:%.*]] = load i8, ptr [[TMP1]], align 1
+// WINDOWS-NEXT:    [[TMP3:%.*]] = and i8 [[TMP2]], 31
+// WINDOWS-NEXT:    store i8 [[TMP3]], ptr [[TMP1]], align 1
+// WINDOWS-NEXT:    ret void
+//
+// ARM-LE-LABEL: define dso_local void @testBitFields(
+// ARM-LE-SAME: ptr noundef [[S:%.*]]) #[[ATTR0]] {
+// ARM-LE-NEXT:  [[ENTRY:.*:]]
+// ARM-LE-NEXT:    [[S_ADDR:%.*]] = alloca ptr, align 4
+// ARM-LE-NEXT:    store ptr [[S]], ptr [[S_ADDR]], align 4
+// ARM-LE-NEXT:    [[TMP0:%.*]] = load ptr, ptr [[S_ADDR]], align 4
+// ARM-LE-NEXT:    [[TMP1:%.*]] = getelementptr i8, ptr [[TMP0]], i32 0
+// ARM-LE-NEXT:    [[TMP2:%.*]] = load i8, ptr [[TMP1]], align 1
+// ARM-LE-NEXT:    [[TMP3:%.*]] = and i8 [[TMP2]], 31
+// ARM-LE-NEXT:    store i8 [[TMP3]], ptr [[TMP1]], align 1
+// ARM-LE-NEXT:    ret void
+//
+// ARM-BE-LABEL: define dso_local void @testBitFields(
+// ARM-BE-SAME: ptr noundef [[S:%.*]]) #[[ATTR0]] {
+// ARM-BE-NEXT:  [[ENTRY:.*:]]
+// ARM-BE-NEXT:    [[S_ADDR:%.*]] = alloca ptr, align 4
+// ARM-BE-NEXT:    store ptr [[S]], ptr [[S_ADDR]], align 4
+// ARM-BE-NEXT:    [[TMP0:%.*]] = load ptr, ptr [[S_ADDR]], align 4
+// ARM-BE-NEXT:    [[TMP1:%.*]] = getelementptr i8, ptr [[TMP0]], i32 0
+// ARM-BE-NEXT:    [[TMP2:%.*]] = load i8, ptr [[TMP1]], align 1
+// ARM-BE-NEXT:    [[TMP3:%.*]] = and i8 [[TMP2]], -8
+// ARM-BE-NEXT:    store i8 [[TMP3]], ptr [[TMP1]], align 1
+// ARM-BE-NEXT:    ret void
+//
+// AARCH64-LE-LABEL: define dso_local void @testBitFields(
+// AARCH64-LE-SAME: ptr noundef [[S:%.*]]) #[[ATTR0]] {
+// AARCH64-LE-NEXT:  [[ENTRY:.*:]]
+// AARCH64-LE-NEXT:    [[S_ADDR:%.*]] = alloca ptr, align 8
+// AARCH64-LE-NEXT:    store ptr [[S]], ptr [[S_ADDR]], align 8
+// AARCH64-LE-NEXT:    [[TMP0:%.*]] = load ptr, ptr [[S_ADDR]], align 8
+// AARCH64-LE-NEXT:    [[TMP1:%.*]] = getelementptr i8, ptr [[TMP0]], i32 0
+// AARCH64-LE-NEXT:    [[TMP2:%.*]] = load i8, ptr [[TMP1]], align 1
+// AARCH64-LE-NEXT:    [[TMP3:%.*]] = and i8 [[TMP2]], 31
+// AARCH64-LE-NEXT:    store i8 [[TMP3]], ptr [[TMP1]], align 1
+// AARCH64-LE-NEXT:    ret void
+//
+// AARCH64-BE-LABEL: define dso_local void @testBitFields(
+// AARCH64-BE-SAME: ptr noundef [[S:%.*]]) #[[ATTR0]] {
+// AARCH64-BE-NEXT:  [[ENTRY:.*:]]
+// AARCH64-BE-NEXT:    [[S_ADDR:%.*]] = alloca ptr, align 8
+// AARCH64-BE-NEXT:    store ptr [[S]], ptr [[S_ADDR]], align 8
+// AARCH64-BE-NEXT:    [[TMP0:%.*]] = load ptr, ptr [[S_ADDR]], align 8
+// AARCH64-BE-NEXT:    [[TMP1:%.*]] = getelementptr i8, ptr [[TMP0]], i32 0
+// AARCH64-BE-NEXT:    [[TMP2:%.*]] = load i8, ptr [[TMP1]], align 1
+// AARCH64-BE-NEXT:    [[TMP3:%.*]] = and i8 [[TMP2]], -8
+// AARCH64-BE-NEXT:    store i8 [[TMP3]], ptr [[TMP1]], align 1
+// AARCH64-BE-NEXT:    ret void
+//
+extern "C" void testBitFields(S11 *s) {
   // "b1" [0-2], "b2" [3-4], PAD [5-7], "b3" [8-13], "b4" [14-15]
-  // to clear 5-7, we should AND 0b00011111 (31)
+  // Little endian: preserve byte 0 with 0x1f. Big endian: use 0xf8.
   __builtin_clear_padding(s);
 }
 
 
-// LINUX-LABEL: define dso_local void @_Z18testArrayNoPaddingv(
-// LINUX-SAME: ) #[[ATTR0]] {
+struct VolatileBitfield {
+  volatile unsigned char a : 5;
+  int b : 7;
+  volatile int c : 27;
+};
+
+// LINUX-LABEL: define dso_local void @testVolatileBitfields(
+// LINUX-SAME: ptr noundef [[S:%.*]]) #[[ATTR0]] {
 // LINUX-NEXT:  [[ENTRY:.*:]]
-// LINUX-NEXT:    [[I:%.*]] = alloca [4 x i32], align 16
+// LINUX-NEXT:    [[S_ADDR:%.*]] = alloca ptr, align 8
+// LINUX-NEXT:    store ptr [[S]], ptr [[S_ADDR]], align 8
+// LINUX-NEXT:    [[TMP0:%.*]] = load ptr, ptr [[S_ADDR]], align 8
+// LINUX-NEXT:    [[TMP1:%.*]] = getelementptr i8, ptr [[TMP0]], i32 1
+// LINUX-NEXT:    [[TMP2:%.*]] = load i8, ptr [[TMP1]], align 1
+// LINUX-NEXT:    [[TMP3:%.*]] = and i8 [[TMP2]], 15
+// LINUX-NEXT:    store i8 [[TMP3]], ptr [[TMP1]], align 1
+// LINUX-NEXT:    [[TMP4:%.*]] = getelementptr i8, ptr [[TMP0]], i32 2
+// LINUX-NEXT:    store i8 0, ptr [[TMP4]], align 2
+// LINUX-NEXT:    [[TMP5:%.*]] = getelementptr i8, ptr [[TMP0]], i32 3
+// LINUX-NEXT:    store i8 0, ptr [[TMP5]], align 1
+// LINUX-NEXT:    [[TMP6:%.*]] = getelementptr i8, ptr [[TMP0]], i32 7
+// LINUX-NEXT:    [[TMP7:%.*]] = load i8, ptr [[TMP6]], align 1
+// LINUX-NEXT:    [[TMP8:%.*]] = and i8 [[TMP7]], 7
+// LINUX-NEXT:    store i8 [[TMP8]], ptr [[TMP6]], align 1
 // LINUX-NEXT:    ret void
 //
-// WINDOWS-LABEL: define dso_local void @"?testArrayNoPadding@@YAXXZ"() #0 {
-// WINDOWS-NEXT: entry:
-// WINDOWS-NEXT:   %i = alloca [4 x i32], align 4
-// WINDOWS-NEXT:   ret void
+// WINDOWS-LABEL: define dso_local void @testVolatileBitfields(
+// WINDOWS-SAME: ptr noundef [[S:%.*]]) #[[ATTR0]] {
+// WINDOWS-NEXT:  [[ENTRY:.*:]]
+// WINDOWS-NEXT:    [[S_ADDR:%.*]] = alloca ptr, align 8
+// WINDOWS-NEXT:    store ptr [[S]], ptr [[S_ADDR]], align 8
+// WINDOWS-NEXT:    [[TMP0:%.*]] = load ptr, ptr [[S_ADDR]], align 8
+// WINDOWS-NEXT:    [[TMP1:%.*]] = getelementptr i8, ptr [[TMP0]], i32 0
+// WINDOWS-NEXT:    [[TMP2:%.*]] = load i8, ptr [[TMP1]], align 4
+// WINDOWS-NEXT:    [[TMP3:%.*]] = and i8 [[TMP2]], 31
+// WINDOWS-NEXT:    store i8 [[TMP3]], ptr [[TMP1]], align 4
+// WINDOWS-NEXT:    [[TMP4:%.*]] = getelementptr i8, ptr [[TMP0]], i32 1
+// WINDOWS-NEXT:    store i8 0, ptr [[TMP4]], align 1
+// WINDOWS-NEXT:    [[TMP5:%.*]] = getelementptr i8, ptr [[TMP0]], i32 2
+// WINDOWS-NEXT:    store i8 0, ptr [[TMP5]], align 2
+// WINDOWS-NEXT:    [[TMP6:%.*]] = getelementptr i8, ptr [[TMP0]], i32 3
+// WINDOWS-NEXT:    store i8 0, ptr [[TMP6]], align 1
+// WINDOWS-NEXT:    [[TMP7:%.*]] = getelementptr i8, ptr [[TMP0]], i32 4
+// WINDOWS-NEXT:    [[TMP8:%.*]] = load i8, ptr [[TMP7]], align 4
+// WINDOWS-NEXT:    [[TMP9:%.*]] = and i8 [[TMP8]], 127
+// WINDOWS-NEXT:    store i8 [[TMP9]], ptr [[TMP7]], align 4
+// WINDOWS-NEXT:    [[TMP10:%.*]] = getelementptr i8, ptr [[TMP0]], i32 5
+// WINDOWS-NEXT:    store i8 0, ptr [[TMP10]], align 1
+// WINDOWS-NEXT:    [[TMP11:%.*]] = getelementptr i8, ptr [[TMP0]], i32 6
+// WINDOWS-NEXT:    store i8 0, ptr [[TMP11]], align 2
+// WINDOWS-NEXT:    [[TMP12:%.*]] = getelementptr i8, ptr [[TMP0]], i32 7
+// WINDOWS-NEXT:    store i8 0, ptr [[TMP12]], align 1
+// WINDOWS-NEXT:    [[TMP13:%.*]] = getelementptr i8, ptr [[TMP0]], i32 11
+// WINDOWS-NEXT:    [[TMP14:%.*]] = load i8, ptr [[TMP13]], align 1
+// WINDOWS-NEXT:    [[TMP15:%.*]] = and i8 [[TMP14]], 7
+// WINDOWS-NEXT:    store i8 [[TMP15]], ptr [[TMP13]], align 1
+// WINDOWS-NEXT:    ret void
 //
-void testArrayNoPadding() {
-  int i[4];
-  // there is no padding in the array.
-  __builtin_clear_padding(&i);
+// ARM-LE-LABEL: define dso_local void @testVolatileBitfields(
+// ARM-LE-SAME: ptr noundef [[S:%.*]]) #[[ATTR0]] {
+// ARM-LE-NEXT:  [[ENTRY:.*:]]
+// ARM-LE-NEXT:    [[S_ADDR:%.*]] = alloca ptr, align 4
+// ARM-LE-NEXT:    store ptr [[S]], ptr [[S_ADDR]], align 4
+// ARM-LE-NEXT:    [[TMP0:%.*]] = load ptr, ptr [[S_ADDR]], align 4
+// ARM-LE-NEXT:    [[TMP1:%.*]] = getelementptr i8, ptr [[TMP0]], i32 1
+// ARM-LE-NEXT:    [[TMP2:%.*]] = load i8, ptr [[TMP1]], align 1
+// ARM-LE-NEXT:    [[TMP3:%.*]] = and i8 [[TMP2]], 15
+// ARM-LE-NEXT:    store i8 [[TMP3]], ptr [[TMP1]], align 1
+// ARM-LE-NEXT:    [[TMP4:%.*]] = getelementptr i8, ptr [[TMP0]], i32 2
+// ARM-LE-NEXT:    store i8 0, ptr [[TMP4]], align 2
+// ARM-LE-NEXT:    [[TMP5:%.*]] = getelementptr i8, ptr [[TMP0]], i32 3
+// ARM-LE-NEXT:    store i8 0, ptr [[TMP5]], align 1
+// ARM-LE-NEXT:    [[TMP6:%.*]] = getelementptr i8, ptr [[TMP0]], i32 7
+// ARM-LE-NEXT:    [[TMP7:%.*]] = load i8, ptr [[TMP6]], align 1
+// ARM-LE-NEXT:    [[TMP8:%.*]] = and i8 [[TMP7]], 7
+// ARM-LE-NEXT:    store i8 [[TMP8]], ptr [[TMP6]], align 1
+// ARM-LE-NEXT:    ret void
+//
+// ARM-BE-LABEL: define dso_local void @testVolatileBitfields(
+// ARM-BE-SAME: ptr noundef [[S:%.*]]) #[[ATTR0]] {
+// ARM-BE-NEXT:  [[ENTRY:.*:]]
+// ARM-BE-NEXT:    [[S_ADDR:%.*]] = alloca ptr, align 4
+// ARM-BE-NEXT:    store ptr [[S]], ptr [[S_ADDR]], align 4
+// ARM-BE-NEXT:    [[TMP0:%.*]] = load ptr, ptr [[S_ADDR]], align 4
+// ARM-BE-NEXT:    [[TMP1:%.*]] = getelementptr i8, ptr [[TMP0]], i32 1
+// ARM-BE-NEXT:    [[TMP2:%.*]] = load i8, ptr [[TMP1]], align 1
+// ARM-BE-NEXT:    [[TMP3:%.*]] = and i8 [[TMP2]], -16
+// ARM-BE-NEXT:    store i8 [[TMP3]], ptr [[TMP1]], align 1
+// ARM-BE-NEXT:    [[TMP4:%.*]] = getelementptr i8, ptr [[TMP0]], i32 2
+// ARM-BE-NEXT:    store i8 0, ptr [[TMP4]], align 2
+// ARM-BE-NEXT:    [[TMP5:%.*]] = getelementptr i8, ptr [[TMP0]], i32 3
+// ARM-BE-NEXT:    store i8 0, ptr [[TMP5]], align 1
+// ARM-BE-NEXT:    [[TMP6:%.*]] = getelementptr i8, ptr [[TMP0]], i32 7
+// ARM-BE-NEXT:    [[TMP7:%.*]] = load i8, ptr [[TMP6]], align 1
+// ARM-BE-NEXT:    [[TMP8:%.*]] = and i8 [[TMP7]], -32
+// ARM-BE-NEXT:    store i8 [[TMP8]], ptr [[TMP6]], align 1
+// ARM-BE-NEXT:    ret void
+//
+// AARCH64-LE-LABEL: define dso_local void @testVolatileBitfields(
+// AARCH64-LE-SAME: ptr noundef [[S:%.*]]) #[[ATTR0]] {
+// AARCH64-LE-NEXT:  [[ENTRY:.*:]]
+// AARCH64-LE-NEXT:    [[S_ADDR:%.*]] = alloca ptr, align 8
+// AARCH64-LE-NEXT:    store ptr [[S]], ptr [[S_ADDR]], align 8
+// AARCH64-LE-NEXT:    [[TMP0:%.*]] = load ptr, ptr [[S_ADDR]], align 8
+// AARCH64-LE-NEXT:    [[TMP1:%.*]] = getelementptr i8, ptr [[TMP0]], i32 1
+// AARCH64-LE-NEXT:    [[TMP2:%.*]] = load i8, ptr [[TMP1]], align 1
+// AARCH64-LE-NEXT:    [[TMP3:%.*]] = and i8 [[TMP2]], 15
+// AARCH64-LE-NEXT:    store i8 [[TMP3]], ptr [[TMP1]], align 1
+// AARCH64-LE-NEXT:    [[TMP4:%.*]] = getelementptr i8, ptr [[TMP0]], i32 2
+// AARCH64-LE-NEXT:    store i8 0, ptr [[TMP4]], align 2
+// AARCH64-LE-NEXT:    [[TMP5:%.*]] = getelementptr i8, ptr [[TMP0]], i32 3
+// AARCH64-LE-NEXT:    store i8 0, ptr [[TMP5]], align 1
+// AARCH64-LE-NEXT:    [[TMP6:%.*]] = getelementptr i8, ptr [[TMP0]], i32 7
+// AARCH64-LE-NEXT:    [[TMP7:%.*]] = load i8, ptr [[TMP6]], align 1
+// AARCH64-LE-NEXT:    [[TMP8:%.*]] = and i8 [[TMP7]], 7
+// AARCH64-LE-NEXT:    store i8 [[TMP8]], ptr [[TMP6]], align 1
+// AARCH64-LE-NEXT:    ret void
+//
+// AARCH64-BE-LABEL: define dso_local void @testVolatileBitfields(
+// AARCH64-BE-SAME: ptr noundef [[S:%.*]]) #[[ATTR0]] {
+// AARCH64-BE-NEXT:  [[ENTRY:.*:]]
+// AARCH64-BE-NEXT:    [[S_ADDR:%.*]] = alloca ptr, align 8
+// AARCH64-BE-NEXT:    store ptr [[S]], ptr [[S_ADDR]], align 8
+// AARCH64-BE-NEXT:    [[TMP0:%.*]] = load ptr, ptr [[S_ADDR]], align 8
+// AARCH64-BE-NEXT:    [[TMP1:%.*]] = getelementptr i8, ptr [[TMP0]], i32 1
+// AARCH64-BE-NEXT:    [[TMP2:%.*]] = load i8, ptr [[TMP1]], align 1
+// AARCH64-BE-NEXT:    [[TMP3:%.*]] = and i8 [[TMP2]], -16
+// AARCH64-BE-NEXT:    store i8 [[TMP3]], ptr [[TMP1]], align 1
+// AARCH64-BE-NEXT:    [[TMP4:%.*]] = getelementptr i8, ptr [[TMP0]], i32 2
+// AARCH64-BE-NEXT:    store i8 0, ptr [[TMP4]], align 2
+// AARCH64-BE-NEXT:    [[TMP5:%.*]] = getelementptr i8, ptr [[TMP0]], i32 3
+// AARCH64-BE-NEXT:    store i8 0, ptr [[TMP5]], align 1
+// AARCH64-BE-NEXT:    [[TMP6:%.*]] = getelementptr i8, ptr [[TMP0]], i32 7
+// AARCH64-BE-NEXT:    [[TMP7:%.*]] = load i8, ptr [[TMP6]], align 1
+// AARCH64-BE-NEXT:    [[TMP8:%.*]] = and i8 [[TMP7]], -32
+// AARCH64-BE-NEXT:    store i8 [[TMP8]], ptr [[TMP6]], align 1
+// AARCH64-BE-NEXT:    ret void
+//
+extern "C" void testVolatileBitfields(struct VolatileBitfield *s) {
+  __builtin_clear_padding(s);
 }
 
-// LINUX-LABEL: define dso_local void @_Z19testArrayLongDoublev(
-// LINUX-SAME: ) #[[ATTR0]] {
+// LINUX-LABEL: define dso_local void @testVolatileBitfieldsVolatileStruct(
+// LINUX-SAME: ptr noundef [[S:%.*]]) #[[ATTR0]] {
 // LINUX-NEXT:  [[ENTRY:.*:]]
-// LINUX-NEXT:    [[LD:%.*]] = alloca [2 x x86_fp80], align 16
-// LINUX-NEXT:    [[TMP0:%.*]] = getelementptr i8, ptr [[LD]], i32 10
-// LINUX-NEXT:    store i8 0, ptr [[TMP0]], align 2
-// LINUX-NEXT:    [[TMP1:%.*]] = getelementptr i8, ptr [[LD]], i32 11
+// LINUX-NEXT:    [[S_ADDR:%.*]] = alloca ptr, align 8
+// LINUX-NEXT:    store ptr [[S]], ptr [[S_ADDR]], align 8
+// LINUX-NEXT:    [[TMP0:%.*]] = load ptr, ptr [[S_ADDR]], align 8
+// LINUX-NEXT:    [[TMP1:%.*]] = getelementptr i8, ptr [[TMP0]], i32 1
+// LINUX-NEXT:    [[TMP2:%.*]] = load i8, ptr [[TMP1]], align 1
+// LINUX-NEXT:    [[TMP3:%.*]] = and i8 [[TMP2]], 15
+// LINUX-NEXT:    store i8 [[TMP3]], ptr [[TMP1]], align 1
+// LINUX-NEXT:    [[TMP4:%.*]] = getelementptr i8, ptr [[TMP0]], i32 2
+// LINUX-NEXT:    store i8 0, ptr [[TMP4]], align 2
+// LINUX-NEXT:    [[TMP5:%.*]] = getelementptr i8, ptr [[TMP0]], i32 3
+// LINUX-NEXT:    store i8 0, ptr [[TMP5]], align 1
+// LINUX-NEXT:    [[TMP6:%.*]] = getelementptr i8, ptr [[TMP0]], i32 7
+// LINUX-NEXT:    [[TMP7:%.*]] = load i8, ptr [[TMP6]], align 1
+// LINUX-NEXT:    [[TMP8:%.*]] = and i8 [[TMP7]], 7
+// LINUX-NEXT:    store i8 [[TMP8]], ptr [[TMP6]], align 1
+// LINUX-NEXT:    ret void
+//
+// WINDOWS-LABEL: define dso_local void @testVolatileBitfieldsVolatileStruct(
+// WINDOWS-SAME: ptr noundef [[S:%.*]]) #[[ATTR0]] {
+// WINDOWS-NEXT:  [[ENTRY:.*:]]
+// WINDOWS-NEXT:    [[S_ADDR:%.*]] = alloca ptr, align 8
+// WINDOWS-NEXT:    store ptr [[S]], ptr [[S_ADDR]], align 8
+// WINDOWS-NEXT:    [[TMP0:%.*]] = load ptr, ptr [[S_ADDR]], align 8
+// WINDOWS-NEXT:    [[TMP1:%.*]] = getelementptr i8, ptr [[TMP0]], i32 0
+// WINDOWS-NEXT:    [[TMP2:%.*]] = load i8, ptr [[TMP1]], align 4
+// WINDOWS-NEXT:    [[TMP3:%.*]] = and i8 [[TMP2]], 31
+// WINDOWS-NEXT:    store i8 [[TMP3]], ptr [[TMP1]], align 4
+// WINDOWS-NEXT:    [[TMP4:%.*]] = getelementptr i8, ptr [[TMP0]], i32 1
+// WINDOWS-NEXT:    store i8 0, ptr [[TMP4]], align 1
+// WINDOWS-NEXT:    [[TMP5:%.*]] = getelementptr i8, ptr [[TMP0]], i32 2
+// WINDOWS-NEXT:    store i8 0, ptr [[TMP5]], align 2
+// WINDOWS-NEXT:    [[TMP6:%.*]] = getelementptr i8, ptr [[TMP0]], i32 3
+// WINDOWS-NEXT:    store i8 0, ptr [[TMP6]], align 1
+// WINDOWS-NEXT:    [[TMP7:%.*]] = getelementptr i8, ptr [[TMP0]], i32 4
+// WINDOWS-NEXT:    [[TMP8:%.*]] = load i8, ptr [[TMP7]], align 4
+// WINDOWS-NEXT:    [[TMP9:%.*]] = and i8 [[TMP8]], 127
+// WINDOWS-NEXT:    store i8 [[TMP9]], ptr [[TMP7]], align 4
+// WINDOWS-NEXT:    [[TMP10:%.*]] = getelementptr i8, ptr [[TMP0]], i32 5
+// WINDOWS-NEXT:    store i8 0, ptr [[TMP10]], align 1
+// WINDOWS-NEXT:    [[TMP11:%.*]] = getelementptr i8, ptr [[TMP0]], i32 6
+// WINDOWS-NEXT:    store i8 0, ptr [[TMP11]], align 2
+// WINDOWS-NEXT:    [[TMP12:%.*]] = getelementptr i8, ptr [[TMP0]], i32 7
+// WINDOWS-NEXT:    store i8 0, ptr [[TMP12]], align 1
+// WINDOWS-NEXT:    [[TMP13:%.*]] = getelementptr i8, ptr [[TMP0]], i32 11
+// WINDOWS-NEXT:    [[TMP14:%.*]] = load i8, ptr [[TMP13]], align 1
+// WINDOWS-NEXT:    [[TMP15:%.*]] = and i8 [[TMP14]], 7
+// WINDOWS-NEXT:    store i8 [[TMP15]], ptr [[TMP13]], align 1
+// WINDOWS-NEXT:    ret void
+//
+// ARM-LE-LABEL: define dso_local void @testVolatileBitfieldsVolatileStruct(
+// ARM-LE-SAME: ptr noundef [[S:%.*]]) #[[ATTR0]] {
+// ARM-LE-NEXT:  [[ENTRY:.*:]]
+// ARM-LE-NEXT:    [[S_ADDR:%.*]] = alloca ptr, align 4
+// ARM-LE-NEXT:    store ptr [[S]], ptr [[S_ADDR]], align 4
+// ARM-LE-NEXT:    [[TMP0:%.*]] = load ptr, ptr [[S_ADDR]], align 4
+// ARM-LE-NEXT:    [[TMP1:%.*]] = getelementptr i8, ptr [[TMP0]], i32 1
+// ARM-LE-NEXT:    [[TMP2:%.*]] = load i8, ptr [[TMP1]], align 1
+// ARM-LE-NEXT:    [[TMP3:%.*]] = and i8 [[TMP2]], 15
+// ARM-LE-NEXT:    store i8 [[TMP3]], ptr [[TMP1]], align 1
+// ARM-LE-NEXT:    [[TMP4:%.*]] = getelementptr i8, ptr [[TMP0]], i32 2
+// ARM-LE-NEXT:    store i8 0, ptr [[TMP4]], align 2
+// ARM-LE-NEXT:    [[TMP5:%.*]] = getelementptr i8, ptr [[TMP0]], i32 3
+// ARM-LE-NEXT:    store i8 0, ptr [[TMP5]], align 1
+// ARM-LE-NEXT:    [[TMP6:%.*]] = getelementptr i8, ptr [[TMP0]], i32 7
+// ARM-LE-NEXT:    [[TMP7:%.*]] = load i8, ptr [[TMP6]], align 1
+// ARM-LE-NEXT:    [[TMP8:%.*]] = and i8 [[TMP7]], 7
+// ARM-LE-NEXT:    store i8 [[TMP8]], ptr [[TMP6]], align 1
+// ARM-LE-NEXT:    ret void
+//
+// ARM-BE-LABEL: define dso_local void @testVolatileBitfieldsVolatileStruct(
+// ARM-BE-SAME: ptr noundef [[S:%.*]]) #[[ATTR0]] {
+// ARM-BE-NEXT:  [[ENTRY:.*:]]
+// ARM-BE-NEXT:    [[S_ADDR:%.*]] = alloca ptr, align 4
+// ARM-BE-NEXT:    store ptr [[S]], ptr [[S_ADDR]], align 4
+// ARM-BE-NEXT:    [[TMP0:%.*]] = load ptr, ptr [[S_ADDR]], align 4
+// ARM-BE-NEXT:    [[TMP1:%.*]] = getelementptr i8, ptr [[TMP0]], i32 1
+// ARM-BE-NEXT:    [[TMP2:%.*]] = load i8, ptr [[TMP1]], align 1
+// ARM-BE-NEXT:    [[TMP3:%.*]] = and i8 [[TMP2]], -16
+// ARM-BE-NEXT:    store i8 [[TMP3]], ptr [[TMP1]], align 1
+// ARM-BE-NEXT:    [[TMP4:%.*]] = getelementptr i8, ptr [[TMP0]], i32 2
+// ARM-BE-NEXT:    store i8 0, ptr [[TMP4]], align 2
+// ARM-BE-NEXT:    [[TMP5:%.*]] = getelementptr i8, ptr [[TMP0]], i32 3
+// ARM-BE-NEXT:    store i8 0, ptr [[TMP5]], align 1
+// ARM-BE-NEXT:    [[TMP6:%.*]] = getelementptr i8, ptr [[TMP0]], i32 7
+// ARM-BE-NEXT:    [[TMP7:%.*]] = load i8, ptr [[TMP6]], align 1
+// ARM-BE-NEXT:    [[TMP8:%.*]] = and i8 [[TMP7]], -32
+// ARM-BE-NEXT:    store i8 [[TMP8]], ptr [[TMP6]], align 1
+// ARM-BE-NEXT:    ret void
+//
+// AARCH64-LE-LABEL: define dso_local void @testVolatileBitfieldsVolatileStruct(
+// AARCH64-LE-SAME: ptr noundef [[S:%.*]]) #[[ATTR0]] {
+// AARCH64-LE-NEXT:  [[ENTRY:.*:]]
+// AARCH64-LE-NEXT:    [[S_ADDR:%.*]] = alloca ptr, align 8
+// AARCH64-LE-NEXT:    store ptr [[S]], ptr [[S_ADDR]], align 8
+// AARCH64-LE-NEXT:    [[TMP0:%.*]] = load ptr, ptr [[S_ADDR]], align 8
+// AARCH64-LE-NEXT:    [[TMP1:%.*]] = getelementptr i8, ptr [[TMP0]], i32 1
+// AARCH64-LE-NEXT:    [[TMP2:%.*]] = load i8, ptr [[TMP1]], align 1
+// AARCH64-LE-NEXT:    [[TMP3:%.*]] = and i8 [[TMP2]], 15
+// AARCH64-LE-NEXT:    store i8 [[TMP3]], ptr [[TMP1]], align 1
+// AARCH64-LE-NEXT:    [[TMP4:%.*]] = getelementptr i8, ptr [[TMP0]], i32 2
+// AARCH64-LE-NEXT:    store i8 0, ptr [[TMP4]], align 2
+// AARCH64-LE-NEXT:    [[TMP5:%.*]] = getelementptr i8, ptr [[TMP0]], i32 3
+// AARCH64-LE-NEXT:    store i8 0, ptr [[TMP5]], align 1
+// AARCH64-LE-NEXT:    [[TMP6:%.*]] = getelementptr i8, ptr [[TMP0]], i32 7
+// AARCH64-LE-NEXT:    [[TMP7:%.*]] = load i8, ptr [[TMP6]], align 1
+// AARCH64-LE-NEXT:    [[TMP8:%.*]] = and i8 [[TMP7]], 7
+// AARCH64-LE-NEXT:    store i8 [[TMP8]], ptr [[TMP6]], align 1
+// AARCH64-LE-NEXT:    ret void
+//
+// AARCH64-BE-LABEL: define dso_local void @testVolatileBitfieldsVolatileStruct(
+// AARCH64-BE-SAME: ptr noundef [[S:%.*]]) #[[ATTR0]] {
+// AARCH64-BE-NEXT:  [[ENTRY:.*:]]
+// AARCH64-BE-NEXT:    [[S_ADDR:%.*]] = alloca ptr, align 8
+// AARCH64-BE-NEXT:    store ptr [[S]], ptr [[S_ADDR]], align 8
+// AARCH64-BE-NEXT:    [[TMP0:%.*]] = load ptr, ptr [[S_ADDR]], align 8
+// AARCH64-BE-NEXT:    [[TMP1:%.*]] = getelementptr i8, ptr [[TMP0]], i32 1
+// AARCH64-BE-NEXT:    [[TMP2:%.*]] = load i8, ptr [[TMP1]], align 1
+// AARCH64-BE-NEXT:    [[TMP3:%.*]] = and i8 [[TMP2]], -16
+// AARCH64-BE-NEXT:    store i8 [[TMP3]], ptr [[TMP1]], align 1
+// AARCH64-BE-NEXT:    [[TMP4:%.*]] = getelementptr i8, ptr [[TMP0]], i32 2
+// AARCH64-BE-NEXT:    store i8 0, ptr [[TMP4]], align 2
+// AARCH64-BE-NEXT:    [[TMP5:%.*]] = getelementptr i8, ptr [[TMP0]], i32 3
+// AARCH64-BE-NEXT:    store i8 0, ptr [[TMP5]], align 1
+// AARCH64-BE-NEXT:    [[TMP6:%.*]] = getelementptr i8, ptr [[TMP0]], i32 7
+// AARCH64-BE-NEXT:    [[TMP7:%.*]] = load i8, ptr [[TMP6]], align 1
+// AARCH64-BE-NEXT:    [[TMP8:%.*]] = and i8 [[TMP7]], -32
+// AARCH64-BE-NEXT:    store i8 [[TMP8]], ptr [[TMP6]], align 1
+// AARCH64-BE-NEXT:    ret void
+//
+extern "C" void testVolatileBitfieldsVolatileStruct(
+    volatile struct VolatileBitfield *s) {
+  __builtin_clear_padding(s);
+}
+
+#ifndef _WIN32
+
+#pragma clang diagnostic push
+#pragma clang diagnostic ignored "-Wbitfield-width"
+
+struct OversizedBitfield {
+  unsigned char a : 10;
+  int b;
+  int c : 35;
+};
+
+// LINUX-LABEL: define dso_local void @testOversizedBitfield(
+// LINUX-SAME: ptr noundef [[S:%.*]]) #[[ATTR0]] {
+// LINUX-NEXT:  [[ENTRY:.*:]]
+// LINUX-NEXT:    [[S_ADDR:%.*]] = alloca ptr, align 8
+// LINUX-NEXT:    store ptr [[S]], ptr [[S_ADDR]], align 8
+// LINUX-NEXT:    [[TMP0:%.*]] = load ptr, ptr [[S_ADDR]], align 8
+// LINUX-NEXT:    [[TMP1:%.*]] = getelementptr i8, ptr [[TMP0]], i32 1
 // LINUX-NEXT:    store i8 0, ptr [[TMP1]], align 1
-// LINUX-NEXT:    [[TMP2:%.*]] = getelementptr i8, ptr [[LD]], i32 12
-// LINUX-NEXT:    store i8 0, ptr [[TMP2]], align 4
-// LINUX-NEXT:    [[TMP3:%.*]] = getelementptr i8, ptr [[LD]], i32 13
+// LINUX-NEXT:    [[TMP2:%.*]] = getelementptr i8, ptr [[TMP0]], i32 2
+// LINUX-NEXT:    store i8 0, ptr [[TMP2]], align 2
+// LINUX-NEXT:    [[TMP3:%.*]] = getelementptr i8, ptr [[TMP0]], i32 3
 // LINUX-NEXT:    store i8 0, ptr [[TMP3]], align 1
-// LINUX-NEXT:    [[TMP4:%.*]] = getelementptr i8, ptr [[LD]], i32 14
-// LINUX-NEXT:    store i8 0, ptr [[TMP4]], align 2
-// LINUX-NEXT:    [[TMP5:%.*]] = getelementptr i8, ptr [[LD]], i32 15
+// LINUX-NEXT:    [[TMP4:%.*]] = getelementptr i8, ptr [[TMP0]], i32 12
+// LINUX-NEXT:    store i8 0, ptr [[TMP4]], align 4
+// LINUX-NEXT:    [[TMP5:%.*]] = getelementptr i8, ptr [[TMP0]], i32 13
 // LINUX-NEXT:    store i8 0, ptr [[TMP5]], align 1
-// LINUX-NEXT:    [[TMP6:%.*]] = getelementptr i8, ptr [[LD]], i32 26
+// LINUX-NEXT:    [[TMP6:%.*]] = getelementptr i8, ptr [[TMP0]], i32 14
 // LINUX-NEXT:    store i8 0, ptr [[TMP6]], align 2
-// LINUX-NEXT:    [[TMP7:%.*]] = getelementptr i8, ptr [[LD]], i32 27
+// LINUX-NEXT:    [[TMP7:%.*]] = getelementptr i8, ptr [[TMP0]], i32 15
 // LINUX-NEXT:    store i8 0, ptr [[TMP7]], align 1
-// LINUX-NEXT:    [[TMP8:%.*]] = getelementptr i8, ptr [[LD]], i32 28
-// LINUX-NEXT:    store i8 0, ptr [[TMP8]], align 4
-// LINUX-NEXT:    [[TMP9:%.*]] = getelementptr i8, ptr [[LD]], i32 29
-// LINUX-NEXT:    store i8 0, ptr [[TMP9]], align 1
-// LINUX-NEXT:    [[TMP10:%.*]] = getelementptr i8, ptr [[LD]], i32 30
-// LINUX-NEXT:    store i8 0, ptr [[TMP10]], align 2
-// LINUX-NEXT:    [[TMP11:%.*]] = getelementptr i8, ptr [[LD]], i32 31
-// LINUX-NEXT:    store i8 0, ptr [[TMP11]], align 1
 // LINUX-NEXT:    ret void
 //
-// WINDOWS-LABEL: define dso_local void @"?testArrayLongDouble@@YAXXZ"() #0 {
-// WINDOWS-NEXT: entry:
-// WINDOWS-NEXT:   %ld = alloca [2 x double], align 8
-// WINDOWS-NEXT:   ret void
+// ARM-LE-LABEL: define dso_local void @testOversizedBitfield(
+// ARM-LE-SAME: ptr noundef [[S:%.*]]) #[[ATTR0]] {
+// ARM-LE-NEXT:  [[ENTRY:.*:]]
+// ARM-LE-NEXT:    [[S_ADDR:%.*]] = alloca ptr, align 4
+// ARM-LE-NEXT:    store ptr [[S]], ptr [[S_ADDR]], align 4
+// ARM-LE-NEXT:    [[TMP0:%.*]] = load ptr, ptr [[S_ADDR]], align 4
+// ARM-LE-NEXT:    [[TMP1:%.*]] = getelementptr i8, ptr [[TMP0]], i32 1
+// ARM-LE-NEXT:    store i8 0, ptr [[TMP1]], align 1
+// ARM-LE-NEXT:    [[TMP2:%.*]] = getelementptr i8, ptr [[TMP0]], i32 2
+// ARM-LE-NEXT:    store i8 0, ptr [[TMP2]], align 2
+// ARM-LE-NEXT:    [[TMP3:%.*]] = getelementptr i8, ptr [[TMP0]], i32 3
+// ARM-LE-NEXT:    store i8 0, ptr [[TMP3]], align 1
+// ARM-LE-NEXT:    [[TMP4:%.*]] = getelementptr i8, ptr [[TMP0]], i32 12
+// ARM-LE-NEXT:    store i8 0, ptr [[TMP4]], align 4
+// ARM-LE-NEXT:    [[TMP5:%.*]] = getelementptr i8, ptr [[TMP0]], i32 13
+// ARM-LE-NEXT:    store i8 0, ptr [[TMP5]], align 1
+// ARM-LE-NEXT:    [[TMP6:%.*]] = getelementptr i8, ptr [[TMP0]], i32 14
+// ARM-LE-NEXT:    store i8 0, ptr [[TMP6]], align 2
+// ARM-LE-NEXT:    [[TMP7:%.*]] = getelementptr i8, ptr [[TMP0]], i32 15
+// ARM-LE-NEXT:    store i8 0, ptr [[TMP7]], align 1
+// ARM-LE-NEXT:    ret void
 //
-void testArrayLongDouble() {
-  // long double 0, [0-9] PAD [10-15]
-  // long double 1, [16-25] PAD [26-31]
-  long double ld[2];
-  __builtin_clear_padding(&ld);
+// ARM-BE-LABEL: define dso_local void @testOversizedBitfield(
+// ARM-BE-SAME: ptr noundef [[S:%.*]]) #[[ATTR0]] {
+// ARM-BE-NEXT:  [[ENTRY:.*:]]
+// ARM-BE-NEXT:    [[S_ADDR:%.*]] = alloca ptr, align 4
+// ARM-BE-NEXT:    store ptr [[S]], ptr [[S_ADDR]], align 4
+// ARM-BE-NEXT:    [[TMP0:%.*]] = load ptr, ptr [[S_ADDR]], align 4
+// ARM-BE-NEXT:    [[TMP1:%.*]] = getelementptr i8, ptr [[TMP0]], i32 0
+// ARM-BE-NEXT:    [[TMP2:%.*]] = load i8, ptr [[TMP1]], align 4
+// ARM-BE-NEXT:    [[TMP3:%.*]] = and i8 [[TMP2]], 63
+// ARM-BE-NEXT:    store i8 [[TMP3]], ptr [[TMP1]], align 4
+// ARM-BE-NEXT:    [[TMP4:%.*]] = getelementptr i8, ptr [[TMP0]], i32 1
+// ARM-BE-NEXT:    [[TMP5:%.*]] = load i8, ptr [[TMP4]], align 1
+// ARM-BE-NEXT:    [[TMP6:%.*]] = and i8 [[TMP5]], -64
+// ARM-BE-NEXT:    store i8 [[TMP6]], ptr [[TMP4]], align 1
+// ARM-BE-NEXT:    [[TMP7:%.*]] = getelementptr i8, ptr [[TMP0]], i32 2
+// ARM-BE-NEXT:    store i8 0, ptr [[TMP7]], align 2
+// ARM-BE-NEXT:    [[TMP8:%.*]] = getelementptr i8, ptr [[TMP0]], i32 3
+// ARM-BE-NEXT:    store i8 0, ptr [[TMP8]], align 1
+// ARM-BE-NEXT:    [[TMP9:%.*]] = getelementptr i8, ptr [[TMP0]], i32 8
+// ARM-BE-NEXT:    [[TMP10:%.*]] = load i8, ptr [[TMP9]], align 4
+// ARM-BE-NEXT:    [[TMP11:%.*]] = and i8 [[TMP10]], 31
+// ARM-BE-NEXT:    store i8 [[TMP11]], ptr [[TMP9]], align 4
+// ARM-BE-NEXT:    [[TMP12:%.*]] = getelementptr i8, ptr [[TMP0]], i32 12
+// ARM-BE-NEXT:    [[TMP13:%.*]] = load i8, ptr [[TMP12]], align 4
+// ARM-BE-NEXT:    [[TMP14:%.*]] = and i8 [[TMP13]], -32
+// ARM-BE-NEXT:    store i8 [[TMP14]], ptr [[TMP12]], align 4
+// ARM-BE-NEXT:    [[TMP15:%.*]] = getelementptr i8, ptr [[TMP0]], i32 13
+// ARM-BE-NEXT:    store i8 0, ptr [[TMP15]], align 1
+// ARM-BE-NEXT:    [[TMP16:%.*]] = getelementptr i8, ptr [[TMP0]], i32 14
+// ARM-BE-NEXT:    store i8 0, ptr [[TMP16]], align 2
+// ARM-BE-NEXT:    [[TMP17:%.*]] = getelementptr i8, ptr [[TMP0]], i32 15
+// ARM-BE-NEXT:    store i8 0, ptr [[TMP17]], align 1
+// ARM-BE-NEXT:    ret void
+//
+// AARCH64-LE-LABEL: define dso_local void @testOversizedBitfield(
+// AARCH64-LE-SAME: ptr noundef [[S:%.*]]) #[[ATTR0]] {
+// AARCH64-LE-NEXT:  [[ENTRY:.*:]]
+// AARCH64-LE-NEXT:    [[S_ADDR:%.*]] = alloca ptr, align 8
+// AARCH64-LE-NEXT:    store ptr [[S]], ptr [[S_ADDR]], align 8
+// AARCH64-LE-NEXT:    [[TMP0:%.*]] = load ptr, ptr [[S_ADDR]], align 8
+// AARCH64-LE-NEXT:    [[TMP1:%.*]] = getelementptr i8, ptr [[TMP0]], i32 1
+// AARCH64-LE-NEXT:    store i8 0, ptr [[TMP1]], align 1
+// AARCH64-LE-NEXT:    [[TMP2:%.*]] = getelementptr i8, ptr [[TMP0]], i32 2
+// AARCH64-LE-NEXT:    store i8 0, ptr [[TMP2]], align 2
+// AARCH64-LE-NEXT:    [[TMP3:%.*]] = getelementptr i8, ptr [[TMP0]], i32 3
+// AARCH64-LE-NEXT:    store i8 0, ptr [[TMP3]], align 1
+// AARCH64-LE-NEXT:    [[TMP4:%.*]] = getelementptr i8, ptr [[TMP0]], i32 12
+// AARCH64-LE-NEXT:    store i8 0, ptr [[TMP4]], align 4
+// AARCH64-LE-NEXT:    [[TMP5:%.*]] = getelementptr i8, ptr [[TMP0]], i32 13
+// AARCH64-LE-NEXT:    store i8 0, ptr [[TMP5]], align 1
+// AARCH64-LE-NEXT:    [[TMP6:%.*]] = getelementptr i8, ptr [[TMP0]], i32 14
+// AARCH64-LE-NEXT:    store i8 0, ptr [[TMP6]], align 2
+// AARCH64-LE-NEXT:    [[TMP7:%.*]] = getelementptr i8, ptr [[TMP0]], i32 15
+// AARCH64-LE-NEXT:    store i8 0, ptr [[TMP7]], align 1
+// AARCH64-LE-NEXT:    ret void
+//
+// AARCH64-BE-LABEL: define dso_local void @testOversizedBitfield(
+// AARCH64-BE-SAME: ptr noundef [[S:%.*]]) #[[ATTR0]] {
+// AARCH64-BE-NEXT:  [[ENTRY:.*:]]
+// AARCH64-BE-NEXT:    [[S_ADDR:%.*]] = alloca ptr, align 8
+// AARCH64-BE-NEXT:    store ptr [[S]], ptr [[S_ADDR]], align 8
+// AARCH64-BE-NEXT:    [[TMP0:%.*]] = load ptr, ptr [[S_ADDR]], align 8
+// AARCH64-BE-NEXT:    [[TMP1:%.*]] = getelementptr i8, ptr [[TMP0]], i32 0
+// AARCH64-BE-NEXT:    [[TMP2:%.*]] = load i8, ptr [[TMP1]], align 4
+// AARCH64-BE-NEXT:    [[TMP3:%.*]] = and i8 [[TMP2]], 63
+// AARCH64-BE-NEXT:    store i8 [[TMP3]], ptr [[TMP1]], align 4
+// AARCH64-BE-NEXT:    [[TMP4:%.*]] = getelementptr i8, ptr [[TMP0]], i32 1
+// AARCH64-BE-NEXT:    [[TMP5:%.*]] = load i8, ptr [[TMP4]], align 1
+// AARCH64-BE-NEXT:    [[TMP6:%.*]] = and i8 [[TMP5]], -64
+// AARCH64-BE-NEXT:    store i8 [[TMP6]], ptr [[TMP4]], align 1
+// AARCH64-BE-NEXT:    [[TMP7:%.*]] = getelementptr i8, ptr [[TMP0]], i32 2
+// AARCH64-BE-NEXT:    store i8 0, ptr [[TMP7]], align 2
+// AARCH64-BE-NEXT:    [[TMP8:%.*]] = getelementptr i8, ptr [[TMP0]], i32 3
+// AARCH64-BE-NEXT:    store i8 0, ptr [[TMP8]], align 1
+// AARCH64-BE-NEXT:    [[TMP9:%.*]] = getelementptr i8, ptr [[TMP0]], i32 8
+// AARCH64-BE-NEXT:    [[TMP10:%.*]] = load i8, ptr [[TMP9]], align 4
+// AARCH64-BE-NEXT:    [[TMP11:%.*]] = and i8 [[TMP10]], 31
+// AARCH64-BE-NEXT:    store i8 [[TMP11]], ptr [[TMP9]], align 4
+// AARCH64-BE-NEXT:    [[TMP12:%.*]] = getelementptr i8, ptr [[TMP0]], i32 12
+// AARCH64-BE-NEXT:    [[TMP13:%.*]] = load i8, ptr [[TMP12]], align 4
+// AARCH64-BE-NEXT:    [[TMP14:%.*]] = and i8 [[TMP13]], -32
+// AARCH64-BE-NEXT:    store i8 [[TMP14]], ptr [[TMP12]], align 4
+// AARCH64-BE-NEXT:    [[TMP15:%.*]] = getelementptr i8, ptr [[TMP0]], i32 13
+// AARCH64-BE-NEXT:    store i8 0, ptr [[TMP15]], align 1
+// AARCH64-BE-NEXT:    [[TMP16:%.*]] = getelementptr i8, ptr [[TMP0]], i32 14
+// AARCH64-BE-NEXT:    store i8 0, ptr [[TMP16]], align 2
+// AARCH64-BE-NEXT:    [[TMP17:%.*]] = getelementptr i8, ptr [[TMP0]], i32 15
+// AARCH64-BE-NEXT:    store i8 0, ptr [[TMP17]], align 1
+// AARCH64-BE-NEXT:    ret void
+//
+extern "C" void testOversizedBitfield(struct OversizedBitfield *s) {
+  __builtin_clear_padding(s);
 }
 
