diff --git a/llvm/lib/Transforms/Instrumentation/HWAddressSanitizer.cpp b/llvm/lib/Transforms/Instrumentation/HWAddressSanitizer.cpp
--- a/llvm/lib/Transforms/Instrumentation/HWAddressSanitizer.cpp
+++ b/llvm/lib/Transforms/Instrumentation/HWAddressSanitizer.cpp
@@ -222,6 +222,10 @@
                                        cl::desc("inline all checks"),
                                        cl::Hidden, cl::init(false));
 
+static cl::opt<bool> ClInlineFastPathChecks("hwasan-inline-fast-path-checks",
+                                            cl::desc("inline all checks"),
+                                            cl::Hidden, cl::init(false));
+
 // Enabled from clang by "-fsanitize-hwaddress-experimental-aliasing".
 static cl::opt<bool> ClUsePageAliases("hwasan-experimental-use-page-aliases",
                                       cl::desc("Use page aliasing in HWASan"),
@@ -372,6 +376,7 @@
   bool CompileKernel;
   bool Recover;
   bool OutlinedChecks;
+  bool InlineFastPath;
   bool UseShortGranules;
   bool InstrumentLandingPads;
   bool InstrumentWithCalls;
@@ -579,6 +584,10 @@
       TargetTriple.isOSBinFormatELF() &&
       (ClInlineAllChecks.getNumOccurrences() ? !ClInlineAllChecks : !Recover);
 
+  InlineFastPath =
+      (ClInlineFastPathChecks.getNumOccurrences() ? ClInlineFastPathChecks
+                                                  : !TargetTriple.isAndroid());
+
   if (ClMatchAllTag.getNumOccurrences()) {
     if (ClMatchAllTag != -1) {
       MatchAllTag = ClMatchAllTag & 0xFF;
@@ -845,6 +854,11 @@
   assert(!UsePageAliases);
   const int64_t AccessInfo = getAccessInfo(IsWrite, AccessSizeIndex);
   IRBuilder<> IRB(InsertBefore);
+
+  if (InlineFastPath) {
+    // TODO.
+  }
+
   Module *M = IRB.GetInsertBlock()->getParent()->getParent();
   Ptr = IRB.CreateBitCast(Ptr, Int8PtrTy);
   IRB.CreateCall(Intrinsic::getDeclaration(
diff --git a/llvm/test/Instrumentation/HWAddressSanitizer/RISCV/basic.ll b/llvm/test/Instrumentation/HWAddressSanitizer/RISCV/basic.ll
--- a/llvm/test/Instrumentation/HWAddressSanitizer/RISCV/basic.ll
+++ b/llvm/test/Instrumentation/HWAddressSanitizer/RISCV/basic.ll
@@ -1,6 +1,9 @@
 ; NOTE: Assertions have been autogenerated by utils/update_test_checks.py UTC_ARGS: --version 2
 ; Test basic address sanitizer instrumentation.
 ;
+; RUN: opt < %s -passes=hwasan -S | FileCheck %s
+; RUN: opt < %s -passes=hwasan -hwasan-inline-fast-path-checks=0 -S | FileCheck %s --check-prefixes=NOFASTPATH
+; RUN: opt < %s -passes=hwasan -hwasan-inline-fast-path-checks=1 -S | FileCheck %s --check-prefixes=FASTPATH
 ; RUN: opt < %s -passes=hwasan -hwasan-recover=0 -hwasan-with-ifunc=1 -hwasan-with-tls=0 -S | FileCheck %s --check-prefixes=ABORT-DYNAMIC-SHADOW
 ; RUN: opt < %s -passes=hwasan -hwasan-recover=1 -hwasan-with-ifunc=1 -hwasan-with-tls=0 -S | FileCheck %s --check-prefixes=RECOVER-DYNAMIC-SHADOW
 ; RUN: opt < %s -passes=hwasan -hwasan-recover=0 -hwasan-mapping-offset=0 -S | FileCheck %s --check-prefixes=ABORT-ZERO-BASED-SHADOW
@@ -13,6 +16,42 @@
 target triple = "riscv64-unknown-linux"
 
 define i8 @test_load8(ptr %a) sanitize_hwaddress {
+; CHECK-LABEL: define i8 @test_load8
+; CHECK-SAME: (ptr [[A:%.*]]) #[[ATTR0:[0-9]+]] {
+; CHECK-NEXT:  entry:
+; CHECK-NEXT:    [[TMP0:%.*]] = load i64, ptr @__hwasan_tls, align 8
+; CHECK-NEXT:    [[TMP1:%.*]] = and i64 [[TMP0]], 72057594037927935
+; CHECK-NEXT:    [[TMP2:%.*]] = or i64 [[TMP1]], 4294967295
+; CHECK-NEXT:    [[HWASAN_SHADOW:%.*]] = add i64 [[TMP2]], 1
+; CHECK-NEXT:    [[TMP3:%.*]] = inttoptr i64 [[HWASAN_SHADOW]] to ptr
+; CHECK-NEXT:    call void @llvm.hwasan.check.memaccess.shortgranules(ptr [[TMP3]], ptr [[A]], i32 0)
+; CHECK-NEXT:    [[B:%.*]] = load i8, ptr [[A]], align 4
+; CHECK-NEXT:    ret i8 [[B]]
+;
+; NOFASTPATH-LABEL: define i8 @test_load8
+; NOFASTPATH-SAME: (ptr [[A:%.*]]) #[[ATTR0:[0-9]+]] {
+; NOFASTPATH-NEXT:  entry:
+; NOFASTPATH-NEXT:    [[TMP0:%.*]] = load i64, ptr @__hwasan_tls, align 8
+; NOFASTPATH-NEXT:    [[TMP1:%.*]] = and i64 [[TMP0]], 72057594037927935
+; NOFASTPATH-NEXT:    [[TMP2:%.*]] = or i64 [[TMP1]], 4294967295
+; NOFASTPATH-NEXT:    [[HWASAN_SHADOW:%.*]] = add i64 [[TMP2]], 1
+; NOFASTPATH-NEXT:    [[TMP3:%.*]] = inttoptr i64 [[HWASAN_SHADOW]] to ptr
+; NOFASTPATH-NEXT:    call void @llvm.hwasan.check.memaccess.shortgranules(ptr [[TMP3]], ptr [[A]], i32 0)
+; NOFASTPATH-NEXT:    [[B:%.*]] = load i8, ptr [[A]], align 4
+; NOFASTPATH-NEXT:    ret i8 [[B]]
+;
+; FASTPATH-LABEL: define i8 @test_load8
+; FASTPATH-SAME: (ptr [[A:%.*]]) #[[ATTR0:[0-9]+]] {
+; FASTPATH-NEXT:  entry:
+; FASTPATH-NEXT:    [[TMP0:%.*]] = load i64, ptr @__hwasan_tls, align 8
+; FASTPATH-NEXT:    [[TMP1:%.*]] = and i64 [[TMP0]], 72057594037927935
+; FASTPATH-NEXT:    [[TMP2:%.*]] = or i64 [[TMP1]], 4294967295