-// LINUX-LABEL: define dso_local void @_Z17testArrayOfStructv(
-// LINUX-SAME: ) #[[ATTR0]] {
+// LINUX-LABEL: define dso_local void @testOversizedBitfieldVolatileStruct(
+// LINUX-SAME: ptr noundef [[S:%.*]]) #[[ATTR0]] {
 // LINUX-NEXT:  [[ENTRY:.*:]]
-// LINUX-NEXT:    [[S:%.*]] = alloca [2 x [[STRUCT_S_0:%.*]]], align 16
-// LINUX-NEXT:    [[TMP0:%.*]] = getelementptr i8, ptr [[S]], i32 5
-// LINUX-NEXT:    store i8 0, ptr [[TMP0]], align 1
-// LINUX-NEXT:    [[TMP1:%.*]] = getelementptr i8, ptr [[S]], i32 6
-// LINUX-NEXT:    store i8 0, ptr [[TMP1]], align 2
-// LINUX-NEXT:    [[TMP2:%.*]] = getelementptr i8, ptr [[S]], i32 7
-// LINUX-NEXT:    store i8 0, ptr [[TMP2]], align 1
-// LINUX-NEXT:    [[TMP3:%.*]] = getelementptr i8, ptr [[S]], i32 13
-// LINUX-NEXT:    store i8 0, ptr [[TMP3]], align 1
-// LINUX-NEXT:    [[TMP4:%.*]] = getelementptr i8, ptr [[S]], i32 14
+// LINUX-NEXT:    [[S_ADDR:%.*]] = alloca ptr, align 8
+// LINUX-NEXT:    store ptr [[S]], ptr [[S_ADDR]], align 8
+// LINUX-NEXT:    [[TMP0:%.*]] = load ptr, ptr [[S_ADDR]], align 8
+// LINUX-NEXT:    [[TMP1:%.*]] = getelementptr i8, ptr [[TMP0]], i32 1
+// LINUX-NEXT:    store i8 0, ptr [[TMP1]], align 1
+// LINUX-NEXT:    [[TMP2:%.*]] = getelementptr i8, ptr [[TMP0]], i32 2
+// LINUX-NEXT:    store i8 0, ptr [[TMP2]], align 2
+// LINUX-NEXT:    [[TMP3:%.*]] = getelementptr i8, ptr [[TMP0]], i32 3
+// LINUX-NEXT:    store i8 0, ptr [[TMP3]], align 1
+// LINUX-NEXT:    [[TMP4:%.*]] = getelementptr i8, ptr [[TMP0]], i32 12
+// LINUX-NEXT:    store i8 0, ptr [[TMP4]], align 4
+// LINUX-NEXT:    [[TMP5:%.*]] = getelementptr i8, ptr [[TMP0]], i32 13
+// LINUX-NEXT:    store i8 0, ptr [[TMP5]], align 1
+// LINUX-NEXT:    [[TMP6:%.*]] = getelementptr i8, ptr [[TMP0]], i32 14
+// LINUX-NEXT:    store i8 0, ptr [[TMP6]], align 2
+// LINUX-NEXT:    [[TMP7:%.*]] = getelementptr i8, ptr [[TMP0]], i32 15
+// LINUX-NEXT:    store i8 0, ptr [[TMP7]], align 1
+// LINUX-NEXT:    ret void
+//
+// ARM-LE-LABEL: define dso_local void @testOversizedBitfieldVolatileStruct(
+// ARM-LE-SAME: ptr noundef [[S:%.*]]) #[[ATTR0]] {
+// ARM-LE-NEXT:  [[ENTRY:.*:]]
+// ARM-LE-NEXT:    [[S_ADDR:%.*]] = alloca ptr, align 4
+// ARM-LE-NEXT:    store ptr [[S]], ptr [[S_ADDR]], align 4
+// ARM-LE-NEXT:    [[TMP0:%.*]] = load ptr, ptr [[S_ADDR]], align 4
+// ARM-LE-NEXT:    [[TMP1:%.*]] = getelementptr i8, ptr [[TMP0]], i32 1
+// ARM-LE-NEXT:    store i8 0, ptr [[TMP1]], align 1
+// ARM-LE-NEXT:    [[TMP2:%.*]] = getelementptr i8, ptr [[TMP0]], i32 2
+// ARM-LE-NEXT:    store i8 0, ptr [[TMP2]], align 2
+// ARM-LE-NEXT:    [[TMP3:%.*]] = getelementptr i8, ptr [[TMP0]], i32 3
+// ARM-LE-NEXT:    store i8 0, ptr [[TMP3]], align 1
+// ARM-LE-NEXT:    [[TMP4:%.*]] = getelementptr i8, ptr [[TMP0]], i32 12
+// ARM-LE-NEXT:    store i8 0, ptr [[TMP4]], align 4
+// ARM-LE-NEXT:    [[TMP5:%.*]] = getelementptr i8, ptr [[TMP0]], i32 13
+// ARM-LE-NEXT:    store i8 0, ptr [[TMP5]], align 1
+// ARM-LE-NEXT:    [[TMP6:%.*]] = getelementptr i8, ptr [[TMP0]], i32 14
+// ARM-LE-NEXT:    store i8 0, ptr [[TMP6]], align 2
+// ARM-LE-NEXT:    [[TMP7:%.*]] = getelementptr i8, ptr [[TMP0]], i32 15
+// ARM-LE-NEXT:    store i8 0, ptr [[TMP7]], align 1
+// ARM-LE-NEXT:    ret void
+//
+// ARM-BE-LABEL: define dso_local void @testOversizedBitfieldVolatileStruct(
+// ARM-BE-SAME: ptr noundef [[S:%.*]]) #[[ATTR0]] {
+// ARM-BE-NEXT:  [[ENTRY:.*:]]
+// ARM-BE-NEXT:    [[S_ADDR:%.*]] = alloca ptr, align 4
+// ARM-BE-NEXT:    store ptr [[S]], ptr [[S_ADDR]], align 4
+// ARM-BE-NEXT:    [[TMP0:%.*]] = load ptr, ptr [[S_ADDR]], align 4
+// ARM-BE-NEXT:    [[TMP1:%.*]] = getelementptr i8, ptr [[TMP0]], i32 0
+// ARM-BE-NEXT:    [[TMP2:%.*]] = load i8, ptr [[TMP1]], align 4
+// ARM-BE-NEXT:    [[TMP3:%.*]] = and i8 [[TMP2]], 63
+// ARM-BE-NEXT:    store i8 [[TMP3]], ptr [[TMP1]], align 4
+// ARM-BE-NEXT:    [[TMP4:%.*]] = getelementptr i8, ptr [[TMP0]], i32 1
+// ARM-BE-NEXT:    [[TMP5:%.*]] = load i8, ptr [[TMP4]], align 1
+// ARM-BE-NEXT:    [[TMP6:%.*]] = and i8 [[TMP5]], -64
+// ARM-BE-NEXT:    store i8 [[TMP6]], ptr [[TMP4]], align 1
+// ARM-BE-NEXT:    [[TMP7:%.*]] = getelementptr i8, ptr [[TMP0]], i32 2
+// ARM-BE-NEXT:    store i8 0, ptr [[TMP7]], align 2
+// ARM-BE-NEXT:    [[TMP8:%.*]] = getelementptr i8, ptr [[TMP0]], i32 3
+// ARM-BE-NEXT:    store i8 0, ptr [[TMP8]], align 1
+// ARM-BE-NEXT:    [[TMP9:%.*]] = getelementptr i8, ptr [[TMP0]], i32 8
+// ARM-BE-NEXT:    [[TMP10:%.*]] = load i8, ptr [[TMP9]], align 4
+// ARM-BE-NEXT:    [[TMP11:%.*]] = and i8 [[TMP10]], 31
+// ARM-BE-NEXT:    store i8 [[TMP11]], ptr [[TMP9]], align 4
+// ARM-BE-NEXT:    [[TMP12:%.*]] = getelementptr i8, ptr [[TMP0]], i32 12
+// ARM-BE-NEXT:    [[TMP13:%.*]] = load i8, ptr [[TMP12]], align 4
+// ARM-BE-NEXT:    [[TMP14:%.*]] = and i8 [[TMP13]], -32
+// ARM-BE-NEXT:    store i8 [[TMP14]], ptr [[TMP12]], align 4
+// ARM-BE-NEXT:    [[TMP15:%.*]] = getelementptr i8, ptr [[TMP0]], i32 13
+// ARM-BE-NEXT:    store i8 0, ptr [[TMP15]], align 1
+// ARM-BE-NEXT:    [[TMP16:%.*]] = getelementptr i8, ptr [[TMP0]], i32 14
+// ARM-BE-NEXT:    store i8 0, ptr [[TMP16]], align 2
+// ARM-BE-NEXT:    [[TMP17:%.*]] = getelementptr i8, ptr [[TMP0]], i32 15
+// ARM-BE-NEXT:    store i8 0, ptr [[TMP17]], align 1
+// ARM-BE-NEXT:    ret void
+//
+// AARCH64-LE-LABEL: define dso_local void @testOversizedBitfieldVolatileStruct(
+// AARCH64-LE-SAME: ptr noundef [[S:%.*]]) #[[ATTR0]] {
+// AARCH64-LE-NEXT:  [[ENTRY:.*:]]
+// AARCH64-LE-NEXT:    [[S_ADDR:%.*]] = alloca ptr, align 8
+// AARCH64-LE-NEXT:    store ptr [[S]], ptr [[S_ADDR]], align 8
+// AARCH64-LE-NEXT:    [[TMP0:%.*]] = load ptr, ptr [[S_ADDR]], align 8
+// AARCH64-LE-NEXT:    [[TMP1:%.*]] = getelementptr i8, ptr [[TMP0]], i32 1
+// AARCH64-LE-NEXT:    store i8 0, ptr [[TMP1]], align 1
+// AARCH64-LE-NEXT:    [[TMP2:%.*]] = getelementptr i8, ptr [[TMP0]], i32 2
+// AARCH64-LE-NEXT:    store i8 0, ptr [[TMP2]], align 2
+// AARCH64-LE-NEXT:    [[TMP3:%.*]] = getelementptr i8, ptr [[TMP0]], i32 3
+// AARCH64-LE-NEXT:    store i8 0, ptr [[TMP3]], align 1
+// AARCH64-LE-NEXT:    [[TMP4:%.*]] = getelementptr i8, ptr [[TMP0]], i32 12
+// AARCH64-LE-NEXT:    store i8 0, ptr [[TMP4]], align 4
+// AARCH64-LE-NEXT:    [[TMP5:%.*]] = getelementptr i8, ptr [[TMP0]], i32 13
+// AARCH64-LE-NEXT:    store i8 0, ptr [[TMP5]], align 1
+// AARCH64-LE-NEXT:    [[TMP6:%.*]] = getelementptr i8, ptr [[TMP0]], i32 14
+// AARCH64-LE-NEXT:    store i8 0, ptr [[TMP6]], align 2
+// AARCH64-LE-NEXT:    [[TMP7:%.*]] = getelementptr i8, ptr [[TMP0]], i32 15
+// AARCH64-LE-NEXT:    store i8 0, ptr [[TMP7]], align 1
+// AARCH64-LE-NEXT:    ret void
+//
+// AARCH64-BE-LABEL: define dso_local void @testOversizedBitfieldVolatileStruct(
+// AARCH64-BE-SAME: ptr noundef [[S:%.*]]) #[[ATTR0]] {
+// AARCH64-BE-NEXT:  [[ENTRY:.*:]]
+// AARCH64-BE-NEXT:    [[S_ADDR:%.*]] = alloca ptr, align 8
+// AARCH64-BE-NEXT:    store ptr [[S]], ptr [[S_ADDR]], align 8
+// AARCH64-BE-NEXT:    [[TMP0:%.*]] = load ptr, ptr [[S_ADDR]], align 8
+// AARCH64-BE-NEXT:    [[TMP1:%.*]] = getelementptr i8, ptr [[TMP0]], i32 0
+// AARCH64-BE-NEXT:    [[TMP2:%.*]] = load i8, ptr [[TMP1]], align 4
+// AARCH64-BE-NEXT:    [[TMP3:%.*]] = and i8 [[TMP2]], 63
+// AARCH64-BE-NEXT:    store i8 [[TMP3]], ptr [[TMP1]], align 4
+// AARCH64-BE-NEXT:    [[TMP4:%.*]] = getelementptr i8, ptr [[TMP0]], i32 1
+// AARCH64-BE-NEXT:    [[TMP5:%.*]] = load i8, ptr [[TMP4]], align 1
+// AARCH64-BE-NEXT:    [[TMP6:%.*]] = and i8 [[TMP5]], -64
+// AARCH64-BE-NEXT:    store i8 [[TMP6]], ptr [[TMP4]], align 1
+// AARCH64-BE-NEXT:    [[TMP7:%.*]] = getelementptr i8, ptr [[TMP0]], i32 2
+// AARCH64-BE-NEXT:    store i8 0, ptr [[TMP7]], align 2
+// AARCH64-BE-NEXT:    [[TMP8:%.*]] = getelementptr i8, ptr [[TMP0]], i32 3
+// AARCH64-BE-NEXT:    store i8 0, ptr [[TMP8]], align 1
+// AARCH64-BE-NEXT:    [[TMP9:%.*]] = getelementptr i8, ptr [[TMP0]], i32 8
+// AARCH64-BE-NEXT:    [[TMP10:%.*]] = load i8, ptr [[TMP9]], align 4
+// AARCH64-BE-NEXT:    [[TMP11:%.*]] = and i8 [[TMP10]], 31
+// AARCH64-BE-NEXT:    store i8 [[TMP11]], ptr [[TMP9]], align 4
+// AARCH64-BE-NEXT:    [[TMP12:%.*]] = getelementptr i8, ptr [[TMP0]], i32 12
+// AARCH64-BE-NEXT:    [[TMP13:%.*]] = load i8, ptr [[TMP12]], align 4
+// AARCH64-BE-NEXT:    [[TMP14:%.*]] = and i8 [[TMP13]], -32
+// AARCH64-BE-NEXT:    store i8 [[TMP14]], ptr [[TMP12]], align 4
+// AARCH64-BE-NEXT:    [[TMP15:%.*]] = getelementptr i8, ptr [[TMP0]], i32 13
+// AARCH64-BE-NEXT:    store i8 0, ptr [[TMP15]], align 1
+// AARCH64-BE-NEXT:    [[TMP16:%.*]] = getelementptr i8, ptr [[TMP0]], i32 14
+// AARCH64-BE-NEXT:    store i8 0, ptr [[TMP16]], align 2
+// AARCH64-BE-NEXT:    [[TMP17:%.*]] = getelementptr i8, ptr [[TMP0]], i32 15
+// AARCH64-BE-NEXT:    store i8 0, ptr [[TMP17]], align 1
+// AARCH64-BE-NEXT:    ret void
+//
+extern "C" void testOversizedBitfieldVolatileStruct(
+    volatile struct OversizedBitfield *s) {
+  __builtin_clear_padding(s);
+}
+
+struct VolatileOversizedBitfield {
+  volatile unsigned char a : 10;
+  int b;
+  volatile int c : 35;
+};
+
+// LINUX-LABEL: define dso_local void @testVolatileOversizedBitfield(
+// LINUX-SAME: ptr noundef [[S:%.*]]) #[[ATTR0]] {
+// LINUX-NEXT:  [[ENTRY:.*:]]
+// LINUX-NEXT:    [[S_ADDR:%.*]] = alloca ptr, align 8
+// LINUX-NEXT:    store ptr [[S]], ptr [[S_ADDR]], align 8
+// LINUX-NEXT:    [[TMP0:%.*]] = load ptr, ptr [[S_ADDR]], align 8
+// LINUX-NEXT:    [[TMP1:%.*]] = getelementptr i8, ptr [[TMP0]], i32 1
+// LINUX-NEXT:    store i8 0, ptr [[TMP1]], align 1
+// LINUX-NEXT:    [[TMP2:%.*]] = getelementptr i8, ptr [[TMP0]], i32 2
+// LINUX-NEXT:    store i8 0, ptr [[TMP2]], align 2
+// LINUX-NEXT:    [[TMP3:%.*]] = getelementptr i8, ptr [[TMP0]], i32 3
+// LINUX-NEXT:    store i8 0, ptr [[TMP3]], align 1
+// LINUX-NEXT:    [[TMP4:%.*]] = getelementptr i8, ptr [[TMP0]], i32 12
+// LINUX-NEXT:    store i8 0, ptr [[TMP4]], align 4
+// LINUX-NEXT:    [[TMP5:%.*]] = getelementptr i8, ptr [[TMP0]], i32 13
+// LINUX-NEXT:    store i8 0, ptr [[TMP5]], align 1
+// LINUX-NEXT:    [[TMP6:%.*]] = getelementptr i8, ptr [[TMP0]], i32 14
+// LINUX-NEXT:    store i8 0, ptr [[TMP6]], align 2
+// LINUX-NEXT:    [[TMP7:%.*]] = getelementptr i8, ptr [[TMP0]], i32 15
+// LINUX-NEXT:    store i8 0, ptr [[TMP7]], align 1
+// LINUX-NEXT:    ret void
+//
+// ARM-LE-LABEL: define dso_local void @testVolatileOversizedBitfield(
+// ARM-LE-SAME: ptr noundef [[S:%.*]]) #[[ATTR0]] {
+// ARM-LE-NEXT:  [[ENTRY:.*:]]
+// ARM-LE-NEXT:    [[S_ADDR:%.*]] = alloca ptr, align 4
+// ARM-LE-NEXT:    store ptr [[S]], ptr [[S_ADDR]], align 4
+// ARM-LE-NEXT:    [[TMP0:%.*]] = load ptr, ptr [[S_ADDR]], align 4
+// ARM-LE-NEXT:    [[TMP1:%.*]] = getelementptr i8, ptr [[TMP0]], i32 1
+// ARM-LE-NEXT:    store i8 0, ptr [[TMP1]], align 1
+// ARM-LE-NEXT:    [[TMP2:%.*]] = getelementptr i8, ptr [[TMP0]], i32 2
+// ARM-LE-NEXT:    store i8 0, ptr [[TMP2]], align 2
+// ARM-LE-NEXT:    [[TMP3:%.*]] = getelementptr i8, ptr [[TMP0]], i32 3
+// ARM-LE-NEXT:    store i8 0, ptr [[TMP3]], align 1
+// ARM-LE-NEXT:    [[TMP4:%.*]] = getelementptr i8, ptr [[TMP0]], i32 12
+// ARM-LE-NEXT:    store i8 0, ptr [[TMP4]], align 4
+// ARM-LE-NEXT:    [[TMP5:%.*]] = getelementptr i8, ptr [[TMP0]], i32 13
+// ARM-LE-NEXT:    store i8 0, ptr [[TMP5]], align 1
+// ARM-LE-NEXT:    [[TMP6:%.*]] = getelementptr i8, ptr [[TMP0]], i32 14
+// ARM-LE-NEXT:    store i8 0, ptr [[TMP6]], align 2
+// ARM-LE-NEXT:    [[TMP7:%.*]] = getelementptr i8, ptr [[TMP0]], i32 15
+// ARM-LE-NEXT:    store i8 0, ptr [[TMP7]], align 1
+// ARM-LE-NEXT:    ret void
+//
+// ARM-BE-LABEL: define dso_local void @testVolatileOversizedBitfield(
+// ARM-BE-SAME: ptr noundef [[S:%.*]]) #[[ATTR0]] {
+// ARM-BE-NEXT:  [[ENTRY:.*:]]
+// ARM-BE-NEXT:    [[S_ADDR:%.*]] = alloca ptr, align 4
+// ARM-BE-NEXT:    store ptr [[S]], ptr [[S_ADDR]], align 4
+// ARM-BE-NEXT:    [[TMP0:%.*]] = load ptr, ptr [[S_ADDR]], align 4
+// ARM-BE-NEXT:    [[TMP1:%.*]] = getelementptr i8, ptr [[TMP0]], i32 0
+// ARM-BE-NEXT:    [[TMP2:%.*]] = load i8, ptr [[TMP1]], align 4
+// ARM-BE-NEXT:    [[TMP3:%.*]] = and i8 [[TMP2]], 63
+// ARM-BE-NEXT:    store i8 [[TMP3]], ptr [[TMP1]], align 4
+// ARM-BE-NEXT:    [[TMP4:%.*]] = getelementptr i8, ptr [[TMP0]], i32 1
+// ARM-BE-NEXT:    [[TMP5:%.*]] = load i8, ptr [[TMP4]], align 1
+// ARM-BE-NEXT:    [[TMP6:%.*]] = and i8 [[TMP5]], -64
+// ARM-BE-NEXT:    store i8 [[TMP6]], ptr [[TMP4]], align 1
+// ARM-BE-NEXT:    [[TMP7:%.*]] = getelementptr i8, ptr [[TMP0]], i32 2
+// ARM-BE-NEXT:    store i8 0, ptr [[TMP7]], align 2
+// ARM-BE-NEXT:    [[TMP8:%.*]] = getelementptr i8, ptr [[TMP0]], i32 3
+// ARM-BE-NEXT:    store i8 0, ptr [[TMP8]], align 1
+// ARM-BE-NEXT:    [[TMP9:%.*]] = getelementptr i8, ptr [[TMP0]], i32 8
+// ARM-BE-NEXT:    [[TMP10:%.*]] = load i8, ptr [[TMP9]], align 4
+// ARM-BE-NEXT:    [[TMP11:%.*]] = and i8 [[TMP10]], 31
+// ARM-BE-NEXT:    store i8 [[TMP11]], ptr [[TMP9]], align 4
+// ARM-BE-NEXT:    [[TMP12:%.*]] = getelementptr i8, ptr [[TMP0]], i32 12
+// ARM-BE-NEXT:    [[TMP13:%.*]] = load i8, ptr [[TMP12]], align 4
+// ARM-BE-NEXT:    [[TMP14:%.*]] = and i8 [[TMP13]], -32
+// ARM-BE-NEXT:    store i8 [[TMP14]], ptr [[TMP12]], align 4
+// ARM-BE-NEXT:    [[TMP15:%.*]] = getelementptr i8, ptr [[TMP0]], i32 13
+// ARM-BE-NEXT:    store i8 0, ptr [[TMP15]], align 1
+// ARM-BE-NEXT:    [[TMP16:%.*]] = getelementptr i8, ptr [[TMP0]], i32 14
+// ARM-BE-NEXT:    store i8 0, ptr [[TMP16]], align 2
+// ARM-BE-NEXT:    [[TMP17:%.*]] = getelementptr i8, ptr [[TMP0]], i32 15
+// ARM-BE-NEXT:    store i8 0, ptr [[TMP17]], align 1
+// ARM-BE-NEXT:    ret void
+//
+// AARCH64-LE-LABEL: define dso_local void @testVolatileOversizedBitfield(
+// AARCH64-LE-SAME: ptr noundef [[S:%.*]]) #[[ATTR0]] {
+// AARCH64-LE-NEXT:  [[ENTRY:.*:]]
+// AARCH64-LE-NEXT:    [[S_ADDR:%.*]] = alloca ptr, align 8
+// AARCH64-LE-NEXT:    store ptr [[S]], ptr [[S_ADDR]], align 8
+// AARCH64-LE-NEXT:    [[TMP0:%.*]] = load ptr, ptr [[S_ADDR]], align 8
+// AARCH64-LE-NEXT:    [[TMP1:%.*]] = getelementptr i8, ptr [[TMP0]], i32 1
+// AARCH64-LE-NEXT:    store i8 0, ptr [[TMP1]], align 1
+// AARCH64-LE-NEXT:    [[TMP2:%.*]] = getelementptr i8, ptr [[TMP0]], i32 2
+// AARCH64-LE-NEXT:    store i8 0, ptr [[TMP2]], align 2
+// AARCH64-LE-NEXT:    [[TMP3:%.*]] = getelementptr i8, ptr [[TMP0]], i32 3
+// AARCH64-LE-NEXT:    store i8 0, ptr [[TMP3]], align 1
+// AARCH64-LE-NEXT:    [[TMP4:%.*]] = getelementptr i8, ptr [[TMP0]], i32 12
+// AARCH64-LE-NEXT:    store i8 0, ptr [[TMP4]], align 4
+// AARCH64-LE-NEXT:    [[TMP5:%.*]] = getelementptr i8, ptr [[TMP0]], i32 13
+// AARCH64-LE-NEXT:    store i8 0, ptr [[TMP5]], align 1
+// AARCH64-LE-NEXT:    [[TMP6:%.*]] = getelementptr i8, ptr [[TMP0]], i32 14
+// AARCH64-LE-NEXT:    store i8 0, ptr [[TMP6]], align 2
+// AARCH64-LE-NEXT:    [[TMP7:%.*]] = getelementptr i8, ptr [[TMP0]], i32 15
+// AARCH64-LE-NEXT:    store i8 0, ptr [[TMP7]], align 1
+// AARCH64-LE-NEXT:    ret void
+//
+// AARCH64-BE-LABEL: define dso_local void @testVolatileOversizedBitfield(
+// AARCH64-BE-SAME: ptr noundef [[S:%.*]]) #[[ATTR0]] {
+// AARCH64-BE-NEXT:  [[ENTRY:.*:]]
+// AARCH64-BE-NEXT:    [[S_ADDR:%.*]] = alloca ptr, align 8
+// AARCH64-BE-NEXT:    store ptr [[S]], ptr [[S_ADDR]], align 8
+// AARCH64-BE-NEXT:    [[TMP0:%.*]] = load ptr, ptr [[S_ADDR]], align 8
+// AARCH64-BE-NEXT:    [[TMP1:%.*]] = getelementptr i8, ptr [[TMP0]], i32 0
+// AARCH64-BE-NEXT:    [[TMP2:%.*]] = load i8, ptr [[TMP1]], align 4
+// AARCH64-BE-NEXT:    [[TMP3:%.*]] = and i8 [[TMP2]], 63
+// AARCH64-BE-NEXT:    store i8 [[TMP3]], ptr [[TMP1]], align 4
+// AARCH64-BE-NEXT:    [[TMP4:%.*]] = getelementptr i8, ptr [[TMP0]], i32 1
+// AARCH64-BE-NEXT:    [[TMP5:%.*]] = load i8, ptr [[TMP4]], align 1
+// AARCH64-BE-NEXT:    [[TMP6:%.*]] = and i8 [[TMP5]], -64
+// AARCH64-BE-NEXT:    store i8 [[TMP6]], ptr [[TMP4]], align 1
+// AARCH64-BE-NEXT:    [[TMP7:%.*]] = getelementptr i8, ptr [[TMP0]], i32 2
+// AARCH64-BE-NEXT:    store i8 0, ptr [[TMP7]], align 2
+// AARCH64-BE-NEXT:    [[TMP8:%.*]] = getelementptr i8, ptr [[TMP0]], i32 3
+// AARCH64-BE-NEXT:    store i8 0, ptr [[TMP8]], align 1
+// AARCH64-BE-NEXT:    [[TMP9:%.*]] = getelementptr i8, ptr [[TMP0]], i32 8
+// AARCH64-BE-NEXT:    [[TMP10:%.*]] = load i8, ptr [[TMP9]], align 4
+// AARCH64-BE-NEXT:    [[TMP11:%.*]] = and i8 [[TMP10]], 31
+// AARCH64-BE-NEXT:    store i8 [[TMP11]], ptr [[TMP9]], align 4
+// AARCH64-BE-NEXT:    [[TMP12:%.*]] = getelementptr i8, ptr [[TMP0]], i32 12
+// AARCH64-BE-NEXT:    [[TMP13:%.*]] = load i8, ptr [[TMP12]], align 4
+// AARCH64-BE-NEXT:    [[TMP14:%.*]] = and i8 [[TMP13]], -32
+// AARCH64-BE-NEXT:    store i8 [[TMP14]], ptr [[TMP12]], align 4
+// AARCH64-BE-NEXT:    [[TMP15:%.*]] = getelementptr i8, ptr [[TMP0]], i32 13
+// AARCH64-BE-NEXT:    store i8 0, ptr [[TMP15]], align 1
+// AARCH64-BE-NEXT:    [[TMP16:%.*]] = getelementptr i8, ptr [[TMP0]], i32 14
+// AARCH64-BE-NEXT:    store i8 0, ptr [[TMP16]], align 2
+// AARCH64-BE-NEXT:    [[TMP17:%.*]] = getelementptr i8, ptr [[TMP0]], i32 15
+// AARCH64-BE-NEXT:    store i8 0, ptr [[TMP17]], align 1
+// AARCH64-BE-NEXT:    ret void
+//
+extern "C" void testVolatileOversizedBitfield(
+    struct VolatileOversizedBitfield *s) {
+  __builtin_clear_padding(s);
+}
+
+// LINUX-LABEL: define dso_local void @testVolatileOversizedBitfieldVolatileStruct(
+// LINUX-SAME: ptr noundef [[S:%.*]]) #[[ATTR0]] {
+// LINUX-NEXT:  [[ENTRY:.*:]]
+// LINUX-NEXT:    [[S_ADDR:%.*]] = alloca ptr, align 8
+// LINUX-NEXT:    store ptr [[S]], ptr [[S_ADDR]], align 8
+// LINUX-NEXT:    [[TMP0:%.*]] = load ptr, ptr [[S_ADDR]], align 8
+// LINUX-NEXT:    [[TMP1:%.*]] = getelementptr i8, ptr [[TMP0]], i32 1
+// LINUX-NEXT:    store i8 0, ptr [[TMP1]], align 1
+// LINUX-NEXT:    [[TMP2:%.*]] = getelementptr i8, ptr [[TMP0]], i32 2
+// LINUX-NEXT:    store i8 0, ptr [[TMP2]], align 2
+// LINUX-NEXT:    [[TMP3:%.*]] = getelementptr i8, ptr [[TMP0]], i32 3
+// LINUX-NEXT:    store i8 0, ptr [[TMP3]], align 1
+// LINUX-NEXT:    [[TMP4:%.*]] = getelementptr i8, ptr [[TMP0]], i32 12
+// LINUX-NEXT:    store i8 0, ptr [[TMP4]], align 4
+// LINUX-NEXT:    [[TMP5:%.*]] = getelementptr i8, ptr [[TMP0]], i32 13
+// LINUX-NEXT:    store i8 0, ptr [[TMP5]], align 1
+// LINUX-NEXT:    [[TMP6:%.*]] = getelementptr i8, ptr [[TMP0]], i32 14
+// LINUX-NEXT:    store i8 0, ptr [[TMP6]], align 2
+// LINUX-NEXT:    [[TMP7:%.*]] = getelementptr i8, ptr [[TMP0]], i32 15
+// LINUX-NEXT:    store i8 0, ptr [[TMP7]], align 1
+// LINUX-NEXT:    ret void
+//
+// ARM-LE-LABEL: define dso_local void @testVolatileOversizedBitfieldVolatileStruct(
+// ARM-LE-SAME: ptr noundef [[S:%.*]]) #[[ATTR0]] {
+// ARM-LE-NEXT:  [[ENTRY:.*:]]
+// ARM-LE-NEXT:    [[S_ADDR:%.*]] = alloca ptr, align 4
+// ARM-LE-NEXT:    store ptr [[S]], ptr [[S_ADDR]], align 4
+// ARM-LE-NEXT:    [[TMP0:%.*]] = load ptr, ptr [[S_ADDR]], align 4
+// ARM-LE-NEXT:    [[TMP1:%.*]] = getelementptr i8, ptr [[TMP0]], i32 1
+// ARM-LE-NEXT:    store i8 0, ptr [[TMP1]], align 1
+// ARM-LE-NEXT:    [[TMP2:%.*]] = getelementptr i8, ptr [[TMP0]], i32 2
+// ARM-LE-NEXT:    store i8 0, ptr [[TMP2]], align 2
+// ARM-LE-NEXT:    [[TMP3:%.*]] = getelementptr i8, ptr [[TMP0]], i32 3
+// ARM-LE-NEXT:    store i8 0, ptr [[TMP3]], align 1
+// ARM-LE-NEXT:    [[TMP4:%.*]] = getelementptr i8, ptr [[TMP0]], i32 12
+// ARM-LE-NEXT:    store i8 0, ptr [[TMP4]], align 4
+// ARM-LE-NEXT:    [[TMP5:%.*]] = getelementptr i8, ptr [[TMP0]], i32 13
+// ARM-LE-NEXT:    store i8 0, ptr [[TMP5]], align 1
+// ARM-LE-NEXT:    [[TMP6:%.*]] = getelementptr i8, ptr [[TMP0]], i32 14
+// ARM-LE-NEXT:    store i8 0, ptr [[TMP6]], align 2
+// ARM-LE-NEXT:    [[TMP7:%.*]] = getelementptr i8, ptr [[TMP0]], i32 15
+// ARM-LE-NEXT:    store i8 0, ptr [[TMP7]], align 1
+// ARM-LE-NEXT:    ret void
+//
+// ARM-BE-LABEL: define dso_local void @testVolatileOversizedBitfieldVolatileStruct(
+// ARM-BE-SAME: ptr noundef [[S:%.*]]) #[[ATTR0]] {
+// ARM-BE-NEXT:  [[ENTRY:.*:]]
+// ARM-BE-NEXT:    [[S_ADDR:%.*]] = alloca ptr, align 4
+// ARM-BE-NEXT:    store ptr [[S]], ptr [[S_ADDR]], align 4
+// ARM-BE-NEXT:    [[TMP0:%.*]] = load ptr, ptr [[S_ADDR]], align 4
+// ARM-BE-NEXT:    [[TMP1:%.*]] = getelementptr i8, ptr [[TMP0]], i32 0
+// ARM-BE-NEXT:    [[TMP2:%.*]] = load i8, ptr [[TMP1]], align 4
+// ARM-BE-NEXT:    [[TMP3:%.*]] = and i8 [[TMP2]], 63
+// ARM-BE-NEXT:    store i8 [[TMP3]], ptr [[TMP1]], align 4
+// ARM-BE-NEXT:    [[TMP4:%.*]] = getelementptr i8, ptr [[TMP0]], i32 1
+// ARM-BE-NEXT:    [[TMP5:%.*]] = load i8, ptr [[TMP4]], align 1
+// ARM-BE-NEXT:    [[TMP6:%.*]] = and i8 [[TMP5]], -64
+// ARM-BE-NEXT:    store i8 [[TMP6]], ptr [[TMP4]], align 1
+// ARM-BE-NEXT:    [[TMP7:%.*]] = getelementptr i8, ptr [[TMP0]], i32 2
+// ARM-BE-NEXT:    store i8 0, ptr [[TMP7]], align 2
+// ARM-BE-NEXT:    [[TMP8:%.*]] = getelementptr i8, ptr [[TMP0]], i32 3
+// ARM-BE-NEXT:    store i8 0, ptr [[TMP8]], align 1
+// ARM-BE-NEXT:    [[TMP9:%.*]] = getelementptr i8, ptr [[TMP0]], i32 8
+// ARM-BE-NEXT:    [[TMP10:%.*]] = load i8, ptr [[TMP9]], align 4
+// ARM-BE-NEXT:    [[TMP11:%.*]] = and i8 [[TMP10]], 31
+// ARM-BE-NEXT:    store i8 [[TMP11]], ptr [[TMP9]], align 4
+// ARM-BE-NEXT:    [[TMP12:%.*]] = getelementptr i8, ptr [[TMP0]], i32 12
+// ARM-BE-NEXT:    [[TMP13:%.*]] = load i8, ptr [[TMP12]], align 4
+// ARM-BE-NEXT:    [[TMP14:%.*]] = and i8 [[TMP13]], -32
+// ARM-BE-NEXT:    store i8 [[TMP14]], ptr [[TMP12]], align 4
+// ARM-BE-NEXT:    [[TMP15:%.*]] = getelementptr i8, ptr [[TMP0]], i32 13
+// ARM-BE-NEXT:    store i8 0, ptr [[TMP15]], align 1
+// ARM-BE-NEXT:    [[TMP16:%.*]] = getelementptr i8, ptr [[TMP0]], i32 14
+// ARM-BE-NEXT:    store i8 0, ptr [[TMP16]], align 2
+// ARM-BE-NEXT:    [[TMP17:%.*]] = getelementptr i8, ptr [[TMP0]], i32 15
+// ARM-BE-NEXT:    store i8 0, ptr [[TMP17]], align 1
+// ARM-BE-NEXT:    ret void
+//
+// AARCH64-LE-LABEL: define dso_local void @testVolatileOversizedBitfieldVolatileStruct(
+// AARCH64-LE-SAME: ptr noundef [[S:%.*]]) #[[ATTR0]] {
+// AARCH64-LE-NEXT:  [[ENTRY:.*:]]
+// AARCH64-LE-NEXT:    [[S_ADDR:%.*]] = alloca ptr, align 8
+// AARCH64-LE-NEXT:    store ptr [[S]], ptr [[S_ADDR]], align 8
+// AARCH64-LE-NEXT:    [[TMP0:%.*]] = load ptr, ptr [[S_ADDR]], align 8
+// AARCH64-LE-NEXT:    [[TMP1:%.*]] = getelementptr i8, ptr [[TMP0]], i32 1
+// AARCH64-LE-NEXT:    store i8 0, ptr [[TMP1]], align 1
+// AARCH64-LE-NEXT:    [[TMP2:%.*]] = getelementptr i8, ptr [[TMP0]], i32 2
+// AARCH64-LE-NEXT:    store i8 0, ptr [[TMP2]], align 2
+// AARCH64-LE-NEXT:    [[TMP3:%.*]] = getelementptr i8, ptr [[TMP0]], i32 3
+// AARCH64-LE-NEXT:    store i8 0, ptr [[TMP3]], align 1
+// AARCH64-LE-NEXT:    [[TMP4:%.*]] = getelementptr i8, ptr [[TMP0]], i32 12
+// AARCH64-LE-NEXT:    store i8 0, ptr [[TMP4]], align 4
+// AARCH64-LE-NEXT:    [[TMP5:%.*]] = getelementptr i8, ptr [[TMP0]], i32 13
+// AARCH64-LE-NEXT:    store i8 0, ptr [[TMP5]], align 1
+// AARCH64-LE-NEXT:    [[TMP6:%.*]] = getelementptr i8, ptr [[TMP0]], i32 14
+// AARCH64-LE-NEXT:    store i8 0, ptr [[TMP6]], align 2
+// AARCH64-LE-NEXT:    [[TMP7:%.*]] = getelementptr i8, ptr [[TMP0]], i32 15
+// AARCH64-LE-NEXT:    store i8 0, ptr [[TMP7]], align 1
+// AARCH64-LE-NEXT:    ret void
+//
+// AARCH64-BE-LABEL: define dso_local void @testVolatileOversizedBitfieldVolatileStruct(
+// AARCH64-BE-SAME: ptr noundef [[S:%.*]]) #[[ATTR0]] {
+// AARCH64-BE-NEXT:  [[ENTRY:.*:]]
+// AARCH64-BE-NEXT:    [[S_ADDR:%.*]] = alloca ptr, align 8
+// AARCH64-BE-NEXT:    store ptr [[S]], ptr [[S_ADDR]], align 8
+// AARCH64-BE-NEXT:    [[TMP0:%.*]] = load ptr, ptr [[S_ADDR]], align 8
+// AARCH64-BE-NEXT:    [[TMP1:%.*]] = getelementptr i8, ptr [[TMP0]], i32 0
+// AARCH64-BE-NEXT:    [[TMP2:%.*]] = load i8, ptr [[TMP1]], align 4
+// AARCH64-BE-NEXT:    [[TMP3:%.*]] = and i8 [[TMP2]], 63
+// AARCH64-BE-NEXT:    store i8 [[TMP3]], ptr [[TMP1]], align 4
+// AARCH64-BE-NEXT:    [[TMP4:%.*]] = getelementptr i8, ptr [[TMP0]], i32 1
+// AARCH64-BE-NEXT:    [[TMP5:%.*]] = load i8, ptr [[TMP4]], align 1
+// AARCH64-BE-NEXT:    [[TMP6:%.*]] = and i8 [[TMP5]], -64
+// AARCH64-BE-NEXT:    store i8 [[TMP6]], ptr [[TMP4]], align 1
+// AARCH64-BE-NEXT:    [[TMP7:%.*]] = getelementptr i8, ptr [[TMP0]], i32 2
+// AARCH64-BE-NEXT:    store i8 0, ptr [[TMP7]], align 2
+// AARCH64-BE-NEXT:    [[TMP8:%.*]] = getelementptr i8, ptr [[TMP0]], i32 3
+// AARCH64-BE-NEXT:    store i8 0, ptr [[TMP8]], align 1
+// AARCH64-BE-NEXT:    [[TMP9:%.*]] = getelementptr i8, ptr [[TMP0]], i32 8
+// AARCH64-BE-NEXT:    [[TMP10:%.*]] = load i8, ptr [[TMP9]], align 4
+// AARCH64-BE-NEXT:    [[TMP11:%.*]] = and i8 [[TMP10]], 31
+// AARCH64-BE-NEXT:    store i8 [[TMP11]], ptr [[TMP9]], align 4
+// AARCH64-BE-NEXT:    [[TMP12:%.*]] = getelementptr i8, ptr [[TMP0]], i32 12
+// AARCH64-BE-NEXT:    [[TMP13:%.*]] = load i8, ptr [[TMP12]], align 4
+// AARCH64-BE-NEXT:    [[TMP14:%.*]] = and i8 [[TMP13]], -32
+// AARCH64-BE-NEXT:    store i8 [[TMP14]], ptr [[TMP12]], align 4
+// AARCH64-BE-NEXT:    [[TMP15:%.*]] = getelementptr i8, ptr [[TMP0]], i32 13
+// AARCH64-BE-NEXT:    store i8 0, ptr [[TMP15]], align 1
+// AARCH64-BE-NEXT:    [[TMP16:%.*]] = getelementptr i8, ptr [[TMP0]], i32 14
+// AARCH64-BE-NEXT:    store i8 0, ptr [[TMP16]], align 2
+// AARCH64-BE-NEXT:    [[TMP17:%.*]] = getelementptr i8, ptr [[TMP0]], i32 15
+// AARCH64-BE-NEXT:    store i8 0, ptr [[TMP17]], align 1
+// AARCH64-BE-NEXT:    ret void
+//
+extern "C" void testVolatileOversizedBitfieldVolatileStruct(
+    volatile struct VolatileOversizedBitfield *s) {
+  __builtin_clear_padding(s);
+}
+
+#pragma clang diagnostic pop
+
+#endif // _WIN32
+
+
+// LINUX-LABEL: define dso_local void @testArrayNoPadding(
+// LINUX-SAME: ) #[[ATTR0]] {
+// LINUX-NEXT:  [[ENTRY:.*:]]
+// LINUX-NEXT:    [[I:%.*]] = alloca [4 x i32], align 16
+// LINUX-NEXT:    ret void
+//
+// WINDOWS-LABEL: define dso_local void @testArrayNoPadding(
+// WINDOWS-SAME: ) #[[ATTR0]] {
+// WINDOWS-NEXT:  [[ENTRY:.*:]]
+// WINDOWS-NEXT:    [[I:%.*]] = alloca [4 x i32], align 4
+// WINDOWS-NEXT:    ret void
+//
+// ARM-LE-LABEL: define dso_local void @testArrayNoPadding(
+// ARM-LE-SAME: ) #[[ATTR0]] {
+// ARM-LE-NEXT:  [[ENTRY:.*:]]
+// ARM-LE-NEXT:    [[I:%.*]] = alloca [4 x i32], align 4
+// ARM-LE-NEXT:    ret void
+//
+// ARM-BE-LABEL: define dso_local void @testArrayNoPadding(
+// ARM-BE-SAME: ) #[[ATTR0]] {
+// ARM-BE-NEXT:  [[ENTRY:.*:]]
+// ARM-BE-NEXT:    [[I:%.*]] = alloca [4 x i32], align 4
+// ARM-BE-NEXT:    ret void
+//
+// AARCH64-LE-LABEL: define dso_local void @testArrayNoPadding(
+// AARCH64-LE-SAME: ) #[[ATTR0]] {
+// AARCH64-LE-NEXT:  [[ENTRY:.*:]]
+// AARCH64-LE-NEXT:    [[I:%.*]] = alloca [4 x i32], align 4
+// AARCH64-LE-NEXT:    ret void
+//
+// AARCH64-BE-LABEL: define dso_local void @testArrayNoPadding(
+// AARCH64-BE-SAME: ) #[[ATTR0]] {
+// AARCH64-BE-NEXT:  [[ENTRY:.*:]]
+// AARCH64-BE-NEXT:    [[I:%.*]] = alloca [4 x i32], align 4
+// AARCH64-BE-NEXT:    ret void
+//
+extern "C" void testArrayNoPadding() {
+  int i[4];
+  // there is no padding in the array.
+  __builtin_clear_padding(&i);
+}
+
+
+// LINUX-LABEL: define dso_local void @testArrayLongDouble(
+// LINUX-SAME: ) #[[ATTR0]] {
+// LINUX-NEXT:  [[ENTRY:.*:]]
+// LINUX-NEXT:    [[LD:%.*]] = alloca [2 x x86_fp80], align 16
+// LINUX-NEXT:    [[TMP0:%.*]] = getelementptr i8, ptr [[LD]], i32 10
+// LINUX-NEXT:    store i8 0, ptr [[TMP0]], align 2
+// LINUX-NEXT:    [[TMP1:%.*]] = getelementptr i8, ptr [[LD]], i32 11
+// LINUX-NEXT:    store i8 0, ptr [[TMP1]], align 1
+// LINUX-NEXT:    [[TMP2:%.*]] = getelementptr i8, ptr [[LD]], i32 12
+// LINUX-NEXT:    store i8 0, ptr [[TMP2]], align 4
+// LINUX-NEXT:    [[TMP3:%.*]] = getelementptr i8, ptr [[LD]], i32 13
+// LINUX-NEXT:    store i8 0, ptr [[TMP3]], align 1
+// LINUX-NEXT:    [[TMP4:%.*]] = getelementptr i8, ptr [[LD]], i32 14
+// LINUX-NEXT:    store i8 0, ptr [[TMP4]], align 2
+// LINUX-NEXT:    [[TMP5:%.*]] = getelementptr i8, ptr [[LD]], i32 15
+// LINUX-NEXT:    store i8 0, ptr [[TMP5]], align 1
+// LINUX-NEXT:    [[TMP6:%.*]] = getelementptr i8, ptr [[LD]], i32 26
+// LINUX-NEXT:    store i8 0, ptr [[TMP6]], align 2
+// LINUX-NEXT:    [[TMP7:%.*]] = getelementptr i8, ptr [[LD]], i32 27
+// LINUX-NEXT:    store i8 0, ptr [[TMP7]], align 1
+// LINUX-NEXT:    [[TMP8:%.*]] = getelementptr i8, ptr [[LD]], i32 28
+// LINUX-NEXT:    store i8 0, ptr [[TMP8]], align 4
+// LINUX-NEXT:    [[TMP9:%.*]] = getelementptr i8, ptr [[LD]], i32 29
+// LINUX-NEXT:    store i8 0, ptr [[TMP9]], align 1
+// LINUX-NEXT:    [[TMP10:%.*]] = getelementptr i8, ptr [[LD]], i32 30
+// LINUX-NEXT:    store i8 0, ptr [[TMP10]], align 2
+// LINUX-NEXT:    [[TMP11:%.*]] = getelementptr i8, ptr [[LD]], i32 31
+// LINUX-NEXT:    store i8 0, ptr [[TMP11]], align 1
+// LINUX-NEXT:    ret void
+//
+// WINDOWS-LABEL: define dso_local void @testArrayLongDouble(
+// WINDOWS-SAME: ) #[[ATTR0]] {
+// WINDOWS-NEXT:  [[ENTRY:.*:]]
+// WINDOWS-NEXT:    [[LD:%.*]] = alloca [2 x double], align 8
+// WINDOWS-NEXT:    ret void
+//
+// ARM-LE-LABEL: define dso_local void @testArrayLongDouble(
+// ARM-LE-SAME: ) #[[ATTR0]] {
+// ARM-LE-NEXT:  [[ENTRY:.*:]]
+// ARM-LE-NEXT:    [[LD:%.*]] = alloca [2 x double], align 8
+// ARM-LE-NEXT:    ret void
+//
+// ARM-BE-LABEL: define dso_local void @testArrayLongDouble(
+// ARM-BE-SAME: ) #[[ATTR0]] {
+// ARM-BE-NEXT:  [[ENTRY:.*:]]
+// ARM-BE-NEXT:    [[LD:%.*]] = alloca [2 x double], align 8
+// ARM-BE-NEXT:    ret void
+//
+// AARCH64-LE-LABEL: define dso_local void @testArrayLongDouble(
+// AARCH64-LE-SAME: ) #[[ATTR0]] {
+// AARCH64-LE-NEXT:  [[ENTRY:.*:]]
+// AARCH64-LE-NEXT:    [[LD:%.*]] = alloca [2 x fp128], align 16
+// AARCH64-LE-NEXT:    ret void
+//
+// AARCH64-BE-LABEL: define dso_local void @testArrayLongDouble(
+// AARCH64-BE-SAME: ) #[[ATTR0]] {
+// AARCH64-BE-NEXT:  [[ENTRY:.*:]]
+// AARCH64-BE-NEXT:    [[LD:%.*]] = alloca [2 x fp128], align 16
+// AARCH64-BE-NEXT:    ret void
+//
+extern "C" void testArrayLongDouble() {
+  // long double 0, [0-9] PAD [10-15]
+  // long double 1, [16-25] PAD [26-31]
+  long double ld[2];
+  __builtin_clear_padding(&ld);
+}
+
+
+// LINUX-LABEL: define dso_local void @testArrayOfStruct(
+// LINUX-SAME: ) #[[ATTR0]] {
+// LINUX-NEXT:  [[ENTRY:.*:]]
+// LINUX-NEXT:    [[S:%.*]] = alloca [2 x [[STRUCT_S_0:%.*]]], align 16
+// LINUX-NEXT:    [[TMP0:%.*]] = getelementptr i8, ptr [[S]], i32 5
+// LINUX-NEXT:    store i8 0, ptr [[TMP0]], align 1
+// LINUX-NEXT:    [[TMP1:%.*]] = getelementptr i8, ptr [[S]], i32 6
+// LINUX-NEXT:    store i8 0, ptr [[TMP1]], align 2
+// LINUX-NEXT:    [[TMP2:%.*]] = getelementptr i8, ptr [[S]], i32 7
+// LINUX-NEXT:    store i8 0, ptr [[TMP2]], align 1
+// LINUX-NEXT:    [[TMP3:%.*]] = getelementptr i8, ptr [[S]], i32 13
+// LINUX-NEXT:    store i8 0, ptr [[TMP3]], align 1
+// LINUX-NEXT:    [[TMP4:%.*]] = getelementptr i8, ptr [[S]], i32 14
 // LINUX-NEXT:    store i8 0, ptr [[TMP4]], align 2
 // LINUX-NEXT:    [[TMP5:%.*]] = getelementptr i8, ptr [[S]], i32 15
 // LINUX-NEXT:    store i8 0, ptr [[TMP5]], align 1
@@ -1242,36 +3560,157 @@ void testArrayLongDouble() {
 // LINUX-NEXT:    store i8 0, ptr [[TMP11]], align 1
 // LINUX-NEXT:    ret void
 //
-// WINDOWS-LABEL: define dso_local void @"?testArrayOfStruct@@YAXXZ"() #0 {
-// WINDOWS-NEXT: entry:
-// WINDOWS-NEXT:   %s = alloca [2 x %struct.S.0], align 4
-// WINDOWS-NEXT:   %0 = getelementptr i8, ptr %s, i32 5
-// WINDOWS-NEXT:   store i8 0, ptr %0, align 1
-// WINDOWS-NEXT:   %1 = getelementptr i8, ptr %s, i32 6
-// WINDOWS-NEXT:   store i8 0, ptr %1, align 2
-// WINDOWS-NEXT:   %2 = getelementptr i8, ptr %s, i32 7
-// WINDOWS-NEXT:   store i8 0, ptr %2, align 1
-// WINDOWS-NEXT:   %3 = getelementptr i8, ptr %s, i32 13
-// WINDOWS-NEXT:   store i8 0, ptr %3, align 1
-// WINDOWS-NEXT:   %4 = getelementptr i8, ptr %s, i32 14
-// WINDOWS-NEXT:   store i8 0, ptr %4, align 2
-// WINDOWS-NEXT:   %5 = getelementptr i8, ptr %s, i32 15
-// WINDOWS-NEXT:   store i8 0, ptr %5, align 1
-// WINDOWS-NEXT:   %6 = getelementptr i8, ptr %s, i32 21
-// WINDOWS-NEXT:   store i8 0, ptr %6, align 1
-// WINDOWS-NEXT:   %7 = getelementptr i8, ptr %s, i32 22
-// WINDOWS-NEXT:   store i8 0, ptr %7, align 2
-// WINDOWS-NEXT:   %8 = getelementptr i8, ptr %s, i32 23
-// WINDOWS-NEXT:   store i8 0, ptr %8, align 1
-// WINDOWS-NEXT:   %9 = getelementptr i8, ptr %s, i32 29
-// WINDOWS-NEXT:   store i8 0, ptr %9, align 1
-// WINDOWS-NEXT:   %10 = getelementptr i8, ptr %s, i32 30
-// WINDOWS-NEXT:   store i8 0, ptr %10, align 2
-// WINDOWS-NEXT:   %11 = getelementptr i8, ptr %s, i32 31
-// WINDOWS-NEXT:   store i8 0, ptr %11, align 1
-// WINDOWS-NEXT:   ret void
-//
-void testArrayOfStruct() {
+// WINDOWS-LABEL: define dso_local void @testArrayOfStruct(
+// WINDOWS-SAME: ) #[[ATTR0]] {
+// WINDOWS-NEXT:  [[ENTRY:.*:]]
+// WINDOWS-NEXT:    [[S:%.*]] = alloca [2 x [[STRUCT_S_0:%.*]]], align 4
+// WINDOWS-NEXT:    [[TMP0:%.*]] = getelementptr i8, ptr [[S]], i32 5
+// WINDOWS-NEXT:    store i8 0, ptr [[TMP0]], align 1
+// WINDOWS-NEXT:    [[TMP1:%.*]] = getelementptr i8, ptr [[S]], i32 6
+// WINDOWS-NEXT:    store i8 0, ptr [[TMP1]], align 2
+// WINDOWS-NEXT:    [[TMP2:%.*]] = getelementptr i8, ptr [[S]], i32 7
+// WINDOWS-NEXT:    store i8 0, ptr [[TMP2]], align 1
+// WINDOWS-NEXT:    [[TMP3:%.*]] = getelementptr i8, ptr [[S]], i32 13
+// WINDOWS-NEXT:    store i8 0, ptr [[TMP3]], align 1
+// WINDOWS-NEXT:    [[TMP4:%.*]] = getelementptr i8, ptr [[S]], i32 14
+// WINDOWS-NEXT:    store i8 0, ptr [[TMP4]], align 2
+// WINDOWS-NEXT:    [[TMP5:%.*]] = getelementptr i8, ptr [[S]], i32 15
+// WINDOWS-NEXT:    store i8 0, ptr [[TMP5]], align 1
+// WINDOWS-NEXT:    [[TMP6:%.*]] = getelementptr i8, ptr [[S]], i32 21
+// WINDOWS-NEXT:    store i8 0, ptr [[TMP6]], align 1
+// WINDOWS-NEXT:    [[TMP7:%.*]] = getelementptr i8, ptr [[S]], i32 22
+// WINDOWS-NEXT:    store i8 0, ptr [[TMP7]], align 2
+// WINDOWS-NEXT:    [[TMP8:%.*]] = getelementptr i8, ptr [[S]], i32 23
+// WINDOWS-NEXT:    store i8 0, ptr [[TMP8]], align 1
+// WINDOWS-NEXT:    [[TMP9:%.*]] = getelementptr i8, ptr [[S]], i32 29
+// WINDOWS-NEXT:    store i8 0, ptr [[TMP9]], align 1
+// WINDOWS-NEXT:    [[TMP10:%.*]] = getelementptr i8, ptr [[S]], i32 30
+// WINDOWS-NEXT:    store i8 0, ptr [[TMP10]], align 2
+// WINDOWS-NEXT:    [[TMP11:%.*]] = getelementptr i8, ptr [[S]], i32 31
+// WINDOWS-NEXT:    store i8 0, ptr [[TMP11]], align 1
+// WINDOWS-NEXT:    ret void
+//
+// ARM-LE-LABEL: define dso_local void @testArrayOfStruct(
+// ARM-LE-SAME: ) #[[ATTR0]] {
+// ARM-LE-NEXT:  [[ENTRY:.*:]]
+// ARM-LE-NEXT:    [[S:%.*]] = alloca [2 x [[STRUCT_S_0:%.*]]], align 4
+// ARM-LE-NEXT:    [[TMP0:%.*]] = getelementptr i8, ptr [[S]], i32 5
+// ARM-LE-NEXT:    store i8 0, ptr [[TMP0]], align 1
+// ARM-LE-NEXT:    [[TMP1:%.*]] = getelementptr i8, ptr [[S]], i32 6
+// ARM-LE-NEXT:    store i8 0, ptr [[TMP1]], align 2
+// ARM-LE-NEXT:    [[TMP2:%.*]] = getelementptr i8, ptr [[S]], i32 7
+// ARM-LE-NEXT:    store i8 0, ptr [[TMP2]], align 1
+// ARM-LE-NEXT:    [[TMP3:%.*]] = getelementptr i8, ptr [[S]], i32 13
+// ARM-LE-NEXT:    store i8 0, ptr [[TMP3]], align 1
+// ARM-LE-NEXT:    [[TMP4:%.*]] = getelementptr i8, ptr [[S]], i32 14
+// ARM-LE-NEXT:    store i8 0, ptr [[TMP4]], align 2
+// ARM-LE-NEXT:    [[TMP5:%.*]] = getelementptr i8, ptr [[S]], i32 15
+// ARM-LE-NEXT:    store i8 0, ptr [[TMP5]], align 1
+// ARM-LE-NEXT:    [[TMP6:%.*]] = getelementptr i8, ptr [[S]], i32 21
+// ARM-LE-NEXT:    store i8 0, ptr [[TMP6]], align 1
+// ARM-LE-NEXT:    [[TMP7:%.*]] = getelementptr i8, ptr [[S]], i32 22
+// ARM-LE-NEXT:    store i8 0, ptr [[TMP7]], align 2
+// ARM-LE-NEXT:    [[TMP8:%.*]] = getelementptr i8, ptr [[S]], i32 23
+// ARM-LE-NEXT:    store i8 0, ptr [[TMP8]], align 1
+// ARM-LE-NEXT:    [[TMP9:%.*]] = getelementptr i8, ptr [[S]], i32 29
+// ARM-LE-NEXT:    store i8 0, ptr [[TMP9]], align 1
+// ARM-LE-NEXT:    [[TMP10:%.*]] = getelementptr i8, ptr [[S]], i32 30
+// ARM-LE-NEXT:    store i8 0, ptr [[TMP10]], align 2
+// ARM-LE-NEXT:    [[TMP11:%.*]] = getelementptr i8, ptr [[S]], i32 31
+// ARM-LE-NEXT:    store i8 0, ptr [[TMP11]], align 1
+// ARM-LE-NEXT:    ret void
+//
+// ARM-BE-LABEL: define dso_local void @testArrayOfStruct(
+// ARM-BE-SAME: ) #[[ATTR0]] {
+// ARM-BE-NEXT:  [[ENTRY:.*:]]
+// ARM-BE-NEXT:    [[S:%.*]] = alloca [2 x [[STRUCT_S_0:%.*]]], align 4
+// ARM-BE-NEXT:    [[TMP0:%.*]] = getelementptr i8, ptr [[S]], i32 5
+// ARM-BE-NEXT:    store i8 0, ptr [[TMP0]], align 1
+// ARM-BE-NEXT:    [[TMP1:%.*]] = getelementptr i8, ptr [[S]], i32 6
+// ARM-BE-NEXT:    store i8 0, ptr [[TMP1]], align 2
+// ARM-BE-NEXT:    [[TMP2:%.*]] = getelementptr i8, ptr [[S]], i32 7
+// ARM-BE-NEXT:    store i8 0, ptr [[TMP2]], align 1
+// ARM-BE-NEXT:    [[TMP3:%.*]] = getelementptr i8, ptr [[S]], i32 13
+// ARM-BE-NEXT:    store i8 0, ptr [[TMP3]], align 1
+// ARM-BE-NEXT:    [[TMP4:%.*]] = getelementptr i8, ptr [[S]], i32 14
+// ARM-BE-NEXT:    store i8 0, ptr [[TMP4]], align 2
+// ARM-BE-NEXT:    [[TMP5:%.*]] = getelementptr i8, ptr [[S]], i32 15
+// ARM-BE-NEXT:    store i8 0, ptr [[TMP5]], align 1
+// ARM-BE-NEXT:    [[TMP6:%.*]] = getelementptr i8, ptr [[S]], i32 21
+// ARM-BE-NEXT:    store i8 0, ptr [[TMP6]], align 1
+// ARM-BE-NEXT:    [[TMP7:%.*]] = getelementptr i8, ptr [[S]], i32 22
+// ARM-BE-NEXT:    store i8 0, ptr [[TMP7]], align 2
+// ARM-BE-NEXT:    [[TMP8:%.*]] = getelementptr i8, ptr [[S]], i32 23
+// ARM-BE-NEXT:    store i8 0, ptr [[TMP8]], align 1
+// ARM-BE-NEXT:    [[TMP9:%.*]] = getelementptr i8, ptr [[S]], i32 29
+// ARM-BE-NEXT:    store i8 0, ptr [[TMP9]], align 1
+// ARM-BE-NEXT:    [[TMP10:%.*]] = getelementptr i8, ptr [[S]], i32 30
+// ARM-BE-NEXT:    store i8 0, ptr [[TMP10]], align 2
+// ARM-BE-NEXT:    [[TMP11:%.*]] = getelementptr i8, ptr [[S]], i32 31
+// ARM-BE-NEXT:    store i8 0, ptr [[TMP11]], align 1
+// ARM-BE-NEXT:    ret void
+//
+// AARCH64-LE-LABEL: define dso_local void @testArrayOfStruct(
+// AARCH64-LE-SAME: ) #[[ATTR0]] {
+// AARCH64-LE-NEXT:  [[ENTRY:.*:]]
+// AARCH64-LE-NEXT:    [[S:%.*]] = alloca [2 x [[STRUCT_S_0:%.*]]], align 4
+// AARCH64-LE-NEXT:    [[TMP0:%.*]] = getelementptr i8, ptr [[S]], i32 5
+// AARCH64-LE-NEXT:    store i8 0, ptr [[TMP0]], align 1
+// AARCH64-LE-NEXT:    [[TMP1:%.*]] = getelementptr i8, ptr [[S]], i32 6
+// AARCH64-LE-NEXT:    store i8 0, ptr [[TMP1]], align 2
+// AARCH64-LE-NEXT:    [[TMP2:%.*]] = getelementptr i8, ptr [[S]], i32 7
+// AARCH64-LE-NEXT:    store i8 0, ptr [[TMP2]], align 1
+// AARCH64-LE-NEXT:    [[TMP3:%.*]] = getelementptr i8, ptr [[S]], i32 13
+// AARCH64-LE-NEXT:    store i8 0, ptr [[TMP3]], align 1
+// AARCH64-LE-NEXT:    [[TMP4:%.*]] = getelementptr i8, ptr [[S]], i32 14
+// AARCH64-LE-NEXT:    store i8 0, ptr [[TMP4]], align 2
+// AARCH64-LE-NEXT:    [[TMP5:%.*]] = getelementptr i8, ptr [[S]], i32 15
+// AARCH64-LE-NEXT:    store i8 0, ptr [[TMP5]], align 1
+// AARCH64-LE-NEXT:    [[TMP6:%.*]] = getelementptr i8, ptr [[S]], i32 21
+// AARCH64-LE-NEXT:    store i8 0, ptr [[TMP6]], align 1
+// AARCH64-LE-NEXT:    [[TMP7:%.*]] = getelementptr i8, ptr [[S]], i32 22
+// AARCH64-LE-NEXT:    store i8 0, ptr [[TMP7]], align 2
+// AARCH64-LE-NEXT:    [[TMP8:%.*]] = getelementptr i8, ptr [[S]], i32 23
+// AARCH64-LE-NEXT:    store i8 0, ptr [[TMP8]], align 1
+// AARCH64-LE-NEXT:    [[TMP9:%.*]] = getelementptr i8, ptr [[S]], i32 29
+// AARCH64-LE-NEXT:    store i8 0, ptr [[TMP9]], align 1
+// AARCH64-LE-NEXT:    [[TMP10:%.*]] = getelementptr i8, ptr [[S]], i32 30
+// AARCH64-LE-NEXT:    store i8 0, ptr [[TMP10]], align 2
+// AARCH64-LE-NEXT:    [[TMP11:%.*]] = getelementptr i8, ptr [[S]], i32 31
+// AARCH64-LE-NEXT:    store i8 0, ptr [[TMP11]], align 1
+// AARCH64-LE-NEXT:    ret void
+//
+// AARCH64-BE-LABEL: define dso_local void @testArrayOfStruct(
+// AARCH64-BE-SAME: ) #[[ATTR0]] {
+// AARCH64-BE-NEXT:  [[ENTRY:.*:]]
+// AARCH64-BE-NEXT:    [[S:%.*]] = alloca [2 x [[STRUCT_S_0:%.*]]], align 4
+// AARCH64-BE-NEXT:    [[TMP0:%.*]] = getelementptr i8, ptr [[S]], i32 5
+// AARCH64-BE-NEXT:    store i8 0, ptr [[TMP0]], align 1
+// AARCH64-BE-NEXT:    [[TMP1:%.*]] = getelementptr i8, ptr [[S]], i32 6
+// AARCH64-BE-NEXT:    store i8 0, ptr [[TMP1]], align 2
+// AARCH64-BE-NEXT:    [[TMP2:%.*]] = getelementptr i8, ptr [[S]], i32 7
+// AARCH64-BE-NEXT:    store i8 0, ptr [[TMP2]], align 1
+// AARCH64-BE-NEXT:    [[TMP3:%.*]] = getelementptr i8, ptr [[S]], i32 13
+// AARCH64-BE-NEXT:    store i8 0, ptr [[TMP3]], align 1
+// AARCH64-BE-NEXT:    [[TMP4:%.*]] = getelementptr i8, ptr [[S]], i32 14
+// AARCH64-BE-NEXT:    store i8 0, ptr [[TMP4]], align 2
+// AARCH64-BE-NEXT:    [[TMP5:%.*]] = getelementptr i8, ptr [[S]], i32 15
+// AARCH64-BE-NEXT:    store i8 0, ptr [[TMP5]], align 1
+// AARCH64-BE-NEXT:    [[TMP6:%.*]] = getelementptr i8, ptr [[S]], i32 21
+// AARCH64-BE-NEXT:    store i8 0, ptr [[TMP6]], align 1
+// AARCH64-BE-NEXT:    [[TMP7:%.*]] = getelementptr i8, ptr [[S]], i32 22
+// AARCH64-BE-NEXT:    store i8 0, ptr [[TMP7]], align 2
+// AARCH64-BE-NEXT:    [[TMP8:%.*]] = getelementptr i8, ptr [[S]], i32 23
+// AARCH64-BE-NEXT:    store i8 0, ptr [[TMP8]], align 1
+// AARCH64-BE-NEXT:    [[TMP9:%.*]] = getelementptr i8, ptr [[S]], i32 29
+// AARCH64-BE-NEXT:    store i8 0, ptr [[TMP9]], align 1
+// AARCH64-BE-NEXT:    [[TMP10:%.*]] = getelementptr i8, ptr [[S]], i32 30
+// AARCH64-BE-NEXT:    store i8 0, ptr [[TMP10]], align 2
+// AARCH64-BE-NEXT:    [[TMP11:%.*]] = getelementptr i8, ptr [[S]], i32 31
+// AARCH64-BE-NEXT:    store i8 0, ptr [[TMP11]], align 1
+// AARCH64-BE-NEXT:    ret void
+//
+extern "C" void testArrayOfStruct() {
   struct S {
     int i1;
     char c1;
@@ -1288,6 +3727,7 @@ void testArrayOfStruct() {
   __builtin_clear_padding(&s);
 }
 
+
 struct ArrOfStructsWithPadding {
   Bar bars[2];
 };
@@ -1296,7 +3736,7 @@ struct ArrOfStructsWithPadding {
 // "c" (1), PAD_1, "d" (1), PAD_2, "c" (2), PAD_3, "d" (2), PAD_4
 // %struct.ArrOfStructsWithPadding = type { [2 x %struct.Bar] }
 
-// LINUX-LABEL: define dso_local void @_Z27testArrOfStructsWithPaddingP23ArrOfStructsWithPadding(
+// LINUX-LABEL: define dso_local void @testArrOfStructsWithPadding(
 // LINUX-SAME: ptr noundef [[ARR:%.*]]) #[[ATTR0]] {
 // LINUX-NEXT:  [[ENTRY:.*:]]
 // LINUX-NEXT:    [[ARR_ADDR:%.*]] = alloca ptr, align 8
@@ -1312,32 +3752,98 @@ struct ArrOfStructsWithPadding {
 // LINUX-NEXT:    store i8 0, ptr [[TMP4]], align 1
 // LINUX-NEXT:    ret void
 //
-// WINDOWS-LABEL: define dso_local void @"?testArrOfStructsWithPadding@@YAXPEAUArrOfStructsWithPadding@@@Z"(ptr noundef %arr) #0 {
-// WINDOWS-NEXT: entry:
-// WINDOWS-NEXT:   %arr.addr = alloca ptr, align 8
-// WINDOWS-NEXT:   store ptr %arr, ptr %arr.addr, align 8
-// WINDOWS-NEXT:   %0 = load ptr, ptr %arr.addr, align 8
-// WINDOWS-NEXT:   %1 = getelementptr i8, ptr %0, i32 1
-// WINDOWS-NEXT:   store i8 0, ptr %1, align 1
-// WINDOWS-NEXT:   %2 = getelementptr i8, ptr %0, i32 3
-// WINDOWS-NEXT:   store i8 0, ptr %2, align 1
-// WINDOWS-NEXT:   %3 = getelementptr i8, ptr %0, i32 5
-// WINDOWS-NEXT:   store i8 0, ptr %3, align 1
-// WINDOWS-NEXT:   %4 = getelementptr i8, ptr %0, i32 7
-// WINDOWS-NEXT:   store i8 0, ptr %4, align 1
-// WINDOWS-NEXT:   ret void
-//
-void testArrOfStructsWithPadding(ArrOfStructsWithPadding *arr) {
+// WINDOWS-LABEL: define dso_local void @testArrOfStructsWithPadding(
+// WINDOWS-SAME: ptr noundef [[ARR:%.*]]) #[[ATTR0]] {
+// WINDOWS-NEXT:  [[ENTRY:.*:]]
+// WINDOWS-NEXT:    [[ARR_ADDR:%.*]] = alloca ptr, align 8
+// WINDOWS-NEXT:    store ptr [[ARR]], ptr [[ARR_ADDR]], align 8
+// WINDOWS-NEXT:    [[TMP0:%.*]] = load ptr, ptr [[ARR_ADDR]], align 8
+// WINDOWS-NEXT:    [[TMP1:%.*]] = getelementptr i8, ptr [[TMP0]], i32 1
+// WINDOWS-NEXT:    store i8 0, ptr [[TMP1]], align 1
+// WINDOWS-NEXT:    [[TMP2:%.*]] = getelementptr i8, ptr [[TMP0]], i32 3
+// WINDOWS-NEXT:    store i8 0, ptr [[TMP2]], align 1
+// WINDOWS-NEXT:    [[TMP3:%.*]] = getelementptr i8, ptr [[TMP0]], i32 5
+// WINDOWS-NEXT:    store i8 0, ptr [[TMP3]], align 1
+// WINDOWS-NEXT:    [[TMP4:%.*]] = getelementptr i8, ptr [[TMP0]], i32 7
+// WINDOWS-NEXT:    store i8 0, ptr [[TMP4]], align 1
+// WINDOWS-NEXT:    ret void
+//
+// ARM-LE-LABEL: define dso_local void @testArrOfStructsWithPadding(
+// ARM-LE-SAME: ptr noundef [[ARR:%.*]]) #[[ATTR0]] {
+// ARM-LE-NEXT:  [[ENTRY:.*:]]
+// ARM-LE-NEXT:    [[ARR_ADDR:%.*]] = alloca ptr, align 4
+// ARM-LE-NEXT:    store ptr [[ARR]], ptr [[ARR_ADDR]], align 4
+// ARM-LE-NEXT:    [[TMP0:%.*]] = load ptr, ptr [[ARR_ADDR]], align 4
+// ARM-LE-NEXT:    [[TMP1:%.*]] = getelementptr i8, ptr [[TMP0]], i32 1
+// ARM-LE-NEXT:    store i8 0, ptr [[TMP1]], align 1
+// ARM-LE-NEXT:    [[TMP2:%.*]] = getelementptr i8, ptr [[TMP0]], i32 3
+// ARM-LE-NEXT:    store i8 0, ptr [[TMP2]], align 1
+// ARM-LE-NEXT:    [[TMP3:%.*]] = getelementptr i8, ptr [[TMP0]], i32 5
+// ARM-LE-NEXT:    store i8 0, ptr [[TMP3]], align 1
+// ARM-LE-NEXT:    [[TMP4:%.*]] = getelementptr i8, ptr [[TMP0]], i32 7
+// ARM-LE-NEXT:    store i8 0, ptr [[TMP4]], align 1
+// ARM-LE-NEXT:    ret void
+//
+// ARM-BE-LABEL: define dso_local void @testArrOfStructsWithPadding(
+// ARM-BE-SAME: ptr noundef [[ARR:%.*]]) #[[ATTR0]] {
+// ARM-BE-NEXT:  [[ENTRY:.*:]]
+// ARM-BE-NEXT:    [[ARR_ADDR:%.*]] = alloca ptr, align 4
+// ARM-BE-NEXT:    store ptr [[ARR]], ptr [[ARR_ADDR]], align 4
+// ARM-BE-NEXT:    [[TMP0:%.*]] = load ptr, ptr [[ARR_ADDR]], align 4
+// ARM-BE-NEXT:    [[TMP1:%.*]] = getelementptr i8, ptr [[TMP0]], i32 1
+// ARM-BE-NEXT:    store i8 0, ptr [[TMP1]], align 1
+// ARM-BE-NEXT:    [[TMP2:%.*]] = getelementptr i8, ptr [[TMP0]], i32 3
+// ARM-BE-NEXT:    store i8 0, ptr [[TMP2]], align 1
+// ARM-BE-NEXT:    [[TMP3:%.*]] = getelementptr i8, ptr [[TMP0]], i32 5
+// ARM-BE-NEXT:    store i8 0, ptr [[TMP3]], align 1
+// ARM-BE-NEXT:    [[TMP4:%.*]] = getelementptr i8, ptr [[TMP0]], i32 7
+// ARM-BE-NEXT:    store i8 0, ptr [[TMP4]], align 1
+// ARM-BE-NEXT:    ret void
+//
+// AARCH64-LE-LABEL: define dso_local void @testArrOfStructsWithPadding(
+// AARCH64-LE-SAME: ptr noundef [[ARR:%.*]]) #[[ATTR0]] {
+// AARCH64-LE-NEXT:  [[ENTRY:.*:]]
+// AARCH64-LE-NEXT:    [[ARR_ADDR:%.*]] = alloca ptr, align 8
+// AARCH64-LE-NEXT:    store ptr [[ARR]], ptr [[ARR_ADDR]], align 8
+// AARCH64-LE-NEXT:    [[TMP0:%.*]] = load ptr, ptr [[ARR_ADDR]], align 8
+// AARCH64-LE-NEXT:    [[TMP1:%.*]] = getelementptr i8, ptr [[TMP0]], i32 1
+// AARCH64-LE-NEXT:    store i8 0, ptr [[TMP1]], align 1
+// AARCH64-LE-NEXT:    [[TMP2:%.*]] = getelementptr i8, ptr [[TMP0]], i32 3
+// AARCH64-LE-NEXT:    store i8 0, ptr [[TMP2]], align 1
+// AARCH64-LE-NEXT:    [[TMP3:%.*]] = getelementptr i8, ptr [[TMP0]], i32 5
+// AARCH64-LE-NEXT:    store i8 0, ptr [[TMP3]], align 1
+// AARCH64-LE-NEXT:    [[TMP4:%.*]] = getelementptr i8, ptr [[TMP0]], i32 7
+// AARCH64-LE-NEXT:    store i8 0, ptr [[TMP4]], align 1
+// AARCH64-LE-NEXT:    ret void
+//
+// AARCH64-BE-LABEL: define dso_local void @testArrOfStructsWithPadding(
+// AARCH64-BE-SAME: ptr noundef [[ARR:%.*]]) #[[ATTR0]] {
+// AARCH64-BE-NEXT:  [[ENTRY:.*:]]
+// AARCH64-BE-NEXT:    [[ARR_ADDR:%.*]] = alloca ptr, align 8
+// AARCH64-BE-NEXT:    store ptr [[ARR]], ptr [[ARR_ADDR]], align 8
+// AARCH64-BE-NEXT:    [[TMP0:%.*]] = load ptr, ptr [[ARR_ADDR]], align 8
+// AARCH64-BE-NEXT:    [[TMP1:%.*]] = getelementptr i8, ptr [[TMP0]], i32 1
+// AARCH64-BE-NEXT:    store i8 0, ptr [[TMP1]], align 1
+// AARCH64-BE-NEXT:    [[TMP2:%.*]] = getelementptr i8, ptr [[TMP0]], i32 3
+// AARCH64-BE-NEXT:    store i8 0, ptr [[TMP2]], align 1
+// AARCH64-BE-NEXT:    [[TMP3:%.*]] = getelementptr i8, ptr [[TMP0]], i32 5
+// AARCH64-BE-NEXT:    store i8 0, ptr [[TMP3]], align 1
+// AARCH64-BE-NEXT:    [[TMP4:%.*]] = getelementptr i8, ptr [[TMP0]], i32 7
+// AARCH64-BE-NEXT:    store i8 0, ptr [[TMP4]], align 1
+// AARCH64-BE-NEXT:    ret void
+//
+extern "C" void testArrOfStructsWithPadding(ArrOfStructsWithPadding *arr) {
   __builtin_clear_padding(arr);
 }
 
+
 template <class T>
 struct S12 {
   T t;
   char c;
 };
 
-// LINUX-LABEL: define dso_local void @_Z18testTemplateStructP3S12IiE(
+// LINUX-LABEL: define dso_local void @testTemplateStruct(
 // LINUX-SAME: ptr noundef [[S12:%.*]]) #[[ATTR0]] {
 // LINUX-NEXT:  [[ENTRY:.*:]]
 // LINUX-NEXT:    [[S12_ADDR:%.*]] = alloca ptr, align 8
@@ -1351,24 +3857,82 @@ struct S12 {
 // LINUX-NEXT:    store i8 0, ptr [[TMP3]], align 1
 // LINUX-NEXT:    ret void
 //
-// WINDOWS-LABEL: define dso_local void @"?testTemplateStruct@@YAXPEAU?$S12 at H@@@Z"(ptr noundef %s12) #0 {
-// WINDOWS-NEXT: entry:
-// WINDOWS-NEXT:   %s12.addr = alloca ptr, align 8
-// WINDOWS-NEXT:   store ptr %s12, ptr %s12.addr, align 8
-// WINDOWS-NEXT:   %0 = load ptr, ptr %s12.addr, align 8
-// WINDOWS-NEXT:   %1 = getelementptr i8, ptr %0, i32 5
-// WINDOWS-NEXT:   store i8 0, ptr %1, align 1
-// WINDOWS-NEXT:   %2 = getelementptr i8, ptr %0, i32 6
-// WINDOWS-NEXT:   store i8 0, ptr %2, align 2
-// WINDOWS-NEXT:   %3 = getelementptr i8, ptr %0, i32 7
-// WINDOWS-NEXT:   store i8 0, ptr %3, align 1
-// WINDOWS-NEXT:   ret void
-//
-void testTemplateStruct(S12<int>* s12) {
+// WINDOWS-LABEL: define dso_local void @testTemplateStruct(
+// WINDOWS-SAME: ptr noundef [[S12:%.*]]) #[[ATTR0]] {
+// WINDOWS-NEXT:  [[ENTRY:.*:]]
+// WINDOWS-NEXT:    [[S12_ADDR:%.*]] = alloca ptr, align 8
+// WINDOWS-NEXT:    store ptr [[S12]], ptr [[S12_ADDR]], align 8
+// WINDOWS-NEXT:    [[TMP0:%.*]] = load ptr, ptr [[S12_ADDR]], align 8
+// WINDOWS-NEXT:    [[TMP1:%.*]] = getelementptr i8, ptr [[TMP0]], i32 5
+// WINDOWS-NEXT:    store i8 0, ptr [[TMP1]], align 1
+// WINDOWS-NEXT:    [[TMP2:%.*]] = getelementptr i8, ptr [[TMP0]], i32 6
+// WINDOWS-NEXT:    store i8 0, ptr [[TMP2]], align 2
+// WINDOWS-NEXT:    [[TMP3:%.*]] = getelementptr i8, ptr [[TMP0]], i32 7
+// WINDOWS-NEXT:    store i8 0, ptr [[TMP3]], align 1
+// WINDOWS-NEXT:    ret void
+//
+// ARM-LE-LABEL: define dso_local void @testTemplateStruct(
+// ARM-LE-SAME: ptr noundef [[S12:%.*]]) #[[ATTR0]] {
+// ARM-LE-NEXT:  [[ENTRY:.*:]]
+// ARM-LE-NEXT:    [[S12_ADDR:%.*]] = alloca ptr, align 4
+// ARM-LE-NEXT:    store ptr [[S12]], ptr [[S12_ADDR]], align 4
+// ARM-LE-NEXT:    [[TMP0:%.*]] = load ptr, ptr [[S12_ADDR]], align 4
+// ARM-LE-NEXT:    [[TMP1:%.*]] = getelementptr i8, ptr [[TMP0]], i32 5
+// ARM-LE-NEXT:    store i8 0, ptr [[TMP1]], align 1
+// ARM-LE-NEXT:    [[TMP2:%.*]] = getelementptr i8, ptr [[TMP0]], i32 6
+// ARM-LE-NEXT:    store i8 0, ptr [[TMP2]], align 2
+// ARM-LE-NEXT:    [[TMP3:%.*]] = getelementptr i8, ptr [[TMP0]], i32 7
+// ARM-LE-NEXT:    store i8 0, ptr [[TMP3]], align 1
+// ARM-LE-NEXT:    ret void
+//
+// ARM-BE-LABEL: define dso_local void @testTemplateStruct(
+// ARM-BE-SAME: ptr noundef [[S12:%.*]]) #[[ATTR0]] {
+// ARM-BE-NEXT:  [[ENTRY:.*:]]
+// ARM-BE-NEXT:    [[S12_ADDR:%.*]] = alloca ptr, align 4
+// ARM-BE-NEXT:    store ptr [[S12]], ptr [[S12_ADDR]], align 4
+// ARM-BE-NEXT:    [[TMP0:%.*]] = load ptr, ptr [[S12_ADDR]], align 4
+// ARM-BE-NEXT:    [[TMP1:%.*]] = getelementptr i8, ptr [[TMP0]], i32 5
+// ARM-BE-NEXT:    store i8 0, ptr [[TMP1]], align 1
+// ARM-BE-NEXT:    [[TMP2:%.*]] = getelementptr i8, ptr [[TMP0]], i32 6
+// ARM-BE-NEXT:    store i8 0, ptr [[TMP2]], align 2
+// ARM-BE-NEXT:    [[TMP3:%.*]] = getelementptr i8, ptr [[TMP0]], i32 7
+// ARM-BE-NEXT:    store i8 0, ptr [[TMP3]], align 1
+// ARM-BE-NEXT:    ret void
+//
+// AARCH64-LE-LABEL: define dso_local void @testTemplateStruct(
+// AARCH64-LE-SAME: ptr noundef [[S12:%.*]]) #[[ATTR0]] {
+// AARCH64-LE-NEXT:  [[ENTRY:.*:]]
+// AARCH64-LE-NEXT:    [[S12_ADDR:%.*]] = alloca ptr, align 8
+// AARCH64-LE-NEXT:    store ptr [[S12]], ptr [[S12_ADDR]], align 8
+// AARCH64-LE-NEXT:    [[TMP0:%.*]] = load ptr, ptr [[S12_ADDR]], align 8
+// AARCH64-LE-NEXT:    [[TMP1:%.*]] = getelementptr i8, ptr [[TMP0]], i32 5
+// AARCH64-LE-NEXT:    store i8 0, ptr [[TMP1]], align 1
+// AARCH64-LE-NEXT:    [[TMP2:%.*]] = getelementptr i8, ptr [[TMP0]], i32 6
+// AARCH64-LE-NEXT:    store i8 0, ptr [[TMP2]], align 2
+// AARCH64-LE-NEXT:    [[TMP3:%.*]] = getelementptr i8, ptr [[TMP0]], i32 7
+// AARCH64-LE-NEXT:    store i8 0, ptr [[TMP3]], align 1
+// AARCH64-LE-NEXT:    ret void
+//
+// AARCH64-BE-LABEL: define dso_local void @testTemplateStruct(
+// AARCH64-BE-SAME: ptr noundef [[S12:%.*]]) #[[ATTR0]] {
+// AARCH64-BE-NEXT:  [[ENTRY:.*:]]
+// AARCH64-BE-NEXT:    [[S12_ADDR:%.*]] = alloca ptr, align 8
+// AARCH64-BE-NEXT:    store ptr [[S12]], ptr [[S12_ADDR]], align 8
+// AARCH64-BE-NEXT:    [[TMP0:%.*]] = load ptr, ptr [[S12_ADDR]], align 8
+// AARCH64-BE-NEXT:    [[TMP1:%.*]] = getelementptr i8, ptr [[TMP0]], i32 5
+// AARCH64-BE-NEXT:    store i8 0, ptr [[TMP1]], align 1
+// AARCH64-BE-NEXT:    [[TMP2:%.*]] = getelementptr i8, ptr [[TMP0]], i32 6
+// AARCH64-BE-NEXT:    store i8 0, ptr [[TMP2]], align 2
+// AARCH64-BE-NEXT:    [[TMP3:%.*]] = getelementptr i8, ptr [[TMP0]], i32 7
+// AARCH64-BE-NEXT:    store i8 0, ptr [[TMP3]], align 1
+// AARCH64-BE-NEXT:    ret void
+//
+extern "C" void testTemplateStruct(S12<int>* s12) {
   __builtin_clear_padding(s12);
 }
 
-// LINUX-LABEL: define dso_local void @_Z10testAtomicPU7_Atomic3Bar(
+
+// LINUX-LABEL: define dso_local void @testAtomic(
 // LINUX-SAME: ptr noundef [[BAR:%.*]]) #[[ATTR0]] {
 // LINUX-NEXT:  [[ENTRY:.*:]]
 // LINUX-NEXT:    [[BAR_ADDR:%.*]] = alloca ptr, align 8
@@ -1380,32 +3944,85 @@ void testTemplateStruct(S12<int>* s12) {
 // LINUX-NEXT:    store i8 0, ptr [[TMP2]], align 1
 // LINUX-NEXT:    ret void
 //
-// WINDOWS-LABEL: define dso_local void @"?testAtomic@@YAXPEAU?$_Atomic at UBar@@@__clang@@@Z"(ptr noundef %bar) #0 {
-// WINDOWS-NEXT: entry:
-// WINDOWS-NEXT:   %bar.addr = alloca ptr, align 8
-// WINDOWS-NEXT:   store ptr %bar, ptr %bar.addr, align 8
-// WINDOWS-NEXT:   %0 = load ptr, ptr %bar.addr, align 8
-// WINDOWS-NEXT:   %1 = getelementptr i8, ptr %0, i32 1
-// WINDOWS-NEXT:   store i8 0, ptr %1, align 1
-// WINDOWS-NEXT:   %2 = getelementptr i8, ptr %0, i32 3
-// WINDOWS-NEXT:   store i8 0, ptr %2, align 1
-// WINDOWS-NEXT:   ret void
-//
-void testAtomic(_Atomic(Bar)* bar) {
+// WINDOWS-LABEL: define dso_local void @testAtomic(
+// WINDOWS-SAME: ptr noundef [[BAR:%.*]]) #[[ATTR0]] {
+// WINDOWS-NEXT:  [[ENTRY:.*:]]
+// WINDOWS-NEXT:    [[BAR_ADDR:%.*]] = alloca ptr, align 8
+// WINDOWS-NEXT:    store ptr [[BAR]], ptr [[BAR_ADDR]], align 8
+// WINDOWS-NEXT:    [[TMP0:%.*]] = load ptr, ptr [[BAR_ADDR]], align 8
+// WINDOWS-NEXT:    [[TMP1:%.*]] = getelementptr i8, ptr [[TMP0]], i32 1
+// WINDOWS-NEXT:    store i8 0, ptr [[TMP1]], align 1
+// WINDOWS-NEXT:    [[TMP2:%.*]] = getelementptr i8, ptr [[TMP0]], i32 3
+// WINDOWS-NEXT:    store i8 0, ptr [[TMP2]], align 1
+// WINDOWS-NEXT:    ret void
+//
+// ARM-LE-LABEL: define dso_local void @testAtomic(
+// ARM-LE-SAME: ptr noundef [[BAR:%.*]]) #[[ATTR0]] {
+// ARM-LE-NEXT:  [[ENTRY:.*:]]
+// ARM-LE-NEXT:    [[BAR_ADDR:%.*]] = alloca ptr, align 4
+// ARM-LE-NEXT:    store ptr [[BAR]], ptr [[BAR_ADDR]], align 4
+// ARM-LE-NEXT:    [[TMP0:%.*]] = load ptr, ptr [[BAR_ADDR]], align 4
+// ARM-LE-NEXT:    [[TMP1:%.*]] = getelementptr i8, ptr [[TMP0]], i32 1
+// ARM-LE-NEXT:    store i8 0, ptr [[TMP1]], align 1
+// ARM-LE-NEXT:    [[TMP2:%.*]] = getelementptr i8, ptr [[TMP0]], i32 3
+// ARM-LE-NEXT:    store i8 0, ptr [[TMP2]], align 1
+// ARM-LE-NEXT:    ret void
+//
+// ARM-BE-LABEL: define dso_local void @testAtomic(
+// ARM-BE-SAME: ptr noundef [[BAR:%.*]]) #[[ATTR0]] {
+// ARM-BE-NEXT:  [[ENTRY:.*:]]
+// ARM-BE-NEXT:    [[BAR_ADDR:%.*]] = alloca ptr, align 4
+// ARM-BE-NEXT:    store ptr [[BAR]], ptr [[BAR_ADDR]], align 4
+// ARM-BE-NEXT:    [[TMP0:%.*]] = load ptr, ptr [[BAR_ADDR]], align 4
+// ARM-BE-NEXT:    [[TMP1:%.*]] = getelementptr i8, ptr [[TMP0]], i32 1
+// ARM-BE-NEXT:    store i8 0, ptr [[TMP1]], align 1
+// ARM-BE-NEXT:    [[TMP2:%.*]] = getelementptr i8, ptr [[TMP0]], i32 3
+// ARM-BE-NEXT:    store i8 0, ptr [[TMP2]], align 1
+// ARM-BE-NEXT:    ret void
+//
+// AARCH64-LE-LABEL: define dso_local void @testAtomic(
+// AARCH64-LE-SAME: ptr noundef [[BAR:%.*]]) #[[ATTR0]] {
+// AARCH64-LE-NEXT:  [[ENTRY:.*:]]
+// AARCH64-LE-NEXT:    [[BAR_ADDR:%.*]] = alloca ptr, align 8
+// AARCH64-LE-NEXT:    store ptr [[BAR]], ptr [[BAR_ADDR]], align 8
+// AARCH64-LE-NEXT:    [[TMP0:%.*]] = load ptr, ptr [[BAR_ADDR]], align 8
+// AARCH64-LE-NEXT:    [[TMP1:%.*]] = getelementptr i8, ptr [[TMP0]], i32 1
+// AARCH64-LE-NEXT:    store i8 0, ptr [[TMP1]], align 1
+// AARCH64-LE-NEXT:    [[TMP2:%.*]] = getelementptr i8, ptr [[TMP0]], i32 3
+// AARCH64-LE-NEXT:    store i8 0, ptr [[TMP2]], align 1
+// AARCH64-LE-NEXT:    ret void
+//
+// AARCH64-BE-LABEL: define dso_local void @testAtomic(
+// AARCH64-BE-SAME: ptr noundef [[BAR:%.*]]) #[[ATTR0]] {
+// AARCH64-BE-NEXT:  [[ENTRY:.*:]]
+// AARCH64-BE-NEXT:    [[BAR_ADDR:%.*]] = alloca ptr, align 8
+// AARCH64-BE-NEXT:    store ptr [[BAR]], ptr [[BAR_ADDR]], align 8
+// AARCH64-BE-NEXT:    [[TMP0:%.*]] = load ptr, ptr [[BAR_ADDR]], align 8
+// AARCH64-BE-NEXT:    [[TMP1:%.*]] = getelementptr i8, ptr [[TMP0]], i32 1
+// AARCH64-BE-NEXT:    store i8 0, ptr [[TMP1]], align 1
+// AARCH64-BE-NEXT:    [[TMP2:%.*]] = getelementptr i8, ptr [[TMP0]], i32 3
+// AARCH64-BE-NEXT:    store i8 0, ptr [[TMP2]], align 1
+// AARCH64-BE-NEXT:    ret void
+//
+extern "C" void testAtomic(_Atomic(Bar)* bar) {
   __builtin_clear_padding(bar);
 }
 
 
+
 struct NonTriviallyCopyable {
   int i;
   char c;
 
+//
   NonTriviallyCopyable(){}
+//
   NonTriviallyCopyable(const NonTriviallyCopyable&) {}
+//
   ~NonTriviallyCopyable() {}
 };
 
-// LINUX-LABEL: define dso_local void @_Z24testNonTriviallyCopyable20NonTriviallyCopyable(
+// LINUX-LABEL: define dso_local void @testNonTriviallyCopyable(
 // LINUX-SAME: ptr nofreeobj noundef align 4 dereferenceable(8) [[NTC:%.*]]) #[[ATTR0]] {
 // LINUX-NEXT:  [[ENTRY:.*:]]
 // LINUX-NEXT:    [[NTC_INDIRECT_ADDR:%.*]] = alloca ptr, align 8
@@ -1418,27 +4035,81 @@ struct NonTriviallyCopyable {
 // LINUX-NEXT:    store i8 0, ptr [[TMP2]], align 1
 // LINUX-NEXT:    ret void
 //
-// WINDOWS-LABEL: define dso_local void @"?testNonTriviallyCopyable@@YAXUNonTriviallyCopyable@@@Z"(ptr nofreeobj noundef align 4 dead_on_return dereferenceable(8) %ntc) #0 {
-// WINDOWS-NEXT: entry:
-// WINDOWS-NEXT:   %ntc.indirect_addr = alloca ptr, align 8
-// WINDOWS-NEXT:   store ptr %ntc, ptr %ntc.indirect_addr, align 8
-// WINDOWS-NEXT:   %0 = getelementptr i8, ptr %ntc, i32 5
-// WINDOWS-NEXT:   store i8 0, ptr %0, align 1
-// WINDOWS-NEXT:   %1 = getelementptr i8, ptr %ntc, i32 6
-// WINDOWS-NEXT:   store i8 0, ptr %1, align 2
-// WINDOWS-NEXT:   %2 = getelementptr i8, ptr %ntc, i32 7
-// WINDOWS-NEXT:   store i8 0, ptr %2, align 1
-// WINDOWS-NEXT:   call void @"??1NonTriviallyCopyable@@QEAA at XZ"(ptr noundef nonnull align 4 dead_on_return(8) dereferenceable(8) %ntc) #1
-// WINDOWS-NEXT:   ret void
-//
-void testNonTriviallyCopyable(NonTriviallyCopyable ntc) {
+// WINDOWS-LABEL: define dso_local void @testNonTriviallyCopyable(
+// WINDOWS-SAME: ptr nofreeobj noundef align 4 dead_on_return dereferenceable(8) [[NTC:%.*]]) #[[ATTR0]] {
+// WINDOWS-NEXT:  [[ENTRY:.*:]]
+// WINDOWS-NEXT:    [[NTC_INDIRECT_ADDR:%.*]] = alloca ptr, align 8
+// WINDOWS-NEXT:    store ptr [[NTC]], ptr [[NTC_INDIRECT_ADDR]], align 8
+// WINDOWS-NEXT:    [[TMP0:%.*]] = getelementptr i8, ptr [[NTC]], i32 5
+// WINDOWS-NEXT:    store i8 0, ptr [[TMP0]], align 1
+// WINDOWS-NEXT:    [[TMP1:%.*]] = getelementptr i8, ptr [[NTC]], i32 6
+// WINDOWS-NEXT:    store i8 0, ptr [[TMP1]], align 2
+// WINDOWS-NEXT:    [[TMP2:%.*]] = getelementptr i8, ptr [[NTC]], i32 7
+// WINDOWS-NEXT:    store i8 0, ptr [[TMP2]], align 1
+// WINDOWS-NEXT:    call void @"??1NonTriviallyCopyable@@QEAA at XZ"(ptr noundef nonnull align 4 dead_on_return(8) dereferenceable(8) [[NTC]]) #[[ATTR1]]
+// WINDOWS-NEXT:    ret void
+//
+// ARM-LE-LABEL: define dso_local void @testNonTriviallyCopyable(
+// ARM-LE-SAME: ptr nofreeobj noundef align 4 dereferenceable(8) [[NTC:%.*]]) #[[ATTR0]] {
+// ARM-LE-NEXT:  [[ENTRY:.*:]]
+// ARM-LE-NEXT:    [[NTC_INDIRECT_ADDR:%.*]] = alloca ptr, align 4
+// ARM-LE-NEXT:    store ptr [[NTC]], ptr [[NTC_INDIRECT_ADDR]], align 4
+// ARM-LE-NEXT:    [[TMP0:%.*]] = getelementptr i8, ptr [[NTC]], i32 5
+// ARM-LE-NEXT:    store i8 0, ptr [[TMP0]], align 1
+// ARM-LE-NEXT:    [[TMP1:%.*]] = getelementptr i8, ptr [[NTC]], i32 6
+// ARM-LE-NEXT:    store i8 0, ptr [[TMP1]], align 2
+// ARM-LE-NEXT:    [[TMP2:%.*]] = getelementptr i8, ptr [[NTC]], i32 7
+// ARM-LE-NEXT:    store i8 0, ptr [[TMP2]], align 1
+// ARM-LE-NEXT:    ret void
+//
+// ARM-BE-LABEL: define dso_local void @testNonTriviallyCopyable(
+// ARM-BE-SAME: ptr nofreeobj noundef align 4 dereferenceable(8) [[NTC:%.*]]) #[[ATTR0]] {
+// ARM-BE-NEXT:  [[ENTRY:.*:]]
+// ARM-BE-NEXT:    [[NTC_INDIRECT_ADDR:%.*]] = alloca ptr, align 4
+// ARM-BE-NEXT:    store ptr [[NTC]], ptr [[NTC_INDIRECT_ADDR]], align 4
+// ARM-BE-NEXT:    [[TMP0:%.*]] = getelementptr i8, ptr [[NTC]], i32 5
+// ARM-BE-NEXT:    store i8 0, ptr [[TMP0]], align 1
+// ARM-BE-NEXT:    [[TMP1:%.*]] = getelementptr i8, ptr [[NTC]], i32 6
+// ARM-BE-NEXT:    store i8 0, ptr [[TMP1]], align 2
+// ARM-BE-NEXT:    [[TMP2:%.*]] = getelementptr i8, ptr [[NTC]], i32 7
+// ARM-BE-NEXT:    store i8 0, ptr [[TMP2]], align 1
+// ARM-BE-NEXT:    ret void
+//
+// AARCH64-LE-LABEL: define dso_local void @testNonTriviallyCopyable(
+// AARCH64-LE-SAME: ptr nofreeobj noundef align 4 dereferenceable(8) [[NTC:%.*]]) #[[ATTR0]] {
+// AARCH64-LE-NEXT:  [[ENTRY:.*:]]
+// AARCH64-LE-NEXT:    [[NTC_INDIRECT_ADDR:%.*]] = alloca ptr, align 8
+// AARCH64-LE-NEXT:    store ptr [[NTC]], ptr [[NTC_INDIRECT_ADDR]], align 8
+// AARCH64-LE-NEXT:    [[TMP0:%.*]] = getelementptr i8, ptr [[NTC]], i32 5
+// AARCH64-LE-NEXT:    store i8 0, ptr [[TMP0]], align 1
+// AARCH64-LE-NEXT:    [[TMP1:%.*]] = getelementptr i8, ptr [[NTC]], i32 6
+// AARCH64-LE-NEXT:    store i8 0, ptr [[TMP1]], align 2
+// AARCH64-LE-NEXT:    [[TMP2:%.*]] = getelementptr i8, ptr [[NTC]], i32 7
+// AARCH64-LE-NEXT:    store i8 0, ptr [[TMP2]], align 1
+// AARCH64-LE-NEXT:    ret void
+//
+// AARCH64-BE-LABEL: define dso_local void @testNonTriviallyCopyable(
+// AARCH64-BE-SAME: ptr nofreeobj noundef align 4 dereferenceable(8) [[NTC:%.*]]) #[[ATTR0]] {
+// AARCH64-BE-NEXT:  [[ENTRY:.*:]]
+// AARCH64-BE-NEXT:    [[NTC_INDIRECT_ADDR:%.*]] = alloca ptr, align 8
+// AARCH64-BE-NEXT:    store ptr [[NTC]], ptr [[NTC_INDIRECT_ADDR]], align 8
+// AARCH64-BE-NEXT:    [[TMP0:%.*]] = getelementptr i8, ptr [[NTC]], i32 5
+// AARCH64-BE-NEXT:    store i8 0, ptr [[TMP0]], align 1
+// AARCH64-BE-NEXT:    [[TMP1:%.*]] = getelementptr i8, ptr [[NTC]], i32 6
+// AARCH64-BE-NEXT:    store i8 0, ptr [[TMP1]], align 2
+// AARCH64-BE-NEXT:    [[TMP2:%.*]] = getelementptr i8, ptr [[NTC]], i32 7
+// AARCH64-BE-NEXT:    store i8 0, ptr [[TMP2]], align 1
+// AARCH64-BE-NEXT:    ret void
+//
+extern "C" void testNonTriviallyCopyable(NonTriviallyCopyable ntc) {
   __builtin_clear_padding(&ntc);
 }
 
+
 typedef float Float3Vec __attribute__((ext_vector_type(3)));
 typedef long double LongDouble3Vec __attribute__((ext_vector_type(3)));
 
-// LINUX-LABEL: define dso_local void @_Z18testAttributedTypePDv3_f(
+// LINUX-LABEL: define dso_local void @testAttributedType(
 // LINUX-SAME: ptr noundef [[V:%.*]]) #[[ATTR0]] {
 // LINUX-NEXT:  [[ENTRY:.*:]]
 // LINUX-NEXT:    [[V_ADDR:%.*]] = alloca ptr, align 8
@@ -1454,26 +4125,92 @@ typedef long double LongDouble3Vec __attribute__((ext_vector_type(3)));
 // LINUX-NEXT:    store i8 0, ptr [[TMP4]], align 1
 // LINUX-NEXT:    ret void
 //
-// WINDOWS-LABEL: define dso_local void @"?testAttributedType@@YAXPEAT?$__vector at M$02 at __clang@@@Z"(ptr noundef %v) #0 {
-// WINDOWS-NEXT: entry:
-// WINDOWS-NEXT:   %v.addr = alloca ptr, align 8
-// WINDOWS-NEXT:   store ptr %v, ptr %v.addr, align 8
-// WINDOWS-NEXT:   %0 = load ptr, ptr %v.addr, align 8
-// WINDOWS-NEXT:   %1 = getelementptr i8, ptr %0, i32 12
-// WINDOWS-NEXT:   store i8 0, ptr %1, align 4
-// WINDOWS-NEXT:   %2 = getelementptr i8, ptr %0, i32 13
-// WINDOWS-NEXT:   store i8 0, ptr %2, align 1
-// WINDOWS-NEXT:   %3 = getelementptr i8, ptr %0, i32 14
-// WINDOWS-NEXT:   store i8 0, ptr %3, align 2
-// WINDOWS-NEXT:   %4 = getelementptr i8, ptr %0, i32 15
-// WINDOWS-NEXT:   store i8 0, ptr %4, align 1
-// WINDOWS-NEXT:   ret void
-//
-void testAttributedType(Float3Vec* v) {
+// WINDOWS-LABEL: define dso_local void @testAttributedType(
+// WINDOWS-SAME: ptr noundef [[V:%.*]]) #[[ATTR0]] {
+// WINDOWS-NEXT:  [[ENTRY:.*:]]
+// WINDOWS-NEXT:    [[V_ADDR:%.*]] = alloca ptr, align 8
+// WINDOWS-NEXT:    store ptr [[V]], ptr [[V_ADDR]], align 8
+// WINDOWS-NEXT:    [[TMP0:%.*]] = load ptr, ptr [[V_ADDR]], align 8
+// WINDOWS-NEXT:    [[TMP1:%.*]] = getelementptr i8, ptr [[TMP0]], i32 12
+// WINDOWS-NEXT:    store i8 0, ptr [[TMP1]], align 4
+// WINDOWS-NEXT:    [[TMP2:%.*]] = getelementptr i8, ptr [[TMP0]], i32 13
+// WINDOWS-NEXT:    store i8 0, ptr [[TMP2]], align 1
+// WINDOWS-NEXT:    [[TMP3:%.*]] = getelementptr i8, ptr [[TMP0]], i32 14
+// WINDOWS-NEXT:    store i8 0, ptr [[TMP3]], align 2
+// WINDOWS-NEXT:    [[TMP4:%.*]] = getelementptr i8, ptr [[TMP0]], i32 15
+// WINDOWS-NEXT:    store i8 0, ptr [[TMP4]], align 1
+// WINDOWS-NEXT:    ret void
+//
+// ARM-LE-LABEL: define dso_local void @testAttributedType(
+// ARM-LE-SAME: ptr noundef [[V:%.*]]) #[[ATTR0]] {
+// ARM-LE-NEXT:  [[ENTRY:.*:]]
+// ARM-LE-NEXT:    [[V_ADDR:%.*]] = alloca ptr, align 4
+// ARM-LE-NEXT:    store ptr [[V]], ptr [[V_ADDR]], align 4
+// ARM-LE-NEXT:    [[TMP0:%.*]] = load ptr, ptr [[V_ADDR]], align 4
+// ARM-LE-NEXT:    [[TMP1:%.*]] = getelementptr i8, ptr [[TMP0]], i32 12
+// ARM-LE-NEXT:    store i8 0, ptr [[TMP1]], align 4
+// ARM-LE-NEXT:    [[TMP2:%.*]] = getelementptr i8, ptr [[TMP0]], i32 13
+// ARM-LE-NEXT:    store i8 0, ptr [[TMP2]], align 1
+// ARM-LE-NEXT:    [[TMP3:%.*]] = getelementptr i8, ptr [[TMP0]], i32 14
+// ARM-LE-NEXT:    store i8 0, ptr [[TMP3]], align 2
+// ARM-LE-NEXT:    [[TMP4:%.*]] = getelementptr i8, ptr [[TMP0]], i32 15
+// ARM-LE-NEXT:    store i8 0, ptr [[TMP4]], align 1
+// ARM-LE-NEXT:    ret void
+//
+// ARM-BE-LABEL: define dso_local void @testAttributedType(
+// ARM-BE-SAME: ptr noundef [[V:%.*]]) #[[ATTR0]] {
+// ARM-BE-NEXT:  [[ENTRY:.*:]]
+// ARM-BE-NEXT:    [[V_ADDR:%.*]] = alloca ptr, align 4
+// ARM-BE-NEXT:    store ptr [[V]], ptr [[V_ADDR]], align 4
+// ARM-BE-NEXT:    [[TMP0:%.*]] = load ptr, ptr [[V_ADDR]], align 4
+// ARM-BE-NEXT:    [[TMP1:%.*]] = getelementptr i8, ptr [[TMP0]], i32 12
+// ARM-BE-NEXT:    store i8 0, ptr [[TMP1]], align 4
+// ARM-BE-NEXT:    [[TMP2:%.*]] = getelementptr i8, ptr [[TMP0]], i32 13
+// ARM-BE-NEXT:    store i8 0, ptr [[TMP2]], align 1
+// ARM-BE-NEXT:    [[TMP3:%.*]] = getelementptr i8, ptr [[TMP0]], i32 14
+// ARM-BE-NEXT:    store i8 0, ptr [[TMP3]], align 2
+// ARM-BE-NEXT:    [[TMP4:%.*]] = getelementptr i8, ptr [[TMP0]], i32 15
+// ARM-BE-NEXT:    store i8 0, ptr [[TMP4]], align 1
+// ARM-BE-NEXT:    ret void
+//
+// AARCH64-LE-LABEL: define dso_local void @testAttributedType(
+// AARCH64-LE-SAME: ptr noundef [[V:%.*]]) #[[ATTR0]] {
+// AARCH64-LE-NEXT:  [[ENTRY:.*:]]
+// AARCH64-LE-NEXT:    [[V_ADDR:%.*]] = alloca ptr, align 8
+// AARCH64-LE-NEXT:    store ptr [[V]], ptr [[V_ADDR]], align 8
+// AARCH64-LE-NEXT:    [[TMP0:%.*]] = load ptr, ptr [[V_ADDR]], align 8
+// AARCH64-LE-NEXT:    [[TMP1:%.*]] = getelementptr i8, ptr [[TMP0]], i32 12
+// AARCH64-LE-NEXT:    store i8 0, ptr [[TMP1]], align 4
+// AARCH64-LE-NEXT:    [[TMP2:%.*]] = getelementptr i8, ptr [[TMP0]], i32 13
+// AARCH64-LE-NEXT:    store i8 0, ptr [[TMP2]], align 1
+// AARCH64-LE-NEXT:    [[TMP3:%.*]] = getelementptr i8, ptr [[TMP0]], i32 14
+// AARCH64-LE-NEXT:    store i8 0, ptr [[TMP3]], align 2
+// AARCH64-LE-NEXT:    [[TMP4:%.*]] = getelementptr i8, ptr [[TMP0]], i32 15
+// AARCH64-LE-NEXT:    store i8 0, ptr [[TMP4]], align 1
+// AARCH64-LE-NEXT:    ret void
+//
+// AARCH64-BE-LABEL: define dso_local void @testAttributedType(
+// AARCH64-BE-SAME: ptr noundef [[V:%.*]]) #[[ATTR0]] {
+// AARCH64-BE-NEXT:  [[ENTRY:.*:]]
+// AARCH64-BE-NEXT:    [[V_ADDR:%.*]] = alloca ptr, align 8
+// AARCH64-BE-NEXT:    store ptr [[V]], ptr [[V_ADDR]], align 8
+// AARCH64-BE-NEXT:    [[TMP0:%.*]] = load ptr, ptr [[V_ADDR]], align 8
+// AARCH64-BE-NEXT:    [[TMP1:%.*]] = getelementptr i8, ptr [[TMP0]], i32 12
+// AARCH64-BE-NEXT:    store i8 0, ptr [[TMP1]], align 4
+// AARCH64-BE-NEXT:    [[TMP2:%.*]] = getelementptr i8, ptr [[TMP0]], i32 13
+// AARCH64-BE-NEXT:    store i8 0, ptr [[TMP2]], align 1
+// AARCH64-BE-NEXT:    [[TMP3:%.*]] = getelementptr i8, ptr [[TMP0]], i32 14
+// AARCH64-BE-NEXT:    store i8 0, ptr [[TMP3]], align 2
+// AARCH64-BE-NEXT:    [[TMP4:%.*]] = getelementptr i8, ptr [[TMP0]], i32 15
+// AARCH64-BE-NEXT:    store i8 0, ptr [[TMP4]], align 1
+// AARCH64-BE-NEXT:    ret void
+//
+extern "C" void testAttributedType(Float3Vec* v) {
   __builtin_clear_padding(v);
 }
 
-// LINUX-LABEL: define dso_local void @_Z28testAttributedLongDoubleTypePDv3_e(
+
+// LINUX-LABEL: define dso_local void @testAttributedLongDoubleType(
 // LINUX-SAME: ptr noundef [[V:%.*]]) #[[ATTR0]] {
 // LINUX-NEXT:  [[ENTRY:.*:]]
 // LINUX-NEXT:    [[V_ADDR:%.*]] = alloca ptr, align 8
@@ -1549,41 +4286,171 @@ void testAttributedType(Float3Vec* v) {
 // LINUX-NEXT:    store i8 0, ptr [[TMP34]], align 1
 // LINUX-NEXT:    ret void
 //
-// WINDOWS-LABEL: define dso_local void @"?testAttributedLongDoubleType@@YAXPEAT?$__vector at O$02 at __clang@@@Z"(ptr noundef %v) #0 {
-// WINDOWS-NEXT: entry:
-// WINDOWS-NEXT:   %v.addr = alloca ptr, align 8
-// WINDOWS-NEXT:   store ptr %v, ptr %v.addr, align 8
-// WINDOWS-NEXT:   %0 = load ptr, ptr %v.addr, align 8
-// WINDOWS-NEXT:   %1 = getelementptr i8, ptr %0, i32 24
-// WINDOWS-NEXT:   store i8 0, ptr %1, align 8
-// WINDOWS-NEXT:   %2 = getelementptr i8, ptr %0, i32 25
-// WINDOWS-NEXT:   store i8 0, ptr %2, align 1
-// WINDOWS-NEXT:   %3 = getelementptr i8, ptr %0, i32 26
-// WINDOWS-NEXT:   store i8 0, ptr %3, align 2
-// WINDOWS-NEXT:   %4 = getelementptr i8, ptr %0, i32 27
-// WINDOWS-NEXT:   store i8 0, ptr %4, align 1
-// WINDOWS-NEXT:   %5 = getelementptr i8, ptr %0, i32 28
-// WINDOWS-NEXT:   store i8 0, ptr %5, align 4
-// WINDOWS-NEXT:   %6 = getelementptr i8, ptr %0, i32 29
-// WINDOWS-NEXT:   store i8 0, ptr %6, align 1
-// WINDOWS-NEXT:   %7 = getelementptr i8, ptr %0, i32 30
-// WINDOWS-NEXT:   store i8 0, ptr %7, align 2
-// WINDOWS-NEXT:   %8 = getelementptr i8, ptr %0, i32 31
-// WINDOWS-NEXT:   store i8 0, ptr %8, align 1
-// WINDOWS-NEXT:   ret void
-//
-void testAttributedLongDoubleType(LongDouble3Vec *v) {
+// WINDOWS-LABEL: define dso_local void @testAttributedLongDoubleType(
+// WINDOWS-SAME: ptr noundef [[V:%.*]]) #[[ATTR0]] {
+// WINDOWS-NEXT:  [[ENTRY:.*:]]
+// WINDOWS-NEXT:    [[V_ADDR:%.*]] = alloca ptr, align 8
+// WINDOWS-NEXT:    store ptr [[V]], ptr [[V_ADDR]], align 8
+// WINDOWS-NEXT:    [[TMP0:%.*]] = load ptr, ptr [[V_ADDR]], align 8
+// WINDOWS-NEXT:    [[TMP1:%.*]] = getelementptr i8, ptr [[TMP0]], i32 24
+// WINDOWS-NEXT:    store i8 0, ptr [[TMP1]], align 8
+// WINDOWS-NEXT:    [[TMP2:%.*]] = getelementptr i8, ptr [[TMP0]], i32 25
+// WINDOWS-NEXT:    store i8 0, ptr [[TMP2]], align 1
+// WINDOWS-NEXT:    [[TMP3:%.*]] = getelementptr i8, ptr [[TMP0]], i32 26
+// WINDOWS-NEXT:    store i8 0, ptr [[TMP3]], align 2
+// WINDOWS-NEXT:    [[TMP4:%.*]] = getelementptr i8, ptr [[TMP0]], i32 27
+// WINDOWS-NEXT:    store i8 0, ptr [[TMP4]], align 1
+// WINDOWS-NEXT:    [[TMP5:%.*]] = getelementptr i8, ptr [[TMP0]], i32 28
+// WINDOWS-NEXT:    store i8 0, ptr [[TMP5]], align 4
+// WINDOWS-NEXT:    [[TMP6:%.*]] = getelementptr i8, ptr [[TMP0]], i32 29
+// WINDOWS-NEXT:    store i8 0, ptr [[TMP6]], align 1
+// WINDOWS-NEXT:    [[TMP7:%.*]] = getelementptr i8, ptr [[TMP0]], i32 30
+// WINDOWS-NEXT:    store i8 0, ptr [[TMP7]], align 2
+// WINDOWS-NEXT:    [[TMP8:%.*]] = getelementptr i8, ptr [[TMP0]], i32 31
+// WINDOWS-NEXT:    store i8 0, ptr [[TMP8]], align 1
+// WINDOWS-NEXT:    ret void
+//
+// ARM-LE-LABEL: define dso_local void @testAttributedLongDoubleType(
+// ARM-LE-SAME: ptr noundef [[V:%.*]]) #[[ATTR0]] {
+// ARM-LE-NEXT:  [[ENTRY:.*:]]
+// ARM-LE-NEXT:    [[V_ADDR:%.*]] = alloca ptr, align 4
+// ARM-LE-NEXT:    store ptr [[V]], ptr [[V_ADDR]], align 4
+// ARM-LE-NEXT:    [[TMP0:%.*]] = load ptr, ptr [[V_ADDR]], align 4
+// ARM-LE-NEXT:    [[TMP1:%.*]] = getelementptr i8, ptr [[TMP0]], i32 24
+// ARM-LE-NEXT:    store i8 0, ptr [[TMP1]], align 8
+// ARM-LE-NEXT:    [[TMP2:%.*]] = getelementptr i8, ptr [[TMP0]], i32 25
+// ARM-LE-NEXT:    store i8 0, ptr [[TMP2]], align 1
+// ARM-LE-NEXT:    [[TMP3:%.*]] = getelementptr i8, ptr [[TMP0]], i32 26
+// ARM-LE-NEXT:    store i8 0, ptr [[TMP3]], align 2
+// ARM-LE-NEXT:    [[TMP4:%.*]] = getelementptr i8, ptr [[TMP0]], i32 27
+// ARM-LE-NEXT:    store i8 0, ptr [[TMP4]], align 1
+// ARM-LE-NEXT:    [[TMP5:%.*]] = getelementptr i8, ptr [[TMP0]], i32 28
+// ARM-LE-NEXT:    store i8 0, ptr [[TMP5]], align 4
+// ARM-LE-NEXT:    [[TMP6:%.*]] = getelementptr i8, ptr [[TMP0]], i32 29
+// ARM-LE-NEXT:    store i8 0, ptr [[TMP6]], align 1
+// ARM-LE-NEXT:    [[TMP7:%.*]] = getelementptr i8, ptr [[TMP0]], i32 30
+// ARM-LE-NEXT:    store i8 0, ptr [[TMP7]], align 2
+// ARM-LE-NEXT:    [[TMP8:%.*]] = getelementptr i8, ptr [[TMP0]], i32 31
+// ARM-LE-NEXT:    store i8 0, ptr [[TMP8]], align 1
+// ARM-LE-NEXT:    ret void
+//
+// ARM-BE-LABEL: define dso_local void @testAttributedLongDoubleType(
+// ARM-BE-SAME: ptr noundef [[V:%.*]]) #[[ATTR0]] {
+// ARM-BE-NEXT:  [[ENTRY:.*:]]
+// ARM-BE-NEXT:    [[V_ADDR:%.*]] = alloca ptr, align 4
+// ARM-BE-NEXT:    store ptr [[V]], ptr [[V_ADDR]], align 4
+// ARM-BE-NEXT:    [[TMP0:%.*]] = load ptr, ptr [[V_ADDR]], align 4
+// ARM-BE-NEXT:    [[TMP1:%.*]] = getelementptr i8, ptr [[TMP0]], i32 24
+// ARM-BE-NEXT:    store i8 0, ptr [[TMP1]], align 8
+// ARM-BE-NEXT:    [[TMP2:%.*]] = getelementptr i8, ptr [[TMP0]], i32 25
+// ARM-BE-NEXT:    store i8 0, ptr [[TMP2]], align 1
+// ARM-BE-NEXT:    [[TMP3:%.*]] = getelementptr i8, ptr [[TMP0]], i32 26
+// ARM-BE-NEXT:    store i8 0, ptr [[TMP3]], align 2
+// ARM-BE-NEXT:    [[TMP4:%.*]] = getelementptr i8, ptr [[TMP0]], i32 27
+// ARM-BE-NEXT:    store i8 0, ptr [[TMP4]], align 1
+// ARM-BE-NEXT:    [[TMP5:%.*]] = getelementptr i8, ptr [[TMP0]], i32 28
+// ARM-BE-NEXT:    store i8 0, ptr [[TMP5]], align 4
+// ARM-BE-NEXT:    [[TMP6:%.*]] = getelementptr i8, ptr [[TMP0]], i32 29
+// ARM-BE-NEXT:    store i8 0, ptr [[TMP6]], align 1
+// ARM-BE-NEXT:    [[TMP7:%.*]] = getelementptr i8, ptr [[TMP0]], i32 30
+// ARM-BE-NEXT:    store i8 0, ptr [[TMP7]], align 2
+// ARM-BE-NEXT:    [[TMP8:%.*]] = getelementptr i8, ptr [[TMP0]], i32 31
+// ARM-BE-NEXT:    store i8 0, ptr [[TMP8]], align 1
+// ARM-BE-NEXT:    ret void
+//
+// AARCH64-LE-LABEL: define dso_local void @testAttributedLongDoubleType(
+// AARCH64-LE-SAME: ptr noundef [[V:%.*]]) #[[ATTR0]] {
+// AARCH64-LE-NEXT:  [[ENTRY:.*:]]
+// AARCH64-LE-NEXT:    [[V_ADDR:%.*]] = alloca ptr, align 8
+// AARCH64-LE-NEXT:    store ptr [[V]], ptr [[V_ADDR]], align 8
+// AARCH64-LE-NEXT:    [[TMP0:%.*]] = load ptr, ptr [[V_ADDR]], align 8
+// AARCH64-LE-NEXT:    [[TMP1:%.*]] = getelementptr i8, ptr [[TMP0]], i32 48
+// AARCH64-LE-NEXT:    store i8 0, ptr [[TMP1]], align 16
+// AARCH64-LE-NEXT:    [[TMP2:%.*]] = getelementptr i8, ptr [[TMP0]], i32 49
+// AARCH64-LE-NEXT:    store i8 0, ptr [[TMP2]], align 1
+// AARCH64-LE-NEXT:    [[TMP3:%.*]] = getelementptr i8, ptr [[TMP0]], i32 50
+// AARCH64-LE-NEXT:    store i8 0, ptr [[TMP3]], align 2
+// AARCH64-LE-NEXT:    [[TMP4:%.*]] = getelementptr i8, ptr [[TMP0]], i32 51
+// AARCH64-LE-NEXT:    store i8 0, ptr [[TMP4]], align 1
+// AARCH64-LE-NEXT:    [[TMP5:%.*]] = getelementptr i8, ptr [[TMP0]], i32 52
+// AARCH64-LE-NEXT:    store i8 0, ptr [[TMP5]], align 4
+// AARCH64-LE-NEXT:    [[TMP6:%.*]] = getelementptr i8, ptr [[TMP0]], i32 53
+// AARCH64-LE-NEXT:    store i8 0, ptr [[TMP6]], align 1
+// AARCH64-LE-NEXT:    [[TMP7:%.*]] = getelementptr i8, ptr [[TMP0]], i32 54
+// AARCH64-LE-NEXT:    store i8 0, ptr [[TMP7]], align 2
+// AARCH64-LE-NEXT:    [[TMP8:%.*]] = getelementptr i8, ptr [[TMP0]], i32 55
+// AARCH64-LE-NEXT:    store i8 0, ptr [[TMP8]], align 1
+// AARCH64-LE-NEXT:    [[TMP9:%.*]] = getelementptr i8, ptr [[TMP0]], i32 56
+// AARCH64-LE-NEXT:    store i8 0, ptr [[TMP9]], align 8
+// AARCH64-LE-NEXT:    [[TMP10:%.*]] = getelementptr i8, ptr [[TMP0]], i32 57
+// AARCH64-LE-NEXT:    store i8 0, ptr [[TMP10]], align 1
+// AARCH64-LE-NEXT:    [[TMP11:%.*]] = getelementptr i8, ptr [[TMP0]], i32 58
+// AARCH64-LE-NEXT:    store i8 0, ptr [[TMP11]], align 2
+// AARCH64-LE-NEXT:    [[TMP12:%.*]] = getelementptr i8, ptr [[TMP0]], i32 59
+// AARCH64-LE-NEXT:    store i8 0, ptr [[TMP12]], align 1
+// AARCH64-LE-NEXT:    [[TMP13:%.*]] = getelementptr i8, ptr [[TMP0]], i32 60
+// AARCH64-LE-NEXT:    store i8 0, ptr [[TMP13]], align 4
+// AARCH64-LE-NEXT:    [[TMP14:%.*]] = getelementptr i8, ptr [[TMP0]], i32 61
+// AARCH64-LE-NEXT:    store i8 0, ptr [[TMP14]], align 1
+// AARCH64-LE-NEXT:    [[TMP15:%.*]] = getelementptr i8, ptr [[TMP0]], i32 62
+// AARCH64-LE-NEXT:    store i8 0, ptr [[TMP15]], align 2
+// AARCH64-LE-NEXT:    [[TMP16:%.*]] = getelementptr i8, ptr [[TMP0]], i32 63
+// AARCH64-LE-NEXT:    store i8 0, ptr [[TMP16]], align 1
+// AARCH64-LE-NEXT:    ret void
+//
+// AARCH64-BE-LABEL: define dso_local void @testAttributedLongDoubleType(
+// AARCH64-BE-SAME: ptr noundef [[V:%.*]]) #[[ATTR0]] {
+// AARCH64-BE-NEXT:  [[ENTRY:.*:]]
+// AARCH64-BE-NEXT:    [[V_ADDR:%.*]] = alloca ptr, align 8
+// AARCH64-BE-NEXT:    store ptr [[V]], ptr [[V_ADDR]], align 8
+// AARCH64-BE-NEXT:    [[TMP0:%.*]] = load ptr, ptr [[V_ADDR]], align 8
+// AARCH64-BE-NEXT:    [[TMP1:%.*]] = getelementptr i8, ptr [[TMP0]], i32 48
+// AARCH64-BE-NEXT:    store i8 0, ptr [[TMP1]], align 16
+// AARCH64-BE-NEXT:    [[TMP2:%.*]] = getelementptr i8, ptr [[TMP0]], i32 49
+// AARCH64-BE-NEXT:    store i8 0, ptr [[TMP2]], align 1
+// AARCH64-BE-NEXT:    [[TMP3:%.*]] = getelementptr i8, ptr [[TMP0]], i32 50
+// AARCH64-BE-NEXT:    store i8 0, ptr [[TMP3]], align 2
+// AARCH64-BE-NEXT:    [[TMP4:%.*]] = getelementptr i8, ptr [[TMP0]], i32 51
+// AARCH64-BE-NEXT:    store i8 0, ptr [[TMP4]], align 1
+// AARCH64-BE-NEXT:    [[TMP5:%.*]] = getelementptr i8, ptr [[TMP0]], i32 52
+// AARCH64-BE-NEXT:    store i8 0, ptr [[TMP5]], align 4
+// AARCH64-BE-NEXT:    [[TMP6:%.*]] = getelementptr i8, ptr [[TMP0]], i32 53
+// AARCH64-BE-NEXT:    store i8 0, ptr [[TMP6]], align 1
+// AARCH64-BE-NEXT:    [[TMP7:%.*]] = getelementptr i8, ptr [[TMP0]], i32 54
+// AARCH64-BE-NEXT:    store i8 0, ptr [[TMP7]], align 2
+// AARCH64-BE-NEXT:    [[TMP8:%.*]] = getelementptr i8, ptr [[TMP0]], i32 55
+// AARCH64-BE-NEXT:    store i8 0, ptr [[TMP8]], align 1
+// AARCH64-BE-NEXT:    [[TMP9:%.*]] = getelementptr i8, ptr [[TMP0]], i32 56
+// AARCH64-BE-NEXT:    store i8 0, ptr [[TMP9]], align 8
+// AARCH64-BE-NEXT:    [[TMP10:%.*]] = getelementptr i8, ptr [[TMP0]], i32 57
+// AARCH64-BE-NEXT:    store i8 0, ptr [[TMP10]], align 1
+// AARCH64-BE-NEXT:    [[TMP11:%.*]] = getelementptr i8, ptr [[TMP0]], i32 58
+// AARCH64-BE-NEXT:    store i8 0, ptr [[TMP11]], align 2
+// AARCH64-BE-NEXT:    [[TMP12:%.*]] = getelementptr i8, ptr [[TMP0]], i32 59
+// AARCH64-BE-NEXT:    store i8 0, ptr [[TMP12]], align 1
+// AARCH64-BE-NEXT:    [[TMP13:%.*]] = getelementptr i8, ptr [[TMP0]], i32 60
+// AARCH64-BE-NEXT:    store i8 0, ptr [[TMP13]], align 4
+// AARCH64-BE-NEXT:    [[TMP14:%.*]] = getelementptr i8, ptr [[TMP0]], i32 61
+// AARCH64-BE-NEXT:    store i8 0, ptr [[TMP14]], align 1
+// AARCH64-BE-NEXT:    [[TMP15:%.*]] = getelementptr i8, ptr [[TMP0]], i32 62
+// AARCH64-BE-NEXT:    store i8 0, ptr [[TMP15]], align 2
+// AARCH64-BE-NEXT:    [[TMP16:%.*]] = getelementptr i8, ptr [[TMP0]], i32 63
+// AARCH64-BE-NEXT:    store i8 0, ptr [[TMP16]], align 1
+// AARCH64-BE-NEXT:    ret void
+//
+extern "C" void testAttributedLongDoubleType(LongDouble3Vec *v) {
   // long double elements occupy [0-9], [16-25], [32-41] on x86.
   __builtin_clear_padding(v);
 }
 