+; FASTPATH-NEXT:    [[HWASAN_SHADOW:%.*]] = add i64 [[TMP2]], 1
+; FASTPATH-NEXT:    [[TMP3:%.*]] = inttoptr i64 [[HWASAN_SHADOW]] to ptr
+; FASTPATH-NEXT:    call void @llvm.hwasan.check.memaccess.shortgranules(ptr [[TMP3]], ptr [[A]], i32 0)
+; FASTPATH-NEXT:    [[B:%.*]] = load i8, ptr [[A]], align 4
+; FASTPATH-NEXT:    ret i8 [[B]]
+;
 ; ABORT-DYNAMIC-SHADOW-LABEL: define i8 @test_load8
 ; ABORT-DYNAMIC-SHADOW-SAME: (ptr [[A:%.*]]) #[[ATTR0:[0-9]+]] {
 ; ABORT-DYNAMIC-SHADOW-NEXT:  entry:
@@ -109,6 +148,42 @@
 }
 
 define i16 @test_load16(ptr %a) sanitize_hwaddress {
+; CHECK-LABEL: define i16 @test_load16
+; CHECK-SAME: (ptr [[A:%.*]]) #[[ATTR0]] {
+; CHECK-NEXT:  entry:
+; CHECK-NEXT:    [[TMP0:%.*]] = load i64, ptr @__hwasan_tls, align 8
+; CHECK-NEXT:    [[TMP1:%.*]] = and i64 [[TMP0]], 72057594037927935
+; CHECK-NEXT:    [[TMP2:%.*]] = or i64 [[TMP1]], 4294967295
+; CHECK-NEXT:    [[HWASAN_SHADOW:%.*]] = add i64 [[TMP2]], 1
+; CHECK-NEXT:    [[TMP3:%.*]] = inttoptr i64 [[HWASAN_SHADOW]] to ptr
+; CHECK-NEXT:    call void @llvm.hwasan.check.memaccess.shortgranules(ptr [[TMP3]], ptr [[A]], i32 1)
+; CHECK-NEXT:    [[B:%.*]] = load i16, ptr [[A]], align 4
+; CHECK-NEXT:    ret i16 [[B]]
+;
+; NOFASTPATH-LABEL: define i16 @test_load16
+; NOFASTPATH-SAME: (ptr [[A:%.*]]) #[[ATTR0]] {
+; NOFASTPATH-NEXT:  entry:
+; NOFASTPATH-NEXT:    [[TMP0:%.*]] = load i64, ptr @__hwasan_tls, align 8
+; NOFASTPATH-NEXT:    [[TMP1:%.*]] = and i64 [[TMP0]], 72057594037927935
+; NOFASTPATH-NEXT:    [[TMP2:%.*]] = or i64 [[TMP1]], 4294967295
+; NOFASTPATH-NEXT:    [[HWASAN_SHADOW:%.*]] = add i64 [[TMP2]], 1
+; NOFASTPATH-NEXT:    [[TMP3:%.*]] = inttoptr i64 [[HWASAN_SHADOW]] to ptr
+; NOFASTPATH-NEXT:    call void @llvm.hwasan.check.memaccess.shortgranules(ptr [[TMP3]], ptr [[A]], i32 1)
+; NOFASTPATH-NEXT:    [[B:%.*]] = load i16, ptr [[A]], align 4
+; NOFASTPATH-NEXT:    ret i16 [[B]]
+;
+; FASTPATH-LABEL: define i16 @test_load16
+; FASTPATH-SAME: (ptr [[A:%.*]]) #[[ATTR0]] {
+; FASTPATH-NEXT:  entry:
+; FASTPATH-NEXT:    [[TMP0:%.*]] = load i64, ptr @__hwasan_tls, align 8
+; FASTPATH-NEXT:    [[TMP1:%.*]] = and i64 [[TMP0]], 72057594037927935
+; FASTPATH-NEXT:    [[TMP2:%.*]] = or i64 [[TMP1]], 4294967295
+; FASTPATH-NEXT:    [[HWASAN_SHADOW:%.*]] = add i64 [[TMP2]], 1
+; FASTPATH-NEXT:    [[TMP3:%.*]] = inttoptr i64 [[HWASAN_SHADOW]] to ptr
+; FASTPATH-NEXT:    call void @llvm.hwasan.check.memaccess.shortgranules(ptr [[TMP3]], ptr [[A]], i32 1)
+; FASTPATH-NEXT:    [[B:%.*]] = load i16, ptr [[A]], align 4
+; FASTPATH-NEXT:    ret i16 [[B]]
+;
 ; ABORT-DYNAMIC-SHADOW-LABEL: define i16 @test_load16
 ; ABORT-DYNAMIC-SHADOW-SAME: (ptr [[A:%.*]]) #[[ATTR0]] {
 ; ABORT-DYNAMIC-SHADOW-NEXT:  entry:
@@ -205,6 +280,42 @@
 }
 
 define i32 @test_load32(ptr %a) sanitize_hwaddress {
+; CHECK-LABEL: define i32 @test_load32
+; CHECK-SAME: (ptr [[A:%.*]]) #[[ATTR0]] {
+; CHECK-NEXT:  entry:
+; CHECK-NEXT:    [[TMP0:%.*]] = load i64, ptr @__hwasan_tls, align 8
+; CHECK-NEXT:    [[TMP1:%.*]] = and i64 [[TMP0]], 72057594037927935
+; CHECK-NEXT:    [[TMP2:%.*]] = or i64 [[TMP1]], 4294967295
+; CHECK-NEXT:    [[HWASAN_SHADOW:%.*]] = add i64 [[TMP2]], 1
+; CHECK-NEXT:    [[TMP3:%.*]] = inttoptr i64 [[HWASAN_SHADOW]] to ptr
+; CHECK-NEXT:    call void @llvm.hwasan.check.memaccess.shortgranules(ptr [[TMP3]], ptr [[A]], i32 2)
+; CHECK-NEXT:    [[B:%.*]] = load i32, ptr [[A]], align 4
+; CHECK-NEXT:    ret i32 [[B]]
+;
+; NOFASTPATH-LABEL: define i32 @test_load32
+; NOFASTPATH-SAME: (ptr [[A:%.*]]) #[[ATTR0]] {
+; NOFASTPATH-NEXT:  entry:
+; NOFASTPATH-NEXT:    [[TMP0:%.*]] = load i64, ptr @__hwasan_tls, align 8
+; NOFASTPATH-NEXT:    [[TMP1:%.*]] = and i64 [[TMP0]], 72057594037927935
+; NOFASTPATH-NEXT:    [[TMP2:%.*]] = or i64 [[TMP1]], 4294967295
+; NOFASTPATH-NEXT:    [[HWASAN_SHADOW:%.*]] = add i64 [[TMP2]], 1
+; NOFASTPATH-NEXT:    [[TMP3:%.*]] = inttoptr i64 [[HWASAN_SHADOW]] to ptr
+; NOFASTPATH-NEXT:    call void @llvm.hwasan.check.memaccess.shortgranules(ptr [[TMP3]], ptr [[A]], i32 2)
+; NOFASTPATH-NEXT:    [[B:%.*]] = load i32, ptr [[A]], align 4
+; NOFASTPATH-NEXT:    ret i32 [[B]]
+;
+; FASTPATH-LABEL: define i32 @test_load32
+; FASTPATH-SAME: (ptr [[A:%.*]]) #[[ATTR0]] {
+; FASTPATH-NEXT:  entry:
+; FASTPATH-NEXT:    [[TMP0:%.*]] = load i64, ptr @__hwasan_tls, align 8
+; FASTPATH-NEXT:    [[TMP1:%.*]] = and i64 [[TMP0]], 72057594037927935
+; FASTPATH-NEXT:    [[TMP2:%.*]] = or i64 [[TMP1]], 4294967295
+; FASTPATH-NEXT:    [[HWASAN_SHADOW:%.*]] = add i64 [[TMP2]], 1
+; FASTPATH-NEXT:    [[TMP3:%.*]] = inttoptr i64 [[HWASAN_SHADOW]] to ptr
+; FASTPATH-NEXT:    call void @llvm.hwasan.check.memaccess.shortgranules(ptr [[TMP3]], ptr [[A]], i32 2)
+; FASTPATH-NEXT:    [[B:%.*]] = load i32, ptr [[A]], align 4
+; FASTPATH-NEXT:    ret i32 [[B]]
+;
 ; ABORT-DYNAMIC-SHADOW-LABEL: define i32 @test_load32
 ; ABORT-DYNAMIC-SHADOW-SAME: (ptr [[A:%.*]]) #[[ATTR0]] {