+
 struct UnnamedBitfieldSingleBit {
   unsigned char a : 3;
   unsigned char   : 1;
   unsigned char b : 4;
 };
 
-// LINUX-LABEL: define dso_local void @_Z28testUnnamedBitfieldSingleBitP24UnnamedBitfieldSingleBit(
+// LINUX-LABEL: define dso_local void @testUnnamedBitfieldSingleBit(
 // LINUX-SAME: ptr noundef [[S:%.*]]) #[[ATTR0]] {
 // LINUX-NEXT:  [[ENTRY:.*:]]
 // LINUX-NEXT:    [[S_ADDR:%.*]] = alloca ptr, align 8
@@ -1595,31 +4462,82 @@ struct UnnamedBitfieldSingleBit {
 // LINUX-NEXT:    store i8 [[TMP3]], ptr [[TMP1]], align 1
 // LINUX-NEXT:    ret void
 //
-// WINDOWS-LABEL: define dso_local void @"?testUnnamedBitfieldSingleBit@@YAXPEAUUnnamedBitfieldSingleBit@@@Z"(ptr noundef %s) #0 {
-// WINDOWS-NEXT: entry:
-// WINDOWS-NEXT:   %s.addr = alloca ptr, align 8
-// WINDOWS-NEXT:   store ptr %s, ptr %s.addr, align 8
-// WINDOWS-NEXT:   %0 = load ptr, ptr %s.addr, align 8
-// WINDOWS-NEXT:   %1 = getelementptr i8, ptr %0, i32 0
-// WINDOWS-NEXT:   %2 = load i8, ptr %1, align 1
-// WINDOWS-NEXT:   %3 = and i8 %2, -9
-// WINDOWS-NEXT:   store i8 %3, ptr %1, align 1
-// WINDOWS-NEXT:   ret void
-//
-void testUnnamedBitfieldSingleBit(struct UnnamedBitfieldSingleBit *s) {
-  // byte 0: a[bits 0-2], unnamed[bit 3], b[bits 4-7]
-  // bit 3 must be cleared
+// WINDOWS-LABEL: define dso_local void @testUnnamedBitfieldSingleBit(
+// WINDOWS-SAME: ptr noundef [[S:%.*]]) #[[ATTR0]] {
+// WINDOWS-NEXT:  [[ENTRY:.*:]]
+// WINDOWS-NEXT:    [[S_ADDR:%.*]] = alloca ptr, align 8
+// WINDOWS-NEXT:    store ptr [[S]], ptr [[S_ADDR]], align 8
+// WINDOWS-NEXT:    [[TMP0:%.*]] = load ptr, ptr [[S_ADDR]], align 8
+// WINDOWS-NEXT:    [[TMP1:%.*]] = getelementptr i8, ptr [[TMP0]], i32 0
+// WINDOWS-NEXT:    [[TMP2:%.*]] = load i8, ptr [[TMP1]], align 1
+// WINDOWS-NEXT:    [[TMP3:%.*]] = and i8 [[TMP2]], -9
+// WINDOWS-NEXT:    store i8 [[TMP3]], ptr [[TMP1]], align 1
+// WINDOWS-NEXT:    ret void
+//
+// ARM-LE-LABEL: define dso_local void @testUnnamedBitfieldSingleBit(
+// ARM-LE-SAME: ptr noundef [[S:%.*]]) #[[ATTR0]] {
+// ARM-LE-NEXT:  [[ENTRY:.*:]]
+// ARM-LE-NEXT:    [[S_ADDR:%.*]] = alloca ptr, align 4
+// ARM-LE-NEXT:    store ptr [[S]], ptr [[S_ADDR]], align 4
+// ARM-LE-NEXT:    [[TMP0:%.*]] = load ptr, ptr [[S_ADDR]], align 4
+// ARM-LE-NEXT:    [[TMP1:%.*]] = getelementptr i8, ptr [[TMP0]], i32 0
+// ARM-LE-NEXT:    [[TMP2:%.*]] = load i8, ptr [[TMP1]], align 1
+// ARM-LE-NEXT:    [[TMP3:%.*]] = and i8 [[TMP2]], -9
+// ARM-LE-NEXT:    store i8 [[TMP3]], ptr [[TMP1]], align 1
+// ARM-LE-NEXT:    ret void
+//
+// ARM-BE-LABEL: define dso_local void @testUnnamedBitfieldSingleBit(
+// ARM-BE-SAME: ptr noundef [[S:%.*]]) #[[ATTR0]] {
+// ARM-BE-NEXT:  [[ENTRY:.*:]]
+// ARM-BE-NEXT:    [[S_ADDR:%.*]] = alloca ptr, align 4
+// ARM-BE-NEXT:    store ptr [[S]], ptr [[S_ADDR]], align 4
+// ARM-BE-NEXT:    [[TMP0:%.*]] = load ptr, ptr [[S_ADDR]], align 4
+// ARM-BE-NEXT:    [[TMP1:%.*]] = getelementptr i8, ptr [[TMP0]], i32 0
+// ARM-BE-NEXT:    [[TMP2:%.*]] = load i8, ptr [[TMP1]], align 1
+// ARM-BE-NEXT:    [[TMP3:%.*]] = and i8 [[TMP2]], -17
+// ARM-BE-NEXT:    store i8 [[TMP3]], ptr [[TMP1]], align 1
+// ARM-BE-NEXT:    ret void
+//
+// AARCH64-LE-LABEL: define dso_local void @testUnnamedBitfieldSingleBit(
+// AARCH64-LE-SAME: ptr noundef [[S:%.*]]) #[[ATTR0]] {
+// AARCH64-LE-NEXT:  [[ENTRY:.*:]]
+// AARCH64-LE-NEXT:    [[S_ADDR:%.*]] = alloca ptr, align 8
+// AARCH64-LE-NEXT:    store ptr [[S]], ptr [[S_ADDR]], align 8
+// AARCH64-LE-NEXT:    [[TMP0:%.*]] = load ptr, ptr [[S_ADDR]], align 8
+// AARCH64-LE-NEXT:    [[TMP1:%.*]] = getelementptr i8, ptr [[TMP0]], i32 0
+// AARCH64-LE-NEXT:    [[TMP2:%.*]] = load i8, ptr [[TMP1]], align 1
+// AARCH64-LE-NEXT:    [[TMP3:%.*]] = and i8 [[TMP2]], -9
+// AARCH64-LE-NEXT:    store i8 [[TMP3]], ptr [[TMP1]], align 1
+// AARCH64-LE-NEXT:    ret void
+//
+// AARCH64-BE-LABEL: define dso_local void @testUnnamedBitfieldSingleBit(
+// AARCH64-BE-SAME: ptr noundef [[S:%.*]]) #[[ATTR0]] {
+// AARCH64-BE-NEXT:  [[ENTRY:.*:]]
+// AARCH64-BE-NEXT:    [[S_ADDR:%.*]] = alloca ptr, align 8
+// AARCH64-BE-NEXT:    store ptr [[S]], ptr [[S_ADDR]], align 8
+// AARCH64-BE-NEXT:    [[TMP0:%.*]] = load ptr, ptr [[S_ADDR]], align 8
+// AARCH64-BE-NEXT:    [[TMP1:%.*]] = getelementptr i8, ptr [[TMP0]], i32 0
+// AARCH64-BE-NEXT:    [[TMP2:%.*]] = load i8, ptr [[TMP1]], align 1
+// AARCH64-BE-NEXT:    [[TMP3:%.*]] = and i8 [[TMP2]], -17
+// AARCH64-BE-NEXT:    store i8 [[TMP3]], ptr [[TMP1]], align 1
+// AARCH64-BE-NEXT:    ret void
+//
+extern "C" void testUnnamedBitfieldSingleBit(
+    struct UnnamedBitfieldSingleBit *s) {
+  // Little endian: a[bits 0-2], unnamed[bit 3], b[bits 4-7].
+  // Clear bit 3 on little endian, or bit 4 on big endian.
   // Mask: 0b11110111 == -9
   __builtin_clear_padding(s);
 }
 
+
 struct UnnamedBitfieldMiddle {
   unsigned char a : 3;
   unsigned char   : 2;
   unsigned char b : 3;
 };
 
-// LINUX-LABEL: define dso_local void @_Z25testUnnamedBitfieldMiddleP21UnnamedBitfieldMiddle(
+// LINUX-LABEL: define dso_local void @testUnnamedBitfieldMiddle(
 // LINUX-SAME: ptr noundef [[S:%.*]]) #[[ATTR0]] {
 // LINUX-NEXT:  [[ENTRY:.*:]]
 // LINUX-NEXT:    [[S_ADDR:%.*]] = alloca ptr, align 8
@@ -1631,31 +4549,81 @@ struct UnnamedBitfieldMiddle {
 // LINUX-NEXT:    store i8 [[TMP3]], ptr [[TMP1]], align 1
 // LINUX-NEXT:    ret void
 //
-// WINDOWS-LABEL: define dso_local void @"?testUnnamedBitfieldMiddle@@YAXPEAUUnnamedBitfieldMiddle@@@Z"(ptr noundef %s) #0 {
-// WINDOWS-NEXT: entry:
-// WINDOWS-NEXT:   %s.addr = alloca ptr, align 8
-// WINDOWS-NEXT:   store ptr %s, ptr %s.addr, align 8
-// WINDOWS-NEXT:   %0 = load ptr, ptr %s.addr, align 8
-// WINDOWS-NEXT:   %1 = getelementptr i8, ptr %0, i32 0
-// WINDOWS-NEXT:   %2 = load i8, ptr %1, align 1
-// WINDOWS-NEXT:   %3 = and i8 %2, -25
-// WINDOWS-NEXT:   store i8 %3, ptr %1, align 1
-// WINDOWS-NEXT:   ret void
-//
-void testUnnamedBitfieldMiddle(struct UnnamedBitfieldMiddle *s) {
+// WINDOWS-LABEL: define dso_local void @testUnnamedBitfieldMiddle(
+// WINDOWS-SAME: ptr noundef [[S:%.*]]) #[[ATTR0]] {
+// WINDOWS-NEXT:  [[ENTRY:.*:]]
+// WINDOWS-NEXT:    [[S_ADDR:%.*]] = alloca ptr, align 8
+// WINDOWS-NEXT:    store ptr [[S]], ptr [[S_ADDR]], align 8
+// WINDOWS-NEXT:    [[TMP0:%.*]] = load ptr, ptr [[S_ADDR]], align 8
+// WINDOWS-NEXT:    [[TMP1:%.*]] = getelementptr i8, ptr [[TMP0]], i32 0
+// WINDOWS-NEXT:    [[TMP2:%.*]] = load i8, ptr [[TMP1]], align 1
+// WINDOWS-NEXT:    [[TMP3:%.*]] = and i8 [[TMP2]], -25
+// WINDOWS-NEXT:    store i8 [[TMP3]], ptr [[TMP1]], align 1
+// WINDOWS-NEXT:    ret void
+//
+// ARM-LE-LABEL: define dso_local void @testUnnamedBitfieldMiddle(
+// ARM-LE-SAME: ptr noundef [[S:%.*]]) #[[ATTR0]] {
+// ARM-LE-NEXT:  [[ENTRY:.*:]]
+// ARM-LE-NEXT:    [[S_ADDR:%.*]] = alloca ptr, align 4
+// ARM-LE-NEXT:    store ptr [[S]], ptr [[S_ADDR]], align 4
+// ARM-LE-NEXT:    [[TMP0:%.*]] = load ptr, ptr [[S_ADDR]], align 4
+// ARM-LE-NEXT:    [[TMP1:%.*]] = getelementptr i8, ptr [[TMP0]], i32 0
+// ARM-LE-NEXT:    [[TMP2:%.*]] = load i8, ptr [[TMP1]], align 1
+// ARM-LE-NEXT:    [[TMP3:%.*]] = and i8 [[TMP2]], -25
+// ARM-LE-NEXT:    store i8 [[TMP3]], ptr [[TMP1]], align 1
+// ARM-LE-NEXT:    ret void
+//
+// ARM-BE-LABEL: define dso_local void @testUnnamedBitfieldMiddle(
+// ARM-BE-SAME: ptr noundef [[S:%.*]]) #[[ATTR0]] {
+// ARM-BE-NEXT:  [[ENTRY:.*:]]
+// ARM-BE-NEXT:    [[S_ADDR:%.*]] = alloca ptr, align 4
+// ARM-BE-NEXT:    store ptr [[S]], ptr [[S_ADDR]], align 4
+// ARM-BE-NEXT:    [[TMP0:%.*]] = load ptr, ptr [[S_ADDR]], align 4
+// ARM-BE-NEXT:    [[TMP1:%.*]] = getelementptr i8, ptr [[TMP0]], i32 0
+// ARM-BE-NEXT:    [[TMP2:%.*]] = load i8, ptr [[TMP1]], align 1
+// ARM-BE-NEXT:    [[TMP3:%.*]] = and i8 [[TMP2]], -25
+// ARM-BE-NEXT:    store i8 [[TMP3]], ptr [[TMP1]], align 1
+// ARM-BE-NEXT:    ret void
+//
+// AARCH64-LE-LABEL: define dso_local void @testUnnamedBitfieldMiddle(
+// AARCH64-LE-SAME: ptr noundef [[S:%.*]]) #[[ATTR0]] {
+// AARCH64-LE-NEXT:  [[ENTRY:.*:]]
+// AARCH64-LE-NEXT:    [[S_ADDR:%.*]] = alloca ptr, align 8
+// AARCH64-LE-NEXT:    store ptr [[S]], ptr [[S_ADDR]], align 8
+// AARCH64-LE-NEXT:    [[TMP0:%.*]] = load ptr, ptr [[S_ADDR]], align 8
+// AARCH64-LE-NEXT:    [[TMP1:%.*]] = getelementptr i8, ptr [[TMP0]], i32 0
+// AARCH64-LE-NEXT:    [[TMP2:%.*]] = load i8, ptr [[TMP1]], align 1
+// AARCH64-LE-NEXT:    [[TMP3:%.*]] = and i8 [[TMP2]], -25
+// AARCH64-LE-NEXT:    store i8 [[TMP3]], ptr [[TMP1]], align 1
+// AARCH64-LE-NEXT:    ret void
+//
+// AARCH64-BE-LABEL: define dso_local void @testUnnamedBitfieldMiddle(
+// AARCH64-BE-SAME: ptr noundef [[S:%.*]]) #[[ATTR0]] {
+// AARCH64-BE-NEXT:  [[ENTRY:.*:]]
+// AARCH64-BE-NEXT:    [[S_ADDR:%.*]] = alloca ptr, align 8
+// AARCH64-BE-NEXT:    store ptr [[S]], ptr [[S_ADDR]], align 8
+// AARCH64-BE-NEXT:    [[TMP0:%.*]] = load ptr, ptr [[S_ADDR]], align 8
+// AARCH64-BE-NEXT:    [[TMP1:%.*]] = getelementptr i8, ptr [[TMP0]], i32 0
+// AARCH64-BE-NEXT:    [[TMP2:%.*]] = load i8, ptr [[TMP1]], align 1
+// AARCH64-BE-NEXT:    [[TMP3:%.*]] = and i8 [[TMP2]], -25
+// AARCH64-BE-NEXT:    store i8 [[TMP3]], ptr [[TMP1]], align 1
+// AARCH64-BE-NEXT:    ret void
+//
+extern "C" void testUnnamedBitfieldMiddle(struct UnnamedBitfieldMiddle *s) {
   // byte 0: a[0-2], unnamed[3-4], b[5-7]
   // bits 3-4 must be cleared
   // Mask: 0b11100111 == -25
   __builtin_clear_padding(s);
 }
 
+
 struct UnnamedBitfieldSurrounding {
   unsigned char   : 2;
   unsigned char a : 4;
   unsigned char   : 2;
 };
 
-// LINUX-LABEL: define dso_local void @_Z30testUnnamedBitfieldSurroundingP26UnnamedBitfieldSurrounding(
+// LINUX-LABEL: define dso_local void @testUnnamedBitfieldSurrounding(
 // LINUX-SAME: ptr noundef [[S:%.*]]) #[[ATTR0]] {
 // LINUX-NEXT:  [[ENTRY:.*:]]
 // LINUX-NEXT:    [[S_ADDR:%.*]] = alloca ptr, align 8
@@ -1671,22 +4639,88 @@ struct UnnamedBitfieldSurrounding {
 // LINUX-NEXT:    store i8 [[TMP6]], ptr [[TMP4]], align 1
 // LINUX-NEXT:    ret void
 //
-// WINDOWS-LABEL: define dso_local void @"?testUnnamedBitfieldSurrounding@@YAXPEAUUnnamedBitfieldSurrounding@@@Z"(ptr noundef %s) #0 {
-// WINDOWS-NEXT: entry:
-// WINDOWS-NEXT:   %s.addr = alloca ptr, align 8
-// WINDOWS-NEXT:   store ptr %s, ptr %s.addr, align 8
-// WINDOWS-NEXT:   %0 = load ptr, ptr %s.addr, align 8
-// WINDOWS-NEXT:   %1 = getelementptr i8, ptr %0, i32 0
-// WINDOWS-NEXT:   %2 = load i8, ptr %1, align 1
-// WINDOWS-NEXT:   %3 = and i8 %2, -4
-// WINDOWS-NEXT:   store i8 %3, ptr %1, align 1
-// WINDOWS-NEXT:   %4 = getelementptr i8, ptr %0, i32 0
-// WINDOWS-NEXT:   %5 = load i8, ptr %4, align 1
-// WINDOWS-NEXT:   %6 = and i8 %5, 63
-// WINDOWS-NEXT:   store i8 %6, ptr %4, align 1
-// WINDOWS-NEXT:   ret void
-//
-void testUnnamedBitfieldSurrounding(struct UnnamedBitfieldSurrounding *s) {
+// WINDOWS-LABEL: define dso_local void @testUnnamedBitfieldSurrounding(
+// WINDOWS-SAME: ptr noundef [[S:%.*]]) #[[ATTR0]] {
+// WINDOWS-NEXT:  [[ENTRY:.*:]]
+// WINDOWS-NEXT:    [[S_ADDR:%.*]] = alloca ptr, align 8
+// WINDOWS-NEXT:    store ptr [[S]], ptr [[S_ADDR]], align 8
+// WINDOWS-NEXT:    [[TMP0:%.*]] = load ptr, ptr [[S_ADDR]], align 8
+// WINDOWS-NEXT:    [[TMP1:%.*]] = getelementptr i8, ptr [[TMP0]], i32 0
+// WINDOWS-NEXT:    [[TMP2:%.*]] = load i8, ptr [[TMP1]], align 1
+// WINDOWS-NEXT:    [[TMP3:%.*]] = and i8 [[TMP2]], -4
+// WINDOWS-NEXT:    store i8 [[TMP3]], ptr [[TMP1]], align 1
+// WINDOWS-NEXT:    [[TMP4:%.*]] = getelementptr i8, ptr [[TMP0]], i32 0
+// WINDOWS-NEXT:    [[TMP5:%.*]] = load i8, ptr [[TMP4]], align 1
+// WINDOWS-NEXT:    [[TMP6:%.*]] = and i8 [[TMP5]], 63
+// WINDOWS-NEXT:    store i8 [[TMP6]], ptr [[TMP4]], align 1
+// WINDOWS-NEXT:    ret void
+//
+// ARM-LE-LABEL: define dso_local void @testUnnamedBitfieldSurrounding(
+// ARM-LE-SAME: ptr noundef [[S:%.*]]) #[[ATTR0]] {
+// ARM-LE-NEXT:  [[ENTRY:.*:]]
+// ARM-LE-NEXT:    [[S_ADDR:%.*]] = alloca ptr, align 4
+// ARM-LE-NEXT:    store ptr [[S]], ptr [[S_ADDR]], align 4
+// ARM-LE-NEXT:    [[TMP0:%.*]] = load ptr, ptr [[S_ADDR]], align 4
+// ARM-LE-NEXT:    [[TMP1:%.*]] = getelementptr i8, ptr [[TMP0]], i32 0
+// ARM-LE-NEXT:    [[TMP2:%.*]] = load i8, ptr [[TMP1]], align 1
+// ARM-LE-NEXT:    [[TMP3:%.*]] = and i8 [[TMP2]], -4
+// ARM-LE-NEXT:    store i8 [[TMP3]], ptr [[TMP1]], align 1
+// ARM-LE-NEXT:    [[TMP4:%.*]] = getelementptr i8, ptr [[TMP0]], i32 0
+// ARM-LE-NEXT:    [[TMP5:%.*]] = load i8, ptr [[TMP4]], align 1
+// ARM-LE-NEXT:    [[TMP6:%.*]] = and i8 [[TMP5]], 63
+// ARM-LE-NEXT:    store i8 [[TMP6]], ptr [[TMP4]], align 1
+// ARM-LE-NEXT:    ret void
+//
+// ARM-BE-LABEL: define dso_local void @testUnnamedBitfieldSurrounding(
+// ARM-BE-SAME: ptr noundef [[S:%.*]]) #[[ATTR0]] {
+// ARM-BE-NEXT:  [[ENTRY:.*:]]
+// ARM-BE-NEXT:    [[S_ADDR:%.*]] = alloca ptr, align 4
+// ARM-BE-NEXT:    store ptr [[S]], ptr [[S_ADDR]], align 4
+// ARM-BE-NEXT:    [[TMP0:%.*]] = load ptr, ptr [[S_ADDR]], align 4
+// ARM-BE-NEXT:    [[TMP1:%.*]] = getelementptr i8, ptr [[TMP0]], i32 0
+// ARM-BE-NEXT:    [[TMP2:%.*]] = load i8, ptr [[TMP1]], align 1
+// ARM-BE-NEXT:    [[TMP3:%.*]] = and i8 [[TMP2]], -4
+// ARM-BE-NEXT:    store i8 [[TMP3]], ptr [[TMP1]], align 1
+// ARM-BE-NEXT:    [[TMP4:%.*]] = getelementptr i8, ptr [[TMP0]], i32 0
+// ARM-BE-NEXT:    [[TMP5:%.*]] = load i8, ptr [[TMP4]], align 1
+// ARM-BE-NEXT:    [[TMP6:%.*]] = and i8 [[TMP5]], 63
+// ARM-BE-NEXT:    store i8 [[TMP6]], ptr [[TMP4]], align 1
+// ARM-BE-NEXT:    ret void
+//
+// AARCH64-LE-LABEL: define dso_local void @testUnnamedBitfieldSurrounding(
+// AARCH64-LE-SAME: ptr noundef [[S:%.*]]) #[[ATTR0]] {
+// AARCH64-LE-NEXT:  [[ENTRY:.*:]]
+// AARCH64-LE-NEXT:    [[S_ADDR:%.*]] = alloca ptr, align 8
+// AARCH64-LE-NEXT:    store ptr [[S]], ptr [[S_ADDR]], align 8
+// AARCH64-LE-NEXT:    [[TMP0:%.*]] = load ptr, ptr [[S_ADDR]], align 8
+// AARCH64-LE-NEXT:    [[TMP1:%.*]] = getelementptr i8, ptr [[TMP0]], i32 0
+// AARCH64-LE-NEXT:    [[TMP2:%.*]] = load i8, ptr [[TMP1]], align 1
+// AARCH64-LE-NEXT:    [[TMP3:%.*]] = and i8 [[TMP2]], -4
+// AARCH64-LE-NEXT:    store i8 [[TMP3]], ptr [[TMP1]], align 1
+// AARCH64-LE-NEXT:    [[TMP4:%.*]] = getelementptr i8, ptr [[TMP0]], i32 0
+// AARCH64-LE-NEXT:    [[TMP5:%.*]] = load i8, ptr [[TMP4]], align 1
+// AARCH64-LE-NEXT:    [[TMP6:%.*]] = and i8 [[TMP5]], 63
+// AARCH64-LE-NEXT:    store i8 [[TMP6]], ptr [[TMP4]], align 1
+// AARCH64-LE-NEXT:    ret void
+//
+// AARCH64-BE-LABEL: define dso_local void @testUnnamedBitfieldSurrounding(
+// AARCH64-BE-SAME: ptr noundef [[S:%.*]]) #[[ATTR0]] {
+// AARCH64-BE-NEXT:  [[ENTRY:.*:]]
+// AARCH64-BE-NEXT:    [[S_ADDR:%.*]] = alloca ptr, align 8
+// AARCH64-BE-NEXT:    store ptr [[S]], ptr [[S_ADDR]], align 8
+// AARCH64-BE-NEXT:    [[TMP0:%.*]] = load ptr, ptr [[S_ADDR]], align 8
+// AARCH64-BE-NEXT:    [[TMP1:%.*]] = getelementptr i8, ptr [[TMP0]], i32 0
+// AARCH64-BE-NEXT:    [[TMP2:%.*]] = load i8, ptr [[TMP1]], align 1
+// AARCH64-BE-NEXT:    [[TMP3:%.*]] = and i8 [[TMP2]], -4
+// AARCH64-BE-NEXT:    store i8 [[TMP3]], ptr [[TMP1]], align 1
+// AARCH64-BE-NEXT:    [[TMP4:%.*]] = getelementptr i8, ptr [[TMP0]], i32 0
+// AARCH64-BE-NEXT:    [[TMP5:%.*]] = load i8, ptr [[TMP4]], align 1
+// AARCH64-BE-NEXT:    [[TMP6:%.*]] = and i8 [[TMP5]], 63
+// AARCH64-BE-NEXT:    store i8 [[TMP6]], ptr [[TMP4]], align 1
+// AARCH64-BE-NEXT:    ret void
+//
+extern "C" void testUnnamedBitfieldSurrounding(
+    struct UnnamedBitfieldSurrounding *s) {
   // byte 0: unnamed[0-1], a[2-5], unnamed[6-7]
   // bits 0-1 and 6-7 must be cleared
   // Masks:
@@ -1695,13 +4729,14 @@ void testUnnamedBitfieldSurrounding(struct UnnamedBitfieldSurrounding *s) {
   __builtin_clear_padding(s);
 }
 
+
 struct UnnamedZeroWidthBitfield {
   unsigned char a : 4;
   unsigned int  : 0;
   unsigned int b : 4;
 };
 
-// LINUX-LABEL: define dso_local void @_Z28testUnnamedZeroWidthBitfieldP24UnnamedZeroWidthBitfield(
+// LINUX-LABEL: define dso_local void @testUnnamedZeroWidthBitfield(
 // LINUX-SAME: ptr noundef [[S:%.*]]) #[[ATTR0]] {
 // LINUX-NEXT:  [[ENTRY:.*:]]
 // LINUX-NEXT:    [[S_ADDR:%.*]] = alloca ptr, align 8
@@ -1729,34 +4764,149 @@ struct UnnamedZeroWidthBitfield {
 // LINUX-NEXT:    store i8 0, ptr [[TMP12]], align 1
 // LINUX-NEXT:    ret void
 //
-// WINDOWS-LABEL: define dso_local void @"?testUnnamedZeroWidthBitfield@@YAXPEAUUnnamedZeroWidthBitfield@@@Z"(ptr noundef %s) #0 {
-// WINDOWS-NEXT: entry:
-// WINDOWS-NEXT:   %s.addr = alloca ptr, align 8
-// WINDOWS-NEXT:   store ptr %s, ptr %s.addr, align 8
-// WINDOWS-NEXT:   %0 = load ptr, ptr %s.addr, align 8
-// WINDOWS-NEXT:   %1 = getelementptr i8, ptr %0, i32 0
-// WINDOWS-NEXT:   %2 = load i8, ptr %1, align 4
-// WINDOWS-NEXT:   %3 = and i8 %2, 15
-// WINDOWS-NEXT:   store i8 %3, ptr %1, align 4
-// WINDOWS-NEXT:   %4 = getelementptr i8, ptr %0, i32 1
-// WINDOWS-NEXT:   store i8 0, ptr %4, align 1
-// WINDOWS-NEXT:   %5 = getelementptr i8, ptr %0, i32 2
-// WINDOWS-NEXT:   store i8 0, ptr %5, align 2
-// WINDOWS-NEXT:   %6 = getelementptr i8, ptr %0, i32 3
-// WINDOWS-NEXT:   store i8 0, ptr %6, align 1
-// WINDOWS-NEXT:   %7 = getelementptr i8, ptr %0, i32 4
-// WINDOWS-NEXT:   %8 = load i8, ptr %7, align 4
-// WINDOWS-NEXT:   %9 = and i8 %8, 15
-// WINDOWS-NEXT:   store i8 %9, ptr %7, align 4
-// WINDOWS-NEXT:   %10 = getelementptr i8, ptr %0, i32 5
-// WINDOWS-NEXT:   store i8 0, ptr %10, align 1
-// WINDOWS-NEXT:   %11 = getelementptr i8, ptr %0, i32 6
-// WINDOWS-NEXT:   store i8 0, ptr %11, align 2
-// WINDOWS-NEXT:   %12 = getelementptr i8, ptr %0, i32 7
-// WINDOWS-NEXT:   store i8 0, ptr %12, align 1
-// WINDOWS-NEXT:   ret void
-//
-void testUnnamedZeroWidthBitfield(struct UnnamedZeroWidthBitfield *s) {
+// WINDOWS-LABEL: define dso_local void @testUnnamedZeroWidthBitfield(
+// WINDOWS-SAME: ptr noundef [[S:%.*]]) #[[ATTR0]] {
+// WINDOWS-NEXT:  [[ENTRY:.*:]]
+// WINDOWS-NEXT:    [[S_ADDR:%.*]] = alloca ptr, align 8
+// WINDOWS-NEXT:    store ptr [[S]], ptr [[S_ADDR]], align 8
+// WINDOWS-NEXT:    [[TMP0:%.*]] = load ptr, ptr [[S_ADDR]], align 8
+// WINDOWS-NEXT:    [[TMP1:%.*]] = getelementptr i8, ptr [[TMP0]], i32 0
+// WINDOWS-NEXT:    [[TMP2:%.*]] = load i8, ptr [[TMP1]], align 4
+// WINDOWS-NEXT:    [[TMP3:%.*]] = and i8 [[TMP2]], 15
+// WINDOWS-NEXT:    store i8 [[TMP3]], ptr [[TMP1]], align 4
+// WINDOWS-NEXT:    [[TMP4:%.*]] = getelementptr i8, ptr [[TMP0]], i32 1
+// WINDOWS-NEXT:    store i8 0, ptr [[TMP4]], align 1
+// WINDOWS-NEXT:    [[TMP5:%.*]] = getelementptr i8, ptr [[TMP0]], i32 2
+// WINDOWS-NEXT:    store i8 0, ptr [[TMP5]], align 2
+// WINDOWS-NEXT:    [[TMP6:%.*]] = getelementptr i8, ptr [[TMP0]], i32 3
+// WINDOWS-NEXT:    store i8 0, ptr [[TMP6]], align 1
+// WINDOWS-NEXT:    [[TMP7:%.*]] = getelementptr i8, ptr [[TMP0]], i32 4
+// WINDOWS-NEXT:    [[TMP8:%.*]] = load i8, ptr [[TMP7]], align 4
+// WINDOWS-NEXT:    [[TMP9:%.*]] = and i8 [[TMP8]], 15
+// WINDOWS-NEXT:    store i8 [[TMP9]], ptr [[TMP7]], align 4
+// WINDOWS-NEXT:    [[TMP10:%.*]] = getelementptr i8, ptr [[TMP0]], i32 5
+// WINDOWS-NEXT:    store i8 0, ptr [[TMP10]], align 1
+// WINDOWS-NEXT:    [[TMP11:%.*]] = getelementptr i8, ptr [[TMP0]], i32 6
+// WINDOWS-NEXT:    store i8 0, ptr [[TMP11]], align 2
+// WINDOWS-NEXT:    [[TMP12:%.*]] = getelementptr i8, ptr [[TMP0]], i32 7
+// WINDOWS-NEXT:    store i8 0, ptr [[TMP12]], align 1
+// WINDOWS-NEXT:    ret void
+//
+// ARM-LE-LABEL: define dso_local void @testUnnamedZeroWidthBitfield(
+// ARM-LE-SAME: ptr noundef [[S:%.*]]) #[[ATTR0]] {
+// ARM-LE-NEXT:  [[ENTRY:.*:]]
+// ARM-LE-NEXT:    [[S_ADDR:%.*]] = alloca ptr, align 4
+// ARM-LE-NEXT:    store ptr [[S]], ptr [[S_ADDR]], align 4
+// ARM-LE-NEXT:    [[TMP0:%.*]] = load ptr, ptr [[S_ADDR]], align 4
+// ARM-LE-NEXT:    [[TMP1:%.*]] = getelementptr i8, ptr [[TMP0]], i32 0
+// ARM-LE-NEXT:    [[TMP2:%.*]] = load i8, ptr [[TMP1]], align 4
+// ARM-LE-NEXT:    [[TMP3:%.*]] = and i8 [[TMP2]], 15
+// ARM-LE-NEXT:    store i8 [[TMP3]], ptr [[TMP1]], align 4
+// ARM-LE-NEXT:    [[TMP4:%.*]] = getelementptr i8, ptr [[TMP0]], i32 1
+// ARM-LE-NEXT:    store i8 0, ptr [[TMP4]], align 1
+// ARM-LE-NEXT:    [[TMP5:%.*]] = getelementptr i8, ptr [[TMP0]], i32 2
+// ARM-LE-NEXT:    store i8 0, ptr [[TMP5]], align 2
+// ARM-LE-NEXT:    [[TMP6:%.*]] = getelementptr i8, ptr [[TMP0]], i32 3
+// ARM-LE-NEXT:    store i8 0, ptr [[TMP6]], align 1
+// ARM-LE-NEXT:    [[TMP7:%.*]] = getelementptr i8, ptr [[TMP0]], i32 4
+// ARM-LE-NEXT:    [[TMP8:%.*]] = load i8, ptr [[TMP7]], align 4
+// ARM-LE-NEXT:    [[TMP9:%.*]] = and i8 [[TMP8]], 15
+// ARM-LE-NEXT:    store i8 [[TMP9]], ptr [[TMP7]], align 4
+// ARM-LE-NEXT:    [[TMP10:%.*]] = getelementptr i8, ptr [[TMP0]], i32 5
+// ARM-LE-NEXT:    store i8 0, ptr [[TMP10]], align 1
+// ARM-LE-NEXT:    [[TMP11:%.*]] = getelementptr i8, ptr [[TMP0]], i32 6
+// ARM-LE-NEXT:    store i8 0, ptr [[TMP11]], align 2
+// ARM-LE-NEXT:    [[TMP12:%.*]] = getelementptr i8, ptr [[TMP0]], i32 7
+// ARM-LE-NEXT:    store i8 0, ptr [[TMP12]], align 1
+// ARM-LE-NEXT:    ret void
+//
+// ARM-BE-LABEL: define dso_local void @testUnnamedZeroWidthBitfield(
+// ARM-BE-SAME: ptr noundef [[S:%.*]]) #[[ATTR0]] {
+// ARM-BE-NEXT:  [[ENTRY:.*:]]
+// ARM-BE-NEXT:    [[S_ADDR:%.*]] = alloca ptr, align 4
+// ARM-BE-NEXT:    store ptr [[S]], ptr [[S_ADDR]], align 4
+// ARM-BE-NEXT:    [[TMP0:%.*]] = load ptr, ptr [[S_ADDR]], align 4
+// ARM-BE-NEXT:    [[TMP1:%.*]] = getelementptr i8, ptr [[TMP0]], i32 0
+// ARM-BE-NEXT:    [[TMP2:%.*]] = load i8, ptr [[TMP1]], align 4
+// ARM-BE-NEXT:    [[TMP3:%.*]] = and i8 [[TMP2]], -16
+// ARM-BE-NEXT:    store i8 [[TMP3]], ptr [[TMP1]], align 4
+// ARM-BE-NEXT:    [[TMP4:%.*]] = getelementptr i8, ptr [[TMP0]], i32 1
+// ARM-BE-NEXT:    store i8 0, ptr [[TMP4]], align 1
+// ARM-BE-NEXT:    [[TMP5:%.*]] = getelementptr i8, ptr [[TMP0]], i32 2
+// ARM-BE-NEXT:    store i8 0, ptr [[TMP5]], align 2
+// ARM-BE-NEXT:    [[TMP6:%.*]] = getelementptr i8, ptr [[TMP0]], i32 3
+// ARM-BE-NEXT:    store i8 0, ptr [[TMP6]], align 1
+// ARM-BE-NEXT:    [[TMP7:%.*]] = getelementptr i8, ptr [[TMP0]], i32 4
+// ARM-BE-NEXT:    [[TMP8:%.*]] = load i8, ptr [[TMP7]], align 4
+// ARM-BE-NEXT:    [[TMP9:%.*]] = and i8 [[TMP8]], -16
+// ARM-BE-NEXT:    store i8 [[TMP9]], ptr [[TMP7]], align 4
+// ARM-BE-NEXT:    [[TMP10:%.*]] = getelementptr i8, ptr [[TMP0]], i32 5
+// ARM-BE-NEXT:    store i8 0, ptr [[TMP10]], align 1
+// ARM-BE-NEXT:    [[TMP11:%.*]] = getelementptr i8, ptr [[TMP0]], i32 6
+// ARM-BE-NEXT:    store i8 0, ptr [[TMP11]], align 2
+// ARM-BE-NEXT:    [[TMP12:%.*]] = getelementptr i8, ptr [[TMP0]], i32 7
+// ARM-BE-NEXT:    store i8 0, ptr [[TMP12]], align 1
+// ARM-BE-NEXT:    ret void
+//
+// AARCH64-LE-LABEL: define dso_local void @testUnnamedZeroWidthBitfield(
+// AARCH64-LE-SAME: ptr noundef [[S:%.*]]) #[[ATTR0]] {
+// AARCH64-LE-NEXT:  [[ENTRY:.*:]]
+// AARCH64-LE-NEXT:    [[S_ADDR:%.*]] = alloca ptr, align 8
+// AARCH64-LE-NEXT:    store ptr [[S]], ptr [[S_ADDR]], align 8
+// AARCH64-LE-NEXT:    [[TMP0:%.*]] = load ptr, ptr [[S_ADDR]], align 8
+// AARCH64-LE-NEXT:    [[TMP1:%.*]] = getelementptr i8, ptr [[TMP0]], i32 0
+// AARCH64-LE-NEXT:    [[TMP2:%.*]] = load i8, ptr [[TMP1]], align 4
+// AARCH64-LE-NEXT:    [[TMP3:%.*]] = and i8 [[TMP2]], 15
+// AARCH64-LE-NEXT:    store i8 [[TMP3]], ptr [[TMP1]], align 4
+// AARCH64-LE-NEXT:    [[TMP4:%.*]] = getelementptr i8, ptr [[TMP0]], i32 1
+// AARCH64-LE-NEXT:    store i8 0, ptr [[TMP4]], align 1
+// AARCH64-LE-NEXT:    [[TMP5:%.*]] = getelementptr i8, ptr [[TMP0]], i32 2
+// AARCH64-LE-NEXT:    store i8 0, ptr [[TMP5]], align 2
+// AARCH64-LE-NEXT:    [[TMP6:%.*]] = getelementptr i8, ptr [[TMP0]], i32 3
+// AARCH64-LE-NEXT:    store i8 0, ptr [[TMP6]], align 1
+// AARCH64-LE-NEXT:    [[TMP7:%.*]] = getelementptr i8, ptr [[TMP0]], i32 4
+// AARCH64-LE-NEXT:    [[TMP8:%.*]] = load i8, ptr [[TMP7]], align 4
+// AARCH64-LE-NEXT:    [[TMP9:%.*]] = and i8 [[TMP8]], 15
+// AARCH64-LE-NEXT:    store i8 [[TMP9]], ptr [[TMP7]], align 4
+// AARCH64-LE-NEXT:    [[TMP10:%.*]] = getelementptr i8, ptr [[TMP0]], i32 5
+// AARCH64-LE-NEXT:    store i8 0, ptr [[TMP10]], align 1
+// AARCH64-LE-NEXT:    [[TMP11:%.*]] = getelementptr i8, ptr [[TMP0]], i32 6
+// AARCH64-LE-NEXT:    store i8 0, ptr [[TMP11]], align 2
+// AARCH64-LE-NEXT:    [[TMP12:%.*]] = getelementptr i8, ptr [[TMP0]], i32 7
+// AARCH64-LE-NEXT:    store i8 0, ptr [[TMP12]], align 1
+// AARCH64-LE-NEXT:    ret void
+//
+// AARCH64-BE-LABEL: define dso_local void @testUnnamedZeroWidthBitfield(
+// AARCH64-BE-SAME: ptr noundef [[S:%.*]]) #[[ATTR0]] {
+// AARCH64-BE-NEXT:  [[ENTRY:.*:]]
+// AARCH64-BE-NEXT:    [[S_ADDR:%.*]] = alloca ptr, align 8
+// AARCH64-BE-NEXT:    store ptr [[S]], ptr [[S_ADDR]], align 8
+// AARCH64-BE-NEXT:    [[TMP0:%.*]] = load ptr, ptr [[S_ADDR]], align 8
+// AARCH64-BE-NEXT:    [[TMP1:%.*]] = getelementptr i8, ptr [[TMP0]], i32 0
+// AARCH64-BE-NEXT:    [[TMP2:%.*]] = load i8, ptr [[TMP1]], align 4
+// AARCH64-BE-NEXT:    [[TMP3:%.*]] = and i8 [[TMP2]], -16
+// AARCH64-BE-NEXT:    store i8 [[TMP3]], ptr [[TMP1]], align 4
+// AARCH64-BE-NEXT:    [[TMP4:%.*]] = getelementptr i8, ptr [[TMP0]], i32 1
+// AARCH64-BE-NEXT:    store i8 0, ptr [[TMP4]], align 1
+// AARCH64-BE-NEXT:    [[TMP5:%.*]] = getelementptr i8, ptr [[TMP0]], i32 2
+// AARCH64-BE-NEXT:    store i8 0, ptr [[TMP5]], align 2
+// AARCH64-BE-NEXT:    [[TMP6:%.*]] = getelementptr i8, ptr [[TMP0]], i32 3
+// AARCH64-BE-NEXT:    store i8 0, ptr [[TMP6]], align 1
+// AARCH64-BE-NEXT:    [[TMP7:%.*]] = getelementptr i8, ptr [[TMP0]], i32 4
+// AARCH64-BE-NEXT:    [[TMP8:%.*]] = load i8, ptr [[TMP7]], align 4
+// AARCH64-BE-NEXT:    [[TMP9:%.*]] = and i8 [[TMP8]], -16
+// AARCH64-BE-NEXT:    store i8 [[TMP9]], ptr [[TMP7]], align 4
+// AARCH64-BE-NEXT:    [[TMP10:%.*]] = getelementptr i8, ptr [[TMP0]], i32 5
+// AARCH64-BE-NEXT:    store i8 0, ptr [[TMP10]], align 1
+// AARCH64-BE-NEXT:    [[TMP11:%.*]] = getelementptr i8, ptr [[TMP0]], i32 6
+// AARCH64-BE-NEXT:    store i8 0, ptr [[TMP11]], align 2
+// AARCH64-BE-NEXT:    [[TMP12:%.*]] = getelementptr i8, ptr [[TMP0]], i32 7
+// AARCH64-BE-NEXT:    store i8 0, ptr [[TMP12]], align 1
+// AARCH64-BE-NEXT:    ret void
+//
+extern "C" void testUnnamedZeroWidthBitfield(
+    struct UnnamedZeroWidthBitfield *s) {
+  // The masks below describe little endian; big endian uses 0xf0.
   // byte 0: a[0-3], unnamed[4-7]
   // bytes 1-3: struct padding
   // bytes 4-7: b[bits 0-3], tail padding[bits 4-31]
@@ -1773,13 +4923,14 @@ void testUnnamedZeroWidthBitfield(struct UnnamedZeroWidthBitfield *s) {
   __builtin_clear_padding(s);
 }
 
+
 struct UnnamedBitfieldMultiByte {
   unsigned short a : 4;
   unsigned short   : 8;
   unsigned short b : 4;
 };
 
-// LINUX-LABEL: define dso_local void @_Z28testUnnamedBitfieldMultiByteP24UnnamedBitfieldMultiByte(
+// LINUX-LABEL: define dso_local void @testUnnamedBitfieldMultiByte(
 // LINUX-SAME: ptr noundef [[S:%.*]]) #[[ATTR0]] {
 // LINUX-NEXT:  [[ENTRY:.*:]]
 // LINUX-NEXT:    [[S_ADDR:%.*]] = alloca ptr, align 8
@@ -1795,22 +4946,89 @@ struct UnnamedBitfieldMultiByte {
 // LINUX-NEXT:    store i8 [[TMP6]], ptr [[TMP4]], align 1
 // LINUX-NEXT:    ret void
 //
-// WINDOWS-LABEL: define dso_local void @"?testUnnamedBitfieldMultiByte@@YAXPEAUUnnamedBitfieldMultiByte@@@Z"(ptr noundef %s) #0 {
-// WINDOWS-NEXT: entry:
-// WINDOWS-NEXT:   %s.addr = alloca ptr, align 8
-// WINDOWS-NEXT:   store ptr %s, ptr %s.addr, align 8
-// WINDOWS-NEXT:   %0 = load ptr, ptr %s.addr, align 8
-// WINDOWS-NEXT:   %1 = getelementptr i8, ptr %0, i32 0
-// WINDOWS-NEXT:   %2 = load i8, ptr %1, align 2
-// WINDOWS-NEXT:   %3 = and i8 %2, 15
-// WINDOWS-NEXT:   store i8 %3, ptr %1, align 2
-// WINDOWS-NEXT:   %4 = getelementptr i8, ptr %0, i32 1
-// WINDOWS-NEXT:   %5 = load i8, ptr %4, align 1
-// WINDOWS-NEXT:   %6 = and i8 %5, -16
-// WINDOWS-NEXT:   store i8 %6, ptr %4, align 1
-// WINDOWS-NEXT:   ret void
-//
-void testUnnamedBitfieldMultiByte(struct UnnamedBitfieldMultiByte *s) {
+// WINDOWS-LABEL: define dso_local void @testUnnamedBitfieldMultiByte(
+// WINDOWS-SAME: ptr noundef [[S:%.*]]) #[[ATTR0]] {
+// WINDOWS-NEXT:  [[ENTRY:.*:]]
+// WINDOWS-NEXT:    [[S_ADDR:%.*]] = alloca ptr, align 8
+// WINDOWS-NEXT:    store ptr [[S]], ptr [[S_ADDR]], align 8
+// WINDOWS-NEXT:    [[TMP0:%.*]] = load ptr, ptr [[S_ADDR]], align 8
+// WINDOWS-NEXT:    [[TMP1:%.*]] = getelementptr i8, ptr [[TMP0]], i32 0
+// WINDOWS-NEXT:    [[TMP2:%.*]] = load i8, ptr [[TMP1]], align 2
+// WINDOWS-NEXT:    [[TMP3:%.*]] = and i8 [[TMP2]], 15
+// WINDOWS-NEXT:    store i8 [[TMP3]], ptr [[TMP1]], align 2
+// WINDOWS-NEXT:    [[TMP4:%.*]] = getelementptr i8, ptr [[TMP0]], i32 1
+// WINDOWS-NEXT:    [[TMP5:%.*]] = load i8, ptr [[TMP4]], align 1
+// WINDOWS-NEXT:    [[TMP6:%.*]] = and i8 [[TMP5]], -16
+// WINDOWS-NEXT:    store i8 [[TMP6]], ptr [[TMP4]], align 1
+// WINDOWS-NEXT:    ret void
+//
+// ARM-LE-LABEL: define dso_local void @testUnnamedBitfieldMultiByte(
+// ARM-LE-SAME: ptr noundef [[S:%.*]]) #[[ATTR0]] {
+// ARM-LE-NEXT:  [[ENTRY:.*:]]
+// ARM-LE-NEXT:    [[S_ADDR:%.*]] = alloca ptr, align 4
+// ARM-LE-NEXT:    store ptr [[S]], ptr [[S_ADDR]], align 4
+// ARM-LE-NEXT:    [[TMP0:%.*]] = load ptr, ptr [[S_ADDR]], align 4
+// ARM-LE-NEXT:    [[TMP1:%.*]] = getelementptr i8, ptr [[TMP0]], i32 0
+// ARM-LE-NEXT:    [[TMP2:%.*]] = load i8, ptr [[TMP1]], align 2
+// ARM-LE-NEXT:    [[TMP3:%.*]] = and i8 [[TMP2]], 15
+// ARM-LE-NEXT:    store i8 [[TMP3]], ptr [[TMP1]], align 2
+// ARM-LE-NEXT:    [[TMP4:%.*]] = getelementptr i8, ptr [[TMP0]], i32 1
+// ARM-LE-NEXT:    [[TMP5:%.*]] = load i8, ptr [[TMP4]], align 1
+// ARM-LE-NEXT:    [[TMP6:%.*]] = and i8 [[TMP5]], -16
+// ARM-LE-NEXT:    store i8 [[TMP6]], ptr [[TMP4]], align 1
+// ARM-LE-NEXT:    ret void
+//
+// ARM-BE-LABEL: define dso_local void @testUnnamedBitfieldMultiByte(
+// ARM-BE-SAME: ptr noundef [[S:%.*]]) #[[ATTR0]] {
+// ARM-BE-NEXT:  [[ENTRY:.*:]]
+// ARM-BE-NEXT:    [[S_ADDR:%.*]] = alloca ptr, align 4
+// ARM-BE-NEXT:    store ptr [[S]], ptr [[S_ADDR]], align 4
+// ARM-BE-NEXT:    [[TMP0:%.*]] = load ptr, ptr [[S_ADDR]], align 4
+// ARM-BE-NEXT:    [[TMP1:%.*]] = getelementptr i8, ptr [[TMP0]], i32 0
+// ARM-BE-NEXT:    [[TMP2:%.*]] = load i8, ptr [[TMP1]], align 2
+// ARM-BE-NEXT:    [[TMP3:%.*]] = and i8 [[TMP2]], -16
+// ARM-BE-NEXT:    store i8 [[TMP3]], ptr [[TMP1]], align 2
+// ARM-BE-NEXT:    [[TMP4:%.*]] = getelementptr i8, ptr [[TMP0]], i32 1
+// ARM-BE-NEXT:    [[TMP5:%.*]] = load i8, ptr [[TMP4]], align 1
+// ARM-BE-NEXT:    [[TMP6:%.*]] = and i8 [[TMP5]], 15
+// ARM-BE-NEXT:    store i8 [[TMP6]], ptr [[TMP4]], align 1
+// ARM-BE-NEXT:    ret void
+//
+// AARCH64-LE-LABEL: define dso_local void @testUnnamedBitfieldMultiByte(
+// AARCH64-LE-SAME: ptr noundef [[S:%.*]]) #[[ATTR0]] {
+// AARCH64-LE-NEXT:  [[ENTRY:.*:]]
+// AARCH64-LE-NEXT:    [[S_ADDR:%.*]] = alloca ptr, align 8
+// AARCH64-LE-NEXT:    store ptr [[S]], ptr [[S_ADDR]], align 8
+// AARCH64-LE-NEXT:    [[TMP0:%.*]] = load ptr, ptr [[S_ADDR]], align 8
+// AARCH64-LE-NEXT:    [[TMP1:%.*]] = getelementptr i8, ptr [[TMP0]], i32 0
+// AARCH64-LE-NEXT:    [[TMP2:%.*]] = load i8, ptr [[TMP1]], align 2
+// AARCH64-LE-NEXT:    [[TMP3:%.*]] = and i8 [[TMP2]], 15
+// AARCH64-LE-NEXT:    store i8 [[TMP3]], ptr [[TMP1]], align 2
+// AARCH64-LE-NEXT:    [[TMP4:%.*]] = getelementptr i8, ptr [[TMP0]], i32 1
+// AARCH64-LE-NEXT:    [[TMP5:%.*]] = load i8, ptr [[TMP4]], align 1
+// AARCH64-LE-NEXT:    [[TMP6:%.*]] = and i8 [[TMP5]], -16
+// AARCH64-LE-NEXT:    store i8 [[TMP6]], ptr [[TMP4]], align 1
+// AARCH64-LE-NEXT:    ret void
+//
+// AARCH64-BE-LABEL: define dso_local void @testUnnamedBitfieldMultiByte(
+// AARCH64-BE-SAME: ptr noundef [[S:%.*]]) #[[ATTR0]] {
+// AARCH64-BE-NEXT:  [[ENTRY:.*:]]
+// AARCH64-BE-NEXT:    [[S_ADDR:%.*]] = alloca ptr, align 8
+// AARCH64-BE-NEXT:    store ptr [[S]], ptr [[S_ADDR]], align 8
+// AARCH64-BE-NEXT:    [[TMP0:%.*]] = load ptr, ptr [[S_ADDR]], align 8
+// AARCH64-BE-NEXT:    [[TMP1:%.*]] = getelementptr i8, ptr [[TMP0]], i32 0
+// AARCH64-BE-NEXT:    [[TMP2:%.*]] = load i8, ptr [[TMP1]], align 2
+// AARCH64-BE-NEXT:    [[TMP3:%.*]] = and i8 [[TMP2]], -16
+// AARCH64-BE-NEXT:    store i8 [[TMP3]], ptr [[TMP1]], align 2
+// AARCH64-BE-NEXT:    [[TMP4:%.*]] = getelementptr i8, ptr [[TMP0]], i32 1
+// AARCH64-BE-NEXT:    [[TMP5:%.*]] = load i8, ptr [[TMP4]], align 1
+// AARCH64-BE-NEXT:    [[TMP6:%.*]] = and i8 [[TMP5]], 15
+// AARCH64-BE-NEXT:    store i8 [[TMP6]], ptr [[TMP4]], align 1
+// AARCH64-BE-NEXT:    ret void
+//
+extern "C" void testUnnamedBitfieldMultiByte(
+    struct UnnamedBitfieldMultiByte *s) {
+  // The masks below describe little endian; big endian swaps them.
   // 2 bytes: a[0-3], unnamed[4-11], b[12-15]
   // byte 0: clear bits 4-7
   // byte 1: clear bits 0-3
@@ -1820,3 +5038,2472 @@ void testUnnamedBitfieldMultiByte(struct UnnamedBitfieldMultiByte *s) {
   //   0b11110000 == -16
   __builtin_clear_padding(s);
 }
+
+
+// Complete bytes require no clearing, including a full 64-bit bitfield.
+struct BitfieldFullBytes {
+  unsigned char first : 8;
+  unsigned char second : 8;
+  unsigned short half : 16;
+  unsigned int word : 32;
+  unsigned long long wide : 64;
+};
+
+// LINUX-LABEL: define dso_local void @testBitfieldFullBytes(
+// LINUX-SAME: ptr noundef [[S:%.*]]) #[[ATTR0]] {
+// LINUX-NEXT:  [[ENTRY:.*:]]
+// LINUX-NEXT:    [[S_ADDR:%.*]] = alloca ptr, align 8
+// LINUX-NEXT:    store ptr [[S]], ptr [[S_ADDR]], align 8
+// LINUX-NEXT:    [[TMP0:%.*]] = load ptr, ptr [[S_ADDR]], align 8
+// LINUX-NEXT:    ret void
+//
+// WINDOWS-LABEL: define dso_local void @testBitfieldFullBytes(
+// WINDOWS-SAME: ptr noundef [[S:%.*]]) #[[ATTR0]] {
+// WINDOWS-NEXT:  [[ENTRY:.*:]]
+// WINDOWS-NEXT:    [[S_ADDR:%.*]] = alloca ptr, align 8
+// WINDOWS-NEXT:    store ptr [[S]], ptr [[S_ADDR]], align 8
+// WINDOWS-NEXT:    [[TMP0:%.*]] = load ptr, ptr [[S_ADDR]], align 8
+// WINDOWS-NEXT:    ret void
+//
+// ARM-LE-LABEL: define dso_local void @testBitfieldFullBytes(
+// ARM-LE-SAME: ptr noundef [[S:%.*]]) #[[ATTR0]] {
+// ARM-LE-NEXT:  [[ENTRY:.*:]]
+// ARM-LE-NEXT:    [[S_ADDR:%.*]] = alloca ptr, align 4
+// ARM-LE-NEXT:    store ptr [[S]], ptr [[S_ADDR]], align 4
+// ARM-LE-NEXT:    [[TMP0:%.*]] = load ptr, ptr [[S_ADDR]], align 4
+// ARM-LE-NEXT:    ret void
+//
+// ARM-BE-LABEL: define dso_local void @testBitfieldFullBytes(
+// ARM-BE-SAME: ptr noundef [[S:%.*]]) #[[ATTR0]] {
+// ARM-BE-NEXT:  [[ENTRY:.*:]]
+// ARM-BE-NEXT:    [[S_ADDR:%.*]] = alloca ptr, align 4
+// ARM-BE-NEXT:    store ptr [[S]], ptr [[S_ADDR]], align 4
+// ARM-BE-NEXT:    [[TMP0:%.*]] = load ptr, ptr [[S_ADDR]], align 4
+// ARM-BE-NEXT:    ret void
+//
+// AARCH64-LE-LABEL: define dso_local void @testBitfieldFullBytes(
+// AARCH64-LE-SAME: ptr noundef [[S:%.*]]) #[[ATTR0]] {
+// AARCH64-LE-NEXT:  [[ENTRY:.*:]]
+// AARCH64-LE-NEXT:    [[S_ADDR:%.*]] = alloca ptr, align 8
+// AARCH64-LE-NEXT:    store ptr [[S]], ptr [[S_ADDR]], align 8
+// AARCH64-LE-NEXT:    [[TMP0:%.*]] = load ptr, ptr [[S_ADDR]], align 8
+// AARCH64-LE-NEXT:    ret void
+//
+// AARCH64-BE-LABEL: define dso_local void @testBitfieldFullBytes(
+// AARCH64-BE-SAME: ptr noundef [[S:%.*]]) #[[ATTR0]] {
+// AARCH64-BE-NEXT:  [[ENTRY:.*:]]
+// AARCH64-BE-NEXT:    [[S_ADDR:%.*]] = alloca ptr, align 8
+// AARCH64-BE-NEXT:    store ptr [[S]], ptr [[S_ADDR]], align 8
+// AARCH64-BE-NEXT:    [[TMP0:%.*]] = load ptr, ptr [[S_ADDR]], align 8
+// AARCH64-BE-NEXT:    ret void
+//
+extern "C" void testBitfieldFullBytes(struct BitfieldFullBytes *s) {
+  __builtin_clear_padding(s);
+}
+
+// The value ends exactly at a byte boundary. Preserve masks are 0xf8, 0xff
+// on little endian and 0x1f, 0xff on big endian.
+struct BitfieldEndAligned {
+  unsigned short : 3;
+  unsigned short value : 13;
+};
+
+// LINUX-LABEL: define dso_local void @testBitfieldEndAligned(
+// LINUX-SAME: ptr noundef [[S:%.*]]) #[[ATTR0]] {
+// LINUX-NEXT:  [[ENTRY:.*:]]
+// LINUX-NEXT:    [[S_ADDR:%.*]] = alloca ptr, align 8
+// LINUX-NEXT:    store ptr [[S]], ptr [[S_ADDR]], align 8
+// LINUX-NEXT:    [[TMP0:%.*]] = load ptr, ptr [[S_ADDR]], align 8
+// LINUX-NEXT:    [[TMP1:%.*]] = getelementptr i8, ptr [[TMP0]], i32 0
+// LINUX-NEXT:    [[TMP2:%.*]] = load i8, ptr [[TMP1]], align 2
+// LINUX-NEXT:    [[TMP3:%.*]] = and i8 [[TMP2]], -8
+// LINUX-NEXT:    store i8 [[TMP3]], ptr [[TMP1]], align 2
+// LINUX-NEXT:    ret void
+//
+// WINDOWS-LABEL: define dso_local void @testBitfieldEndAligned(
+// WINDOWS-SAME: ptr noundef [[S:%.*]]) #[[ATTR0]] {
+// WINDOWS-NEXT:  [[ENTRY:.*:]]
+// WINDOWS-NEXT:    [[S_ADDR:%.*]] = alloca ptr, align 8
+// WINDOWS-NEXT:    store ptr [[S]], ptr [[S_ADDR]], align 8
+// WINDOWS-NEXT:    [[TMP0:%.*]] = load ptr, ptr [[S_ADDR]], align 8
+// WINDOWS-NEXT:    [[TMP1:%.*]] = getelementptr i8, ptr [[TMP0]], i32 0
+// WINDOWS-NEXT:    [[TMP2:%.*]] = load i8, ptr [[TMP1]], align 2
+// WINDOWS-NEXT:    [[TMP3:%.*]] = and i8 [[TMP2]], -8
+// WINDOWS-NEXT:    store i8 [[TMP3]], ptr [[TMP1]], align 2
+// WINDOWS-NEXT:    ret void
+//
+// ARM-LE-LABEL: define dso_local void @testBitfieldEndAligned(
+// ARM-LE-SAME: ptr noundef [[S:%.*]]) #[[ATTR0]] {
+// ARM-LE-NEXT:  [[ENTRY:.*:]]
+// ARM-LE-NEXT:    [[S_ADDR:%.*]] = alloca ptr, align 4
+// ARM-LE-NEXT:    store ptr [[S]], ptr [[S_ADDR]], align 4
+// ARM-LE-NEXT:    [[TMP0:%.*]] = load ptr, ptr [[S_ADDR]], align 4
+// ARM-LE-NEXT:    [[TMP1:%.*]] = getelementptr i8, ptr [[TMP0]], i32 0
+// ARM-LE-NEXT:    [[TMP2:%.*]] = load i8, ptr [[TMP1]], align 2
+// ARM-LE-NEXT:    [[TMP3:%.*]] = and i8 [[TMP2]], -8
+// ARM-LE-NEXT:    store i8 [[TMP3]], ptr [[TMP1]], align 2
+// ARM-LE-NEXT:    ret void
+//
+// ARM-BE-LABEL: define dso_local void @testBitfieldEndAligned(
+// ARM-BE-SAME: ptr noundef [[S:%.*]]) #[[ATTR0]] {
+// ARM-BE-NEXT:  [[ENTRY:.*:]]
+// ARM-BE-NEXT:    [[S_ADDR:%.*]] = alloca ptr, align 4
+// ARM-BE-NEXT:    store ptr [[S]], ptr [[S_ADDR]], align 4
+// ARM-BE-NEXT:    [[TMP0:%.*]] = load ptr, ptr [[S_ADDR]], align 4
+// ARM-BE-NEXT:    [[TMP1:%.*]] = getelementptr i8, ptr [[TMP0]], i32 0
+// ARM-BE-NEXT:    [[TMP2:%.*]] = load i8, ptr [[TMP1]], align 2
+// ARM-BE-NEXT:    [[TMP3:%.*]] = and i8 [[TMP2]], 31
+// ARM-BE-NEXT:    store i8 [[TMP3]], ptr [[TMP1]], align 2
+// ARM-BE-NEXT:    ret void
+//
+// AARCH64-LE-LABEL: define dso_local void @testBitfieldEndAligned(
+// AARCH64-LE-SAME: ptr noundef [[S:%.*]]) #[[ATTR0]] {
+// AARCH64-LE-NEXT:  [[ENTRY:.*:]]
+// AARCH64-LE-NEXT:    [[S_ADDR:%.*]] = alloca ptr, align 8
+// AARCH64-LE-NEXT:    store ptr [[S]], ptr [[S_ADDR]], align 8
+// AARCH64-LE-NEXT:    [[TMP0:%.*]] = load ptr, ptr [[S_ADDR]], align 8
+// AARCH64-LE-NEXT:    [[TMP1:%.*]] = getelementptr i8, ptr [[TMP0]], i32 0
+// AARCH64-LE-NEXT:    [[TMP2:%.*]] = load i8, ptr [[TMP1]], align 2
+// AARCH64-LE-NEXT:    [[TMP3:%.*]] = and i8 [[TMP2]], -8
+// AARCH64-LE-NEXT:    store i8 [[TMP3]], ptr [[TMP1]], align 2
+// AARCH64-LE-NEXT:    ret void
+//
+// AARCH64-BE-LABEL: define dso_local void @testBitfieldEndAligned(
+// AARCH64-BE-SAME: ptr noundef [[S:%.*]]) #[[ATTR0]] {
+// AARCH64-BE-NEXT:  [[ENTRY:.*:]]
+// AARCH64-BE-NEXT:    [[S_ADDR:%.*]] = alloca ptr, align 8
+// AARCH64-BE-NEXT:    store ptr [[S]], ptr [[S_ADDR]], align 8
+// AARCH64-BE-NEXT:    [[TMP0:%.*]] = load ptr, ptr [[S_ADDR]], align 8
+// AARCH64-BE-NEXT:    [[TMP1:%.*]] = getelementptr i8, ptr [[TMP0]], i32 0
+// AARCH64-BE-NEXT:    [[TMP2:%.*]] = load i8, ptr [[TMP1]], align 2
+// AARCH64-BE-NEXT:    [[TMP3:%.*]] = and i8 [[TMP2]], 31
+// AARCH64-BE-NEXT:    store i8 [[TMP3]], ptr [[TMP1]], align 2
+// AARCH64-BE-NEXT:    ret void
+//
+extern "C" void testBitfieldEndAligned(struct BitfieldEndAligned *s) {
+  __builtin_clear_padding(s);
+}
+
+// A normal field with two partial endpoints and a complete middle byte.
+// Preserve masks are 0xf8, 0xff, 0x1f, 0x00 on little endian and
+// 0x1f, 0xff, 0xf8, 0x00 on big endian.
+struct BitfieldPartialEnds {
+  unsigned int : 3;
+  unsigned int value : 18;
+  unsigned int : 11;
+};
+
+// LINUX-LABEL: define dso_local void @testBitfieldPartialEnds(
+// LINUX-SAME: ptr noundef [[S:%.*]]) #[[ATTR0]] {
+// LINUX-NEXT:  [[ENTRY:.*:]]
+// LINUX-NEXT:    [[S_ADDR:%.*]] = alloca ptr, align 8
+// LINUX-NEXT:    store ptr [[S]], ptr [[S_ADDR]], align 8
+// LINUX-NEXT:    [[TMP0:%.*]] = load ptr, ptr [[S_ADDR]], align 8
+// LINUX-NEXT:    [[TMP1:%.*]] = getelementptr i8, ptr [[TMP0]], i32 0
+// LINUX-NEXT:    [[TMP2:%.*]] = load i8, ptr [[TMP1]], align 4
+// LINUX-NEXT:    [[TMP3:%.*]] = and i8 [[TMP2]], -8
+// LINUX-NEXT:    store i8 [[TMP3]], ptr [[TMP1]], align 4
+// LINUX-NEXT:    [[TMP4:%.*]] = getelementptr i8, ptr [[TMP0]], i32 2
+// LINUX-NEXT:    [[TMP5:%.*]] = load i8, ptr [[TMP4]], align 2
+// LINUX-NEXT:    [[TMP6:%.*]] = and i8 [[TMP5]], 31
+// LINUX-NEXT:    store i8 [[TMP6]], ptr [[TMP4]], align 2
+// LINUX-NEXT:    [[TMP7:%.*]] = getelementptr i8, ptr [[TMP0]], i32 3
+// LINUX-NEXT:    store i8 0, ptr [[TMP7]], align 1
+// LINUX-NEXT:    ret void
+//
+// WINDOWS-LABEL: define dso_local void @testBitfieldPartialEnds(
+// WINDOWS-SAME: ptr noundef [[S:%.*]]) #[[ATTR0]] {
+// WINDOWS-NEXT:  [[ENTRY:.*:]]
+// WINDOWS-NEXT:    [[S_ADDR:%.*]] = alloca ptr, align 8
+// WINDOWS-NEXT:    store ptr [[S]], ptr [[S_ADDR]], align 8
+// WINDOWS-NEXT:    [[TMP0:%.*]] = load ptr, ptr [[S_ADDR]], align 8
+// WINDOWS-NEXT:    [[TMP1:%.*]] = getelementptr i8, ptr [[TMP0]], i32 0
+// WINDOWS-NEXT:    [[TMP2:%.*]] = load i8, ptr [[TMP1]], align 4
+// WINDOWS-NEXT:    [[TMP3:%.*]] = and i8 [[TMP2]], -8
+// WINDOWS-NEXT:    store i8 [[TMP3]], ptr [[TMP1]], align 4
+// WINDOWS-NEXT:    [[TMP4:%.*]] = getelementptr i8, ptr [[TMP0]], i32 2
+// WINDOWS-NEXT:    [[TMP5:%.*]] = load i8, ptr [[TMP4]], align 2
+// WINDOWS-NEXT:    [[TMP6:%.*]] = and i8 [[TMP5]], 31
+// WINDOWS-NEXT:    store i8 [[TMP6]], ptr [[TMP4]], align 2
+// WINDOWS-NEXT:    [[TMP7:%.*]] = getelementptr i8, ptr [[TMP0]], i32 3
+// WINDOWS-NEXT:    store i8 0, ptr [[TMP7]], align 1
+// WINDOWS-NEXT:    ret void
+//
+// ARM-LE-LABEL: define dso_local void @testBitfieldPartialEnds(
+// ARM-LE-SAME: ptr noundef [[S:%.*]]) #[[ATTR0]] {
+// ARM-LE-NEXT:  [[ENTRY:.*:]]
+// ARM-LE-NEXT:    [[S_ADDR:%.*]] = alloca ptr, align 4
+// ARM-LE-NEXT:    store ptr [[S]], ptr [[S_ADDR]], align 4
+// ARM-LE-NEXT:    [[TMP0:%.*]] = load ptr, ptr [[S_ADDR]], align 4
+// ARM-LE-NEXT:    [[TMP1:%.*]] = getelementptr i8, ptr [[TMP0]], i32 0
+// ARM-LE-NEXT:    [[TMP2:%.*]] = load i8, ptr [[TMP1]], align 4
+// ARM-LE-NEXT:    [[TMP3:%.*]] = and i8 [[TMP2]], -8
+// ARM-LE-NEXT:    store i8 [[TMP3]], ptr [[TMP1]], align 4
+// ARM-LE-NEXT:    [[TMP4:%.*]] = getelementptr i8, ptr [[TMP0]], i32 2
+// ARM-LE-NEXT:    [[TMP5:%.*]] = load i8, ptr [[TMP4]], align 2
+// ARM-LE-NEXT:    [[TMP6:%.*]] = and i8 [[TMP5]], 31
+// ARM-LE-NEXT:    store i8 [[TMP6]], ptr [[TMP4]], align 2
+// ARM-LE-NEXT:    [[TMP7:%.*]] = getelementptr i8, ptr [[TMP0]], i32 3
+// ARM-LE-NEXT:    store i8 0, ptr [[TMP7]], align 1
+// ARM-LE-NEXT:    ret void
+//
+// ARM-BE-LABEL: define dso_local void @testBitfieldPartialEnds(
+// ARM-BE-SAME: ptr noundef [[S:%.*]]) #[[ATTR0]] {
+// ARM-BE-NEXT:  [[ENTRY:.*:]]
+// ARM-BE-NEXT:    [[S_ADDR:%.*]] = alloca ptr, align 4
+// ARM-BE-NEXT:    store ptr [[S]], ptr [[S_ADDR]], align 4
+// ARM-BE-NEXT:    [[TMP0:%.*]] = load ptr, ptr [[S_ADDR]], align 4
+// ARM-BE-NEXT:    [[TMP1:%.*]] = getelementptr i8, ptr [[TMP0]], i32 0
+// ARM-BE-NEXT:    [[TMP2:%.*]] = load i8, ptr [[TMP1]], align 4
+// ARM-BE-NEXT:    [[TMP3:%.*]] = and i8 [[TMP2]], 31
+// ARM-BE-NEXT:    store i8 [[TMP3]], ptr [[TMP1]], align 4
+// ARM-BE-NEXT:    [[TMP4:%.*]] = getelementptr i8, ptr [[TMP0]], i32 2
+// ARM-BE-NEXT:    [[TMP5:%.*]] = load i8, ptr [[TMP4]], align 2
+// ARM-BE-NEXT:    [[TMP6:%.*]] = and i8 [[TMP5]], -8
+// ARM-BE-NEXT:    store i8 [[TMP6]], ptr [[TMP4]], align 2
+// ARM-BE-NEXT:    [[TMP7:%.*]] = getelementptr i8, ptr [[TMP0]], i32 3
+// ARM-BE-NEXT:    store i8 0, ptr [[TMP7]], align 1
+// ARM-BE-NEXT:    ret void
+//
+// AARCH64-LE-LABEL: define dso_local void @testBitfieldPartialEnds(
+// AARCH64-LE-SAME: ptr noundef [[S:%.*]]) #[[ATTR0]] {
+// AARCH64-LE-NEXT:  [[ENTRY:.*:]]
+// AARCH64-LE-NEXT:    [[S_ADDR:%.*]] = alloca ptr, align 8
+// AARCH64-LE-NEXT:    store ptr [[S]], ptr [[S_ADDR]], align 8
+// AARCH64-LE-NEXT:    [[TMP0:%.*]] = load ptr, ptr [[S_ADDR]], align 8
+// AARCH64-LE-NEXT:    [[TMP1:%.*]] = getelementptr i8, ptr [[TMP0]], i32 0
+// AARCH64-LE-NEXT:    [[TMP2:%.*]] = load i8, ptr [[TMP1]], align 4
+// AARCH64-LE-NEXT:    [[TMP3:%.*]] = and i8 [[TMP2]], -8
+// AARCH64-LE-NEXT:    store i8 [[TMP3]], ptr [[TMP1]], align 4
+// AARCH64-LE-NEXT:    [[TMP4:%.*]] = getelementptr i8, ptr [[TMP0]], i32 2
+// AARCH64-LE-NEXT:    [[TMP5:%.*]] = load i8, ptr [[TMP4]], align 2
+// AARCH64-LE-NEXT:    [[TMP6:%.*]] = and i8 [[TMP5]], 31
+// AARCH64-LE-NEXT:    store i8 [[TMP6]], ptr [[TMP4]], align 2
+// AARCH64-LE-NEXT:    [[TMP7:%.*]] = getelementptr i8, ptr [[TMP0]], i32 3
+// AARCH64-LE-NEXT:    store i8 0, ptr [[TMP7]], align 1
+// AARCH64-LE-NEXT:    ret void
+//
+// AARCH64-BE-LABEL: define dso_local void @testBitfieldPartialEnds(
+// AARCH64-BE-SAME: ptr noundef [[S:%.*]]) #[[ATTR0]] {
+// AARCH64-BE-NEXT:  [[ENTRY:.*:]]
+// AARCH64-BE-NEXT:    [[S_ADDR:%.*]] = alloca ptr, align 8
+// AARCH64-BE-NEXT:    store ptr [[S]], ptr [[S_ADDR]], align 8
+// AARCH64-BE-NEXT:    [[TMP0:%.*]] = load ptr, ptr [[S_ADDR]], align 8
+// AARCH64-BE-NEXT:    [[TMP1:%.*]] = getelementptr i8, ptr [[TMP0]], i32 0
+// AARCH64-BE-NEXT:    [[TMP2:%.*]] = load i8, ptr [[TMP1]], align 4
+// AARCH64-BE-NEXT:    [[TMP3:%.*]] = and i8 [[TMP2]], 31
+// AARCH64-BE-NEXT:    store i8 [[TMP3]], ptr [[TMP1]], align 4
+// AARCH64-BE-NEXT:    [[TMP4:%.*]] = getelementptr i8, ptr [[TMP0]], i32 2
+// AARCH64-BE-NEXT:    [[TMP5:%.*]] = load i8, ptr [[TMP4]], align 2
+// AARCH64-BE-NEXT:    [[TMP6:%.*]] = and i8 [[TMP5]], -8
+// AARCH64-BE-NEXT:    store i8 [[TMP6]], ptr [[TMP4]], align 2
+// AARCH64-BE-NEXT:    [[TMP7:%.*]] = getelementptr i8, ptr [[TMP0]], i32 3
+// AARCH64-BE-NEXT:    store i8 0, ptr [[TMP7]], align 1
+// AARCH64-BE-NEXT:    ret void
+//
+extern "C" void testBitfieldPartialEnds(struct BitfieldPartialEnds *s) {
+  __builtin_clear_padding(s);
+}
+
+// An asymmetric single-bit interval catches endpoint and reflection errors.
+struct BitfieldSingleValueBit {
+  unsigned char : 2;
+  unsigned char value : 1;
+  unsigned char : 5;
+};
+
+// LINUX-LABEL: define dso_local void @testBitfieldSingleValueBit(
+// LINUX-SAME: ptr noundef [[S:%.*]]) #[[ATTR0]] {
+// LINUX-NEXT:  [[ENTRY:.*:]]
+// LINUX-NEXT:    [[S_ADDR:%.*]] = alloca ptr, align 8
+// LINUX-NEXT:    store ptr [[S]], ptr [[S_ADDR]], align 8
+// LINUX-NEXT:    [[TMP0:%.*]] = load ptr, ptr [[S_ADDR]], align 8
+// LINUX-NEXT:    [[TMP1:%.*]] = getelementptr i8, ptr [[TMP0]], i32 0
+// LINUX-NEXT:    [[TMP2:%.*]] = load i8, ptr [[TMP1]], align 1
+// LINUX-NEXT:    [[TMP3:%.*]] = and i8 [[TMP2]], -4
+// LINUX-NEXT:    store i8 [[TMP3]], ptr [[TMP1]], align 1
+// LINUX-NEXT:    [[TMP4:%.*]] = getelementptr i8, ptr [[TMP0]], i32 0
+// LINUX-NEXT:    [[TMP5:%.*]] = load i8, ptr [[TMP4]], align 1
+// LINUX-NEXT:    [[TMP6:%.*]] = and i8 [[TMP5]], 7
+// LINUX-NEXT:    store i8 [[TMP6]], ptr [[TMP4]], align 1
+// LINUX-NEXT:    ret void
+//
+// WINDOWS-LABEL: define dso_local void @testBitfieldSingleValueBit(
+// WINDOWS-SAME: ptr noundef [[S:%.*]]) #[[ATTR0]] {
+// WINDOWS-NEXT:  [[ENTRY:.*:]]
+// WINDOWS-NEXT:    [[S_ADDR:%.*]] = alloca ptr, align 8
+// WINDOWS-NEXT:    store ptr [[S]], ptr [[S_ADDR]], align 8
+// WINDOWS-NEXT:    [[TMP0:%.*]] = load ptr, ptr [[S_ADDR]], align 8
+// WINDOWS-NEXT:    [[TMP1:%.*]] = getelementptr i8, ptr [[TMP0]], i32 0
+// WINDOWS-NEXT:    [[TMP2:%.*]] = load i8, ptr [[TMP1]], align 1
+// WINDOWS-NEXT:    [[TMP3:%.*]] = and i8 [[TMP2]], -4
+// WINDOWS-NEXT:    store i8 [[TMP3]], ptr [[TMP1]], align 1
+// WINDOWS-NEXT:    [[TMP4:%.*]] = getelementptr i8, ptr [[TMP0]], i32 0
+// WINDOWS-NEXT:    [[TMP5:%.*]] = load i8, ptr [[TMP4]], align 1
+// WINDOWS-NEXT:    [[TMP6:%.*]] = and i8 [[TMP5]], 7
+// WINDOWS-NEXT:    store i8 [[TMP6]], ptr [[TMP4]], align 1
+// WINDOWS-NEXT:    ret void
+//
+// ARM-LE-LABEL: define dso_local void @testBitfieldSingleValueBit(
+// ARM-LE-SAME: ptr noundef [[S:%.*]]) #[[ATTR0]] {
+// ARM-LE-NEXT:  [[ENTRY:.*:]]
+// ARM-LE-NEXT:    [[S_ADDR:%.*]] = alloca ptr, align 4
+// ARM-LE-NEXT:    store ptr [[S]], ptr [[S_ADDR]], align 4
+// ARM-LE-NEXT:    [[TMP0:%.*]] = load ptr, ptr [[S_ADDR]], align 4
+// ARM-LE-NEXT:    [[TMP1:%.*]] = getelementptr i8, ptr [[TMP0]], i32 0
+// ARM-LE-NEXT:    [[TMP2:%.*]] = load i8, ptr [[TMP1]], align 1
+// ARM-LE-NEXT:    [[TMP3:%.*]] = and i8 [[TMP2]], -4
+// ARM-LE-NEXT:    store i8 [[TMP3]], ptr [[TMP1]], align 1
+// ARM-LE-NEXT:    [[TMP4:%.*]] = getelementptr i8, ptr [[TMP0]], i32 0
+// ARM-LE-NEXT:    [[TMP5:%.*]] = load i8, ptr [[TMP4]], align 1
+// ARM-LE-NEXT:    [[TMP6:%.*]] = and i8 [[TMP5]], 7
+// ARM-LE-NEXT:    store i8 [[TMP6]], ptr [[TMP4]], align 1
+// ARM-LE-NEXT:    ret void
+//
+// ARM-BE-LABEL: define dso_local void @testBitfieldSingleValueBit(
+// ARM-BE-SAME: ptr noundef [[S:%.*]]) #[[ATTR0]] {
+// ARM-BE-NEXT:  [[ENTRY:.*:]]
+// ARM-BE-NEXT:    [[S_ADDR:%.*]] = alloca ptr, align 4
+// ARM-BE-NEXT:    store ptr [[S]], ptr [[S_ADDR]], align 4
+// ARM-BE-NEXT:    [[TMP0:%.*]] = load ptr, ptr [[S_ADDR]], align 4
+// ARM-BE-NEXT:    [[TMP1:%.*]] = getelementptr i8, ptr [[TMP0]], i32 0
+// ARM-BE-NEXT:    [[TMP2:%.*]] = load i8, ptr [[TMP1]], align 1
+// ARM-BE-NEXT:    [[TMP3:%.*]] = and i8 [[TMP2]], -32
+// ARM-BE-NEXT:    store i8 [[TMP3]], ptr [[TMP1]], align 1
+// ARM-BE-NEXT:    [[TMP4:%.*]] = getelementptr i8, ptr [[TMP0]], i32 0
+// ARM-BE-NEXT:    [[TMP5:%.*]] = load i8, ptr [[TMP4]], align 1
+// ARM-BE-NEXT:    [[TMP6:%.*]] = and i8 [[TMP5]], 63
+// ARM-BE-NEXT:    store i8 [[TMP6]], ptr [[TMP4]], align 1
+// ARM-BE-NEXT:    ret void
+//
+// AARCH64-LE-LABEL: define dso_local void @testBitfieldSingleValueBit(
+// AARCH64-LE-SAME: ptr noundef [[S:%.*]]) #[[ATTR0]] {
+// AARCH64-LE-NEXT:  [[ENTRY:.*:]]
+// AARCH64-LE-NEXT:    [[S_ADDR:%.*]] = alloca ptr, align 8
+// AARCH64-LE-NEXT:    store ptr [[S]], ptr [[S_ADDR]], align 8
+// AARCH64-LE-NEXT:    [[TMP0:%.*]] = load ptr, ptr [[S_ADDR]], align 8
+// AARCH64-LE-NEXT:    [[TMP1:%.*]] = getelementptr i8, ptr [[TMP0]], i32 0
+// AARCH64-LE-NEXT:    [[TMP2:%.*]] = load i8, ptr [[TMP1]], align 1
+// AARCH64-LE-NEXT:    [[TMP3:%.*]] = and i8 [[TMP2]], -4
+// AARCH64-LE-NEXT:    store i8 [[TMP3]], ptr [[TMP1]], align 1
+// AARCH64-LE-NEXT:    [[TMP4:%.*]] = getelementptr i8, ptr [[TMP0]], i32 0
+// AARCH64-LE-NEXT:    [[TMP5:%.*]] = load i8, ptr [[TMP4]], align 1
+// AARCH64-LE-NEXT:    [[TMP6:%.*]] = and i8 [[TMP5]], 7
+// AARCH64-LE-NEXT:    store i8 [[TMP6]], ptr [[TMP4]], align 1
+// AARCH64-LE-NEXT:    ret void
+//
+// AARCH64-BE-LABEL: define dso_local void @testBitfieldSingleValueBit(
+// AARCH64-BE-SAME: ptr noundef [[S:%.*]]) #[[ATTR0]] {
+// AARCH64-BE-NEXT:  [[ENTRY:.*:]]
+// AARCH64-BE-NEXT:    [[S_ADDR:%.*]] = alloca ptr, align 8
+// AARCH64-BE-NEXT:    store ptr [[S]], ptr [[S_ADDR]], align 8
+// AARCH64-BE-NEXT:    [[TMP0:%.*]] = load ptr, ptr [[S_ADDR]], align 8
+// AARCH64-BE-NEXT:    [[TMP1:%.*]] = getelementptr i8, ptr [[TMP0]], i32 0
+// AARCH64-BE-NEXT:    [[TMP2:%.*]] = load i8, ptr [[TMP1]], align 1
+// AARCH64-BE-NEXT:    [[TMP3:%.*]] = and i8 [[TMP2]], -32
+// AARCH64-BE-NEXT:    store i8 [[TMP3]], ptr [[TMP1]], align 1
+// AARCH64-BE-NEXT:    [[TMP4:%.*]] = getelementptr i8, ptr [[TMP0]], i32 0
+// AARCH64-BE-NEXT:    [[TMP5:%.*]] = load i8, ptr [[TMP4]], align 1
+// AARCH64-BE-NEXT:    [[TMP6:%.*]] = and i8 [[TMP5]], 63
+// AARCH64-BE-NEXT:    store i8 [[TMP6]], ptr [[TMP4]], align 1
+// AARCH64-BE-NEXT:    ret void
+//
+extern "C" void testBitfieldSingleValueBit(struct BitfieldSingleValueBit *s) {
+  __builtin_clear_padding(s);
+}
+
+// Ordinary members share bitfield containers on Arm, and separate two runs of
+// bitfields. Clearing padding must preserve every ordinary member.
+struct BitfieldMixedMembers {
+  unsigned char prefix;
+  unsigned int first : 3;
+  unsigned char separator;
+  unsigned int second : 5;
+  unsigned short suffix;
+};
+
+// LINUX-LABEL: define dso_local void @testBitfieldMixedMembers(
+// LINUX-SAME: ptr noundef [[S:%.*]]) #[[ATTR0]] {
+// LINUX-NEXT:  [[ENTRY:.*:]]
+// LINUX-NEXT:    [[S_ADDR:%.*]] = alloca ptr, align 8
+// LINUX-NEXT:    store ptr [[S]], ptr [[S_ADDR]], align 8
+// LINUX-NEXT:    [[TMP0:%.*]] = load ptr, ptr [[S_ADDR]], align 8
+// LINUX-NEXT:    [[TMP1:%.*]] = getelementptr i8, ptr [[TMP0]], i32 1
+// LINUX-NEXT:    [[TMP2:%.*]] = load i8, ptr [[TMP1]], align 1
+// LINUX-NEXT:    [[TMP3:%.*]] = and i8 [[TMP2]], 7
+// LINUX-NEXT:    store i8 [[TMP3]], ptr [[TMP1]], align 1
+// LINUX-NEXT:    [[TMP4:%.*]] = getelementptr i8, ptr [[TMP0]], i32 3
+// LINUX-NEXT:    [[TMP5:%.*]] = load i8, ptr [[TMP4]], align 1
+// LINUX-NEXT:    [[TMP6:%.*]] = and i8 [[TMP5]], 31
+// LINUX-NEXT:    store i8 [[TMP6]], ptr [[TMP4]], align 1
+// LINUX-NEXT:    [[TMP7:%.*]] = getelementptr i8, ptr [[TMP0]], i32 6
+// LINUX-NEXT:    store i8 0, ptr [[TMP7]], align 2
+// LINUX-NEXT:    [[TMP8:%.*]] = getelementptr i8, ptr [[TMP0]], i32 7
+// LINUX-NEXT:    store i8 0, ptr [[TMP8]], align 1
+// LINUX-NEXT:    ret void
+//
+// WINDOWS-LABEL: define dso_local void @testBitfieldMixedMembers(
+// WINDOWS-SAME: ptr noundef [[S:%.*]]) #[[ATTR0]] {
+// WINDOWS-NEXT:  [[ENTRY:.*:]]
+// WINDOWS-NEXT:    [[S_ADDR:%.*]] = alloca ptr, align 8
+// WINDOWS-NEXT:    store ptr [[S]], ptr [[S_ADDR]], align 8
+// WINDOWS-NEXT:    [[TMP0:%.*]] = load ptr, ptr [[S_ADDR]], align 8
+// WINDOWS-NEXT:    [[TMP1:%.*]] = getelementptr i8, ptr [[TMP0]], i32 1
+// WINDOWS-NEXT:    store i8 0, ptr [[TMP1]], align 1
+// WINDOWS-NEXT:    [[TMP2:%.*]] = getelementptr i8, ptr [[TMP0]], i32 2
+// WINDOWS-NEXT:    store i8 0, ptr [[TMP2]], align 2
+// WINDOWS-NEXT:    [[TMP3:%.*]] = getelementptr i8, ptr [[TMP0]], i32 3
+// WINDOWS-NEXT:    store i8 0, ptr [[TMP3]], align 1
+// WINDOWS-NEXT:    [[TMP4:%.*]] = getelementptr i8, ptr [[TMP0]], i32 4
+// WINDOWS-NEXT:    [[TMP5:%.*]] = load i8, ptr [[TMP4]], align 4
+// WINDOWS-NEXT:    [[TMP6:%.*]] = and i8 [[TMP5]], 7
+// WINDOWS-NEXT:    store i8 [[TMP6]], ptr [[TMP4]], align 4
+// WINDOWS-NEXT:    [[TMP7:%.*]] = getelementptr i8, ptr [[TMP0]], i32 5
+// WINDOWS-NEXT:    store i8 0, ptr [[TMP7]], align 1
+// WINDOWS-NEXT:    [[TMP8:%.*]] = getelementptr i8, ptr [[TMP0]], i32 6
+// WINDOWS-NEXT:    store i8 0, ptr [[TMP8]], align 2
+// WINDOWS-NEXT:    [[TMP9:%.*]] = getelementptr i8, ptr [[TMP0]], i32 7
+// WINDOWS-NEXT:    store i8 0, ptr [[TMP9]], align 1
+// WINDOWS-NEXT:    [[TMP10:%.*]] = getelementptr i8, ptr [[TMP0]], i32 9
+// WINDOWS-NEXT:    store i8 0, ptr [[TMP10]], align 1
+// WINDOWS-NEXT:    [[TMP11:%.*]] = getelementptr i8, ptr [[TMP0]], i32 10
+// WINDOWS-NEXT:    store i8 0, ptr [[TMP11]], align 2
+// WINDOWS-NEXT:    [[TMP12:%.*]] = getelementptr i8, ptr [[TMP0]], i32 11
+// WINDOWS-NEXT:    store i8 0, ptr [[TMP12]], align 1
+// WINDOWS-NEXT:    [[TMP13:%.*]] = getelementptr i8, ptr [[TMP0]], i32 12
+// WINDOWS-NEXT:    [[TMP14:%.*]] = load i8, ptr [[TMP13]], align 4
+// WINDOWS-NEXT:    [[TMP15:%.*]] = and i8 [[TMP14]], 31
+// WINDOWS-NEXT:    store i8 [[TMP15]], ptr [[TMP13]], align 4
+// WINDOWS-NEXT:    [[TMP16:%.*]] = getelementptr i8, ptr [[TMP0]], i32 13
+// WINDOWS-NEXT:    store i8 0, ptr [[TMP16]], align 1
+// WINDOWS-NEXT:    [[TMP17:%.*]] = getelementptr i8, ptr [[TMP0]], i32 14
+// WINDOWS-NEXT:    store i8 0, ptr [[TMP17]], align 2
+// WINDOWS-NEXT:    [[TMP18:%.*]] = getelementptr i8, ptr [[TMP0]], i32 15
+// WINDOWS-NEXT:    store i8 0, ptr [[TMP18]], align 1
+// WINDOWS-NEXT:    [[TMP19:%.*]] = getelementptr i8, ptr [[TMP0]], i32 18
+// WINDOWS-NEXT:    store i8 0, ptr [[TMP19]], align 2
+// WINDOWS-NEXT:    [[TMP20:%.*]] = getelementptr i8, ptr [[TMP0]], i32 19
+// WINDOWS-NEXT:    store i8 0, ptr [[TMP20]], align 1
+// WINDOWS-NEXT:    ret void
+//
+// ARM-LE-LABEL: define dso_local void @testBitfieldMixedMembers(
+// ARM-LE-SAME: ptr noundef [[S:%.*]]) #[[ATTR0]] {
+// ARM-LE-NEXT:  [[ENTRY:.*:]]
+// ARM-LE-NEXT:    [[S_ADDR:%.*]] = alloca ptr, align 4
+// ARM-LE-NEXT:    store ptr [[S]], ptr [[S_ADDR]], align 4
+// ARM-LE-NEXT:    [[TMP0:%.*]] = load ptr, ptr [[S_ADDR]], align 4
+// ARM-LE-NEXT:    [[TMP1:%.*]] = getelementptr i8, ptr [[TMP0]], i32 1
+// ARM-LE-NEXT:    [[TMP2:%.*]] = load i8, ptr [[TMP1]], align 1
+// ARM-LE-NEXT:    [[TMP3:%.*]] = and i8 [[TMP2]], 7
+// ARM-LE-NEXT:    store i8 [[TMP3]], ptr [[TMP1]], align 1
+// ARM-LE-NEXT:    [[TMP4:%.*]] = getelementptr i8, ptr [[TMP0]], i32 3
+// ARM-LE-NEXT:    [[TMP5:%.*]] = load i8, ptr [[TMP4]], align 1
+// ARM-LE-NEXT:    [[TMP6:%.*]] = and i8 [[TMP5]], 31
+// ARM-LE-NEXT:    store i8 [[TMP6]], ptr [[TMP4]], align 1
+// ARM-LE-NEXT:    [[TMP7:%.*]] = getelementptr i8, ptr [[TMP0]], i32 6
+// ARM-LE-NEXT:    store i8 0, ptr [[TMP7]], align 2
+// ARM-LE-NEXT:    [[TMP8:%.*]] = getelementptr i8, ptr [[TMP0]], i32 7
+// ARM-LE-NEXT:    store i8 0, ptr [[TMP8]], align 1
+// ARM-LE-NEXT:    ret void
+//
+// ARM-BE-LABEL: define dso_local void @testBitfieldMixedMembers(
+// ARM-BE-SAME: ptr noundef [[S:%.*]]) #[[ATTR0]] {
+// ARM-BE-NEXT:  [[ENTRY:.*:]]
+// ARM-BE-NEXT:    [[S_ADDR:%.*]] = alloca ptr, align 4
+// ARM-BE-NEXT:    store ptr [[S]], ptr [[S_ADDR]], align 4
+// ARM-BE-NEXT:    [[TMP0:%.*]] = load ptr, ptr [[S_ADDR]], align 4
+// ARM-BE-NEXT:    [[TMP1:%.*]] = getelementptr i8, ptr [[TMP0]], i32 1
+// ARM-BE-NEXT:    [[TMP2:%.*]] = load i8, ptr [[TMP1]], align 1
+// ARM-BE-NEXT:    [[TMP3:%.*]] = and i8 [[TMP2]], -32
+// ARM-BE-NEXT:    store i8 [[TMP3]], ptr [[TMP1]], align 1
+// ARM-BE-NEXT:    [[TMP4:%.*]] = getelementptr i8, ptr [[TMP0]], i32 3
+// ARM-BE-NEXT:    [[TMP5:%.*]] = load i8, ptr [[TMP4]], align 1
+// ARM-BE-NEXT:    [[TMP6:%.*]] = and i8 [[TMP5]], -8
+// ARM-BE-NEXT:    store i8 [[TMP6]], ptr [[TMP4]], align 1
+// ARM-BE-NEXT:    [[TMP7:%.*]] = getelementptr i8, ptr [[TMP0]], i32 6
+// ARM-BE-NEXT:    store i8 0, ptr [[TMP7]], align 2
+// ARM-BE-NEXT:    [[TMP8:%.*]] = getelementptr i8, ptr [[TMP0]], i32 7
+// ARM-BE-NEXT:    store i8 0, ptr [[TMP8]], align 1
+// ARM-BE-NEXT:    ret void
+//
+// AARCH64-LE-LABEL: define dso_local void @testBitfieldMixedMembers(
+// AARCH64-LE-SAME: ptr noundef [[S:%.*]]) #[[ATTR0]] {
+// AARCH64-LE-NEXT:  [[ENTRY:.*:]]
+// AARCH64-LE-NEXT:    [[S_ADDR:%.*]] = alloca ptr, align 8
+// AARCH64-LE-NEXT:    store ptr [[S]], ptr [[S_ADDR]], align 8
+// AARCH64-LE-NEXT:    [[TMP0:%.*]] = load ptr, ptr [[S_ADDR]], align 8
+// AARCH64-LE-NEXT:    [[TMP1:%.*]] = getelementptr i8, ptr [[TMP0]], i32 1
+// AARCH64-LE-NEXT:    [[TMP2:%.*]] = load i8, ptr [[TMP1]], align 1
+// AARCH64-LE-NEXT:    [[TMP3:%.*]] = and i8 [[TMP2]], 7
+// AARCH64-LE-NEXT:    store i8 [[TMP3]], ptr [[TMP1]], align 1
+// AARCH64-LE-NEXT:    [[TMP4:%.*]] = getelementptr i8, ptr [[TMP0]], i32 3
+// AARCH64-LE-NEXT:    [[TMP5:%.*]] = load i8, ptr [[TMP4]], align 1
+// AARCH64-LE-NEXT:    [[TMP6:%.*]] = and i8 [[TMP5]], 31
+// AARCH64-LE-NEXT:    store i8 [[TMP6]], ptr [[TMP4]], align 1
+// AARCH64-LE-NEXT:    [[TMP7:%.*]] = getelementptr i8, ptr [[TMP0]], i32 6
+// AARCH64-LE-NEXT:    store i8 0, ptr [[TMP7]], align 2
+// AARCH64-LE-NEXT:    [[TMP8:%.*]] = getelementptr i8, ptr [[TMP0]], i32 7
+// AARCH64-LE-NEXT:    store i8 0, ptr [[TMP8]], align 1
+// AARCH64-LE-NEXT:    ret void
+//
+// AARCH64-BE-LABEL: define dso_local void @testBitfieldMixedMembers(
+// AARCH64-BE-SAME: ptr noundef [[S:%.*]]) #[[ATTR0]] {
+// AARCH64-BE-NEXT:  [[ENTRY:.*:]]
+// AARCH64-BE-NEXT:    [[S_ADDR:%.*]] = alloca ptr, align 8
+// AARCH64-BE-NEXT:    store ptr [[S]], ptr [[S_ADDR]], align 8
+// AARCH64-BE-NEXT:    [[TMP0:%.*]] = load ptr, ptr [[S_ADDR]], align 8
+// AARCH64-BE-NEXT:    [[TMP1:%.*]] = getelementptr i8, ptr [[TMP0]], i32 1
+// AARCH64-BE-NEXT:    [[TMP2:%.*]] = load i8, ptr [[TMP1]], align 1
+// AARCH64-BE-NEXT:    [[TMP3:%.*]] = and i8 [[TMP2]], -32
+// AARCH64-BE-NEXT:    store i8 [[TMP3]], ptr [[TMP1]], align 1
+// AARCH64-BE-NEXT:    [[TMP4:%.*]] = getelementptr i8, ptr [[TMP0]], i32 3
+// AARCH64-BE-NEXT:    [[TMP5:%.*]] = load i8, ptr [[TMP4]], align 1
+// AARCH64-BE-NEXT:    [[TMP6:%.*]] = and i8 [[TMP5]], -8
+// AARCH64-BE-NEXT:    store i8 [[TMP6]], ptr [[TMP4]], align 1
+// AARCH64-BE-NEXT:    [[TMP7:%.*]] = getelementptr i8, ptr [[TMP0]], i32 6
+// AARCH64-BE-NEXT:    store i8 0, ptr [[TMP7]], align 2
+// AARCH64-BE-NEXT:    [[TMP8:%.*]] = getelementptr i8, ptr [[TMP0]], i32 7
+// AARCH64-BE-NEXT:    store i8 0, ptr [[TMP8]], align 1
+// AARCH64-BE-NEXT:    ret void
+//
+extern "C" void testBitfieldMixedMembers(struct BitfieldMixedMembers *s) {
+  __builtin_clear_padding(s);
+}
+
+// LINUX-LABEL: define dso_local void @testBitfieldMixedMembersVolatileObject(
+// LINUX-SAME: ptr noundef [[S:%.*]]) #[[ATTR0]] {
+// LINUX-NEXT:  [[ENTRY:.*:]]
+// LINUX-NEXT:    [[S_ADDR:%.*]] = alloca ptr, align 8
+// LINUX-NEXT:    store ptr [[S]], ptr [[S_ADDR]], align 8
+// LINUX-NEXT:    [[TMP0:%.*]] = load ptr, ptr [[S_ADDR]], align 8
+// LINUX-NEXT:    [[TMP1:%.*]] = getelementptr i8, ptr [[TMP0]], i32 1
+// LINUX-NEXT:    [[TMP2:%.*]] = load i8, ptr [[TMP1]], align 1
+// LINUX-NEXT:    [[TMP3:%.*]] = and i8 [[TMP2]], 7
+// LINUX-NEXT:    store i8 [[TMP3]], ptr [[TMP1]], align 1
+// LINUX-NEXT:    [[TMP4:%.*]] = getelementptr i8, ptr [[TMP0]], i32 3
+// LINUX-NEXT:    [[TMP5:%.*]] = load i8, ptr [[TMP4]], align 1
+// LINUX-NEXT:    [[TMP6:%.*]] = and i8 [[TMP5]], 31
+// LINUX-NEXT:    store i8 [[TMP6]], ptr [[TMP4]], align 1
+// LINUX-NEXT:    [[TMP7:%.*]] = getelementptr i8, ptr [[TMP0]], i32 6
+// LINUX-NEXT:    store i8 0, ptr [[TMP7]], align 2
+// LINUX-NEXT:    [[TMP8:%.*]] = getelementptr i8, ptr [[TMP0]], i32 7
+// LINUX-NEXT:    store i8 0, ptr [[TMP8]], align 1
+// LINUX-NEXT:    ret void
+//
+// WINDOWS-LABEL: define dso_local void @testBitfieldMixedMembersVolatileObject(
+// WINDOWS-SAME: ptr noundef [[S:%.*]]) #[[ATTR0]] {
+// WINDOWS-NEXT:  [[ENTRY:.*:]]
+// WINDOWS-NEXT:    [[S_ADDR:%.*]] = alloca ptr, align 8
+// WINDOWS-NEXT:    store ptr [[S]], ptr [[S_ADDR]], align 8
+// WINDOWS-NEXT:    [[TMP0:%.*]] = load ptr, ptr [[S_ADDR]], align 8
+// WINDOWS-NEXT:    [[TMP1:%.*]] = getelementptr i8, ptr [[TMP0]], i32 1
+// WINDOWS-NEXT:    store i8 0, ptr [[TMP1]], align 1
+// WINDOWS-NEXT:    [[TMP2:%.*]] = getelementptr i8, ptr [[TMP0]], i32 2
+// WINDOWS-NEXT:    store i8 0, ptr [[TMP2]], align 2
+// WINDOWS-NEXT:    [[TMP3:%.*]] = getelementptr i8, ptr [[TMP0]], i32 3
+// WINDOWS-NEXT:    store i8 0, ptr [[TMP3]], align 1
+// WINDOWS-NEXT:    [[TMP4:%.*]] = getelementptr i8, ptr [[TMP0]], i32 4
+// WINDOWS-NEXT:    [[TMP5:%.*]] = load i8, ptr [[TMP4]], align 4
+// WINDOWS-NEXT:    [[TMP6:%.*]] = and i8 [[TMP5]], 7
+// WINDOWS-NEXT:    store i8 [[TMP6]], ptr [[TMP4]], align 4
+// WINDOWS-NEXT:    [[TMP7:%.*]] = getelementptr i8, ptr [[TMP0]], i32 5
+// WINDOWS-NEXT:    store i8 0, ptr [[TMP7]], align 1
+// WINDOWS-NEXT:    [[TMP8:%.*]] = getelementptr i8, ptr [[TMP0]], i32 6
+// WINDOWS-NEXT:    store i8 0, ptr [[TMP8]], align 2
+// WINDOWS-NEXT:    [[TMP9:%.*]] = getelementptr i8, ptr [[TMP0]], i32 7
+// WINDOWS-NEXT:    store i8 0, ptr [[TMP9]], align 1
+// WINDOWS-NEXT:    [[TMP10:%.*]] = getelementptr i8, ptr [[TMP0]], i32 9
+// WINDOWS-NEXT:    store i8 0, ptr [[TMP10]], align 1
+// WINDOWS-NEXT:    [[TMP11:%.*]] = getelementptr i8, ptr [[TMP0]], i32 10
+// WINDOWS-NEXT:    store i8 0, ptr [[TMP11]], align 2
+// WINDOWS-NEXT:    [[TMP12:%.*]] = getelementptr i8, ptr [[TMP0]], i32 11
+// WINDOWS-NEXT:    store i8 0, ptr [[TMP12]], align 1
+// WINDOWS-NEXT:    [[TMP13:%.*]] = getelementptr i8, ptr [[TMP0]], i32 12
+// WINDOWS-NEXT:    [[TMP14:%.*]] = load i8, ptr [[TMP13]], align 4
+// WINDOWS-NEXT:    [[TMP15:%.*]] = and i8 [[TMP14]], 31
+// WINDOWS-NEXT:    store i8 [[TMP15]], ptr [[TMP13]], align 4
+// WINDOWS-NEXT:    [[TMP16:%.*]] = getelementptr i8, ptr [[TMP0]], i32 13
+// WINDOWS-NEXT:    store i8 0, ptr [[TMP16]], align 1
+// WINDOWS-NEXT:    [[TMP17:%.*]] = getelementptr i8, ptr [[TMP0]], i32 14
+// WINDOWS-NEXT:    store i8 0, ptr [[TMP17]], align 2
+// WINDOWS-NEXT:    [[TMP18:%.*]] = getelementptr i8, ptr [[TMP0]], i32 15
+// WINDOWS-NEXT:    store i8 0, ptr [[TMP18]], align 1
+// WINDOWS-NEXT:    [[TMP19:%.*]] = getelementptr i8, ptr [[TMP0]], i32 18
+// WINDOWS-NEXT:    store i8 0, ptr [[TMP19]], align 2
+// WINDOWS-NEXT:    [[TMP20:%.*]] = getelementptr i8, ptr [[TMP0]], i32 19
+// WINDOWS-NEXT:    store i8 0, ptr [[TMP20]], align 1
+// WINDOWS-NEXT:    ret void
+//
+// ARM-LE-LABEL: define dso_local void @testBitfieldMixedMembersVolatileObject(
+// ARM-LE-SAME: ptr noundef [[S:%.*]]) #[[ATTR0]] {
+// ARM-LE-NEXT:  [[ENTRY:.*:]]
+// ARM-LE-NEXT:    [[S_ADDR:%.*]] = alloca ptr, align 4
+// ARM-LE-NEXT:    store ptr [[S]], ptr [[S_ADDR]], align 4
+// ARM-LE-NEXT:    [[TMP0:%.*]] = load ptr, ptr [[S_ADDR]], align 4
+// ARM-LE-NEXT:    [[TMP1:%.*]] = getelementptr i8, ptr [[TMP0]], i32 1
+// ARM-LE-NEXT:    [[TMP2:%.*]] = load i8, ptr [[TMP1]], align 1
+// ARM-LE-NEXT:    [[TMP3:%.*]] = and i8 [[TMP2]], 7
+// ARM-LE-NEXT:    store i8 [[TMP3]], ptr [[TMP1]], align 1
+// ARM-LE-NEXT:    [[TMP4:%.*]] = getelementptr i8, ptr [[TMP0]], i32 3
+// ARM-LE-NEXT:    [[TMP5:%.*]] = load i8, ptr [[TMP4]], align 1
+// ARM-LE-NEXT:    [[TMP6:%.*]] = and i8 [[TMP5]], 31
+// ARM-LE-NEXT:    store i8 [[TMP6]], ptr [[TMP4]], align 1
+// ARM-LE-NEXT:    [[TMP7:%.*]] = getelementptr i8, ptr [[TMP0]], i32 6
+// ARM-LE-NEXT:    store i8 0, ptr [[TMP7]], align 2
+// ARM-LE-NEXT:    [[TMP8:%.*]] = getelementptr i8, ptr [[TMP0]], i32 7
+// ARM-LE-NEXT:    store i8 0, ptr [[TMP8]], align 1
+// ARM-LE-NEXT:    ret void
+//
+// ARM-BE-LABEL: define dso_local void @testBitfieldMixedMembersVolatileObject(
+// ARM-BE-SAME: ptr noundef [[S:%.*]]) #[[ATTR0]] {
+// ARM-BE-NEXT:  [[ENTRY:.*:]]
+// ARM-BE-NEXT:    [[S_ADDR:%.*]] = alloca ptr, align 4
+// ARM-BE-NEXT:    store ptr [[S]], ptr [[S_ADDR]], align 4
+// ARM-BE-NEXT:    [[TMP0:%.*]] = load ptr, ptr [[S_ADDR]], align 4
+// ARM-BE-NEXT:    [[TMP1:%.*]] = getelementptr i8, ptr [[TMP0]], i32 1
+// ARM-BE-NEXT:    [[TMP2:%.*]] = load i8, ptr [[TMP1]], align 1
+// ARM-BE-NEXT:    [[TMP3:%.*]] = and i8 [[TMP2]], -32
+// ARM-BE-NEXT:    store i8 [[TMP3]], ptr [[TMP1]], align 1
+// ARM-BE-NEXT:    [[TMP4:%.*]] = getelementptr i8, ptr [[TMP0]], i32 3
+// ARM-BE-NEXT:    [[TMP5:%.*]] = load i8, ptr [[TMP4]], align 1
+// ARM-BE-NEXT:    [[TMP6:%.*]] = and i8 [[TMP5]], -8
+// ARM-BE-NEXT:    store i8 [[TMP6]], ptr [[TMP4]], align 1
+// ARM-BE-NEXT:    [[TMP7:%.*]] = getelementptr i8, ptr [[TMP0]], i32 6
+// ARM-BE-NEXT:    store i8 0, ptr [[TMP7]], align 2
+// ARM-BE-NEXT:    [[TMP8:%.*]] = getelementptr i8, ptr [[TMP0]], i32 7
+// ARM-BE-NEXT:    store i8 0, ptr [[TMP8]], align 1
+// ARM-BE-NEXT:    ret void
+//
+// AARCH64-LE-LABEL: define dso_local void @testBitfieldMixedMembersVolatileObject(
+// AARCH64-LE-SAME: ptr noundef [[S:%.*]]) #[[ATTR0]] {
+// AARCH64-LE-NEXT:  [[ENTRY:.*:]]
+// AARCH64-LE-NEXT:    [[S_ADDR:%.*]] = alloca ptr, align 8
+// AARCH64-LE-NEXT:    store ptr [[S]], ptr [[S_ADDR]], align 8
+// AARCH64-LE-NEXT:    [[TMP0:%.*]] = load ptr, ptr [[S_ADDR]], align 8
+// AARCH64-LE-NEXT:    [[TMP1:%.*]] = getelementptr i8, ptr [[TMP0]], i32 1
+// AARCH64-LE-NEXT:    [[TMP2:%.*]] = load i8, ptr [[TMP1]], align 1
+// AARCH64-LE-NEXT:    [[TMP3:%.*]] = and i8 [[TMP2]], 7
+// AARCH64-LE-NEXT:    store i8 [[TMP3]], ptr [[TMP1]], align 1
+// AARCH64-LE-NEXT:    [[TMP4:%.*]] = getelementptr i8, ptr [[TMP0]], i32 3
+// AARCH64-LE-NEXT:    [[TMP5:%.*]] = load i8, ptr [[TMP4]], align 1
+// AARCH64-LE-NEXT:    [[TMP6:%.*]] = and i8 [[TMP5]], 31
+// AARCH64-LE-NEXT:    store i8 [[TMP6]], ptr [[TMP4]], align 1
+// AARCH64-LE-NEXT:    [[TMP7:%.*]] = getelementptr i8, ptr [[TMP0]], i32 6
+// AARCH64-LE-NEXT:    store i8 0, ptr [[TMP7]], align 2
+// AARCH64-LE-NEXT:    [[TMP8:%.*]] = getelementptr i8, ptr [[TMP0]], i32 7
+// AARCH64-LE-NEXT:    store i8 0, ptr [[TMP8]], align 1
+// AARCH64-LE-NEXT:    ret void
+//
+// AARCH64-BE-LABEL: define dso_local void @testBitfieldMixedMembersVolatileObject(
+// AARCH64-BE-SAME: ptr noundef [[S:%.*]]) #[[ATTR0]] {
+// AARCH64-BE-NEXT:  [[ENTRY:.*:]]
+// AARCH64-BE-NEXT:    [[S_ADDR:%.*]] = alloca ptr, align 8
+// AARCH64-BE-NEXT:    store ptr [[S]], ptr [[S_ADDR]], align 8
+// AARCH64-BE-NEXT:    [[TMP0:%.*]] = load ptr, ptr [[S_ADDR]], align 8
+// AARCH64-BE-NEXT:    [[TMP1:%.*]] = getelementptr i8, ptr [[TMP0]], i32 1
+// AARCH64-BE-NEXT:    [[TMP2:%.*]] = load i8, ptr [[TMP1]], align 1
+// AARCH64-BE-NEXT:    [[TMP3:%.*]] = and i8 [[TMP2]], -32
+// AARCH64-BE-NEXT:    store i8 [[TMP3]], ptr [[TMP1]], align 1
+// AARCH64-BE-NEXT:    [[TMP4:%.*]] = getelementptr i8, ptr [[TMP0]], i32 3
+// AARCH64-BE-NEXT:    [[TMP5:%.*]] = load i8, ptr [[TMP4]], align 1
+// AARCH64-BE-NEXT:    [[TMP6:%.*]] = and i8 [[TMP5]], -8
+// AARCH64-BE-NEXT:    store i8 [[TMP6]], ptr [[TMP4]], align 1
+// AARCH64-BE-NEXT:    [[TMP7:%.*]] = getelementptr i8, ptr [[TMP0]], i32 6
+// AARCH64-BE-NEXT:    store i8 0, ptr [[TMP7]], align 2
+// AARCH64-BE-NEXT:    [[TMP8:%.*]] = getelementptr i8, ptr [[TMP0]], i32 7
+// AARCH64-BE-NEXT:    store i8 0, ptr [[TMP8]], align 1
+// AARCH64-BE-NEXT:    ret void
+//
+extern "C" void testBitfieldMixedMembersVolatileObject(
+    volatile struct BitfieldMixedMembers *s) {
+  __builtin_clear_padding(s);
+}
+
+struct BitfieldMixedVolatileMembers {
+  unsigned char prefix;
+  volatile unsigned int first : 3;
+  unsigned char separator;
+  volatile unsigned int second : 5;
+  unsigned short suffix;
+};
+
+// LINUX-LABEL: define dso_local void @testBitfieldMixedVolatileMembers(
+// LINUX-SAME: ptr noundef [[S:%.*]]) #[[ATTR0]] {
+// LINUX-NEXT:  [[ENTRY:.*:]]
+// LINUX-NEXT:    [[S_ADDR:%.*]] = alloca ptr, align 8
+// LINUX-NEXT:    store ptr [[S]], ptr [[S_ADDR]], align 8
+// LINUX-NEXT:    [[TMP0:%.*]] = load ptr, ptr [[S_ADDR]], align 8
+// LINUX-NEXT:    [[TMP1:%.*]] = getelementptr i8, ptr [[TMP0]], i32 1
+// LINUX-NEXT:    [[TMP2:%.*]] = load i8, ptr [[TMP1]], align 1
+// LINUX-NEXT:    [[TMP3:%.*]] = and i8 [[TMP2]], 7
+// LINUX-NEXT:    store i8 [[TMP3]], ptr [[TMP1]], align 1
+// LINUX-NEXT:    [[TMP4:%.*]] = getelementptr i8, ptr [[TMP0]], i32 3
+// LINUX-NEXT:    [[TMP5:%.*]] = load i8, ptr [[TMP4]], align 1
+// LINUX-NEXT:    [[TMP6:%.*]] = and i8 [[TMP5]], 31
+// LINUX-NEXT:    store i8 [[TMP6]], ptr [[TMP4]], align 1
+// LINUX-NEXT:    [[TMP7:%.*]] = getelementptr i8, ptr [[TMP0]], i32 6
+// LINUX-NEXT:    store i8 0, ptr [[TMP7]], align 2
+// LINUX-NEXT:    [[TMP8:%.*]] = getelementptr i8, ptr [[TMP0]], i32 7
+// LINUX-NEXT:    store i8 0, ptr [[TMP8]], align 1
+// LINUX-NEXT:    ret void
+//
+// WINDOWS-LABEL: define dso_local void @testBitfieldMixedVolatileMembers(
+// WINDOWS-SAME: ptr noundef [[S:%.*]]) #[[ATTR0]] {
+// WINDOWS-NEXT:  [[ENTRY:.*:]]
+// WINDOWS-NEXT:    [[S_ADDR:%.*]] = alloca ptr, align 8
+// WINDOWS-NEXT:    store ptr [[S]], ptr [[S_ADDR]], align 8
+// WINDOWS-NEXT:    [[TMP0:%.*]] = load ptr, ptr [[S_ADDR]], align 8
+// WINDOWS-NEXT:    [[TMP1:%.*]] = getelementptr i8, ptr [[TMP0]], i32 1
+// WINDOWS-NEXT:    store i8 0, ptr [[TMP1]], align 1
+// WINDOWS-NEXT:    [[TMP2:%.*]] = getelementptr i8, ptr [[TMP0]], i32 2
+// WINDOWS-NEXT:    store i8 0, ptr [[TMP2]], align 2
+// WINDOWS-NEXT:    [[TMP3:%.*]] = getelementptr i8, ptr [[TMP0]], i32 3
+// WINDOWS-NEXT:    store i8 0, ptr [[TMP3]], align 1
+// WINDOWS-NEXT:    [[TMP4:%.*]] = getelementptr i8, ptr [[TMP0]], i32 4
+// WINDOWS-NEXT:    [[TMP5:%.*]] = load i8, ptr [[TMP4]], align 4
+// WINDOWS-NEXT:    [[TMP6:%.*]] = and i8 [[TMP5]], 7
+// WINDOWS-NEXT:    store i8 [[TMP6]], ptr [[TMP4]], align 4
+// WINDOWS-NEXT:    [[TMP7:%.*]] = getelementptr i8, ptr [[TMP0]], i32 5
+// WINDOWS-NEXT:    store i8 0, ptr [[TMP7]], align 1
+// WINDOWS-NEXT:    [[TMP8:%.*]] = getelementptr i8, ptr [[TMP0]], i32 6
+// WINDOWS-NEXT:    store i8 0, ptr [[TMP8]], align 2
+// WINDOWS-NEXT:    [[TMP9:%.*]] = getelementptr i8, ptr [[TMP0]], i32 7
+// WINDOWS-NEXT:    store i8 0, ptr [[TMP9]], align 1
+// WINDOWS-NEXT:    [[TMP10:%.*]] = getelementptr i8, ptr [[TMP0]], i32 9
+// WINDOWS-NEXT:    store i8 0, ptr [[TMP10]], align 1
+// WINDOWS-NEXT:    [[TMP11:%.*]] = getelementptr i8, ptr [[TMP0]], i32 10
+// WINDOWS-NEXT:    store i8 0, ptr [[TMP11]], align 2
+// WINDOWS-NEXT:    [[TMP12:%.*]] = getelementptr i8, ptr [[TMP0]], i32 11
+// WINDOWS-NEXT:    store i8 0, ptr [[TMP12]], align 1
+// WINDOWS-NEXT:    [[TMP13:%.*]] = getelementptr i8, ptr [[TMP0]], i32 12
+// WINDOWS-NEXT:    [[TMP14:%.*]] = load i8, ptr [[TMP13]], align 4
+// WINDOWS-NEXT:    [[TMP15:%.*]] = and i8 [[TMP14]], 31
+// WINDOWS-NEXT:    store i8 [[TMP15]], ptr [[TMP13]], align 4
+// WINDOWS-NEXT:    [[TMP16:%.*]] = getelementptr i8, ptr [[TMP0]], i32 13
+// WINDOWS-NEXT:    store i8 0, ptr [[TMP16]], align 1
+// WINDOWS-NEXT:    [[TMP17:%.*]] = getelementptr i8, ptr [[TMP0]], i32 14
+// WINDOWS-NEXT:    store i8 0, ptr [[TMP17]], align 2
+// WINDOWS-NEXT:    [[TMP18:%.*]] = getelementptr i8, ptr [[TMP0]], i32 15
+// WINDOWS-NEXT:    store i8 0, ptr [[TMP18]], align 1
+// WINDOWS-NEXT:    [[TMP19:%.*]] = getelementptr i8, ptr [[TMP0]], i32 18
+// WINDOWS-NEXT:    store i8 0, ptr [[TMP19]], align 2
+// WINDOWS-NEXT:    [[TMP20:%.*]] = getelementptr i8, ptr [[TMP0]], i32 19
+// WINDOWS-NEXT:    store i8 0, ptr [[TMP20]], align 1
+// WINDOWS-NEXT:    ret void
+//
+// ARM-LE-LABEL: define dso_local void @testBitfieldMixedVolatileMembers(
+// ARM-LE-SAME: ptr noundef [[S:%.*]]) #[[ATTR0]] {
+// ARM-LE-NEXT:  [[ENTRY:.*:]]
+// ARM-LE-NEXT:    [[S_ADDR:%.*]] = alloca ptr, align 4
+// ARM-LE-NEXT:    store ptr [[S]], ptr [[S_ADDR]], align 4
+// ARM-LE-NEXT:    [[TMP0:%.*]] = load ptr, ptr [[S_ADDR]], align 4
+// ARM-LE-NEXT:    [[TMP1:%.*]] = getelementptr i8, ptr [[TMP0]], i32 1
+// ARM-LE-NEXT:    [[TMP2:%.*]] = load i8, ptr [[TMP1]], align 1
+// ARM-LE-NEXT:    [[TMP3:%.*]] = and i8 [[TMP2]], 7
+// ARM-LE-NEXT:    store i8 [[TMP3]], ptr [[TMP1]], align 1
+// ARM-LE-NEXT:    [[TMP4:%.*]] = getelementptr i8, ptr [[TMP0]], i32 3
+// ARM-LE-NEXT:    [[TMP5:%.*]] = load i8, ptr [[TMP4]], align 1
+// ARM-LE-NEXT:    [[TMP6:%.*]] = and i8 [[TMP5]], 31
+// ARM-LE-NEXT:    store i8 [[TMP6]], ptr [[TMP4]], align 1
+// ARM-LE-NEXT:    [[TMP7:%.*]] = getelementptr i8, ptr [[TMP0]], i32 6
+// ARM-LE-NEXT:    store i8 0, ptr [[TMP7]], align 2
+// ARM-LE-NEXT:    [[TMP8:%.*]] = getelementptr i8, ptr [[TMP0]], i32 7
+// ARM-LE-NEXT:    store i8 0, ptr [[TMP8]], align 1
+// ARM-LE-NEXT:    ret void
+//
+// ARM-BE-LABEL: define dso_local void @testBitfieldMixedVolatileMembers(
+// ARM-BE-SAME: ptr noundef [[S:%.*]]) #[[ATTR0]] {
+// ARM-BE-NEXT:  [[ENTRY:.*:]]
+// ARM-BE-NEXT:    [[S_ADDR:%.*]] = alloca ptr, align 4
+// ARM-BE-NEXT:    store ptr [[S]], ptr [[S_ADDR]], align 4
+// ARM-BE-NEXT:    [[TMP0:%.*]] = load ptr, ptr [[S_ADDR]], align 4
+// ARM-BE-NEXT:    [[TMP1:%.*]] = getelementptr i8, ptr [[TMP0]], i32 1
+// ARM-BE-NEXT:    [[TMP2:%.*]] = load i8, ptr [[TMP1]], align 1
+// ARM-BE-NEXT:    [[TMP3:%.*]] = and i8 [[TMP2]], -32
+// ARM-BE-NEXT:    store i8 [[TMP3]], ptr [[TMP1]], align 1
+// ARM-BE-NEXT:    [[TMP4:%.*]] = getelementptr i8, ptr [[TMP0]], i32 3
+// ARM-BE-NEXT:    [[TMP5:%.*]] = load i8, ptr [[TMP4]], align 1
+// ARM-BE-NEXT:    [[TMP6:%.*]] = and i8 [[TMP5]], -8
+// ARM-BE-NEXT:    store i8 [[TMP6]], ptr [[TMP4]], align 1
+// ARM-BE-NEXT:    [[TMP7:%.*]] = getelementptr i8, ptr [[TMP0]], i32 6
+// ARM-BE-NEXT:    store i8 0, ptr [[TMP7]], align 2
+// ARM-BE-NEXT:    [[TMP8:%.*]] = getelementptr i8, ptr [[TMP0]], i32 7
+// ARM-BE-NEXT:    store i8 0, ptr [[TMP8]], align 1
+// ARM-BE-NEXT:    ret void
+//
+// AARCH64-LE-LABEL: define dso_local void @testBitfieldMixedVolatileMembers(
+// AARCH64-LE-SAME: ptr noundef [[S:%.*]]) #[[ATTR0]] {
+// AARCH64-LE-NEXT:  [[ENTRY:.*:]]
+// AARCH64-LE-NEXT:    [[S_ADDR:%.*]] = alloca ptr, align 8
+// AARCH64-LE-NEXT:    store ptr [[S]], ptr [[S_ADDR]], align 8
+// AARCH64-LE-NEXT:    [[TMP0:%.*]] = load ptr, ptr [[S_ADDR]], align 8
+// AARCH64-LE-NEXT:    [[TMP1:%.*]] = getelementptr i8, ptr [[TMP0]], i32 1
+// AARCH64-LE-NEXT:    [[TMP2:%.*]] = load i8, ptr [[TMP1]], align 1
+// AARCH64-LE-NEXT:    [[TMP3:%.*]] = and i8 [[TMP2]], 7
+// AARCH64-LE-NEXT:    store i8 [[TMP3]], ptr [[TMP1]], align 1
+// AARCH64-LE-NEXT:    [[TMP4:%.*]] = getelementptr i8, ptr [[TMP0]], i32 3
+// AARCH64-LE-NEXT:    [[TMP5:%.*]] = load i8, ptr [[TMP4]], align 1
+// AARCH64-LE-NEXT:    [[TMP6:%.*]] = and i8 [[TMP5]], 31
+// AARCH64-LE-NEXT:    store i8 [[TMP6]], ptr [[TMP4]], align 1
+// AARCH64-LE-NEXT:    [[TMP7:%.*]] = getelementptr i8, ptr [[TMP0]], i32 6
+// AARCH64-LE-NEXT:    store i8 0, ptr [[TMP7]], align 2
+// AARCH64-LE-NEXT:    [[TMP8:%.*]] = getelementptr i8, ptr [[TMP0]], i32 7
+// AARCH64-LE-NEXT:    store i8 0, ptr [[TMP8]], align 1
+// AARCH64-LE-NEXT:    ret void
+//
+// AARCH64-BE-LABEL: define dso_local void @testBitfieldMixedVolatileMembers(
+// AARCH64-BE-SAME: ptr noundef [[S:%.*]]) #[[ATTR0]] {
+// AARCH64-BE-NEXT:  [[ENTRY:.*:]]
+// AARCH64-BE-NEXT:    [[S_ADDR:%.*]] = alloca ptr, align 8
+// AARCH64-BE-NEXT:    store ptr [[S]], ptr [[S_ADDR]], align 8
+// AARCH64-BE-NEXT:    [[TMP0:%.*]] = load ptr, ptr [[S_ADDR]], align 8
+// AARCH64-BE-NEXT:    [[TMP1:%.*]] = getelementptr i8, ptr [[TMP0]], i32 1
+// AARCH64-BE-NEXT:    [[TMP2:%.*]] = load i8, ptr [[TMP1]], align 1
+// AARCH64-BE-NEXT:    [[TMP3:%.*]] = and i8 [[TMP2]], -32
+// AARCH64-BE-NEXT:    store i8 [[TMP3]], ptr [[TMP1]], align 1
+// AARCH64-BE-NEXT:    [[TMP4:%.*]] = getelementptr i8, ptr [[TMP0]], i32 3
+// AARCH64-BE-NEXT:    [[TMP5:%.*]] = load i8, ptr [[TMP4]], align 1
+// AARCH64-BE-NEXT:    [[TMP6:%.*]] = and i8 [[TMP5]], -8
+// AARCH64-BE-NEXT:    store i8 [[TMP6]], ptr [[TMP4]], align 1
+// AARCH64-BE-NEXT:    [[TMP7:%.*]] = getelementptr i8, ptr [[TMP0]], i32 6
+// AARCH64-BE-NEXT:    store i8 0, ptr [[TMP7]], align 2
+// AARCH64-BE-NEXT:    [[TMP8:%.*]] = getelementptr i8, ptr [[TMP0]], i32 7
+// AARCH64-BE-NEXT:    store i8 0, ptr [[TMP8]], align 1
+// AARCH64-BE-NEXT:    ret void
+//
+extern "C" void testBitfieldMixedVolatileMembers(
+    struct BitfieldMixedVolatileMembers *s) {
+  __builtin_clear_padding(s);
+}
+
+// LINUX-LABEL: define dso_local void @testBitfieldMixedVolatileMembersAndObject(
+// LINUX-SAME: ptr noundef [[S:%.*]]) #[[ATTR0]] {
+// LINUX-NEXT:  [[ENTRY:.*:]]
+// LINUX-NEXT:    [[S_ADDR:%.*]] = alloca ptr, align 8
+// LINUX-NEXT:    store ptr [[S]], ptr [[S_ADDR]], align 8
+// LINUX-NEXT:    [[TMP0:%.*]] = load ptr, ptr [[S_ADDR]], align 8
+// LINUX-NEXT:    [[TMP1:%.*]] = getelementptr i8, ptr [[TMP0]], i32 1
+// LINUX-NEXT:    [[TMP2:%.*]] = load i8, ptr [[TMP1]], align 1
+// LINUX-NEXT:    [[TMP3:%.*]] = and i8 [[TMP2]], 7
+// LINUX-NEXT:    store i8 [[TMP3]], ptr [[TMP1]], align 1
+// LINUX-NEXT:    [[TMP4:%.*]] = getelementptr i8, ptr [[TMP0]], i32 3
+// LINUX-NEXT:    [[TMP5:%.*]] = load i8, ptr [[TMP4]], align 1
+// LINUX-NEXT:    [[TMP6:%.*]] = and i8 [[TMP5]], 31
+// LINUX-NEXT:    store i8 [[TMP6]], ptr [[TMP4]], align 1
+// LINUX-NEXT:    [[TMP7:%.*]] = getelementptr i8, ptr [[TMP0]], i32 6
+// LINUX-NEXT:    store i8 0, ptr [[TMP7]], align 2
+// LINUX-NEXT:    [[TMP8:%.*]] = getelementptr i8, ptr [[TMP0]], i32 7
+// LINUX-NEXT:    store i8 0, ptr [[TMP8]], align 1
+// LINUX-NEXT:    ret void
+//
+// WINDOWS-LABEL: define dso_local void @testBitfieldMixedVolatileMembersAndObject(
+// WINDOWS-SAME: ptr noundef [[S:%.*]]) #[[ATTR0]] {
+// WINDOWS-NEXT:  [[ENTRY:.*:]]
+// WINDOWS-NEXT:    [[S_ADDR:%.*]] = alloca ptr, align 8
+// WINDOWS-NEXT:    store ptr [[S]], ptr [[S_ADDR]], align 8
+// WINDOWS-NEXT:    [[TMP0:%.*]] = load ptr, ptr [[S_ADDR]], align 8
+// WINDOWS-NEXT:    [[TMP1:%.*]] = getelementptr i8, ptr [[TMP0]], i32 1
+// WINDOWS-NEXT:    store i8 0, ptr [[TMP1]], align 1
+// WINDOWS-NEXT:    [[TMP2:%.*]] = getelementptr i8, ptr [[TMP0]], i32 2
+// WINDOWS-NEXT:    store i8 0, ptr [[TMP2]], align 2
+// WINDOWS-NEXT:    [[TMP3:%.*]] = getelementptr i8, ptr [[TMP0]], i32 3
+// WINDOWS-NEXT:    store i8 0, ptr [[TMP3]], align 1
+// WINDOWS-NEXT:    [[TMP4:%.*]] = getelementptr i8, ptr [[TMP0]], i32 4
+// WINDOWS-NEXT:    [[TMP5:%.*]] = load i8, ptr [[TMP4]], align 4
+// WINDOWS-NEXT:    [[TMP6:%.*]] = and i8 [[TMP5]], 7
+// WINDOWS-NEXT:    store i8 [[TMP6]], ptr [[TMP4]], align 4
+// WINDOWS-NEXT:    [[TMP7:%.*]] = getelementptr i8, ptr [[TMP0]], i32 5
+// WINDOWS-NEXT:    store i8 0, ptr [[TMP7]], align 1
+// WINDOWS-NEXT:    [[TMP8:%.*]] = getelementptr i8, ptr [[TMP0]], i32 6
+// WINDOWS-NEXT:    store i8 0, ptr [[TMP8]], align 2
+// WINDOWS-NEXT:    [[TMP9:%.*]] = getelementptr i8, ptr [[TMP0]], i32 7
+// WINDOWS-NEXT:    store i8 0, ptr [[TMP9]], align 1
+// WINDOWS-NEXT:    [[TMP10:%.*]] = getelementptr i8, ptr [[TMP0]], i32 9
+// WINDOWS-NEXT:    store i8 0, ptr [[TMP10]], align 1
+// WINDOWS-NEXT:    [[TMP11:%.*]] = getelementptr i8, ptr [[TMP0]], i32 10
+// WINDOWS-NEXT:    store i8 0, ptr [[TMP11]], align 2
+// WINDOWS-NEXT:    [[TMP12:%.*]] = getelementptr i8, ptr [[TMP0]], i32 11
+// WINDOWS-NEXT:    store i8 0, ptr [[TMP12]], align 1
+// WINDOWS-NEXT:    [[TMP13:%.*]] = getelementptr i8, ptr [[TMP0]], i32 12
+// WINDOWS-NEXT:    [[TMP14:%.*]] = load i8, ptr [[TMP13]], align 4
+// WINDOWS-NEXT:    [[TMP15:%.*]] = and i8 [[TMP14]], 31
+// WINDOWS-NEXT:    store i8 [[TMP15]], ptr [[TMP13]], align 4
+// WINDOWS-NEXT:    [[TMP16:%.*]] = getelementptr i8, ptr [[TMP0]], i32 13
+// WINDOWS-NEXT:    store i8 0, ptr [[TMP16]], align 1
+// WINDOWS-NEXT:    [[TMP17:%.*]] = getelementptr i8, ptr [[TMP0]], i32 14
+// WINDOWS-NEXT:    store i8 0, ptr [[TMP17]], align 2
+// WINDOWS-NEXT:    [[TMP18:%.*]] = getelementptr i8, ptr [[TMP0]], i32 15
+// WINDOWS-NEXT:    store i8 0, ptr [[TMP18]], align 1
+// WINDOWS-NEXT:    [[TMP19:%.*]] = getelementptr i8, ptr [[TMP0]], i32 18
+// WINDOWS-NEXT:    store i8 0, ptr [[TMP19]], align 2
+// WINDOWS-NEXT:    [[TMP20:%.*]] = getelementptr i8, ptr [[TMP0]], i32 19
+// WINDOWS-NEXT:    store i8 0, ptr [[TMP20]], align 1
+// WINDOWS-NEXT:    ret void
+//
+// ARM-LE-LABEL: define dso_local void @testBitfieldMixedVolatileMembersAndObject(
+// ARM-LE-SAME: ptr noundef [[S:%.*]]) #[[ATTR0]] {
+// ARM-LE-NEXT:  [[ENTRY:.*:]]
+// ARM-LE-NEXT:    [[S_ADDR:%.*]] = alloca ptr, align 4
+// ARM-LE-NEXT:    store ptr [[S]], ptr [[S_ADDR]], align 4
+// ARM-LE-NEXT:    [[TMP0:%.*]] = load ptr, ptr [[S_ADDR]], align 4
+// ARM-LE-NEXT:    [[TMP1:%.*]] = getelementptr i8, ptr [[TMP0]], i32 1
+// ARM-LE-NEXT:    [[TMP2:%.*]] = load i8, ptr [[TMP1]], align 1
+// ARM-LE-NEXT:    [[TMP3:%.*]] = and i8 [[TMP2]], 7
+// ARM-LE-NEXT:    store i8 [[TMP3]], ptr [[TMP1]], align 1
+// ARM-LE-NEXT:    [[TMP4:%.*]] = getelementptr i8, ptr [[TMP0]], i32 3
+// ARM-LE-NEXT:    [[TMP5:%.*]] = load i8, ptr [[TMP4]], align 1
+// ARM-LE-NEXT:    [[TMP6:%.*]] = and i8 [[TMP5]], 31
+// ARM-LE-NEXT:    store i8 [[TMP6]], ptr [[TMP4]], align 1
+// ARM-LE-NEXT:    [[TMP7:%.*]] = getelementptr i8, ptr [[TMP0]], i32 6
+// ARM-LE-NEXT:    store i8 0, ptr [[TMP7]], align 2
+// ARM-LE-NEXT:    [[TMP8:%.*]] = getelementptr i8, ptr [[TMP0]], i32 7
+// ARM-LE-NEXT:    store i8 0, ptr [[TMP8]], align 1
+// ARM-LE-NEXT:    ret void
+//
+// ARM-BE-LABEL: define dso_local void @testBitfieldMixedVolatileMembersAndObject(
+// ARM-BE-SAME: ptr noundef [[S:%.*]]) #[[ATTR0]] {
+// ARM-BE-NEXT:  [[ENTRY:.*:]]
+// ARM-BE-NEXT:    [[S_ADDR:%.*]] = alloca ptr, align 4
+// ARM-BE-NEXT:    store ptr [[S]], ptr [[S_ADDR]], align 4
+// ARM-BE-NEXT:    [[TMP0:%.*]] = load ptr, ptr [[S_ADDR]], align 4
+// ARM-BE-NEXT:    [[TMP1:%.*]] = getelementptr i8, ptr [[TMP0]], i32 1
+// ARM-BE-NEXT:    [[TMP2:%.*]] = load i8, ptr [[TMP1]], align 1
+// ARM-BE-NEXT:    [[TMP3:%.*]] = and i8 [[TMP2]], -32
+// ARM-BE-NEXT:    store i8 [[TMP3]], ptr [[TMP1]], align 1
+// ARM-BE-NEXT:    [[TMP4:%.*]] = getelementptr i8, ptr [[TMP0]], i32 3
+// ARM-BE-NEXT:    [[TMP5:%.*]] = load i8, ptr [[TMP4]], align 1
+// ARM-BE-NEXT:    [[TMP6:%.*]] = and i8 [[TMP5]], -8
+// ARM-BE-NEXT:    store i8 [[TMP6]], ptr [[TMP4]], align 1
+// ARM-BE-NEXT:    [[TMP7:%.*]] = getelementptr i8, ptr [[TMP0]], i32 6
+// ARM-BE-NEXT:    store i8 0, ptr [[TMP7]], align 2
+// ARM-BE-NEXT:    [[TMP8:%.*]] = getelementptr i8, ptr [[TMP0]], i32 7
+// ARM-BE-NEXT:    store i8 0, ptr [[TMP8]], align 1
+// ARM-BE-NEXT:    ret void
+//
+// AARCH64-LE-LABEL: define dso_local void @testBitfieldMixedVolatileMembersAndObject(
+// AARCH64-LE-SAME: ptr noundef [[S:%.*]]) #[[ATTR0]] {
+// AARCH64-LE-NEXT:  [[ENTRY:.*:]]
+// AARCH64-LE-NEXT:    [[S_ADDR:%.*]] = alloca ptr, align 8
+// AARCH64-LE-NEXT:    store ptr [[S]], ptr [[S_ADDR]], align 8
+// AARCH64-LE-NEXT:    [[TMP0:%.*]] = load ptr, ptr [[S_ADDR]], align 8
+// AARCH64-LE-NEXT:    [[TMP1:%.*]] = getelementptr i8, ptr [[TMP0]], i32 1
+// AARCH64-LE-NEXT:    [[TMP2:%.*]] = load i8, ptr [[TMP1]], align 1
+// AARCH64-LE-NEXT:    [[TMP3:%.*]] = and i8 [[TMP2]], 7
+// AARCH64-LE-NEXT:    store i8 [[TMP3]], ptr [[TMP1]], align 1
+// AARCH64-LE-NEXT:    [[TMP4:%.*]] = getelementptr i8, ptr [[TMP0]], i32 3
+// AARCH64-LE-NEXT:    [[TMP5:%.*]] = load i8, ptr [[TMP4]], align 1
+// AARCH64-LE-NEXT:    [[TMP6:%.*]] = and i8 [[TMP5]], 31
+// AARCH64-LE-NEXT:    store i8 [[TMP6]], ptr [[TMP4]], align 1
+// AARCH64-LE-NEXT:    [[TMP7:%.*]] = getelementptr i8, ptr [[TMP0]], i32 6
+// AARCH64-LE-NEXT:    store i8 0, ptr [[TMP7]], align 2
+// AARCH64-LE-NEXT:    [[TMP8:%.*]] = getelementptr i8, ptr [[TMP0]], i32 7
+// AARCH64-LE-NEXT:    store i8 0, ptr [[TMP8]], align 1
+// AARCH64-LE-NEXT:    ret void
+//
+// AARCH64-BE-LABEL: define dso_local void @testBitfieldMixedVolatileMembersAndObject(
+// AARCH64-BE-SAME: ptr noundef [[S:%.*]]) #[[ATTR0]] {
+// AARCH64-BE-NEXT:  [[ENTRY:.*:]]
+// AARCH64-BE-NEXT:    [[S_ADDR:%.*]] = alloca ptr, align 8
+// AARCH64-BE-NEXT:    store ptr [[S]], ptr [[S_ADDR]], align 8
+// AARCH64-BE-NEXT:    [[TMP0:%.*]] = load ptr, ptr [[S_ADDR]], align 8
+// AARCH64-BE-NEXT:    [[TMP1:%.*]] = getelementptr i8, ptr [[TMP0]], i32 1
+// AARCH64-BE-NEXT:    [[TMP2:%.*]] = load i8, ptr [[TMP1]], align 1
+// AARCH64-BE-NEXT:    [[TMP3:%.*]] = and i8 [[TMP2]], -32
+// AARCH64-BE-NEXT:    store i8 [[TMP3]], ptr [[TMP1]], align 1
+// AARCH64-BE-NEXT:    [[TMP4:%.*]] = getelementptr i8, ptr [[TMP0]], i32 3
+// AARCH64-BE-NEXT:    [[TMP5:%.*]] = load i8, ptr [[TMP4]], align 1
+// AARCH64-BE-NEXT:    [[TMP6:%.*]] = and i8 [[TMP5]], -8
+// AARCH64-BE-NEXT:    store i8 [[TMP6]], ptr [[TMP4]], align 1
+// AARCH64-BE-NEXT:    [[TMP7:%.*]] = getelementptr i8, ptr [[TMP0]], i32 6
+// AARCH64-BE-NEXT:    store i8 0, ptr [[TMP7]], align 2
+// AARCH64-BE-NEXT:    [[TMP8:%.*]] = getelementptr i8, ptr [[TMP0]], i32 7
+// AARCH64-BE-NEXT:    store i8 0, ptr [[TMP8]], align 1
+// AARCH64-BE-NEXT:    ret void
+//
+extern "C" void testBitfieldMixedVolatileMembersAndObject(
+    volatile struct BitfieldMixedVolatileMembers *s) {
+  __builtin_clear_padding(s);
+}
+
+// Both array elements have partial-byte padding, and neither starts at offset
+// zero in the enclosing object. Preserve the prefix and suffix too.
+struct BitfieldNestedArray {
+  unsigned char prefix[3];
+  struct BitfieldPartialEnds elements[2];
+  unsigned char suffix;
+};
+
+// LINUX-LABEL: define dso_local void @testBitfieldNestedArray(
+// LINUX-SAME: ptr noundef [[S:%.*]]) #[[ATTR0]] {
+// LINUX-NEXT:  [[ENTRY:.*:]]
+// LINUX-NEXT:    [[S_ADDR:%.*]] = alloca ptr, align 8
+// LINUX-NEXT:    store ptr [[S]], ptr [[S_ADDR]], align 8
+// LINUX-NEXT:    [[TMP0:%.*]] = load ptr, ptr [[S_ADDR]], align 8
+// LINUX-NEXT:    [[TMP1:%.*]] = getelementptr i8, ptr [[TMP0]], i32 3
+// LINUX-NEXT:    store i8 0, ptr [[TMP1]], align 1
+// LINUX-NEXT:    [[TMP2:%.*]] = getelementptr i8, ptr [[TMP0]], i32 4
+// LINUX-NEXT:    [[TMP3:%.*]] = load i8, ptr [[TMP2]], align 4
+// LINUX-NEXT:    [[TMP4:%.*]] = and i8 [[TMP3]], -8
+// LINUX-NEXT:    store i8 [[TMP4]], ptr [[TMP2]], align 4
+// LINUX-NEXT:    [[TMP5:%.*]] = getelementptr i8, ptr [[TMP0]], i32 6
+// LINUX-NEXT:    [[TMP6:%.*]] = load i8, ptr [[TMP5]], align 2
+// LINUX-NEXT:    [[TMP7:%.*]] = and i8 [[TMP6]], 31
+// LINUX-NEXT:    store i8 [[TMP7]], ptr [[TMP5]], align 2
+// LINUX-NEXT:    [[TMP8:%.*]] = getelementptr i8, ptr [[TMP0]], i32 7
+// LINUX-NEXT:    store i8 0, ptr [[TMP8]], align 1
+// LINUX-NEXT:    [[TMP9:%.*]] = getelementptr i8, ptr [[TMP0]], i32 8
+// LINUX-NEXT:    [[TMP10:%.*]] = load i8, ptr [[TMP9]], align 4
+// LINUX-NEXT:    [[TMP11:%.*]] = and i8 [[TMP10]], -8
+// LINUX-NEXT:    store i8 [[TMP11]], ptr [[TMP9]], align 4
+// LINUX-NEXT:    [[TMP12:%.*]] = getelementptr i8, ptr [[TMP0]], i32 10
+// LINUX-NEXT:    [[TMP13:%.*]] = load i8, ptr [[TMP12]], align 2
+// LINUX-NEXT:    [[TMP14:%.*]] = and i8 [[TMP13]], 31
+// LINUX-NEXT:    store i8 [[TMP14]], ptr [[TMP12]], align 2
+// LINUX-NEXT:    [[TMP15:%.*]] = getelementptr i8, ptr [[TMP0]], i32 11
+// LINUX-NEXT:    store i8 0, ptr [[TMP15]], align 1
+// LINUX-NEXT:    [[TMP16:%.*]] = getelementptr i8, ptr [[TMP0]], i32 13
+// LINUX-NEXT:    store i8 0, ptr [[TMP16]], align 1
+// LINUX-NEXT:    [[TMP17:%.*]] = getelementptr i8, ptr [[TMP0]], i32 14
+// LINUX-NEXT:    store i8 0, ptr [[TMP17]], align 2
+// LINUX-NEXT:    [[TMP18:%.*]] = getelementptr i8, ptr [[TMP0]], i32 15
+// LINUX-NEXT:    store i8 0, ptr [[TMP18]], align 1
+// LINUX-NEXT:    ret void
+//
+// WINDOWS-LABEL: define dso_local void @testBitfieldNestedArray(
+// WINDOWS-SAME: ptr noundef [[S:%.*]]) #[[ATTR0]] {
+// WINDOWS-NEXT:  [[ENTRY:.*:]]
+// WINDOWS-NEXT:    [[S_ADDR:%.*]] = alloca ptr, align 8
+// WINDOWS-NEXT:    store ptr [[S]], ptr [[S_ADDR]], align 8
+// WINDOWS-NEXT:    [[TMP0:%.*]] = load ptr, ptr [[S_ADDR]], align 8
+// WINDOWS-NEXT:    [[TMP1:%.*]] = getelementptr i8, ptr [[TMP0]], i32 3
+// WINDOWS-NEXT:    store i8 0, ptr [[TMP1]], align 1
+// WINDOWS-NEXT:    [[TMP2:%.*]] = getelementptr i8, ptr [[TMP0]], i32 4
+// WINDOWS-NEXT:    [[TMP3:%.*]] = load i8, ptr [[TMP2]], align 4
+// WINDOWS-NEXT:    [[TMP4:%.*]] = and i8 [[TMP3]], -8
+// WINDOWS-NEXT:    store i8 [[TMP4]], ptr [[TMP2]], align 4
+// WINDOWS-NEXT:    [[TMP5:%.*]] = getelementptr i8, ptr [[TMP0]], i32 6
+// WINDOWS-NEXT:    [[TMP6:%.*]] = load i8, ptr [[TMP5]], align 2
+// WINDOWS-NEXT:    [[TMP7:%.*]] = and i8 [[TMP6]], 31
+// WINDOWS-NEXT:    store i8 [[TMP7]], ptr [[TMP5]], align 2
+// WINDOWS-NEXT:    [[TMP8:%.*]] = getelementptr i8, ptr [[TMP0]], i32 7
+// WINDOWS-NEXT:    store i8 0, ptr [[TMP8]], align 1
+// WINDOWS-NEXT:    [[TMP9:%.*]] = getelementptr i8, ptr [[TMP0]], i32 8
+// WINDOWS-NEXT:    [[TMP10:%.*]] = load i8, ptr [[TMP9]], align 4
+// WINDOWS-NEXT:    [[TMP11:%.*]] = and i8 [[TMP10]], -8
+// WINDOWS-NEXT:    store i8 [[TMP11]], ptr [[TMP9]], align 4
+// WINDOWS-NEXT:    [[TMP12:%.*]] = getelementptr i8, ptr [[TMP0]], i32 10
+// WINDOWS-NEXT:    [[TMP13:%.*]] = load i8, ptr [[TMP12]], align 2
+// WINDOWS-NEXT:    [[TMP14:%.*]] = and i8 [[TMP13]], 31
+// WINDOWS-NEXT:    store i8 [[TMP14]], ptr [[TMP12]], align 2
+// WINDOWS-NEXT:    [[TMP15:%.*]] = getelementptr i8, ptr [[TMP0]], i32 11
+// WINDOWS-NEXT:    store i8 0, ptr [[TMP15]], align 1
+// WINDOWS-NEXT:    [[TMP16:%.*]] = getelementptr i8, ptr [[TMP0]], i32 13
+// WINDOWS-NEXT:    store i8 0, ptr [[TMP16]], align 1
+// WINDOWS-NEXT:    [[TMP17:%.*]] = getelementptr i8, ptr [[TMP0]], i32 14
+// WINDOWS-NEXT:    store i8 0, ptr [[TMP17]], align 2
+// WINDOWS-NEXT:    [[TMP18:%.*]] = getelementptr i8, ptr [[TMP0]], i32 15
+// WINDOWS-NEXT:    store i8 0, ptr [[TMP18]], align 1
+// WINDOWS-NEXT:    ret void
+//
+// ARM-LE-LABEL: define dso_local void @testBitfieldNestedArray(
+// ARM-LE-SAME: ptr noundef [[S:%.*]]) #[[ATTR0]] {
+// ARM-LE-NEXT:  [[ENTRY:.*:]]
+// ARM-LE-NEXT:    [[S_ADDR:%.*]] = alloca ptr, align 4
+// ARM-LE-NEXT:    store ptr [[S]], ptr [[S_ADDR]], align 4
+// ARM-LE-NEXT:    [[TMP0:%.*]] = load ptr, ptr [[S_ADDR]], align 4
+// ARM-LE-NEXT:    [[TMP1:%.*]] = getelementptr i8, ptr [[TMP0]], i32 3
+// ARM-LE-NEXT:    store i8 0, ptr [[TMP1]], align 1
+// ARM-LE-NEXT:    [[TMP2:%.*]] = getelementptr i8, ptr [[TMP0]], i32 4
+// ARM-LE-NEXT:    [[TMP3:%.*]] = load i8, ptr [[TMP2]], align 4
+// ARM-LE-NEXT:    [[TMP4:%.*]] = and i8 [[TMP3]], -8
+// ARM-LE-NEXT:    store i8 [[TMP4]], ptr [[TMP2]], align 4
+// ARM-LE-NEXT:    [[TMP5:%.*]] = getelementptr i8, ptr [[TMP0]], i32 6
+// ARM-LE-NEXT:    [[TMP6:%.*]] = load i8, ptr [[TMP5]], align 2
+// ARM-LE-NEXT:    [[TMP7:%.*]] = and i8 [[TMP6]], 31
+// ARM-LE-NEXT:    store i8 [[TMP7]], ptr [[TMP5]], align 2
+// ARM-LE-NEXT:    [[TMP8:%.*]] = getelementptr i8, ptr [[TMP0]], i32 7
+// ARM-LE-NEXT:    store i8 0, ptr [[TMP8]], align 1
+// ARM-LE-NEXT:    [[TMP9:%.*]] = getelementptr i8, ptr [[TMP0]], i32 8
+// ARM-LE-NEXT:    [[TMP10:%.*]] = load i8, ptr [[TMP9]], align 4
+// ARM-LE-NEXT:    [[TMP11:%.*]] = and i8 [[TMP10]], -8
+// ARM-LE-NEXT:    store i8 [[TMP11]], ptr [[TMP9]], align 4
+// ARM-LE-NEXT:    [[TMP12:%.*]] = getelementptr i8, ptr [[TMP0]], i32 10
+// ARM-LE-NEXT:    [[TMP13:%.*]] = load i8, ptr [[TMP12]], align 2
+// ARM-LE-NEXT:    [[TMP14:%.*]] = and i8 [[TMP13]], 31
+// ARM-LE-NEXT:    store i8 [[TMP14]], ptr [[TMP12]], align 2
+// ARM-LE-NEXT:    [[TMP15:%.*]] = getelementptr i8, ptr [[TMP0]], i32 11
+// ARM-LE-NEXT:    store i8 0, ptr [[TMP15]], align 1
+// ARM-LE-NEXT:    [[TMP16:%.*]] = getelementptr i8, ptr [[TMP0]], i32 13
+// ARM-LE-NEXT:    store i8 0, ptr [[TMP16]], align 1
+// ARM-LE-NEXT:    [[TMP17:%.*]] = getelementptr i8, ptr [[TMP0]], i32 14
+// ARM-LE-NEXT:    store i8 0, ptr [[TMP17]], align 2
+// ARM-LE-NEXT:    [[TMP18:%.*]] = getelementptr i8, ptr [[TMP0]], i32 15
+// ARM-LE-NEXT:    store i8 0, ptr [[TMP18]], align 1
+// ARM-LE-NEXT:    ret void
+//
+// ARM-BE-LABEL: define dso_local void @testBitfieldNestedArray(
+// ARM-BE-SAME: ptr noundef [[S:%.*]]) #[[ATTR0]] {
+// ARM-BE-NEXT:  [[ENTRY:.*:]]
+// ARM-BE-NEXT:    [[S_ADDR:%.*]] = alloca ptr, align 4
+// ARM-BE-NEXT:    store ptr [[S]], ptr [[S_ADDR]], align 4
+// ARM-BE-NEXT:    [[TMP0:%.*]] = load ptr, ptr [[S_ADDR]], align 4
+// ARM-BE-NEXT:    [[TMP1:%.*]] = getelementptr i8, ptr [[TMP0]], i32 3
+// ARM-BE-NEXT:    store i8 0, ptr [[TMP1]], align 1
+// ARM-BE-NEXT:    [[TMP2:%.*]] = getelementptr i8, ptr [[TMP0]], i32 4
+// ARM-BE-NEXT:    [[TMP3:%.*]] = load i8, ptr [[TMP2]], align 4
+// ARM-BE-NEXT:    [[TMP4:%.*]] = and i8 [[TMP3]], 31
+// ARM-BE-NEXT:    store i8 [[TMP4]], ptr [[TMP2]], align 4
+// ARM-BE-NEXT:    [[TMP5:%.*]] = getelementptr i8, ptr [[TMP0]], i32 6
+// ARM-BE-NEXT:    [[TMP6:%.*]] = load i8, ptr [[TMP5]], align 2
+// ARM-BE-NEXT:    [[TMP7:%.*]] = and i8 [[TMP6]], -8
+// ARM-BE-NEXT:    store i8 [[TMP7]], ptr [[TMP5]], align 2
+// ARM-BE-NEXT:    [[TMP8:%.*]] = getelementptr i8, ptr [[TMP0]], i32 7
+// ARM-BE-NEXT:    store i8 0, ptr [[TMP8]], align 1
+// ARM-BE-NEXT:    [[TMP9:%.*]] = getelementptr i8, ptr [[TMP0]], i32 8
+// ARM-BE-NEXT:    [[TMP10:%.*]] = load i8, ptr [[TMP9]], align 4
+// ARM-BE-NEXT:    [[TMP11:%.*]] = and i8 [[TMP10]], 31
+// ARM-BE-NEXT:    store i8 [[TMP11]], ptr [[TMP9]], align 4
+// ARM-BE-NEXT:    [[TMP12:%.*]] = getelementptr i8, ptr [[TMP0]], i32 10
+// ARM-BE-NEXT:    [[TMP13:%.*]] = load i8, ptr [[TMP12]], align 2
+// ARM-BE-NEXT:    [[TMP14:%.*]] = and i8 [[TMP13]], -8
+// ARM-BE-NEXT:    store i8 [[TMP14]], ptr [[TMP12]], align 2
+// ARM-BE-NEXT:    [[TMP15:%.*]] = getelementptr i8, ptr [[TMP0]], i32 11
+// ARM-BE-NEXT:    store i8 0, ptr [[TMP15]], align 1
+// ARM-BE-NEXT:    [[TMP16:%.*]] = getelementptr i8, ptr [[TMP0]], i32 13
+// ARM-BE-NEXT:    store i8 0, ptr [[TMP16]], align 1
+// ARM-BE-NEXT:    [[TMP17:%.*]] = getelementptr i8, ptr [[TMP0]], i32 14
+// ARM-BE-NEXT:    store i8 0, ptr [[TMP17]], align 2
+// ARM-BE-NEXT:    [[TMP18:%.*]] = getelementptr i8, ptr [[TMP0]], i32 15
+// ARM-BE-NEXT:    store i8 0, ptr [[TMP18]], align 1
+// ARM-BE-NEXT:    ret void
+//
+// AARCH64-LE-LABEL: define dso_local void @testBitfieldNestedArray(
+// AARCH64-LE-SAME: ptr noundef [[S:%.*]]) #[[ATTR0]] {
+// AARCH64-LE-NEXT:  [[ENTRY:.*:]]
+// AARCH64-LE-NEXT:    [[S_ADDR:%.*]] = alloca ptr, align 8
+// AARCH64-LE-NEXT:    store ptr [[S]], ptr [[S_ADDR]], align 8
+// AARCH64-LE-NEXT:    [[TMP0:%.*]] = load ptr, ptr [[S_ADDR]], align 8
+// AARCH64-LE-NEXT:    [[TMP1:%.*]] = getelementptr i8, ptr [[TMP0]], i32 3
+// AARCH64-LE-NEXT:    store i8 0, ptr [[TMP1]], align 1
+// AARCH64-LE-NEXT:    [[TMP2:%.*]] = getelementptr i8, ptr [[TMP0]], i32 4
+// AARCH64-LE-NEXT:    [[TMP3:%.*]] = load i8, ptr [[TMP2]], align 4
+// AARCH64-LE-NEXT:    [[TMP4:%.*]] = and i8 [[TMP3]], -8
+// AARCH64-LE-NEXT:    store i8 [[TMP4]], ptr [[TMP2]], align 4
+// AARCH64-LE-NEXT:    [[TMP5:%.*]] = getelementptr i8, ptr [[TMP0]], i32 6
+// AARCH64-LE-NEXT:    [[TMP6:%.*]] = load i8, ptr [[TMP5]], align 2
+// AARCH64-LE-NEXT:    [[TMP7:%.*]] = and i8 [[TMP6]], 31
+// AARCH64-LE-NEXT:    store i8 [[TMP7]], ptr [[TMP5]], align 2
+// AARCH64-LE-NEXT:    [[TMP8:%.*]] = getelementptr i8, ptr [[TMP0]], i32 7
+// AARCH64-LE-NEXT:    store i8 0, ptr [[TMP8]], align 1
+// AARCH64-LE-NEXT:    [[TMP9:%.*]] = getelementptr i8, ptr [[TMP0]], i32 8
+// AARCH64-LE-NEXT:    [[TMP10:%.*]] = load i8, ptr [[TMP9]], align 4
+// AARCH64-LE-NEXT:    [[TMP11:%.*]] = and i8 [[TMP10]], -8
+// AARCH64-LE-NEXT:    store i8 [[TMP11]], ptr [[TMP9]], align 4
+// AARCH64-LE-NEXT:    [[TMP12:%.*]] = getelementptr i8, ptr [[TMP0]], i32 10
+// AARCH64-LE-NEXT:    [[TMP13:%.*]] = load i8, ptr [[TMP12]], align 2
+// AARCH64-LE-NEXT:    [[TMP14:%.*]] = and i8 [[TMP13]], 31
+// AARCH64-LE-NEXT:    store i8 [[TMP14]], ptr [[TMP12]], align 2
+// AARCH64-LE-NEXT:    [[TMP15:%.*]] = getelementptr i8, ptr [[TMP0]], i32 11
+// AARCH64-LE-NEXT:    store i8 0, ptr [[TMP15]], align 1
+// AARCH64-LE-NEXT:    [[TMP16:%.*]] = getelementptr i8, ptr [[TMP0]], i32 13
+// AARCH64-LE-NEXT:    store i8 0, ptr [[TMP16]], align 1
+// AARCH64-LE-NEXT:    [[TMP17:%.*]] = getelementptr i8, ptr [[TMP0]], i32 14
+// AARCH64-LE-NEXT:    store i8 0, ptr [[TMP17]], align 2
+// AARCH64-LE-NEXT:    [[TMP18:%.*]] = getelementptr i8, ptr [[TMP0]], i32 15
+// AARCH64-LE-NEXT:    store i8 0, ptr [[TMP18]], align 1
+// AARCH64-LE-NEXT:    ret void
+//
+// AARCH64-BE-LABEL: define dso_local void @testBitfieldNestedArray(
+// AARCH64-BE-SAME: ptr noundef [[S:%.*]]) #[[ATTR0]] {
+// AARCH64-BE-NEXT:  [[ENTRY:.*:]]
+// AARCH64-BE-NEXT:    [[S_ADDR:%.*]] = alloca ptr, align 8
+// AARCH64-BE-NEXT:    store ptr [[S]], ptr [[S_ADDR]], align 8
+// AARCH64-BE-NEXT:    [[TMP0:%.*]] = load ptr, ptr [[S_ADDR]], align 8
+// AARCH64-BE-NEXT:    [[TMP1:%.*]] = getelementptr i8, ptr [[TMP0]], i32 3
+// AARCH64-BE-NEXT:    store i8 0, ptr [[TMP1]], align 1
+// AARCH64-BE-NEXT:    [[TMP2:%.*]] = getelementptr i8, ptr [[TMP0]], i32 4
+// AARCH64-BE-NEXT:    [[TMP3:%.*]] = load i8, ptr [[TMP2]], align 4
+// AARCH64-BE-NEXT:    [[TMP4:%.*]] = and i8 [[TMP3]], 31
+// AARCH64-BE-NEXT:    store i8 [[TMP4]], ptr [[TMP2]], align 4
+// AARCH64-BE-NEXT:    [[TMP5:%.*]] = getelementptr i8, ptr [[TMP0]], i32 6
+// AARCH64-BE-NEXT:    [[TMP6:%.*]] = load i8, ptr [[TMP5]], align 2
+// AARCH64-BE-NEXT:    [[TMP7:%.*]] = and i8 [[TMP6]], -8
+// AARCH64-BE-NEXT:    store i8 [[TMP7]], ptr [[TMP5]], align 2
+// AARCH64-BE-NEXT:    [[TMP8:%.*]] = getelementptr i8, ptr [[TMP0]], i32 7
+// AARCH64-BE-NEXT:    store i8 0, ptr [[TMP8]], align 1
+// AARCH64-BE-NEXT:    [[TMP9:%.*]] = getelementptr i8, ptr [[TMP0]], i32 8
+// AARCH64-BE-NEXT:    [[TMP10:%.*]] = load i8, ptr [[TMP9]], align 4
+// AARCH64-BE-NEXT:    [[TMP11:%.*]] = and i8 [[TMP10]], 31
+// AARCH64-BE-NEXT:    store i8 [[TMP11]], ptr [[TMP9]], align 4
+// AARCH64-BE-NEXT:    [[TMP12:%.*]] = getelementptr i8, ptr [[TMP0]], i32 10
+// AARCH64-BE-NEXT:    [[TMP13:%.*]] = load i8, ptr [[TMP12]], align 2
+// AARCH64-BE-NEXT:    [[TMP14:%.*]] = and i8 [[TMP13]], -8
+// AARCH64-BE-NEXT:    store i8 [[TMP14]], ptr [[TMP12]], align 2
+// AARCH64-BE-NEXT:    [[TMP15:%.*]] = getelementptr i8, ptr [[TMP0]], i32 11
+// AARCH64-BE-NEXT:    store i8 0, ptr [[TMP15]], align 1
+// AARCH64-BE-NEXT:    [[TMP16:%.*]] = getelementptr i8, ptr [[TMP0]], i32 13
+// AARCH64-BE-NEXT:    store i8 0, ptr [[TMP16]], align 1
+// AARCH64-BE-NEXT:    [[TMP17:%.*]] = getelementptr i8, ptr [[TMP0]], i32 14
+// AARCH64-BE-NEXT:    store i8 0, ptr [[TMP17]], align 2
+// AARCH64-BE-NEXT:    [[TMP18:%.*]] = getelementptr i8, ptr [[TMP0]], i32 15
+// AARCH64-BE-NEXT:    store i8 0, ptr [[TMP18]], align 1
+// AARCH64-BE-NEXT:    ret void
+//
+extern "C" void testBitfieldNestedArray(struct BitfieldNestedArray *s) {
+  __builtin_clear_padding(s);
+}
+
+// The raw member occupies the first byte, including padding in the fields
+// member. Only padding shared by both union alternatives may be cleared.
+union BitfieldUnionOverlay {
+  struct BitfieldPartialEnds fields;
+  unsigned char raw;
+};
+
+// LINUX-LABEL: define dso_local void @testBitfieldUnionOverlay(
+// LINUX-SAME: ptr noundef [[S:%.*]]) #[[ATTR0]] {
+// LINUX-NEXT:  [[ENTRY:.*:]]
+// LINUX-NEXT:    [[S_ADDR:%.*]] = alloca ptr, align 8
+// LINUX-NEXT:    store ptr [[S]], ptr [[S_ADDR]], align 8
+// LINUX-NEXT:    [[TMP0:%.*]] = load ptr, ptr [[S_ADDR]], align 8
+// LINUX-NEXT:    [[TMP1:%.*]] = getelementptr i8, ptr [[TMP0]], i32 2
+// LINUX-NEXT:    [[TMP2:%.*]] = load i8, ptr [[TMP1]], align 2
+// LINUX-NEXT:    [[TMP3:%.*]] = and i8 [[TMP2]], 31
+// LINUX-NEXT:    store i8 [[TMP3]], ptr [[TMP1]], align 2
+// LINUX-NEXT:    [[TMP4:%.*]] = getelementptr i8, ptr [[TMP0]], i32 3
+// LINUX-NEXT:    store i8 0, ptr [[TMP4]], align 1
+// LINUX-NEXT:    ret void
+//
+// WINDOWS-LABEL: define dso_local void @testBitfieldUnionOverlay(
+// WINDOWS-SAME: ptr noundef [[S:%.*]]) #[[ATTR0]] {
+// WINDOWS-NEXT:  [[ENTRY:.*:]]
+// WINDOWS-NEXT:    [[S_ADDR:%.*]] = alloca ptr, align 8
+// WINDOWS-NEXT:    store ptr [[S]], ptr [[S_ADDR]], align 8
+// WINDOWS-NEXT:    [[TMP0:%.*]] = load ptr, ptr [[S_ADDR]], align 8
+// WINDOWS-NEXT:    [[TMP1:%.*]] = getelementptr i8, ptr [[TMP0]], i32 2
+// WINDOWS-NEXT:    [[TMP2:%.*]] = load i8, ptr [[TMP1]], align 2
+// WINDOWS-NEXT:    [[TMP3:%.*]] = and i8 [[TMP2]], 31
+// WINDOWS-NEXT:    store i8 [[TMP3]], ptr [[TMP1]], align 2
+// WINDOWS-NEXT:    [[TMP4:%.*]] = getelementptr i8, ptr [[TMP0]], i32 3
+// WINDOWS-NEXT:    store i8 0, ptr [[TMP4]], align 1
+// WINDOWS-NEXT:    ret void
+//
+// ARM-LE-LABEL: define dso_local void @testBitfieldUnionOverlay(
+// ARM-LE-SAME: ptr noundef [[S:%.*]]) #[[ATTR0]] {
+// ARM-LE-NEXT:  [[ENTRY:.*:]]
+// ARM-LE-NEXT:    [[S_ADDR:%.*]] = alloca ptr, align 4
+// ARM-LE-NEXT:    store ptr [[S]], ptr [[S_ADDR]], align 4
+// ARM-LE-NEXT:    [[TMP0:%.*]] = load ptr, ptr [[S_ADDR]], align 4
+// ARM-LE-NEXT:    [[TMP1:%.*]] = getelementptr i8, ptr [[TMP0]], i32 2
+// ARM-LE-NEXT:    [[TMP2:%.*]] = load i8, ptr [[TMP1]], align 2
+// ARM-LE-NEXT:    [[TMP3:%.*]] = and i8 [[TMP2]], 31
+// ARM-LE-NEXT:    store i8 [[TMP3]], ptr [[TMP1]], align 2
+// ARM-LE-NEXT:    [[TMP4:%.*]] = getelementptr i8, ptr [[TMP0]], i32 3
+// ARM-LE-NEXT:    store i8 0, ptr [[TMP4]], align 1
+// ARM-LE-NEXT:    ret void
+//
+// ARM-BE-LABEL: define dso_local void @testBitfieldUnionOverlay(
+// ARM-BE-SAME: ptr noundef [[S:%.*]]) #[[ATTR0]] {
+// ARM-BE-NEXT:  [[ENTRY:.*:]]
+// ARM-BE-NEXT:    [[S_ADDR:%.*]] = alloca ptr, align 4
+// ARM-BE-NEXT:    store ptr [[S]], ptr [[S_ADDR]], align 4
+// ARM-BE-NEXT:    [[TMP0:%.*]] = load ptr, ptr [[S_ADDR]], align 4
+// ARM-BE-NEXT:    [[TMP1:%.*]] = getelementptr i8, ptr [[TMP0]], i32 2
+// ARM-BE-NEXT:    [[TMP2:%.*]] = load i8, ptr [[TMP1]], align 2
+// ARM-BE-NEXT:    [[TMP3:%.*]] = and i8 [[TMP2]], -8
+// ARM-BE-NEXT:    store i8 [[TMP3]], ptr [[TMP1]], align 2
+// ARM-BE-NEXT:    [[TMP4:%.*]] = getelementptr i8, ptr [[TMP0]], i32 3
+// ARM-BE-NEXT:    store i8 0, ptr [[TMP4]], align 1
+// ARM-BE-NEXT:    ret void
+//
+// AARCH64-LE-LABEL: define dso_local void @testBitfieldUnionOverlay(
+// AARCH64-LE-SAME: ptr noundef [[S:%.*]]) #[[ATTR0]] {
+// AARCH64-LE-NEXT:  [[ENTRY:.*:]]
+// AARCH64-LE-NEXT:    [[S_ADDR:%.*]] = alloca ptr, align 8
+// AARCH64-LE-NEXT:    store ptr [[S]], ptr [[S_ADDR]], align 8
+// AARCH64-LE-NEXT:    [[TMP0:%.*]] = load ptr, ptr [[S_ADDR]], align 8
+// AARCH64-LE-NEXT:    [[TMP1:%.*]] = getelementptr i8, ptr [[TMP0]], i32 2
+// AARCH64-LE-NEXT:    [[TMP2:%.*]] = load i8, ptr [[TMP1]], align 2
+// AARCH64-LE-NEXT:    [[TMP3:%.*]] = and i8 [[TMP2]], 31
+// AARCH64-LE-NEXT:    store i8 [[TMP3]], ptr [[TMP1]], align 2
+// AARCH64-LE-NEXT:    [[TMP4:%.*]] = getelementptr i8, ptr [[TMP0]], i32 3
+// AARCH64-LE-NEXT:    store i8 0, ptr [[TMP4]], align 1
+// AARCH64-LE-NEXT:    ret void
+//
+// AARCH64-BE-LABEL: define dso_local void @testBitfieldUnionOverlay(
+// AARCH64-BE-SAME: ptr noundef [[S:%.*]]) #[[ATTR0]] {
+// AARCH64-BE-NEXT:  [[ENTRY:.*:]]
+// AARCH64-BE-NEXT:    [[S_ADDR:%.*]] = alloca ptr, align 8
+// AARCH64-BE-NEXT:    store ptr [[S]], ptr [[S_ADDR]], align 8
+// AARCH64-BE-NEXT:    [[TMP0:%.*]] = load ptr, ptr [[S_ADDR]], align 8
+// AARCH64-BE-NEXT:    [[TMP1:%.*]] = getelementptr i8, ptr [[TMP0]], i32 2
+// AARCH64-BE-NEXT:    [[TMP2:%.*]] = load i8, ptr [[TMP1]], align 2
+// AARCH64-BE-NEXT:    [[TMP3:%.*]] = and i8 [[TMP2]], -8
+// AARCH64-BE-NEXT:    store i8 [[TMP3]], ptr [[TMP1]], align 2
+// AARCH64-BE-NEXT:    [[TMP4:%.*]] = getelementptr i8, ptr [[TMP0]], i32 3
+// AARCH64-BE-NEXT:    store i8 0, ptr [[TMP4]], align 1
+// AARCH64-BE-NEXT:    ret void
+//
+extern "C" void testBitfieldUnionOverlay(union BitfieldUnionOverlay *s) {
+  __builtin_clear_padding(s);
+}
+
+enum BitfieldState { StateZero, StateOne, StateTwo, StateThree };
+struct BitfieldBoolEnum {
+  bool flag : 1;
+  enum BitfieldState state : 2;
+  unsigned char payload : 3;
+};
+
+// LINUX-LABEL: define dso_local void @testBitfieldBoolEnum(
+// LINUX-SAME: ptr noundef [[S:%.*]]) #[[ATTR0]] {
+// LINUX-NEXT:  [[ENTRY:.*:]]
+// LINUX-NEXT:    [[S_ADDR:%.*]] = alloca ptr, align 8
+// LINUX-NEXT:    store ptr [[S]], ptr [[S_ADDR]], align 8
+// LINUX-NEXT:    [[TMP0:%.*]] = load ptr, ptr [[S_ADDR]], align 8
+// LINUX-NEXT:    [[TMP1:%.*]] = getelementptr i8, ptr [[TMP0]], i32 0
+// LINUX-NEXT:    [[TMP2:%.*]] = load i8, ptr [[TMP1]], align 4
+// LINUX-NEXT:    [[TMP3:%.*]] = and i8 [[TMP2]], 63
+// LINUX-NEXT:    store i8 [[TMP3]], ptr [[TMP1]], align 4
+// LINUX-NEXT:    [[TMP4:%.*]] = getelementptr i8, ptr [[TMP0]], i32 1
+// LINUX-NEXT:    store i8 0, ptr [[TMP4]], align 1
+// LINUX-NEXT:    [[TMP5:%.*]] = getelementptr i8, ptr [[TMP0]], i32 2
+// LINUX-NEXT:    store i8 0, ptr [[TMP5]], align 2
+// LINUX-NEXT:    [[TMP6:%.*]] = getelementptr i8, ptr [[TMP0]], i32 3
+// LINUX-NEXT:    store i8 0, ptr [[TMP6]], align 1
+// LINUX-NEXT:    ret void
+//
+// WINDOWS-LABEL: define dso_local void @testBitfieldBoolEnum(
+// WINDOWS-SAME: ptr noundef [[S:%.*]]) #[[ATTR0]] {
+// WINDOWS-NEXT:  [[ENTRY:.*:]]
+// WINDOWS-NEXT:    [[S_ADDR:%.*]] = alloca ptr, align 8
+// WINDOWS-NEXT:    store ptr [[S]], ptr [[S_ADDR]], align 8
+// WINDOWS-NEXT:    [[TMP0:%.*]] = load ptr, ptr [[S_ADDR]], align 8
+// WINDOWS-NEXT:    [[TMP1:%.*]] = getelementptr i8, ptr [[TMP0]], i32 0
+// WINDOWS-NEXT:    [[TMP2:%.*]] = load i8, ptr [[TMP1]], align 4
+// WINDOWS-NEXT:    [[TMP3:%.*]] = and i8 [[TMP2]], 1
+// WINDOWS-NEXT:    store i8 [[TMP3]], ptr [[TMP1]], align 4
+// WINDOWS-NEXT:    [[TMP4:%.*]] = getelementptr i8, ptr [[TMP0]], i32 1
+// WINDOWS-NEXT:    store i8 0, ptr [[TMP4]], align 1
+// WINDOWS-NEXT:    [[TMP5:%.*]] = getelementptr i8, ptr [[TMP0]], i32 2
+// WINDOWS-NEXT:    store i8 0, ptr [[TMP5]], align 2
+// WINDOWS-NEXT:    [[TMP6:%.*]] = getelementptr i8, ptr [[TMP0]], i32 3
+// WINDOWS-NEXT:    store i8 0, ptr [[TMP6]], align 1
+// WINDOWS-NEXT:    [[TMP7:%.*]] = getelementptr i8, ptr [[TMP0]], i32 4
+// WINDOWS-NEXT:    [[TMP8:%.*]] = load i8, ptr [[TMP7]], align 4
+// WINDOWS-NEXT:    [[TMP9:%.*]] = and i8 [[TMP8]], 3
+// WINDOWS-NEXT:    store i8 [[TMP9]], ptr [[TMP7]], align 4
+// WINDOWS-NEXT:    [[TMP10:%.*]] = getelementptr i8, ptr [[TMP0]], i32 5
+// WINDOWS-NEXT:    store i8 0, ptr [[TMP10]], align 1
+// WINDOWS-NEXT:    [[TMP11:%.*]] = getelementptr i8, ptr [[TMP0]], i32 6
+// WINDOWS-NEXT:    store i8 0, ptr [[TMP11]], align 2
+// WINDOWS-NEXT:    [[TMP12:%.*]] = getelementptr i8, ptr [[TMP0]], i32 7
+// WINDOWS-NEXT:    store i8 0, ptr [[TMP12]], align 1
+// WINDOWS-NEXT:    [[TMP13:%.*]] = getelementptr i8, ptr [[TMP0]], i32 8
+// WINDOWS-NEXT:    [[TMP14:%.*]] = load i8, ptr [[TMP13]], align 4
+// WINDOWS-NEXT:    [[TMP15:%.*]] = and i8 [[TMP14]], 7
+// WINDOWS-NEXT:    store i8 [[TMP15]], ptr [[TMP13]], align 4
+// WINDOWS-NEXT:    [[TMP16:%.*]] = getelementptr i8, ptr [[TMP0]], i32 9
+// WINDOWS-NEXT:    store i8 0, ptr [[TMP16]], align 1
+// WINDOWS-NEXT:    [[TMP17:%.*]] = getelementptr i8, ptr [[TMP0]], i32 10
+// WINDOWS-NEXT:    store i8 0, ptr [[TMP17]], align 2
+// WINDOWS-NEXT:    [[TMP18:%.*]] = getelementptr i8, ptr [[TMP0]], i32 11
+// WINDOWS-NEXT:    store i8 0, ptr [[TMP18]], align 1
+// WINDOWS-NEXT:    ret void
+//
+// ARM-LE-LABEL: define dso_local void @testBitfieldBoolEnum(
+// ARM-LE-SAME: ptr noundef [[S:%.*]]) #[[ATTR0]] {
+// ARM-LE-NEXT:  [[ENTRY:.*:]]
+// ARM-LE-NEXT:    [[S_ADDR:%.*]] = alloca ptr, align 4
+// ARM-LE-NEXT:    store ptr [[S]], ptr [[S_ADDR]], align 4
+// ARM-LE-NEXT:    [[TMP0:%.*]] = load ptr, ptr [[S_ADDR]], align 4
+// ARM-LE-NEXT:    [[TMP1:%.*]] = getelementptr i8, ptr [[TMP0]], i32 0
+// ARM-LE-NEXT:    [[TMP2:%.*]] = load i8, ptr [[TMP1]], align 4
+// ARM-LE-NEXT:    [[TMP3:%.*]] = and i8 [[TMP2]], 63
+// ARM-LE-NEXT:    store i8 [[TMP3]], ptr [[TMP1]], align 4
+// ARM-LE-NEXT:    [[TMP4:%.*]] = getelementptr i8, ptr [[TMP0]], i32 1
+// ARM-LE-NEXT:    store i8 0, ptr [[TMP4]], align 1
+// ARM-LE-NEXT:    [[TMP5:%.*]] = getelementptr i8, ptr [[TMP0]], i32 2
+// ARM-LE-NEXT:    store i8 0, ptr [[TMP5]], align 2
+// ARM-LE-NEXT:    [[TMP6:%.*]] = getelementptr i8, ptr [[TMP0]], i32 3
+// ARM-LE-NEXT:    store i8 0, ptr [[TMP6]], align 1
+// ARM-LE-NEXT:    ret void
+//
+// ARM-BE-LABEL: define dso_local void @testBitfieldBoolEnum(
+// ARM-BE-SAME: ptr noundef [[S:%.*]]) #[[ATTR0]] {
+// ARM-BE-NEXT:  [[ENTRY:.*:]]
+// ARM-BE-NEXT:    [[S_ADDR:%.*]] = alloca ptr, align 4
+// ARM-BE-NEXT:    store ptr [[S]], ptr [[S_ADDR]], align 4
+// ARM-BE-NEXT:    [[TMP0:%.*]] = load ptr, ptr [[S_ADDR]], align 4
+// ARM-BE-NEXT:    [[TMP1:%.*]] = getelementptr i8, ptr [[TMP0]], i32 0
+// ARM-BE-NEXT:    [[TMP2:%.*]] = load i8, ptr [[TMP1]], align 4
+// ARM-BE-NEXT:    [[TMP3:%.*]] = and i8 [[TMP2]], -4
+// ARM-BE-NEXT:    store i8 [[TMP3]], ptr [[TMP1]], align 4
+// ARM-BE-NEXT:    [[TMP4:%.*]] = getelementptr i8, ptr [[TMP0]], i32 1
+// ARM-BE-NEXT:    store i8 0, ptr [[TMP4]], align 1
+// ARM-BE-NEXT:    [[TMP5:%.*]] = getelementptr i8, ptr [[TMP0]], i32 2
+// ARM-BE-NEXT:    store i8 0, ptr [[TMP5]], align 2
+// ARM-BE-NEXT:    [[TMP6:%.*]] = getelementptr i8, ptr [[TMP0]], i32 3
+// ARM-BE-NEXT:    store i8 0, ptr [[TMP6]], align 1
+// ARM-BE-NEXT:    ret void
+//
+// AARCH64-LE-LABEL: define dso_local void @testBitfieldBoolEnum(
+// AARCH64-LE-SAME: ptr noundef [[S:%.*]]) #[[ATTR0]] {
+// AARCH64-LE-NEXT:  [[ENTRY:.*:]]
+// AARCH64-LE-NEXT:    [[S_ADDR:%.*]] = alloca ptr, align 8
+// AARCH64-LE-NEXT:    store ptr [[S]], ptr [[S_ADDR]], align 8
+// AARCH64-LE-NEXT:    [[TMP0:%.*]] = load ptr, ptr [[S_ADDR]], align 8
+// AARCH64-LE-NEXT:    [[TMP1:%.*]] = getelementptr i8, ptr [[TMP0]], i32 0
+// AARCH64-LE-NEXT:    [[TMP2:%.*]] = load i8, ptr [[TMP1]], align 4
+// AARCH64-LE-NEXT:    [[TMP3:%.*]] = and i8 [[TMP2]], 63
+// AARCH64-LE-NEXT:    store i8 [[TMP3]], ptr [[TMP1]], align 4
+// AARCH64-LE-NEXT:    [[TMP4:%.*]] = getelementptr i8, ptr [[TMP0]], i32 1
+// AARCH64-LE-NEXT:    store i8 0, ptr [[TMP4]], align 1
+// AARCH64-LE-NEXT:    [[TMP5:%.*]] = getelementptr i8, ptr [[TMP0]], i32 2
+// AARCH64-LE-NEXT:    store i8 0, ptr [[TMP5]], align 2
+// AARCH64-LE-NEXT:    [[TMP6:%.*]] = getelementptr i8, ptr [[TMP0]], i32 3
+// AARCH64-LE-NEXT:    store i8 0, ptr [[TMP6]], align 1
+// AARCH64-LE-NEXT:    ret void
+//
+// AARCH64-BE-LABEL: define dso_local void @testBitfieldBoolEnum(
+// AARCH64-BE-SAME: ptr noundef [[S:%.*]]) #[[ATTR0]] {
+// AARCH64-BE-NEXT:  [[ENTRY:.*:]]
+// AARCH64-BE-NEXT:    [[S_ADDR:%.*]] = alloca ptr, align 8
+// AARCH64-BE-NEXT:    store ptr [[S]], ptr [[S_ADDR]], align 8
+// AARCH64-BE-NEXT:    [[TMP0:%.*]] = load ptr, ptr [[S_ADDR]], align 8
+// AARCH64-BE-NEXT:    [[TMP1:%.*]] = getelementptr i8, ptr [[TMP0]], i32 0
+// AARCH64-BE-NEXT:    [[TMP2:%.*]] = load i8, ptr [[TMP1]], align 4
+// AARCH64-BE-NEXT:    [[TMP3:%.*]] = and i8 [[TMP2]], -4
+// AARCH64-BE-NEXT:    store i8 [[TMP3]], ptr [[TMP1]], align 4
+// AARCH64-BE-NEXT:    [[TMP4:%.*]] = getelementptr i8, ptr [[TMP0]], i32 1
+// AARCH64-BE-NEXT:    store i8 0, ptr [[TMP4]], align 1
+// AARCH64-BE-NEXT:    [[TMP5:%.*]] = getelementptr i8, ptr [[TMP0]], i32 2
+// AARCH64-BE-NEXT:    store i8 0, ptr [[TMP5]], align 2
+// AARCH64-BE-NEXT:    [[TMP6:%.*]] = getelementptr i8, ptr [[TMP0]], i32 3
+// AARCH64-BE-NEXT:    store i8 0, ptr [[TMP6]], align 1
+// AARCH64-BE-NEXT:    ret void
+//
+extern "C" void testBitfieldBoolEnum(struct BitfieldBoolEnum *s) {
+  __builtin_clear_padding(s);
+}
+
+// Exercise bit-precise declared types and an occupied width greater than 32.
+struct BitfieldBitIntMembers {
+  unsigned _BitInt(9) small : 5;
+  signed _BitInt(17) medium : 13;
+  unsigned _BitInt(65) wide : 61;
+};
+
+// LINUX-LABEL: define dso_local void @testBitfieldBitIntMembers(
+// LINUX-SAME: ptr noundef [[S:%.*]]) #[[ATTR0]] {
+// LINUX-NEXT:  [[ENTRY:.*:]]
+// LINUX-NEXT:    [[S_ADDR:%.*]] = alloca ptr, align 8
+// LINUX-NEXT:    store ptr [[S]], ptr [[S_ADDR]], align 8
+// LINUX-NEXT:    [[TMP0:%.*]] = load ptr, ptr [[S_ADDR]], align 8
+// LINUX-NEXT:    [[TMP1:%.*]] = getelementptr i8, ptr [[TMP0]], i32 9
+// LINUX-NEXT:    [[TMP2:%.*]] = load i8, ptr [[TMP1]], align 1
+// LINUX-NEXT:    [[TMP3:%.*]] = and i8 [[TMP2]], 127
+// LINUX-NEXT:    store i8 [[TMP3]], ptr [[TMP1]], align 1
+// LINUX-NEXT:    [[TMP4:%.*]] = getelementptr i8, ptr [[TMP0]], i32 10
+// LINUX-NEXT:    store i8 0, ptr [[TMP4]], align 2
+// LINUX-NEXT:    [[TMP5:%.*]] = getelementptr i8, ptr [[TMP0]], i32 11
+// LINUX-NEXT:    store i8 0, ptr [[TMP5]], align 1
+// LINUX-NEXT:    [[TMP6:%.*]] = getelementptr i8, ptr [[TMP0]], i32 12
+// LINUX-NEXT:    store i8 0, ptr [[TMP6]], align 4
+// LINUX-NEXT:    [[TMP7:%.*]] = getelementptr i8, ptr [[TMP0]], i32 13
+// LINUX-NEXT:    store i8 0, ptr [[TMP7]], align 1
+// LINUX-NEXT:    [[TMP8:%.*]] = getelementptr i8, ptr [[TMP0]], i32 14
+// LINUX-NEXT:    store i8 0, ptr [[TMP8]], align 2
+// LINUX-NEXT:    [[TMP9:%.*]] = getelementptr i8, ptr [[TMP0]], i32 15
+// LINUX-NEXT:    store i8 0, ptr [[TMP9]], align 1
+// LINUX-NEXT:    ret void
+//
+// WINDOWS-LABEL: define dso_local void @testBitfieldBitIntMembers(
+// WINDOWS-SAME: ptr noundef [[S:%.*]]) #[[ATTR0]] {
+// WINDOWS-NEXT:  [[ENTRY:.*:]]
+// WINDOWS-NEXT:    [[S_ADDR:%.*]] = alloca ptr, align 8
+// WINDOWS-NEXT:    store ptr [[S]], ptr [[S_ADDR]], align 8
+// WINDOWS-NEXT:    [[TMP0:%.*]] = load ptr, ptr [[S_ADDR]], align 8
+// WINDOWS-NEXT:    [[TMP1:%.*]] = getelementptr i8, ptr [[TMP0]], i32 0
+// WINDOWS-NEXT:    [[TMP2:%.*]] = load i8, ptr [[TMP1]], align 8
+// WINDOWS-NEXT:    [[TMP3:%.*]] = and i8 [[TMP2]], 31
+// WINDOWS-NEXT:    store i8 [[TMP3]], ptr [[TMP1]], align 8
+// WINDOWS-NEXT:    [[TMP4:%.*]] = getelementptr i8, ptr [[TMP0]], i32 1
+// WINDOWS-NEXT:    store i8 0, ptr [[TMP4]], align 1
+// WINDOWS-NEXT:    [[TMP5:%.*]] = getelementptr i8, ptr [[TMP0]], i32 2
+// WINDOWS-NEXT:    store i8 0, ptr [[TMP5]], align 2
+// WINDOWS-NEXT:    [[TMP6:%.*]] = getelementptr i8, ptr [[TMP0]], i32 3
+// WINDOWS-NEXT:    store i8 0, ptr [[TMP6]], align 1
+// WINDOWS-NEXT:    [[TMP7:%.*]] = getelementptr i8, ptr [[TMP0]], i32 5
+// WINDOWS-NEXT:    [[TMP8:%.*]] = load i8, ptr [[TMP7]], align 1
+// WINDOWS-NEXT:    [[TMP9:%.*]] = and i8 [[TMP8]], 31
+// WINDOWS-NEXT:    store i8 [[TMP9]], ptr [[TMP7]], align 1
+// WINDOWS-NEXT:    [[TMP10:%.*]] = getelementptr i8, ptr [[TMP0]], i32 6
+// WINDOWS-NEXT:    store i8 0, ptr [[TMP10]], align 2
+// WINDOWS-NEXT:    [[TMP11:%.*]] = getelementptr i8, ptr [[TMP0]], i32 7
+// WINDOWS-NEXT:    store i8 0, ptr [[TMP11]], align 1
+// WINDOWS-NEXT:    [[TMP12:%.*]] = getelementptr i8, ptr [[TMP0]], i32 15
+// WINDOWS-NEXT:    [[TMP13:%.*]] = load i8, ptr [[TMP12]], align 1
+// WINDOWS-NEXT:    [[TMP14:%.*]] = and i8 [[TMP13]], 31
+// WINDOWS-NEXT:    store i8 [[TMP14]], ptr [[TMP12]], align 1
+// WINDOWS-NEXT:    [[TMP15:%.*]] = getelementptr i8, ptr [[TMP0]], i32 16
+// WINDOWS-NEXT:    store i8 0, ptr [[TMP15]], align 8
+// WINDOWS-NEXT:    [[TMP16:%.*]] = getelementptr i8, ptr [[TMP0]], i32 17
+// WINDOWS-NEXT:    store i8 0, ptr [[TMP16]], align 1
+// WINDOWS-NEXT:    [[TMP17:%.*]] = getelementptr i8, ptr [[TMP0]], i32 18
+// WINDOWS-NEXT:    store i8 0, ptr [[TMP17]], align 2
+// WINDOWS-NEXT:    [[TMP18:%.*]] = getelementptr i8, ptr [[TMP0]], i32 19
+// WINDOWS-NEXT:    store i8 0, ptr [[TMP18]], align 1
+// WINDOWS-NEXT:    [[TMP19:%.*]] = getelementptr i8, ptr [[TMP0]], i32 20
+// WINDOWS-NEXT:    store i8 0, ptr [[TMP19]], align 4
+// WINDOWS-NEXT:    [[TMP20:%.*]] = getelementptr i8, ptr [[TMP0]], i32 21
+// WINDOWS-NEXT:    store i8 0, ptr [[TMP20]], align 1
+// WINDOWS-NEXT:    [[TMP21:%.*]] = getelementptr i8, ptr [[TMP0]], i32 22
+// WINDOWS-NEXT:    store i8 0, ptr [[TMP21]], align 2
+// WINDOWS-NEXT:    [[TMP22:%.*]] = getelementptr i8, ptr [[TMP0]], i32 23
+// WINDOWS-NEXT:    store i8 0, ptr [[TMP22]], align 1
+// WINDOWS-NEXT:    ret void
+//
+// ARM-LE-LABEL: define dso_local void @testBitfieldBitIntMembers(
+// ARM-LE-SAME: ptr noundef [[S:%.*]]) #[[ATTR0]] {
+// ARM-LE-NEXT:  [[ENTRY:.*:]]
+// ARM-LE-NEXT:    [[S_ADDR:%.*]] = alloca ptr, align 4
+// ARM-LE-NEXT:    store ptr [[S]], ptr [[S_ADDR]], align 4
+// ARM-LE-NEXT:    [[TMP0:%.*]] = load ptr, ptr [[S_ADDR]], align 4
+// ARM-LE-NEXT:    [[TMP1:%.*]] = getelementptr i8, ptr [[TMP0]], i32 9
+// ARM-LE-NEXT:    [[TMP2:%.*]] = load i8, ptr [[TMP1]], align 1
+// ARM-LE-NEXT:    [[TMP3:%.*]] = and i8 [[TMP2]], 127
+// ARM-LE-NEXT:    store i8 [[TMP3]], ptr [[TMP1]], align 1
+// ARM-LE-NEXT:    [[TMP4:%.*]] = getelementptr i8, ptr [[TMP0]], i32 10
+// ARM-LE-NEXT:    store i8 0, ptr [[TMP4]], align 2
+// ARM-LE-NEXT:    [[TMP5:%.*]] = getelementptr i8, ptr [[TMP0]], i32 11
+// ARM-LE-NEXT:    store i8 0, ptr [[TMP5]], align 1
+// ARM-LE-NEXT:    [[TMP6:%.*]] = getelementptr i8, ptr [[TMP0]], i32 12
+// ARM-LE-NEXT:    store i8 0, ptr [[TMP6]], align 4
+// ARM-LE-NEXT:    [[TMP7:%.*]] = getelementptr i8, ptr [[TMP0]], i32 13
+// ARM-LE-NEXT:    store i8 0, ptr [[TMP7]], align 1
+// ARM-LE-NEXT:    [[TMP8:%.*]] = getelementptr i8, ptr [[TMP0]], i32 14
+// ARM-LE-NEXT:    store i8 0, ptr [[TMP8]], align 2
+// ARM-LE-NEXT:    [[TMP9:%.*]] = getelementptr i8, ptr [[TMP0]], i32 15
+// ARM-LE-NEXT:    store i8 0, ptr [[TMP9]], align 1
+// ARM-LE-NEXT:    ret void
+//
+// ARM-BE-LABEL: define dso_local void @testBitfieldBitIntMembers(
+// ARM-BE-SAME: ptr noundef [[S:%.*]]) #[[ATTR0]] {
+// ARM-BE-NEXT:  [[ENTRY:.*:]]
+// ARM-BE-NEXT:    [[S_ADDR:%.*]] = alloca ptr, align 4
+// ARM-BE-NEXT:    store ptr [[S]], ptr [[S_ADDR]], align 4
+// ARM-BE-NEXT:    [[TMP0:%.*]] = load ptr, ptr [[S_ADDR]], align 4
+// ARM-BE-NEXT:    [[TMP1:%.*]] = getelementptr i8, ptr [[TMP0]], i32 9
+// ARM-BE-NEXT:    [[TMP2:%.*]] = load i8, ptr [[TMP1]], align 1
+// ARM-BE-NEXT:    [[TMP3:%.*]] = and i8 [[TMP2]], -2
+// ARM-BE-NEXT:    store i8 [[TMP3]], ptr [[TMP1]], align 1
+// ARM-BE-NEXT:    [[TMP4:%.*]] = getelementptr i8, ptr [[TMP0]], i32 10
+// ARM-BE-NEXT:    store i8 0, ptr [[TMP4]], align 2
+// ARM-BE-NEXT:    [[TMP5:%.*]] = getelementptr i8, ptr [[TMP0]], i32 11
+// ARM-BE-NEXT:    store i8 0, ptr [[TMP5]], align 1
+// ARM-BE-NEXT:    [[TMP6:%.*]] = getelementptr i8, ptr [[TMP0]], i32 12
+// ARM-BE-NEXT:    store i8 0, ptr [[TMP6]], align 4
+// ARM-BE-NEXT:    [[TMP7:%.*]] = getelementptr i8, ptr [[TMP0]], i32 13
+// ARM-BE-NEXT:    store i8 0, ptr [[TMP7]], align 1
+// ARM-BE-NEXT:    [[TMP8:%.*]] = getelementptr i8, ptr [[TMP0]], i32 14
+// ARM-BE-NEXT:    store i8 0, ptr [[TMP8]], align 2
+// ARM-BE-NEXT:    [[TMP9:%.*]] = getelementptr i8, ptr [[TMP0]], i32 15
+// ARM-BE-NEXT:    store i8 0, ptr [[TMP9]], align 1
+// ARM-BE-NEXT:    ret void
+//
+// AARCH64-LE-LABEL: define dso_local void @testBitfieldBitIntMembers(
+// AARCH64-LE-SAME: ptr noundef [[S:%.*]]) #[[ATTR0]] {
+// AARCH64-LE-NEXT:  [[ENTRY:.*:]]
+// AARCH64-LE-NEXT:    [[S_ADDR:%.*]] = alloca ptr, align 8
+// AARCH64-LE-NEXT:    store ptr [[S]], ptr [[S_ADDR]], align 8
+// AARCH64-LE-NEXT:    [[TMP0:%.*]] = load ptr, ptr [[S_ADDR]], align 8
+// AARCH64-LE-NEXT:    [[TMP1:%.*]] = getelementptr i8, ptr [[TMP0]], i32 9
+// AARCH64-LE-NEXT:    [[TMP2:%.*]] = load i8, ptr [[TMP1]], align 1
+// AARCH64-LE-NEXT:    [[TMP3:%.*]] = and i8 [[TMP2]], 127
+// AARCH64-LE-NEXT:    store i8 [[TMP3]], ptr [[TMP1]], align 1
+// AARCH64-LE-NEXT:    [[TMP4:%.*]] = getelementptr i8, ptr [[TMP0]], i32 10
+// AARCH64-LE-NEXT:    store i8 0, ptr [[TMP4]], align 2
+// AARCH64-LE-NEXT:    [[TMP5:%.*]] = getelementptr i8, ptr [[TMP0]], i32 11
+// AARCH64-LE-NEXT:    store i8 0, ptr [[TMP5]], align 1
+// AARCH64-LE-NEXT:    [[TMP6:%.*]] = getelementptr i8, ptr [[TMP0]], i32 12
+// AARCH64-LE-NEXT:    store i8 0, ptr [[TMP6]], align 4
+// AARCH64-LE-NEXT:    [[TMP7:%.*]] = getelementptr i8, ptr [[TMP0]], i32 13
+// AARCH64-LE-NEXT:    store i8 0, ptr [[TMP7]], align 1
+// AARCH64-LE-NEXT:    [[TMP8:%.*]] = getelementptr i8, ptr [[TMP0]], i32 14
+// AARCH64-LE-NEXT:    store i8 0, ptr [[TMP8]], align 2
+// AARCH64-LE-NEXT:    [[TMP9:%.*]] = getelementptr i8, ptr [[TMP0]], i32 15
+// AARCH64-LE-NEXT:    store i8 0, ptr [[TMP9]], align 1
+// AARCH64-LE-NEXT:    ret void
+//
+// AARCH64-BE-LABEL: define dso_local void @testBitfieldBitIntMembers(
+// AARCH64-BE-SAME: ptr noundef [[S:%.*]]) #[[ATTR0]] {
+// AARCH64-BE-NEXT:  [[ENTRY:.*:]]
+// AARCH64-BE-NEXT:    [[S_ADDR:%.*]] = alloca ptr, align 8
+// AARCH64-BE-NEXT:    store ptr [[S]], ptr [[S_ADDR]], align 8
+// AARCH64-BE-NEXT:    [[TMP0:%.*]] = load ptr, ptr [[S_ADDR]], align 8
+// AARCH64-BE-NEXT:    [[TMP1:%.*]] = getelementptr i8, ptr [[TMP0]], i32 9
+// AARCH64-BE-NEXT:    [[TMP2:%.*]] = load i8, ptr [[TMP1]], align 1
+// AARCH64-BE-NEXT:    [[TMP3:%.*]] = and i8 [[TMP2]], -2
+// AARCH64-BE-NEXT:    store i8 [[TMP3]], ptr [[TMP1]], align 1
+// AARCH64-BE-NEXT:    [[TMP4:%.*]] = getelementptr i8, ptr [[TMP0]], i32 10
+// AARCH64-BE-NEXT:    store i8 0, ptr [[TMP4]], align 2
+// AARCH64-BE-NEXT:    [[TMP5:%.*]] = getelementptr i8, ptr [[TMP0]], i32 11
+// AARCH64-BE-NEXT:    store i8 0, ptr [[TMP5]], align 1
+// AARCH64-BE-NEXT:    [[TMP6:%.*]] = getelementptr i8, ptr [[TMP0]], i32 12
+// AARCH64-BE-NEXT:    store i8 0, ptr [[TMP6]], align 4
+// AARCH64-BE-NEXT:    [[TMP7:%.*]] = getelementptr i8, ptr [[TMP0]], i32 13
+// AARCH64-BE-NEXT:    store i8 0, ptr [[TMP7]], align 1
+// AARCH64-BE-NEXT:    [[TMP8:%.*]] = getelementptr i8, ptr [[TMP0]], i32 14
+// AARCH64-BE-NEXT:    store i8 0, ptr [[TMP8]], align 2
+// AARCH64-BE-NEXT:    [[TMP9:%.*]] = getelementptr i8, ptr [[TMP0]], i32 15
+// AARCH64-BE-NEXT:    store i8 0, ptr [[TMP9]], align 1
+// AARCH64-BE-NEXT:    ret void
+//
+extern "C" void testBitfieldBitIntMembers(struct BitfieldBitIntMembers *s) {
+  __builtin_clear_padding(s);
+}
+
+// Packing permits the wide field to straddle its declared type's usual
+// storage-unit boundary on Arm. Ordinary members surround the bitfield run.
+struct __attribute__((packed)) BitfieldPackedMembers {
+  unsigned char prefix;
+  unsigned int first : 5;
+  unsigned long long wide : 60;
+  unsigned char suffix;
+};
+
+// LINUX-LABEL: define dso_local void @testBitfieldPackedMembers(
+// LINUX-SAME: ptr noundef [[S:%.*]]) #[[ATTR0]] {
+// LINUX-NEXT:  [[ENTRY:.*:]]
+// LINUX-NEXT:    [[S_ADDR:%.*]] = alloca ptr, align 8
+// LINUX-NEXT:    store ptr [[S]], ptr [[S_ADDR]], align 8
+// LINUX-NEXT:    [[TMP0:%.*]] = load ptr, ptr [[S_ADDR]], align 8
+// LINUX-NEXT:    [[TMP1:%.*]] = getelementptr i8, ptr [[TMP0]], i32 9
+// LINUX-NEXT:    [[TMP2:%.*]] = load i8, ptr [[TMP1]], align 1
+// LINUX-NEXT:    [[TMP3:%.*]] = and i8 [[TMP2]], 1
+// LINUX-NEXT:    store i8 [[TMP3]], ptr [[TMP1]], align 1
+// LINUX-NEXT:    ret void
+//
+// WINDOWS-LABEL: define dso_local void @testBitfieldPackedMembers(
+// WINDOWS-SAME: ptr noundef [[S:%.*]]) #[[ATTR0]] {
+// WINDOWS-NEXT:  [[ENTRY:.*:]]
+// WINDOWS-NEXT:    [[S_ADDR:%.*]] = alloca ptr, align 8
+// WINDOWS-NEXT:    store ptr [[S]], ptr [[S_ADDR]], align 8
+// WINDOWS-NEXT:    [[TMP0:%.*]] = load ptr, ptr [[S_ADDR]], align 8
+// WINDOWS-NEXT:    [[TMP1:%.*]] = getelementptr i8, ptr [[TMP0]], i32 1
+// WINDOWS-NEXT:    [[TMP2:%.*]] = load i8, ptr [[TMP1]], align 1
+// WINDOWS-NEXT:    [[TMP3:%.*]] = and i8 [[TMP2]], 31
+// WINDOWS-NEXT:    store i8 [[TMP3]], ptr [[TMP1]], align 1
+// WINDOWS-NEXT:    [[TMP4:%.*]] = getelementptr i8, ptr [[TMP0]], i32 2
+// WINDOWS-NEXT:    store i8 0, ptr [[TMP4]], align 1
+// WINDOWS-NEXT:    [[TMP5:%.*]] = getelementptr i8, ptr [[TMP0]], i32 3
+// WINDOWS-NEXT:    store i8 0, ptr [[TMP5]], align 1
+// WINDOWS-NEXT:    [[TMP6:%.*]] = getelementptr i8, ptr [[TMP0]], i32 4
+// WINDOWS-NEXT:    store i8 0, ptr [[TMP6]], align 1
+// WINDOWS-NEXT:    [[TMP7:%.*]] = getelementptr i8, ptr [[TMP0]], i32 12
+// WINDOWS-NEXT:    [[TMP8:%.*]] = load i8, ptr [[TMP7]], align 1
+// WINDOWS-NEXT:    [[TMP9:%.*]] = and i8 [[TMP8]], 15
+// WINDOWS-NEXT:    store i8 [[TMP9]], ptr [[TMP7]], align 1
+// WINDOWS-NEXT:    ret void
+//
+// ARM-LE-LABEL: define dso_local void @testBitfieldPackedMembers(
+// ARM-LE-SAME: ptr noundef [[S:%.*]]) #[[ATTR0]] {
+// ARM-LE-NEXT:  [[ENTRY:.*:]]
+// ARM-LE-NEXT:    [[S_ADDR:%.*]] = alloca ptr, align 4
+// ARM-LE-NEXT:    store ptr [[S]], ptr [[S_ADDR]], align 4
+// ARM-LE-NEXT:    [[TMP0:%.*]] = load ptr, ptr [[S_ADDR]], align 4
+// ARM-LE-NEXT:    [[TMP1:%.*]] = getelementptr i8, ptr [[TMP0]], i32 9
+// ARM-LE-NEXT:    [[TMP2:%.*]] = load i8, ptr [[TMP1]], align 1
+// ARM-LE-NEXT:    [[TMP3:%.*]] = and i8 [[TMP2]], 1
+// ARM-LE-NEXT:    store i8 [[TMP3]], ptr [[TMP1]], align 1
+// ARM-LE-NEXT:    ret void
+//
+// ARM-BE-LABEL: define dso_local void @testBitfieldPackedMembers(
+// ARM-BE-SAME: ptr noundef [[S:%.*]]) #[[ATTR0]] {
+// ARM-BE-NEXT:  [[ENTRY:.*:]]
+// ARM-BE-NEXT:    [[S_ADDR:%.*]] = alloca ptr, align 4
+// ARM-BE-NEXT:    store ptr [[S]], ptr [[S_ADDR]], align 4
+// ARM-BE-NEXT:    [[TMP0:%.*]] = load ptr, ptr [[S_ADDR]], align 4
+// ARM-BE-NEXT:    [[TMP1:%.*]] = getelementptr i8, ptr [[TMP0]], i32 9
+// ARM-BE-NEXT:    [[TMP2:%.*]] = load i8, ptr [[TMP1]], align 1
+// ARM-BE-NEXT:    [[TMP3:%.*]] = and i8 [[TMP2]], -128
+// ARM-BE-NEXT:    store i8 [[TMP3]], ptr [[TMP1]], align 1
+// ARM-BE-NEXT:    ret void
+//
+// AARCH64-LE-LABEL: define dso_local void @testBitfieldPackedMembers(
+// AARCH64-LE-SAME: ptr noundef [[S:%.*]]) #[[ATTR0]] {
+// AARCH64-LE-NEXT:  [[ENTRY:.*:]]
+// AARCH64-LE-NEXT:    [[S_ADDR:%.*]] = alloca ptr, align 8
+// AARCH64-LE-NEXT:    store ptr [[S]], ptr [[S_ADDR]], align 8
+// AARCH64-LE-NEXT:    [[TMP0:%.*]] = load ptr, ptr [[S_ADDR]], align 8
+// AARCH64-LE-NEXT:    [[TMP1:%.*]] = getelementptr i8, ptr [[TMP0]], i32 9
+// AARCH64-LE-NEXT:    [[TMP2:%.*]] = load i8, ptr [[TMP1]], align 1
+// AARCH64-LE-NEXT:    [[TMP3:%.*]] = and i8 [[TMP2]], 1
+// AARCH64-LE-NEXT:    store i8 [[TMP3]], ptr [[TMP1]], align 1
+// AARCH64-LE-NEXT:    ret void
+//
+// AARCH64-BE-LABEL: define dso_local void @testBitfieldPackedMembers(
+// AARCH64-BE-SAME: ptr noundef [[S:%.*]]) #[[ATTR0]] {
+// AARCH64-BE-NEXT:  [[ENTRY:.*:]]
+// AARCH64-BE-NEXT:    [[S_ADDR:%.*]] = alloca ptr, align 8
+// AARCH64-BE-NEXT:    store ptr [[S]], ptr [[S_ADDR]], align 8
+// AARCH64-BE-NEXT:    [[TMP0:%.*]] = load ptr, ptr [[S_ADDR]], align 8
+// AARCH64-BE-NEXT:    [[TMP1:%.*]] = getelementptr i8, ptr [[TMP0]], i32 9
+// AARCH64-BE-NEXT:    [[TMP2:%.*]] = load i8, ptr [[TMP1]], align 1
+// AARCH64-BE-NEXT:    [[TMP3:%.*]] = and i8 [[TMP2]], -128
+// AARCH64-BE-NEXT:    store i8 [[TMP3]], ptr [[TMP1]], align 1
+// AARCH64-BE-NEXT:    ret void
+//
+extern "C" void testBitfieldPackedMembers(struct BitfieldPackedMembers *s) {
+  __builtin_clear_padding(s);
+}
+
+// There are no occupied intervals: clear the entire object representation.
+struct BitfieldUnnamedOnly {
+  unsigned int : 3;
+  unsigned int : 0;
+  unsigned char : 2;
+};
+
+// LINUX-LABEL: define dso_local void @testBitfieldUnnamedOnly(
+// LINUX-SAME: ptr noundef [[S:%.*]]) #[[ATTR0]] {
+// LINUX-NEXT:  [[ENTRY:.*:]]
+// LINUX-NEXT:    [[S_ADDR:%.*]] = alloca ptr, align 8
+// LINUX-NEXT:    store ptr [[S]], ptr [[S_ADDR]], align 8
+// LINUX-NEXT:    [[TMP0:%.*]] = load ptr, ptr [[S_ADDR]], align 8
+// LINUX-NEXT:    [[TMP1:%.*]] = getelementptr i8, ptr [[TMP0]], i32 0
+// LINUX-NEXT:    store i8 0, ptr [[TMP1]], align 1
+// LINUX-NEXT:    [[TMP2:%.*]] = getelementptr i8, ptr [[TMP0]], i32 1
+// LINUX-NEXT:    store i8 0, ptr [[TMP2]], align 1
+// LINUX-NEXT:    [[TMP3:%.*]] = getelementptr i8, ptr [[TMP0]], i32 2
+// LINUX-NEXT:    store i8 0, ptr [[TMP3]], align 1
+// LINUX-NEXT:    [[TMP4:%.*]] = getelementptr i8, ptr [[TMP0]], i32 3
+// LINUX-NEXT:    store i8 0, ptr [[TMP4]], align 1
+// LINUX-NEXT:    [[TMP5:%.*]] = getelementptr i8, ptr [[TMP0]], i32 4
+// LINUX-NEXT:    store i8 0, ptr [[TMP5]], align 1
+// LINUX-NEXT:    ret void
+//
+// WINDOWS-LABEL: define dso_local void @testBitfieldUnnamedOnly(
+// WINDOWS-SAME: ptr noundef [[S:%.*]]) #[[ATTR0]] {
+// WINDOWS-NEXT:  [[ENTRY:.*:]]
+// WINDOWS-NEXT:    [[S_ADDR:%.*]] = alloca ptr, align 8
+// WINDOWS-NEXT:    store ptr [[S]], ptr [[S_ADDR]], align 8
+// WINDOWS-NEXT:    [[TMP0:%.*]] = load ptr, ptr [[S_ADDR]], align 8
+// WINDOWS-NEXT:    [[TMP1:%.*]] = getelementptr i8, ptr [[TMP0]], i32 0
+// WINDOWS-NEXT:    store i8 0, ptr [[TMP1]], align 4
+// WINDOWS-NEXT:    [[TMP2:%.*]] = getelementptr i8, ptr [[TMP0]], i32 1
+// WINDOWS-NEXT:    store i8 0, ptr [[TMP2]], align 1
+// WINDOWS-NEXT:    [[TMP3:%.*]] = getelementptr i8, ptr [[TMP0]], i32 2
+// WINDOWS-NEXT:    store i8 0, ptr [[TMP3]], align 2
+// WINDOWS-NEXT:    [[TMP4:%.*]] = getelementptr i8, ptr [[TMP0]], i32 3
+// WINDOWS-NEXT:    store i8 0, ptr [[TMP4]], align 1
+// WINDOWS-NEXT:    [[TMP5:%.*]] = getelementptr i8, ptr [[TMP0]], i32 4
+// WINDOWS-NEXT:    store i8 0, ptr [[TMP5]], align 4
+// WINDOWS-NEXT:    [[TMP6:%.*]] = getelementptr i8, ptr [[TMP0]], i32 5
+// WINDOWS-NEXT:    store i8 0, ptr [[TMP6]], align 1
+// WINDOWS-NEXT:    [[TMP7:%.*]] = getelementptr i8, ptr [[TMP0]], i32 6
+// WINDOWS-NEXT:    store i8 0, ptr [[TMP7]], align 2
+// WINDOWS-NEXT:    [[TMP8:%.*]] = getelementptr i8, ptr [[TMP0]], i32 7
+// WINDOWS-NEXT:    store i8 0, ptr [[TMP8]], align 1
+// WINDOWS-NEXT:    ret void
+//
+// ARM-LE-LABEL: define dso_local void @testBitfieldUnnamedOnly(
+// ARM-LE-SAME: ptr noundef [[S:%.*]]) #[[ATTR0]] {
+// ARM-LE-NEXT:  [[ENTRY:.*:]]
+// ARM-LE-NEXT:    [[S_ADDR:%.*]] = alloca ptr, align 4
+// ARM-LE-NEXT:    store ptr [[S]], ptr [[S_ADDR]], align 4
+// ARM-LE-NEXT:    [[TMP0:%.*]] = load ptr, ptr [[S_ADDR]], align 4
+// ARM-LE-NEXT:    [[TMP1:%.*]] = getelementptr i8, ptr [[TMP0]], i32 0
+// ARM-LE-NEXT:    store i8 0, ptr [[TMP1]], align 4
+// ARM-LE-NEXT:    [[TMP2:%.*]] = getelementptr i8, ptr [[TMP0]], i32 1
+// ARM-LE-NEXT:    store i8 0, ptr [[TMP2]], align 1
+// ARM-LE-NEXT:    [[TMP3:%.*]] = getelementptr i8, ptr [[TMP0]], i32 2
+// ARM-LE-NEXT:    store i8 0, ptr [[TMP3]], align 2
+// ARM-LE-NEXT:    [[TMP4:%.*]] = getelementptr i8, ptr [[TMP0]], i32 3
+// ARM-LE-NEXT:    store i8 0, ptr [[TMP4]], align 1
+// ARM-LE-NEXT:    [[TMP5:%.*]] = getelementptr i8, ptr [[TMP0]], i32 4
+// ARM-LE-NEXT:    store i8 0, ptr [[TMP5]], align 4
+// ARM-LE-NEXT:    [[TMP6:%.*]] = getelementptr i8, ptr [[TMP0]], i32 5
+// ARM-LE-NEXT:    store i8 0, ptr [[TMP6]], align 1
+// ARM-LE-NEXT:    [[TMP7:%.*]] = getelementptr i8, ptr [[TMP0]], i32 6
+// ARM-LE-NEXT:    store i8 0, ptr [[TMP7]], align 2
+// ARM-LE-NEXT:    [[TMP8:%.*]] = getelementptr i8, ptr [[TMP0]], i32 7
+// ARM-LE-NEXT:    store i8 0, ptr [[TMP8]], align 1
+// ARM-LE-NEXT:    ret void
+//
+// ARM-BE-LABEL: define dso_local void @testBitfieldUnnamedOnly(
+// ARM-BE-SAME: ptr noundef [[S:%.*]]) #[[ATTR0]] {
+// ARM-BE-NEXT:  [[ENTRY:.*:]]
+// ARM-BE-NEXT:    [[S_ADDR:%.*]] = alloca ptr, align 4
+// ARM-BE-NEXT:    store ptr [[S]], ptr [[S_ADDR]], align 4
+// ARM-BE-NEXT:    [[TMP0:%.*]] = load ptr, ptr [[S_ADDR]], align 4
+// ARM-BE-NEXT:    [[TMP1:%.*]] = getelementptr i8, ptr [[TMP0]], i32 0
+// ARM-BE-NEXT:    store i8 0, ptr [[TMP1]], align 4
+// ARM-BE-NEXT:    [[TMP2:%.*]] = getelementptr i8, ptr [[TMP0]], i32 1
+// ARM-BE-NEXT:    store i8 0, ptr [[TMP2]], align 1
+// ARM-BE-NEXT:    [[TMP3:%.*]] = getelementptr i8, ptr [[TMP0]], i32 2
+// ARM-BE-NEXT:    store i8 0, ptr [[TMP3]], align 2
+// ARM-BE-NEXT:    [[TMP4:%.*]] = getelementptr i8, ptr [[TMP0]], i32 3
+// ARM-BE-NEXT:    store i8 0, ptr [[TMP4]], align 1
+// ARM-BE-NEXT:    [[TMP5:%.*]] = getelementptr i8, ptr [[TMP0]], i32 4
+// ARM-BE-NEXT:    store i8 0, ptr [[TMP5]], align 4
+// ARM-BE-NEXT:    [[TMP6:%.*]] = getelementptr i8, ptr [[TMP0]], i32 5
+// ARM-BE-NEXT:    store i8 0, ptr [[TMP6]], align 1
+// ARM-BE-NEXT:    [[TMP7:%.*]] = getelementptr i8, ptr [[TMP0]], i32 6
+// ARM-BE-NEXT:    store i8 0, ptr [[TMP7]], align 2
+// ARM-BE-NEXT:    [[TMP8:%.*]] = getelementptr i8, ptr [[TMP0]], i32 7
+// ARM-BE-NEXT:    store i8 0, ptr [[TMP8]], align 1
+// ARM-BE-NEXT:    ret void
+//
+// AARCH64-LE-LABEL: define dso_local void @testBitfieldUnnamedOnly(
+// AARCH64-LE-SAME: ptr noundef [[S:%.*]]) #[[ATTR0]] {
+// AARCH64-LE-NEXT:  [[ENTRY:.*:]]
+// AARCH64-LE-NEXT:    [[S_ADDR:%.*]] = alloca ptr, align 8
+// AARCH64-LE-NEXT:    store ptr [[S]], ptr [[S_ADDR]], align 8
+// AARCH64-LE-NEXT:    [[TMP0:%.*]] = load ptr, ptr [[S_ADDR]], align 8
+// AARCH64-LE-NEXT:    [[TMP1:%.*]] = getelementptr i8, ptr [[TMP0]], i32 0
+// AARCH64-LE-NEXT:    store i8 0, ptr [[TMP1]], align 4
+// AARCH64-LE-NEXT:    [[TMP2:%.*]] = getelementptr i8, ptr [[TMP0]], i32 1
+// AARCH64-LE-NEXT:    store i8 0, ptr [[TMP2]], align 1
+// AARCH64-LE-NEXT:    [[TMP3:%.*]] = getelementptr i8, ptr [[TMP0]], i32 2
+// AARCH64-LE-NEXT:    store i8 0, ptr [[TMP3]], align 2
+// AARCH64-LE-NEXT:    [[TMP4:%.*]] = getelementptr i8, ptr [[TMP0]], i32 3
+// AARCH64-LE-NEXT:    store i8 0, ptr [[TMP4]], align 1
+// AARCH64-LE-NEXT:    [[TMP5:%.*]] = getelementptr i8, ptr [[TMP0]], i32 4
+// AARCH64-LE-NEXT:    store i8 0, ptr [[TMP5]], align 4
+// AARCH64-LE-NEXT:    [[TMP6:%.*]] = getelementptr i8, ptr [[TMP0]], i32 5
+// AARCH64-LE-NEXT:    store i8 0, ptr [[TMP6]], align 1
+// AARCH64-LE-NEXT:    [[TMP7:%.*]] = getelementptr i8, ptr [[TMP0]], i32 6
+// AARCH64-LE-NEXT:    store i8 0, ptr [[TMP7]], align 2
+// AARCH64-LE-NEXT:    [[TMP8:%.*]] = getelementptr i8, ptr [[TMP0]], i32 7
+// AARCH64-LE-NEXT:    store i8 0, ptr [[TMP8]], align 1
+// AARCH64-LE-NEXT:    ret void
+//
+// AARCH64-BE-LABEL: define dso_local void @testBitfieldUnnamedOnly(
+// AARCH64-BE-SAME: ptr noundef [[S:%.*]]) #[[ATTR0]] {
+// AARCH64-BE-NEXT:  [[ENTRY:.*:]]
+// AARCH64-BE-NEXT:    [[S_ADDR:%.*]] = alloca ptr, align 8
+// AARCH64-BE-NEXT:    store ptr [[S]], ptr [[S_ADDR]], align 8
+// AARCH64-BE-NEXT:    [[TMP0:%.*]] = load ptr, ptr [[S_ADDR]], align 8
+// AARCH64-BE-NEXT:    [[TMP1:%.*]] = getelementptr i8, ptr [[TMP0]], i32 0
+// AARCH64-BE-NEXT:    store i8 0, ptr [[TMP1]], align 4
+// AARCH64-BE-NEXT:    [[TMP2:%.*]] = getelementptr i8, ptr [[TMP0]], i32 1
+// AARCH64-BE-NEXT:    store i8 0, ptr [[TMP2]], align 1
+// AARCH64-BE-NEXT:    [[TMP3:%.*]] = getelementptr i8, ptr [[TMP0]], i32 2
+// AARCH64-BE-NEXT:    store i8 0, ptr [[TMP3]], align 2
+// AARCH64-BE-NEXT:    [[TMP4:%.*]] = getelementptr i8, ptr [[TMP0]], i32 3
+// AARCH64-BE-NEXT:    store i8 0, ptr [[TMP4]], align 1
+// AARCH64-BE-NEXT:    [[TMP5:%.*]] = getelementptr i8, ptr [[TMP0]], i32 4
+// AARCH64-BE-NEXT:    store i8 0, ptr [[TMP5]], align 4
+// AARCH64-BE-NEXT:    [[TMP6:%.*]] = getelementptr i8, ptr [[TMP0]], i32 5
+// AARCH64-BE-NEXT:    store i8 0, ptr [[TMP6]], align 1
+// AARCH64-BE-NEXT:    [[TMP7:%.*]] = getelementptr i8, ptr [[TMP0]], i32 6
+// AARCH64-BE-NEXT:    store i8 0, ptr [[TMP7]], align 2
+// AARCH64-BE-NEXT:    [[TMP8:%.*]] = getelementptr i8, ptr [[TMP0]], i32 7
+// AARCH64-BE-NEXT:    store i8 0, ptr [[TMP8]], align 1
+// AARCH64-BE-NEXT:    ret void
+//
+extern "C" void testBitfieldUnnamedOnly(struct BitfieldUnnamedOnly *s) {
+  __builtin_clear_padding(s);
+}
+
+#ifndef _WIN32
+#pragma clang diagnostic push
+#pragma clang diagnostic ignored "-Wbitfield-width"
+
+// Match Clang's existing oversized-field representation. On big endian the
+// occupied byte is byte 1, so clearing padding must leave that byte intact.
+struct BitfieldOversizedAligned {
+  unsigned char value : 16;
+};
+
+// LINUX-LABEL: define dso_local void @testBitfieldOversizedAligned(
+// LINUX-SAME: ptr noundef [[S:%.*]]) #[[ATTR0]] {
+// LINUX-NEXT:  [[ENTRY:.*:]]
+// LINUX-NEXT:    [[S_ADDR:%.*]] = alloca ptr, align 8
+// LINUX-NEXT:    store ptr [[S]], ptr [[S_ADDR]], align 8
+// LINUX-NEXT:    [[TMP0:%.*]] = load ptr, ptr [[S_ADDR]], align 8
+// LINUX-NEXT:    [[TMP1:%.*]] = getelementptr i8, ptr [[TMP0]], i32 1
+// LINUX-NEXT:    store i8 0, ptr [[TMP1]], align 1
+// LINUX-NEXT:    ret void
+//
+// ARM-LE-LABEL: define dso_local void @testBitfieldOversizedAligned(
+// ARM-LE-SAME: ptr noundef [[S:%.*]]) #[[ATTR0]] {
+// ARM-LE-NEXT:  [[ENTRY:.*:]]
+// ARM-LE-NEXT:    [[S_ADDR:%.*]] = alloca ptr, align 4
+// ARM-LE-NEXT:    store ptr [[S]], ptr [[S_ADDR]], align 4
+// ARM-LE-NEXT:    [[TMP0:%.*]] = load ptr, ptr [[S_ADDR]], align 4
+// ARM-LE-NEXT:    [[TMP1:%.*]] = getelementptr i8, ptr [[TMP0]], i32 1
+// ARM-LE-NEXT:    store i8 0, ptr [[TMP1]], align 1
+// ARM-LE-NEXT:    ret void
+//
+// ARM-BE-LABEL: define dso_local void @testBitfieldOversizedAligned(
+// ARM-BE-SAME: ptr noundef [[S:%.*]]) #[[ATTR0]] {
+// ARM-BE-NEXT:  [[ENTRY:.*:]]
+// ARM-BE-NEXT:    [[S_ADDR:%.*]] = alloca ptr, align 4
+// ARM-BE-NEXT:    store ptr [[S]], ptr [[S_ADDR]], align 4
+// ARM-BE-NEXT:    [[TMP0:%.*]] = load ptr, ptr [[S_ADDR]], align 4
+// ARM-BE-NEXT:    [[TMP1:%.*]] = getelementptr i8, ptr [[TMP0]], i32 0
+// ARM-BE-NEXT:    store i8 0, ptr [[TMP1]], align 2
+// ARM-BE-NEXT:    ret void
+//
+// AARCH64-LE-LABEL: define dso_local void @testBitfieldOversizedAligned(
+// AARCH64-LE-SAME: ptr noundef [[S:%.*]]) #[[ATTR0]] {
+// AARCH64-LE-NEXT:  [[ENTRY:.*:]]
+// AARCH64-LE-NEXT:    [[S_ADDR:%.*]] = alloca ptr, align 8
+// AARCH64-LE-NEXT:    store ptr [[S]], ptr [[S_ADDR]], align 8
+// AARCH64-LE-NEXT:    [[TMP0:%.*]] = load ptr, ptr [[S_ADDR]], align 8
+// AARCH64-LE-NEXT:    [[TMP1:%.*]] = getelementptr i8, ptr [[TMP0]], i32 1
+// AARCH64-LE-NEXT:    store i8 0, ptr [[TMP1]], align 1
+// AARCH64-LE-NEXT:    ret void
+//
+// AARCH64-BE-LABEL: define dso_local void @testBitfieldOversizedAligned(
+// AARCH64-BE-SAME: ptr noundef [[S:%.*]]) #[[ATTR0]] {
+// AARCH64-BE-NEXT:  [[ENTRY:.*:]]
+// AARCH64-BE-NEXT:    [[S_ADDR:%.*]] = alloca ptr, align 8
+// AARCH64-BE-NEXT:    store ptr [[S]], ptr [[S_ADDR]], align 8
+// AARCH64-BE-NEXT:    [[TMP0:%.*]] = load ptr, ptr [[S_ADDR]], align 8
+// AARCH64-BE-NEXT:    [[TMP1:%.*]] = getelementptr i8, ptr [[TMP0]], i32 0
+// AARCH64-BE-NEXT:    store i8 0, ptr [[TMP1]], align 2
+// AARCH64-BE-NEXT:    ret void
+//
+extern "C" void testBitfieldOversizedAligned(BitfieldOversizedAligned *s) {
+  __builtin_clear_padding(s);
+}
+
+// The occupied bits cross a byte boundary after more than one byte of leading
+// padding on big endian. The declared width also exceeds the allocation type.
+struct BitfieldOversizedBeyondBoundary {
+  unsigned char value : 17;
+};
+
+// LINUX-LABEL: define dso_local void @testBitfieldOversizedBeyondBoundary(
+// LINUX-SAME: ptr noundef [[S:%.*]]) #[[ATTR0]] {
+// LINUX-NEXT:  [[ENTRY:.*:]]
+// LINUX-NEXT:    [[S_ADDR:%.*]] = alloca ptr, align 8
+// LINUX-NEXT:    store ptr [[S]], ptr [[S_ADDR]], align 8
+// LINUX-NEXT:    [[TMP0:%.*]] = load ptr, ptr [[S_ADDR]], align 8
+// LINUX-NEXT:    [[TMP1:%.*]] = getelementptr i8, ptr [[TMP0]], i32 1
+// LINUX-NEXT:    store i8 0, ptr [[TMP1]], align 1
+// LINUX-NEXT:    [[TMP2:%.*]] = getelementptr i8, ptr [[TMP0]], i32 2
+// LINUX-NEXT:    store i8 0, ptr [[TMP2]], align 2
+// LINUX-NEXT:    [[TMP3:%.*]] = getelementptr i8, ptr [[TMP0]], i32 3
+// LINUX-NEXT:    store i8 0, ptr [[TMP3]], align 1
+// LINUX-NEXT:    ret void
+//
+// ARM-LE-LABEL: define dso_local void @testBitfieldOversizedBeyondBoundary(
+// ARM-LE-SAME: ptr noundef [[S:%.*]]) #[[ATTR0]] {
+// ARM-LE-NEXT:  [[ENTRY:.*:]]
+// ARM-LE-NEXT:    [[S_ADDR:%.*]] = alloca ptr, align 4
+// ARM-LE-NEXT:    store ptr [[S]], ptr [[S_ADDR]], align 4
+// ARM-LE-NEXT:    [[TMP0:%.*]] = load ptr, ptr [[S_ADDR]], align 4
+// ARM-LE-NEXT:    [[TMP1:%.*]] = getelementptr i8, ptr [[TMP0]], i32 1
+// ARM-LE-NEXT:    store i8 0, ptr [[TMP1]], align 1
+// ARM-LE-NEXT:    [[TMP2:%.*]] = getelementptr i8, ptr [[TMP0]], i32 2
+// ARM-LE-NEXT:    store i8 0, ptr [[TMP2]], align 2
+// ARM-LE-NEXT:    [[TMP3:%.*]] = getelementptr i8, ptr [[TMP0]], i32 3
+// ARM-LE-NEXT:    store i8 0, ptr [[TMP3]], align 1
+// ARM-LE-NEXT:    ret void
+//
+// ARM-BE-LABEL: define dso_local void @testBitfieldOversizedBeyondBoundary(
+// ARM-BE-SAME: ptr noundef [[S:%.*]]) #[[ATTR0]] {
+// ARM-BE-NEXT:  [[ENTRY:.*:]]
+// ARM-BE-NEXT:    [[S_ADDR:%.*]] = alloca ptr, align 4
+// ARM-BE-NEXT:    store ptr [[S]], ptr [[S_ADDR]], align 4
+// ARM-BE-NEXT:    [[TMP0:%.*]] = load ptr, ptr [[S_ADDR]], align 4
+// ARM-BE-NEXT:    [[TMP1:%.*]] = getelementptr i8, ptr [[TMP0]], i32 0
+// ARM-BE-NEXT:    store i8 0, ptr [[TMP1]], align 2
+// ARM-BE-NEXT:    [[TMP2:%.*]] = getelementptr i8, ptr [[TMP0]], i32 1
+// ARM-BE-NEXT:    [[TMP3:%.*]] = load i8, ptr [[TMP2]], align 1
+// ARM-BE-NEXT:    [[TMP4:%.*]] = and i8 [[TMP3]], 127
+// ARM-BE-NEXT:    store i8 [[TMP4]], ptr [[TMP2]], align 1
+// ARM-BE-NEXT:    [[TMP5:%.*]] = getelementptr i8, ptr [[TMP0]], i32 2
+// ARM-BE-NEXT:    [[TMP6:%.*]] = load i8, ptr [[TMP5]], align 2
+// ARM-BE-NEXT:    [[TMP7:%.*]] = and i8 [[TMP6]], -128
+// ARM-BE-NEXT:    store i8 [[TMP7]], ptr [[TMP5]], align 2
+// ARM-BE-NEXT:    [[TMP8:%.*]] = getelementptr i8, ptr [[TMP0]], i32 3
+// ARM-BE-NEXT:    store i8 0, ptr [[TMP8]], align 1
+// ARM-BE-NEXT:    ret void
+//
+// AARCH64-LE-LABEL: define dso_local void @testBitfieldOversizedBeyondBoundary(
+// AARCH64-LE-SAME: ptr noundef [[S:%.*]]) #[[ATTR0]] {
+// AARCH64-LE-NEXT:  [[ENTRY:.*:]]
+// AARCH64-LE-NEXT:    [[S_ADDR:%.*]] = alloca ptr, align 8
+// AARCH64-LE-NEXT:    store ptr [[S]], ptr [[S_ADDR]], align 8
+// AARCH64-LE-NEXT:    [[TMP0:%.*]] = load ptr, ptr [[S_ADDR]], align 8
+// AARCH64-LE-NEXT:    [[TMP1:%.*]] = getelementptr i8, ptr [[TMP0]], i32 1
+// AARCH64-LE-NEXT:    store i8 0, ptr [[TMP1]], align 1
+// AARCH64-LE-NEXT:    [[TMP2:%.*]] = getelementptr i8, ptr [[TMP0]], i32 2
+// AARCH64-LE-NEXT:    store i8 0, ptr [[TMP2]], align 2
+// AARCH64-LE-NEXT:    [[TMP3:%.*]] = getelementptr i8, ptr [[TMP0]], i32 3
+// AARCH64-LE-NEXT:    store i8 0, ptr [[TMP3]], align 1
+// AARCH64-LE-NEXT:    ret void
+//
+// AARCH64-BE-LABEL: define dso_local void @testBitfieldOversizedBeyondBoundary(
+// AARCH64-BE-SAME: ptr noundef [[S:%.*]]) #[[ATTR0]] {
+// AARCH64-BE-NEXT:  [[ENTRY:.*:]]
+// AARCH64-BE-NEXT:    [[S_ADDR:%.*]] = alloca ptr, align 8
+// AARCH64-BE-NEXT:    store ptr [[S]], ptr [[S_ADDR]], align 8
+// AARCH64-BE-NEXT:    [[TMP0:%.*]] = load ptr, ptr [[S_ADDR]], align 8
+// AARCH64-BE-NEXT:    [[TMP1:%.*]] = getelementptr i8, ptr [[TMP0]], i32 0
+// AARCH64-BE-NEXT:    store i8 0, ptr [[TMP1]], align 2
+// AARCH64-BE-NEXT:    [[TMP2:%.*]] = getelementptr i8, ptr [[TMP0]], i32 1
+// AARCH64-BE-NEXT:    [[TMP3:%.*]] = load i8, ptr [[TMP2]], align 1
+// AARCH64-BE-NEXT:    [[TMP4:%.*]] = and i8 [[TMP3]], 127
+// AARCH64-BE-NEXT:    store i8 [[TMP4]], ptr [[TMP2]], align 1
+// AARCH64-BE-NEXT:    [[TMP5:%.*]] = getelementptr i8, ptr [[TMP0]], i32 2
+// AARCH64-BE-NEXT:    [[TMP6:%.*]] = load i8, ptr [[TMP5]], align 2
+// AARCH64-BE-NEXT:    [[TMP7:%.*]] = and i8 [[TMP6]], -128
+// AARCH64-BE-NEXT:    store i8 [[TMP7]], ptr [[TMP5]], align 2
+// AARCH64-BE-NEXT:    [[TMP8:%.*]] = getelementptr i8, ptr [[TMP0]], i32 3
+// AARCH64-BE-NEXT:    store i8 0, ptr [[TMP8]], align 1
+// AARCH64-BE-NEXT:    ret void
+//
+extern "C" void testBitfieldOversizedBeyondBoundary(
+    BitfieldOversizedBeyondBoundary *s) {
+  __builtin_clear_padding(s);
+}
+
+// Value widths differ from both the declared bitfield widths and, for bool and
+// _BitInt, their storage sizes.
+enum class BitfieldSmallEnum : unsigned char { Zero, One };
+struct BitfieldOversizedValueTypes {
+  bool flag : 3;
+  BitfieldSmallEnum state : 12;
+  unsigned _BitInt(9) precise : 12;
+};
+
+// LINUX-LABEL: define dso_local void @testBitfieldOversizedValueTypes(
+// LINUX-SAME: ptr noundef [[S:%.*]]) #[[ATTR0]] {
+// LINUX-NEXT:  [[ENTRY:.*:]]
+// LINUX-NEXT:    [[S_ADDR:%.*]] = alloca ptr, align 8
+// LINUX-NEXT:    store ptr [[S]], ptr [[S_ADDR]], align 8
+// LINUX-NEXT:    [[TMP0:%.*]] = load ptr, ptr [[S_ADDR]], align 8
+// LINUX-NEXT:    [[TMP1:%.*]] = getelementptr i8, ptr [[TMP0]], i32 0
+// LINUX-NEXT:    [[TMP2:%.*]] = load i8, ptr [[TMP1]], align 2
+// LINUX-NEXT:    [[TMP3:%.*]] = and i8 [[TMP2]], 1
+// LINUX-NEXT:    store i8 [[TMP3]], ptr [[TMP1]], align 2
+// LINUX-NEXT:    [[TMP4:%.*]] = getelementptr i8, ptr [[TMP0]], i32 2
+// LINUX-NEXT:    [[TMP5:%.*]] = load i8, ptr [[TMP4]], align 2
+// LINUX-NEXT:    [[TMP6:%.*]] = and i8 [[TMP5]], -16
+// LINUX-NEXT:    store i8 [[TMP6]], ptr [[TMP4]], align 2
+// LINUX-NEXT:    [[TMP7:%.*]] = getelementptr i8, ptr [[TMP0]], i32 3
+// LINUX-NEXT:    [[TMP8:%.*]] = load i8, ptr [[TMP7]], align 1
+// LINUX-NEXT:    [[TMP9:%.*]] = and i8 [[TMP8]], 31
+// LINUX-NEXT:    store i8 [[TMP9]], ptr [[TMP7]], align 1
+// LINUX-NEXT:    ret void
+//
+// ARM-LE-LABEL: define dso_local void @testBitfieldOversizedValueTypes(
+// ARM-LE-SAME: ptr noundef [[S:%.*]]) #[[ATTR0]] {
+// ARM-LE-NEXT:  [[ENTRY:.*:]]
+// ARM-LE-NEXT:    [[S_ADDR:%.*]] = alloca ptr, align 4
+// ARM-LE-NEXT:    store ptr [[S]], ptr [[S_ADDR]], align 4
+// ARM-LE-NEXT:    [[TMP0:%.*]] = load ptr, ptr [[S_ADDR]], align 4
+// ARM-LE-NEXT:    [[TMP1:%.*]] = getelementptr i8, ptr [[TMP0]], i32 0
+// ARM-LE-NEXT:    [[TMP2:%.*]] = load i8, ptr [[TMP1]], align 2
+// ARM-LE-NEXT:    [[TMP3:%.*]] = and i8 [[TMP2]], 1
+// ARM-LE-NEXT:    store i8 [[TMP3]], ptr [[TMP1]], align 2
+// ARM-LE-NEXT:    [[TMP4:%.*]] = getelementptr i8, ptr [[TMP0]], i32 2
+// ARM-LE-NEXT:    [[TMP5:%.*]] = load i8, ptr [[TMP4]], align 2
+// ARM-LE-NEXT:    [[TMP6:%.*]] = and i8 [[TMP5]], -16
+// ARM-LE-NEXT:    store i8 [[TMP6]], ptr [[TMP4]], align 2
+// ARM-LE-NEXT:    [[TMP7:%.*]] = getelementptr i8, ptr [[TMP0]], i32 3
+// ARM-LE-NEXT:    [[TMP8:%.*]] = load i8, ptr [[TMP7]], align 1
+// ARM-LE-NEXT:    [[TMP9:%.*]] = and i8 [[TMP8]], 31
+// ARM-LE-NEXT:    store i8 [[TMP9]], ptr [[TMP7]], align 1
+// ARM-LE-NEXT:    ret void
+//
+// ARM-BE-LABEL: define dso_local void @testBitfieldOversizedValueTypes(
+// ARM-BE-SAME: ptr noundef [[S:%.*]]) #[[ATTR0]] {
+// ARM-BE-NEXT:  [[ENTRY:.*:]]
+// ARM-BE-NEXT:    [[S_ADDR:%.*]] = alloca ptr, align 4
+// ARM-BE-NEXT:    store ptr [[S]], ptr [[S_ADDR]], align 4
+// ARM-BE-NEXT:    [[TMP0:%.*]] = load ptr, ptr [[S_ADDR]], align 4
+// ARM-BE-NEXT:    [[TMP1:%.*]] = getelementptr i8, ptr [[TMP0]], i32 0
+// ARM-BE-NEXT:    [[TMP2:%.*]] = load i8, ptr [[TMP1]], align 2
+// ARM-BE-NEXT:    [[TMP3:%.*]] = and i8 [[TMP2]], -32
+// ARM-BE-NEXT:    store i8 [[TMP3]], ptr [[TMP1]], align 2
+// ARM-BE-NEXT:    [[TMP4:%.*]] = getelementptr i8, ptr [[TMP0]], i32 0
+// ARM-BE-NEXT:    [[TMP5:%.*]] = load i8, ptr [[TMP4]], align 2
+// ARM-BE-NEXT:    [[TMP6:%.*]] = and i8 [[TMP5]], 63
+// ARM-BE-NEXT:    store i8 [[TMP6]], ptr [[TMP4]], align 2
+// ARM-BE-NEXT:    [[TMP7:%.*]] = getelementptr i8, ptr [[TMP0]], i32 1
+// ARM-BE-NEXT:    [[TMP8:%.*]] = load i8, ptr [[TMP7]], align 1
+// ARM-BE-NEXT:    [[TMP9:%.*]] = and i8 [[TMP8]], 15
+// ARM-BE-NEXT:    store i8 [[TMP9]], ptr [[TMP7]], align 1
+// ARM-BE-NEXT:    [[TMP10:%.*]] = getelementptr i8, ptr [[TMP0]], i32 2
+// ARM-BE-NEXT:    [[TMP11:%.*]] = load i8, ptr [[TMP10]], align 2
+// ARM-BE-NEXT:    [[TMP12:%.*]] = and i8 [[TMP11]], -15
+// ARM-BE-NEXT:    store i8 [[TMP12]], ptr [[TMP10]], align 2
+// ARM-BE-NEXT:    ret void
+//
+// AARCH64-LE-LABEL: define dso_local void @testBitfieldOversizedValueTypes(
+// AARCH64-LE-SAME: ptr noundef [[S:%.*]]) #[[ATTR0]] {
+// AARCH64-LE-NEXT:  [[ENTRY:.*:]]
+// AARCH64-LE-NEXT:    [[S_ADDR:%.*]] = alloca ptr, align 8
+// AARCH64-LE-NEXT:    store ptr [[S]], ptr [[S_ADDR]], align 8
+// AARCH64-LE-NEXT:    [[TMP0:%.*]] = load ptr, ptr [[S_ADDR]], align 8
+// AARCH64-LE-NEXT:    [[TMP1:%.*]] = getelementptr i8, ptr [[TMP0]], i32 0
+// AARCH64-LE-NEXT:    [[TMP2:%.*]] = load i8, ptr [[TMP1]], align 2
+// AARCH64-LE-NEXT:    [[TMP3:%.*]] = and i8 [[TMP2]], 1
+// AARCH64-LE-NEXT:    store i8 [[TMP3]], ptr [[TMP1]], align 2
+// AARCH64-LE-NEXT:    [[TMP4:%.*]] = getelementptr i8, ptr [[TMP0]], i32 2
+// AARCH64-LE-NEXT:    [[TMP5:%.*]] = load i8, ptr [[TMP4]], align 2
+// AARCH64-LE-NEXT:    [[TMP6:%.*]] = and i8 [[TMP5]], -16
+// AARCH64-LE-NEXT:    store i8 [[TMP6]], ptr [[TMP4]], align 2
+// AARCH64-LE-NEXT:    [[TMP7:%.*]] = getelementptr i8, ptr [[TMP0]], i32 3
+// AARCH64-LE-NEXT:    [[TMP8:%.*]] = load i8, ptr [[TMP7]], align 1
+// AARCH64-LE-NEXT:    [[TMP9:%.*]] = and i8 [[TMP8]], 31
+// AARCH64-LE-NEXT:    store i8 [[TMP9]], ptr [[TMP7]], align 1
+// AARCH64-LE-NEXT:    ret void
+//
+// AARCH64-BE-LABEL: define dso_local void @testBitfieldOversizedValueTypes(
+// AARCH64-BE-SAME: ptr noundef [[S:%.*]]) #[[ATTR0]] {
+// AARCH64-BE-NEXT:  [[ENTRY:.*:]]
+// AARCH64-BE-NEXT:    [[S_ADDR:%.*]] = alloca ptr, align 8
+// AARCH64-BE-NEXT:    store ptr [[S]], ptr [[S_ADDR]], align 8
+// AARCH64-BE-NEXT:    [[TMP0:%.*]] = load ptr, ptr [[S_ADDR]], align 8
+// AARCH64-BE-NEXT:    [[TMP1:%.*]] = getelementptr i8, ptr [[TMP0]], i32 0
+// AARCH64-BE-NEXT:    [[TMP2:%.*]] = load i8, ptr [[TMP1]], align 2
+// AARCH64-BE-NEXT:    [[TMP3:%.*]] = and i8 [[TMP2]], -32
+// AARCH64-BE-NEXT:    store i8 [[TMP3]], ptr [[TMP1]], align 2
+// AARCH64-BE-NEXT:    [[TMP4:%.*]] = getelementptr i8, ptr [[TMP0]], i32 0
+// AARCH64-BE-NEXT:    [[TMP5:%.*]] = load i8, ptr [[TMP4]], align 2
+// AARCH64-BE-NEXT:    [[TMP6:%.*]] = and i8 [[TMP5]], 63
+// AARCH64-BE-NEXT:    store i8 [[TMP6]], ptr [[TMP4]], align 2
+// AARCH64-BE-NEXT:    [[TMP7:%.*]] = getelementptr i8, ptr [[TMP0]], i32 1
+// AARCH64-BE-NEXT:    [[TMP8:%.*]] = load i8, ptr [[TMP7]], align 1
+// AARCH64-BE-NEXT:    [[TMP9:%.*]] = and i8 [[TMP8]], 15
+// AARCH64-BE-NEXT:    store i8 [[TMP9]], ptr [[TMP7]], align 1
+// AARCH64-BE-NEXT:    [[TMP10:%.*]] = getelementptr i8, ptr [[TMP0]], i32 2
+// AARCH64-BE-NEXT:    [[TMP11:%.*]] = load i8, ptr [[TMP10]], align 2
+// AARCH64-BE-NEXT:    [[TMP12:%.*]] = and i8 [[TMP11]], -15
+// AARCH64-BE-NEXT:    store i8 [[TMP12]], ptr [[TMP10]], align 2
+// AARCH64-BE-NEXT:    ret void
+//
+extern "C" void testBitfieldOversizedValueTypes(
+    BitfieldOversizedValueTypes *s) {
+  __builtin_clear_padding(s);
+}
+
+#pragma clang diagnostic pop
+#endif
+
+struct BitfieldPrefixBase {
+  unsigned int prefix;
+};
+
+struct BitfieldDerived : BitfieldPrefixBase, BitfieldPartialEnds {
+  unsigned char suffix;
+};
+
+// LINUX-LABEL: define dso_local void @testBitfieldBaseOffset(
+// LINUX-SAME: ptr noundef [[S:%.*]]) #[[ATTR0]] {
+// LINUX-NEXT:  [[ENTRY:.*:]]
+// LINUX-NEXT:    [[S_ADDR:%.*]] = alloca ptr, align 8
+// LINUX-NEXT:    store ptr [[S]], ptr [[S_ADDR]], align 8
+// LINUX-NEXT:    [[TMP0:%.*]] = load ptr, ptr [[S_ADDR]], align 8
+// LINUX-NEXT:    [[TMP1:%.*]] = getelementptr i8, ptr [[TMP0]], i32 4
+// LINUX-NEXT:    [[TMP2:%.*]] = load i8, ptr [[TMP1]], align 4
+// LINUX-NEXT:    [[TMP3:%.*]] = and i8 [[TMP2]], -8
+// LINUX-NEXT:    store i8 [[TMP3]], ptr [[TMP1]], align 4
+// LINUX-NEXT:    [[TMP4:%.*]] = getelementptr i8, ptr [[TMP0]], i32 6
+// LINUX-NEXT:    [[TMP5:%.*]] = load i8, ptr [[TMP4]], align 2
+// LINUX-NEXT:    [[TMP6:%.*]] = and i8 [[TMP5]], 31
+// LINUX-NEXT:    store i8 [[TMP6]], ptr [[TMP4]], align 2
+// LINUX-NEXT:    [[TMP7:%.*]] = getelementptr i8, ptr [[TMP0]], i32 7
+// LINUX-NEXT:    store i8 0, ptr [[TMP7]], align 1
+// LINUX-NEXT:    [[TMP8:%.*]] = getelementptr i8, ptr [[TMP0]], i32 9
+// LINUX-NEXT:    store i8 0, ptr [[TMP8]], align 1
+// LINUX-NEXT:    [[TMP9:%.*]] = getelementptr i8, ptr [[TMP0]], i32 10
+// LINUX-NEXT:    store i8 0, ptr [[TMP9]], align 2
+// LINUX-NEXT:    [[TMP10:%.*]] = getelementptr i8, ptr [[TMP0]], i32 11
+// LINUX-NEXT:    store i8 0, ptr [[TMP10]], align 1
+// LINUX-NEXT:    ret void
+//
+// WINDOWS-LABEL: define dso_local void @testBitfieldBaseOffset(
+// WINDOWS-SAME: ptr noundef [[S:%.*]]) #[[ATTR0]] {
+// WINDOWS-NEXT:  [[ENTRY:.*:]]
+// WINDOWS-NEXT:    [[S_ADDR:%.*]] = alloca ptr, align 8
+// WINDOWS-NEXT:    store ptr [[S]], ptr [[S_ADDR]], align 8
+// WINDOWS-NEXT:    [[TMP0:%.*]] = load ptr, ptr [[S_ADDR]], align 8
+// WINDOWS-NEXT:    [[TMP1:%.*]] = getelementptr i8, ptr [[TMP0]], i32 4
+// WINDOWS-NEXT:    [[TMP2:%.*]] = load i8, ptr [[TMP1]], align 4
+// WINDOWS-NEXT:    [[TMP3:%.*]] = and i8 [[TMP2]], -8
+// WINDOWS-NEXT:    store i8 [[TMP3]], ptr [[TMP1]], align 4
+// WINDOWS-NEXT:    [[TMP4:%.*]] = getelementptr i8, ptr [[TMP0]], i32 6
+// WINDOWS-NEXT:    [[TMP5:%.*]] = load i8, ptr [[TMP4]], align 2
+// WINDOWS-NEXT:    [[TMP6:%.*]] = and i8 [[TMP5]], 31
+// WINDOWS-NEXT:    store i8 [[TMP6]], ptr [[TMP4]], align 2
+// WINDOWS-NEXT:    [[TMP7:%.*]] = getelementptr i8, ptr [[TMP0]], i32 7
+// WINDOWS-NEXT:    store i8 0, ptr [[TMP7]], align 1
+// WINDOWS-NEXT:    [[TMP8:%.*]] = getelementptr i8, ptr [[TMP0]], i32 9
+// WINDOWS-NEXT:    store i8 0, ptr [[TMP8]], align 1
+// WINDOWS-NEXT:    [[TMP9:%.*]] = getelementptr i8, ptr [[TMP0]], i32 10
+// WINDOWS-NEXT:    store i8 0, ptr [[TMP9]], align 2
+// WINDOWS-NEXT:    [[TMP10:%.*]] = getelementptr i8, ptr [[TMP0]], i32 11
+// WINDOWS-NEXT:    store i8 0, ptr [[TMP10]], align 1
+// WINDOWS-NEXT:    ret void
+//
+// ARM-LE-LABEL: define dso_local void @testBitfieldBaseOffset(
+// ARM-LE-SAME: ptr noundef [[S:%.*]]) #[[ATTR0]] {
+// ARM-LE-NEXT:  [[ENTRY:.*:]]
+// ARM-LE-NEXT:    [[S_ADDR:%.*]] = alloca ptr, align 4
+// ARM-LE-NEXT:    store ptr [[S]], ptr [[S_ADDR]], align 4
+// ARM-LE-NEXT:    [[TMP0:%.*]] = load ptr, ptr [[S_ADDR]], align 4
+// ARM-LE-NEXT:    [[TMP1:%.*]] = getelementptr i8, ptr [[TMP0]], i32 4
+// ARM-LE-NEXT:    [[TMP2:%.*]] = load i8, ptr [[TMP1]], align 4
+// ARM-LE-NEXT:    [[TMP3:%.*]] = and i8 [[TMP2]], -8
+// ARM-LE-NEXT:    store i8 [[TMP3]], ptr [[TMP1]], align 4
+// ARM-LE-NEXT:    [[TMP4:%.*]] = getelementptr i8, ptr [[TMP0]], i32 6
+// ARM-LE-NEXT:    [[TMP5:%.*]] = load i8, ptr [[TMP4]], align 2
+// ARM-LE-NEXT:    [[TMP6:%.*]] = and i8 [[TMP5]], 31
+// ARM-LE-NEXT:    store i8 [[TMP6]], ptr [[TMP4]], align 2
+// ARM-LE-NEXT:    [[TMP7:%.*]] = getelementptr i8, ptr [[TMP0]], i32 7
+// ARM-LE-NEXT:    store i8 0, ptr [[TMP7]], align 1
+// ARM-LE-NEXT:    [[TMP8:%.*]] = getelementptr i8, ptr [[TMP0]], i32 9
+// ARM-LE-NEXT:    store i8 0, ptr [[TMP8]], align 1
+// ARM-LE-NEXT:    [[TMP9:%.*]] = getelementptr i8, ptr [[TMP0]], i32 10
+// ARM-LE-NEXT:    store i8 0, ptr [[TMP9]], align 2
+// ARM-LE-NEXT:    [[TMP10:%.*]] = getelementptr i8, ptr [[TMP0]], i32 11
+// ARM-LE-NEXT:    store i8 0, ptr [[TMP10]], align 1
+// ARM-LE-NEXT:    ret void
+//
+// ARM-BE-LABEL: define dso_local void @testBitfieldBaseOffset(
+// ARM-BE-SAME: ptr noundef [[S:%.*]]) #[[ATTR0]] {
+// ARM-BE-NEXT:  [[ENTRY:.*:]]
+// ARM-BE-NEXT:    [[S_ADDR:%.*]] = alloca ptr, align 4
+// ARM-BE-NEXT:    store ptr [[S]], ptr [[S_ADDR]], align 4
+// ARM-BE-NEXT:    [[TMP0:%.*]] = load ptr, ptr [[S_ADDR]], align 4
+// ARM-BE-NEXT:    [[TMP1:%.*]] = getelementptr i8, ptr [[TMP0]], i32 4
+// ARM-BE-NEXT:    [[TMP2:%.*]] = load i8, ptr [[TMP1]], align 4
+// ARM-BE-NEXT:    [[TMP3:%.*]] = and i8 [[TMP2]], 31
+// ARM-BE-NEXT:    store i8 [[TMP3]], ptr [[TMP1]], align 4
+// ARM-BE-NEXT:    [[TMP4:%.*]] = getelementptr i8, ptr [[TMP0]], i32 6
+// ARM-BE-NEXT:    [[TMP5:%.*]] = load i8, ptr [[TMP4]], align 2
+// ARM-BE-NEXT:    [[TMP6:%.*]] = and i8 [[TMP5]], -8
+// ARM-BE-NEXT:    store i8 [[TMP6]], ptr [[TMP4]], align 2
+// ARM-BE-NEXT:    [[TMP7:%.*]] = getelementptr i8, ptr [[TMP0]], i32 7
+// ARM-BE-NEXT:    store i8 0, ptr [[TMP7]], align 1
+// ARM-BE-NEXT:    [[TMP8:%.*]] = getelementptr i8, ptr [[TMP0]], i32 9
+// ARM-BE-NEXT:    store i8 0, ptr [[TMP8]], align 1
+// ARM-BE-NEXT:    [[TMP9:%.*]] = getelementptr i8, ptr [[TMP0]], i32 10
+// ARM-BE-NEXT:    store i8 0, ptr [[TMP9]], align 2
+// ARM-BE-NEXT:    [[TMP10:%.*]] = getelementptr i8, ptr [[TMP0]], i32 11
+// ARM-BE-NEXT:    store i8 0, ptr [[TMP10]], align 1
+// ARM-BE-NEXT:    ret void
+//
+// AARCH64-LE-LABEL: define dso_local void @testBitfieldBaseOffset(
+// AARCH64-LE-SAME: ptr noundef [[S:%.*]]) #[[ATTR0]] {
+// AARCH64-LE-NEXT:  [[ENTRY:.*:]]
+// AARCH64-LE-NEXT:    [[S_ADDR:%.*]] = alloca ptr, align 8
+// AARCH64-LE-NEXT:    store ptr [[S]], ptr [[S_ADDR]], align 8
+// AARCH64-LE-NEXT:    [[TMP0:%.*]] = load ptr, ptr [[S_ADDR]], align 8
+// AARCH64-LE-NEXT:    [[TMP1:%.*]] = getelementptr i8, ptr [[TMP0]], i32 4
+// AARCH64-LE-NEXT:    [[TMP2:%.*]] = load i8, ptr [[TMP1]], align 4
+// AARCH64-LE-NEXT:    [[TMP3:%.*]] = and i8 [[TMP2]], -8
+// AARCH64-LE-NEXT:    store i8 [[TMP3]], ptr [[TMP1]], align 4
+// AARCH64-LE-NEXT:    [[TMP4:%.*]] = getelementptr i8, ptr [[TMP0]], i32 6
+// AARCH64-LE-NEXT:    [[TMP5:%.*]] = load i8, ptr [[TMP4]], align 2
+// AARCH64-LE-NEXT:    [[TMP6:%.*]] = and i8 [[TMP5]], 31
+// AARCH64-LE-NEXT:    store i8 [[TMP6]], ptr [[TMP4]], align 2
+// AARCH64-LE-NEXT:    [[TMP7:%.*]] = getelementptr i8, ptr [[TMP0]], i32 7
+// AARCH64-LE-NEXT:    store i8 0, ptr [[TMP7]], align 1
+// AARCH64-LE-NEXT:    [[TMP8:%.*]] = getelementptr i8, ptr [[TMP0]], i32 9
+// AARCH64-LE-NEXT:    store i8 0, ptr [[TMP8]], align 1
+// AARCH64-LE-NEXT:    [[TMP9:%.*]] = getelementptr i8, ptr [[TMP0]], i32 10
+// AARCH64-LE-NEXT:    store i8 0, ptr [[TMP9]], align 2
+// AARCH64-LE-NEXT:    [[TMP10:%.*]] = getelementptr i8, ptr [[TMP0]], i32 11
+// AARCH64-LE-NEXT:    store i8 0, ptr [[TMP10]], align 1
+// AARCH64-LE-NEXT:    ret void
+//
+// AARCH64-BE-LABEL: define dso_local void @testBitfieldBaseOffset(
+// AARCH64-BE-SAME: ptr noundef [[S:%.*]]) #[[ATTR0]] {
+// AARCH64-BE-NEXT:  [[ENTRY:.*:]]
+// AARCH64-BE-NEXT:    [[S_ADDR:%.*]] = alloca ptr, align 8
+// AARCH64-BE-NEXT:    store ptr [[S]], ptr [[S_ADDR]], align 8
+// AARCH64-BE-NEXT:    [[TMP0:%.*]] = load ptr, ptr [[S_ADDR]], align 8
+// AARCH64-BE-NEXT:    [[TMP1:%.*]] = getelementptr i8, ptr [[TMP0]], i32 4
+// AARCH64-BE-NEXT:    [[TMP2:%.*]] = load i8, ptr [[TMP1]], align 4
+// AARCH64-BE-NEXT:    [[TMP3:%.*]] = and i8 [[TMP2]], 31
+// AARCH64-BE-NEXT:    store i8 [[TMP3]], ptr [[TMP1]], align 4
+// AARCH64-BE-NEXT:    [[TMP4:%.*]] = getelementptr i8, ptr [[TMP0]], i32 6
+// AARCH64-BE-NEXT:    [[TMP5:%.*]] = load i8, ptr [[TMP4]], align 2
+// AARCH64-BE-NEXT:    [[TMP6:%.*]] = and i8 [[TMP5]], -8
+// AARCH64-BE-NEXT:    store i8 [[TMP6]], ptr [[TMP4]], align 2
+// AARCH64-BE-NEXT:    [[TMP7:%.*]] = getelementptr i8, ptr [[TMP0]], i32 7
+// AARCH64-BE-NEXT:    store i8 0, ptr [[TMP7]], align 1
+// AARCH64-BE-NEXT:    [[TMP8:%.*]] = getelementptr i8, ptr [[TMP0]], i32 9
+// AARCH64-BE-NEXT:    store i8 0, ptr [[TMP8]], align 1
+// AARCH64-BE-NEXT:    [[TMP9:%.*]] = getelementptr i8, ptr [[TMP0]], i32 10
+// AARCH64-BE-NEXT:    store i8 0, ptr [[TMP9]], align 2
+// AARCH64-BE-NEXT:    [[TMP10:%.*]] = getelementptr i8, ptr [[TMP0]], i32 11
+// AARCH64-BE-NEXT:    store i8 0, ptr [[TMP10]], align 1
+// AARCH64-BE-NEXT:    ret void
+//
+extern "C" void testBitfieldBaseOffset(BitfieldDerived *s) {
+  __builtin_clear_padding(s);
+}
+
+struct BitfieldVirtualDerived : virtual BitfieldPartialEnds {
+  unsigned char tag;
+};
+
+// LINUX-LABEL: define dso_local void @testBitfieldVirtualBase(
+// LINUX-SAME: ptr nofreeobj noundef align 8 dead_on_return dereferenceable(16) [[S:%.*]]) #[[ATTR0]] {
+// LINUX-NEXT:  [[ENTRY:.*:]]
+// LINUX-NEXT:    [[S_INDIRECT_ADDR:%.*]] = alloca ptr, align 8
+// LINUX-NEXT:    store ptr [[S]], ptr [[S_INDIRECT_ADDR]], align 8
+// LINUX-NEXT:    [[TMP0:%.*]] = getelementptr i8, ptr [[S]], i32 9
+// LINUX-NEXT:    store i8 0, ptr [[TMP0]], align 1
+// LINUX-NEXT:    [[TMP1:%.*]] = getelementptr i8, ptr [[S]], i32 10
+// LINUX-NEXT:    store i8 0, ptr [[TMP1]], align 2
+// LINUX-NEXT:    [[TMP2:%.*]] = getelementptr i8, ptr [[S]], i32 11
+// LINUX-NEXT:    store i8 0, ptr [[TMP2]], align 1
+// LINUX-NEXT:    [[TMP3:%.*]] = getelementptr i8, ptr [[S]], i32 12
+// LINUX-NEXT:    [[TMP4:%.*]] = load i8, ptr [[TMP3]], align 4
+// LINUX-NEXT:    [[TMP5:%.*]] = and i8 [[TMP4]], -8
+// LINUX-NEXT:    store i8 [[TMP5]], ptr [[TMP3]], align 4
+// LINUX-NEXT:    [[TMP6:%.*]] = getelementptr i8, ptr [[S]], i32 14
+// LINUX-NEXT:    [[TMP7:%.*]] = load i8, ptr [[TMP6]], align 2
+// LINUX-NEXT:    [[TMP8:%.*]] = and i8 [[TMP7]], 31
+// LINUX-NEXT:    store i8 [[TMP8]], ptr [[TMP6]], align 2
+// LINUX-NEXT:    [[TMP9:%.*]] = getelementptr i8, ptr [[S]], i32 15
+// LINUX-NEXT:    store i8 0, ptr [[TMP9]], align 1
+// LINUX-NEXT:    ret void
+//
+// WINDOWS-LABEL: define dso_local void @testBitfieldVirtualBase(
+// WINDOWS-SAME: ptr nofreeobj noundef align 8 dead_on_return dereferenceable(24) [[S:%.*]]) #[[ATTR0]] {
+// WINDOWS-NEXT:  [[ENTRY:.*:]]
+// WINDOWS-NEXT:    [[S_INDIRECT_ADDR:%.*]] = alloca ptr, align 8
+// WINDOWS-NEXT:    store ptr [[S]], ptr [[S_INDIRECT_ADDR]], align 8
+// WINDOWS-NEXT:    [[TMP0:%.*]] = getelementptr i8, ptr [[S]], i32 9
+// WINDOWS-NEXT:    store i8 0, ptr [[TMP0]], align 1
+// WINDOWS-NEXT:    [[TMP1:%.*]] = getelementptr i8, ptr [[S]], i32 10
+// WINDOWS-NEXT:    store i8 0, ptr [[TMP1]], align 2
+// WINDOWS-NEXT:    [[TMP2:%.*]] = getelementptr i8, ptr [[S]], i32 11
+// WINDOWS-NEXT:    store i8 0, ptr [[TMP2]], align 1
+// WINDOWS-NEXT:    [[TMP3:%.*]] = getelementptr i8, ptr [[S]], i32 12
+// WINDOWS-NEXT:    store i8 0, ptr [[TMP3]], align 4
+// WINDOWS-NEXT:    [[TMP4:%.*]] = getelementptr i8, ptr [[S]], i32 13
+// WINDOWS-NEXT:    store i8 0, ptr [[TMP4]], align 1
+// WINDOWS-NEXT:    [[TMP5:%.*]] = getelementptr i8, ptr [[S]], i32 14
+// WINDOWS-NEXT:    store i8 0, ptr [[TMP5]], align 2
+// WINDOWS-NEXT:    [[TMP6:%.*]] = getelementptr i8, ptr [[S]], i32 15
+// WINDOWS-NEXT:    store i8 0, ptr [[TMP6]], align 1
+// WINDOWS-NEXT:    [[TMP7:%.*]] = getelementptr i8, ptr [[S]], i32 16
+// WINDOWS-NEXT:    [[TMP8:%.*]] = load i8, ptr [[TMP7]], align 8
+// WINDOWS-NEXT:    [[TMP9:%.*]] = and i8 [[TMP8]], -8
+// WINDOWS-NEXT:    store i8 [[TMP9]], ptr [[TMP7]], align 8
+// WINDOWS-NEXT:    [[TMP10:%.*]] = getelementptr i8, ptr [[S]], i32 18
+// WINDOWS-NEXT:    [[TMP11:%.*]] = load i8, ptr [[TMP10]], align 2
+// WINDOWS-NEXT:    [[TMP12:%.*]] = and i8 [[TMP11]], 31
+// WINDOWS-NEXT:    store i8 [[TMP12]], ptr [[TMP10]], align 2
+// WINDOWS-NEXT:    [[TMP13:%.*]] = getelementptr i8, ptr [[S]], i32 19
+// WINDOWS-NEXT:    store i8 0, ptr [[TMP13]], align 1
+// WINDOWS-NEXT:    [[TMP14:%.*]] = getelementptr i8, ptr [[S]], i32 20
+// WINDOWS-NEXT:    store i8 0, ptr [[TMP14]], align 4
+// WINDOWS-NEXT:    [[TMP15:%.*]] = getelementptr i8, ptr [[S]], i32 21
+// WINDOWS-NEXT:    store i8 0, ptr [[TMP15]], align 1
+// WINDOWS-NEXT:    [[TMP16:%.*]] = getelementptr i8, ptr [[S]], i32 22
+// WINDOWS-NEXT:    store i8 0, ptr [[TMP16]], align 2
+// WINDOWS-NEXT:    [[TMP17:%.*]] = getelementptr i8, ptr [[S]], i32 23
+// WINDOWS-NEXT:    store i8 0, ptr [[TMP17]], align 1
+// WINDOWS-NEXT:    ret void
+//
+// ARM-LE-LABEL: define dso_local void @testBitfieldVirtualBase(
+// ARM-LE-SAME: ptr nofreeobj noundef align 4 dead_on_return dereferenceable(12) [[S:%.*]]) #[[ATTR0]] {
+// ARM-LE-NEXT:  [[ENTRY:.*:]]
+// ARM-LE-NEXT:    [[S_INDIRECT_ADDR:%.*]] = alloca ptr, align 4
+// ARM-LE-NEXT:    store ptr [[S]], ptr [[S_INDIRECT_ADDR]], align 4
+// ARM-LE-NEXT:    [[TMP0:%.*]] = getelementptr i8, ptr [[S]], i32 5
+// ARM-LE-NEXT:    store i8 0, ptr [[TMP0]], align 1
+// ARM-LE-NEXT:    [[TMP1:%.*]] = getelementptr i8, ptr [[S]], i32 6
+// ARM-LE-NEXT:    store i8 0, ptr [[TMP1]], align 2
+// ARM-LE-NEXT:    [[TMP2:%.*]] = getelementptr i8, ptr [[S]], i32 7
+// ARM-LE-NEXT:    store i8 0, ptr [[TMP2]], align 1
+// ARM-LE-NEXT:    [[TMP3:%.*]] = getelementptr i8, ptr [[S]], i32 8
+// ARM-LE-NEXT:    [[TMP4:%.*]] = load i8, ptr [[TMP3]], align 4
+// ARM-LE-NEXT:    [[TMP5:%.*]] = and i8 [[TMP4]], -8
+// ARM-LE-NEXT:    store i8 [[TMP5]], ptr [[TMP3]], align 4
+// ARM-LE-NEXT:    [[TMP6:%.*]] = getelementptr i8, ptr [[S]], i32 10
+// ARM-LE-NEXT:    [[TMP7:%.*]] = load i8, ptr [[TMP6]], align 2
+// ARM-LE-NEXT:    [[TMP8:%.*]] = and i8 [[TMP7]], 31
+// ARM-LE-NEXT:    store i8 [[TMP8]], ptr [[TMP6]], align 2
+// ARM-LE-NEXT:    [[TMP9:%.*]] = getelementptr i8, ptr [[S]], i32 11
+// ARM-LE-NEXT:    store i8 0, ptr [[TMP9]], align 1
+// ARM-LE-NEXT:    ret void
+//
+// ARM-BE-LABEL: define dso_local void @testBitfieldVirtualBase(
+// ARM-BE-SAME: ptr nofreeobj noundef align 4 dead_on_return dereferenceable(12) [[S:%.*]]) #[[ATTR0]] {
+// ARM-BE-NEXT:  [[ENTRY:.*:]]
+// ARM-BE-NEXT:    [[S_INDIRECT_ADDR:%.*]] = alloca ptr, align 4
+// ARM-BE-NEXT:    store ptr [[S]], ptr [[S_INDIRECT_ADDR]], align 4
+// ARM-BE-NEXT:    [[TMP0:%.*]] = getelementptr i8, ptr [[S]], i32 5
+// ARM-BE-NEXT:    store i8 0, ptr [[TMP0]], align 1
+// ARM-BE-NEXT:    [[TMP1:%.*]] = getelementptr i8, ptr [[S]], i32 6
+// ARM-BE-NEXT:    store i8 0, ptr [[TMP1]], align 2
+// ARM-BE-NEXT:    [[TMP2:%.*]] = getelementptr i8, ptr [[S]], i32 7
+// ARM-BE-NEXT:    store i8 0, ptr [[TMP2]], align 1
+// ARM-BE-NEXT:    [[TMP3:%.*]] = getelementptr i8, ptr [[S]], i32 8
+// ARM-BE-NEXT:    [[TMP4:%.*]] = load i8, ptr [[TMP3]], align 4
+// ARM-BE-NEXT:    [[TMP5:%.*]] = and i8 [[TMP4]], 31
+// ARM-BE-NEXT:    store i8 [[TMP5]], ptr [[TMP3]], align 4
+// ARM-BE-NEXT:    [[TMP6:%.*]] = getelementptr i8, ptr [[S]], i32 10
+// ARM-BE-NEXT:    [[TMP7:%.*]] = load i8, ptr [[TMP6]], align 2
+// ARM-BE-NEXT:    [[TMP8:%.*]] = and i8 [[TMP7]], -8
+// ARM-BE-NEXT:    store i8 [[TMP8]], ptr [[TMP6]], align 2
+// ARM-BE-NEXT:    [[TMP9:%.*]] = getelementptr i8, ptr [[S]], i32 11
+// ARM-BE-NEXT:    store i8 0, ptr [[TMP9]], align 1
+// ARM-BE-NEXT:    ret void
+//
+// AARCH64-LE-LABEL: define dso_local void @testBitfieldVirtualBase(
+// AARCH64-LE-SAME: ptr nofreeobj noundef align 8 dead_on_return dereferenceable(16) [[S:%.*]]) #[[ATTR0]] {
+// AARCH64-LE-NEXT:  [[ENTRY:.*:]]
+// AARCH64-LE-NEXT:    [[S_INDIRECT_ADDR:%.*]] = alloca ptr, align 8
+// AARCH64-LE-NEXT:    store ptr [[S]], ptr [[S_INDIRECT_ADDR]], align 8
+// AARCH64-LE-NEXT:    [[TMP0:%.*]] = getelementptr i8, ptr [[S]], i32 9
+// AARCH64-LE-NEXT:    store i8 0, ptr [[TMP0]], align 1
+// AARCH64-LE-NEXT:    [[TMP1:%.*]] = getelementptr i8, ptr [[S]], i32 10
+// AARCH64-LE-NEXT:    store i8 0, ptr [[TMP1]], align 2
+// AARCH64-LE-NEXT:    [[TMP2:%.*]] = getelementptr i8, ptr [[S]], i32 11
+// AARCH64-LE-NEXT:    store i8 0, ptr [[TMP2]], align 1
+// AARCH64-LE-NEXT:    [[TMP3:%.*]] = getelementptr i8, ptr [[S]], i32 12
+// AARCH64-LE-NEXT:    [[TMP4:%.*]] = load i8, ptr [[TMP3]], align 4
+// AARCH64-LE-NEXT:    [[TMP5:%.*]] = and i8 [[TMP4]], -8
+// AARCH64-LE-NEXT:    store i8 [[TMP5]], ptr [[TMP3]], align 4
+// AARCH64-LE-NEXT:    [[TMP6:%.*]] = getelementptr i8, ptr [[S]], i32 14
+// AARCH64-LE-NEXT:    [[TMP7:%.*]] = load i8, ptr [[TMP6]], align 2
+// AARCH64-LE-NEXT:    [[TMP8:%.*]] = and i8 [[TMP7]], 31
+// AARCH64-LE-NEXT:    store i8 [[TMP8]], ptr [[TMP6]], align 2
+// AARCH64-LE-NEXT:    [[TMP9:%.*]] = getelementptr i8, ptr [[S]], i32 15
+// AARCH64-LE-NEXT:    store i8 0, ptr [[TMP9]], align 1
+// AARCH64-LE-NEXT:    ret void
+//
+// AARCH64-BE-LABEL: define dso_local void @testBitfieldVirtualBase(
+// AARCH64-BE-SAME: ptr nofreeobj noundef align 8 dead_on_return dereferenceable(16) [[S:%.*]]) #[[ATTR0]] {
+// AARCH64-BE-NEXT:  [[ENTRY:.*:]]
+// AARCH64-BE-NEXT:    [[S_INDIRECT_ADDR:%.*]] = alloca ptr, align 8
+// AARCH64-BE-NEXT:    store ptr [[S]], ptr [[S_INDIRECT_ADDR]], align 8
+// AARCH64-BE-NEXT:    [[TMP0:%.*]] = getelementptr i8, ptr [[S]], i32 9
+// AARCH64-BE-NEXT:    store i8 0, ptr [[TMP0]], align 1
+// AARCH64-BE-NEXT:    [[TMP1:%.*]] = getelementptr i8, ptr [[S]], i32 10
+// AARCH64-BE-NEXT:    store i8 0, ptr [[TMP1]], align 2
+// AARCH64-BE-NEXT:    [[TMP2:%.*]] = getelementptr i8, ptr [[S]], i32 11
+// AARCH64-BE-NEXT:    store i8 0, ptr [[TMP2]], align 1
+// AARCH64-BE-NEXT:    [[TMP3:%.*]] = getelementptr i8, ptr [[S]], i32 12
+// AARCH64-BE-NEXT:    [[TMP4:%.*]] = load i8, ptr [[TMP3]], align 4
+// AARCH64-BE-NEXT:    [[TMP5:%.*]] = and i8 [[TMP4]], 31
+// AARCH64-BE-NEXT:    store i8 [[TMP5]], ptr [[TMP3]], align 4
+// AARCH64-BE-NEXT:    [[TMP6:%.*]] = getelementptr i8, ptr [[S]], i32 14
+// AARCH64-BE-NEXT:    [[TMP7:%.*]] = load i8, ptr [[TMP6]], align 2
+// AARCH64-BE-NEXT:    [[TMP8:%.*]] = and i8 [[TMP7]], -8
+// AARCH64-BE-NEXT:    store i8 [[TMP8]], ptr [[TMP6]], align 2
+// AARCH64-BE-NEXT:    [[TMP9:%.*]] = getelementptr i8, ptr [[S]], i32 15
+// AARCH64-BE-NEXT:    store i8 0, ptr [[TMP9]], align 1
+// AARCH64-BE-NEXT:    ret void
+//
+extern "C" void testBitfieldVirtualBase(BitfieldVirtualDerived s) {
+  __builtin_clear_padding(&s);
+}

>From e73aff989a57af697a5cc74b57d877ecde9087b2 Mon Sep 17 00:00:00 2001
From: Victor Campos <victor.campos at arm.com>
Date: Wed, 16 Sep 2026 11:42:37 +0100
Subject: [PATCH 2/2] Add citation to the Itanium C++ ABI

---
 clang/lib/AST/ASTContext.cpp | 7 ++++---
 1 file changed, 4 insertions(+), 3 deletions(-)

diff --git a/clang/lib/AST/ASTContext.cpp b/clang/lib/AST/ASTContext.cpp
index 61936e44f6278..d89a40856ca1e 100644
--- a/clang/lib/AST/ASTContext.cpp
+++ b/clang/lib/AST/ASTContext.cpp
@@ -15902,9 +15902,10 @@ struct PaddingCalculator {
     //
     // FIXME: For over-sized bitfields in BE, Clang allocates padding bits
     // before the occupied bits. This violates the ABI rules, which say that
-    // padding should be allocated after, regardless of endianness. The current
-    // code accommodates for Clang's current behaviour though, and bumps Start
-    // forward to skip the leading padding bits.
+    // padding should be allocated after, regardless of endianness (Itanium C++
+    // ABI ยง2.4, II.1(b)). The current code accommodates for Clang's current
+    // behaviour though, and bumps Start forward to skip the leading padding
+    // bits.
     const uint64_t Start =
         StartBitOffset + DeclaredSizeInBits - OccupiedSizeInBits;
     const uint64_t End = Start + OccupiedSizeInBits;



More information about the cfe-commits mailing list