 ; ABORT-DYNAMIC-SHADOW-NEXT:  entry:
@@ -301,6 +412,42 @@
 }
 
 define i64 @test_load64(ptr %a) sanitize_hwaddress {
+; CHECK-LABEL: define i64 @test_load64
+; CHECK-SAME: (ptr [[A:%.*]]) #[[ATTR0]] {
+; CHECK-NEXT:  entry:
+; CHECK-NEXT:    [[TMP0:%.*]] = load i64, ptr @__hwasan_tls, align 8
+; CHECK-NEXT:    [[TMP1:%.*]] = and i64 [[TMP0]], 72057594037927935
+; CHECK-NEXT:    [[TMP2:%.*]] = or i64 [[TMP1]], 4294967295
+; CHECK-NEXT:    [[HWASAN_SHADOW:%.*]] = add i64 [[TMP2]], 1
+; CHECK-NEXT:    [[TMP3:%.*]] = inttoptr i64 [[HWASAN_SHADOW]] to ptr
+; CHECK-NEXT:    call void @llvm.hwasan.check.memaccess.shortgranules(ptr [[TMP3]], ptr [[A]], i32 3)
+; CHECK-NEXT:    [[B:%.*]] = load i64, ptr [[A]], align 8
+; CHECK-NEXT:    ret i64 [[B]]
+;
+; NOFASTPATH-LABEL: define i64 @test_load64
+; NOFASTPATH-SAME: (ptr [[A:%.*]]) #[[ATTR0]] {
+; NOFASTPATH-NEXT:  entry:
+; NOFASTPATH-NEXT:    [[TMP0:%.*]] = load i64, ptr @__hwasan_tls, align 8
+; NOFASTPATH-NEXT:    [[TMP1:%.*]] = and i64 [[TMP0]], 72057594037927935
+; NOFASTPATH-NEXT:    [[TMP2:%.*]] = or i64 [[TMP1]], 4294967295
+; NOFASTPATH-NEXT:    [[HWASAN_SHADOW:%.*]] = add i64 [[TMP2]], 1
+; NOFASTPATH-NEXT:    [[TMP3:%.*]] = inttoptr i64 [[HWASAN_SHADOW]] to ptr
+; NOFASTPATH-NEXT:    call void @llvm.hwasan.check.memaccess.shortgranules(ptr [[TMP3]], ptr [[A]], i32 3)
+; NOFASTPATH-NEXT:    [[B:%.*]] = load i64, ptr [[A]], align 8
+; NOFASTPATH-NEXT:    ret i64 [[B]]
+;
+; FASTPATH-LABEL: define i64 @test_load64
+; FASTPATH-SAME: (ptr [[A:%.*]]) #[[ATTR0]] {
+; FASTPATH-NEXT:  entry:
+; FASTPATH-NEXT:    [[TMP0:%.*]] = load i64, ptr @__hwasan_tls, align 8
+; FASTPATH-NEXT:    [[TMP1:%.*]] = and i64 [[TMP0]], 72057594037927935
+; FASTPATH-NEXT:    [[TMP2:%.*]] = or i64 [[TMP1]], 4294967295
+; FASTPATH-NEXT:    [[HWASAN_SHADOW:%.*]] = add i64 [[TMP2]], 1
+; FASTPATH-NEXT:    [[TMP3:%.*]] = inttoptr i64 [[HWASAN_SHADOW]] to ptr
+; FASTPATH-NEXT:    call void @llvm.hwasan.check.memaccess.shortgranules(ptr [[TMP3]], ptr [[A]], i32 3)
+; FASTPATH-NEXT:    [[B:%.*]] = load i64, ptr [[A]], align 8
+; FASTPATH-NEXT:    ret i64 [[B]]
+;
 ; ABORT-DYNAMIC-SHADOW-LABEL: define i64 @test_load64
 ; ABORT-DYNAMIC-SHADOW-SAME: (ptr [[A:%.*]]) #[[ATTR0]] {
 ; ABORT-DYNAMIC-SHADOW-NEXT:  entry:
@@ -397,6 +544,42 @@
 }
 
 define i128 @test_load128(ptr %a) sanitize_hwaddress {
+; CHECK-LABEL: define i128 @test_load128
+; CHECK-SAME: (ptr [[A:%.*]]) #[[ATTR0]] {
+; CHECK-NEXT:  entry:
+; CHECK-NEXT:    [[TMP0:%.*]] = load i64, ptr @__hwasan_tls, align 8
+; CHECK-NEXT:    [[TMP1:%.*]] = and i64 [[TMP0]], 72057594037927935
+; CHECK-NEXT:    [[TMP2:%.*]] = or i64 [[TMP1]], 4294967295
+; CHECK-NEXT:    [[HWASAN_SHADOW:%.*]] = add i64 [[TMP2]], 1
+; CHECK-NEXT:    [[TMP3:%.*]] = inttoptr i64 [[HWASAN_SHADOW]] to ptr
+; CHECK-NEXT:    call void @llvm.hwasan.check.memaccess.shortgranules(ptr [[TMP3]], ptr [[A]], i32 4)
+; CHECK-NEXT:    [[B:%.*]] = load i128, ptr [[A]], align 16
+; CHECK-NEXT:    ret i128 [[B]]
+;
+; NOFASTPATH-LABEL: define i128 @test_load128
+; NOFASTPATH-SAME: (ptr [[A:%.*]]) #[[ATTR0]] {
+; NOFASTPATH-NEXT:  entry:
+; NOFASTPATH-NEXT:    [[TMP0:%.*]] = load i64, ptr @__hwasan_tls, align 8
+; NOFASTPATH-NEXT:    [[TMP1:%.*]] = and i64 [[TMP0]], 72057594037927935
+; NOFASTPATH-NEXT:    [[TMP2:%.*]] = or i64 [[TMP1]], 4294967295
+; NOFASTPATH-NEXT:    [[HWASAN_SHADOW:%.*]] = add i64 [[TMP2]], 1
+; NOFASTPATH-NEXT:    [[TMP3:%.*]] = inttoptr i64 [[HWASAN_SHADOW]] to ptr
+; NOFASTPATH-NEXT:    call void @llvm.hwasan.check.memaccess.shortgranules(ptr [[TMP3]], ptr [[A]], i32 4)
+; NOFASTPATH-NEXT:    [[B:%.*]] = load i128, ptr [[A]], align 16
+; NOFASTPATH-NEXT:    ret i128 [[B]]
+;
+; FASTPATH-LABEL: define i128 @test_load128
+; FASTPATH-SAME: (ptr [[A:%.*]]) #[[ATTR0]] {
+; FASTPATH-NEXT:  entry:
+; FASTPATH-NEXT:    [[TMP0:%.*]] = load i64, ptr @__hwasan_tls, align 8
+; FASTPATH-NEXT:    [[TMP1:%.*]] = and i64 [[TMP0]], 72057594037927935
+; FASTPATH-NEXT:    [[TMP2:%.*]] = or i64 [[TMP1]], 4294967295
+; FASTPATH-NEXT:    [[HWASAN_SHADOW:%.*]] = add i64 [[TMP2]], 1
+; FASTPATH-NEXT:    [[TMP3:%.*]] = inttoptr i64 [[HWASAN_SHADOW]] to ptr
+; FASTPATH-NEXT:    call void @llvm.hwasan.check.memaccess.shortgranules(ptr [[TMP3]], ptr [[A]], i32 4)
+; FASTPATH-NEXT:    [[B:%.*]] = load i128, ptr [[A]], align 16
+; FASTPATH-NEXT:    ret i128 [[B]]
+;
 ; ABORT-DYNAMIC-SHADOW-LABEL: define i128 @test_load128
 ; ABORT-DYNAMIC-SHADOW-SAME: (ptr [[A:%.*]]) #[[ATTR0]] {
 ; ABORT-DYNAMIC-SHADOW-NEXT:  entry:
@@ -493,6 +676,45 @@
 }
 
 define i40 @test_load40(ptr %a) sanitize_hwaddress {
+; CHECK-LABEL: define i40 @test_load40
+; CHECK-SAME: (ptr [[A:%.*]]) #[[ATTR0]] {
+; CHECK-NEXT:  entry:
+; CHECK-NEXT:    [[TMP0:%.*]] = load i64, ptr @__hwasan_tls, align 8
+; CHECK-NEXT:    [[TMP1:%.*]] = and i64 [[TMP0]], 72057594037927935
+; CHECK-NEXT:    [[TMP2:%.*]] = or i64 [[TMP1]], 4294967295
+; CHECK-NEXT:    [[HWASAN_SHADOW:%.*]] = add i64 [[TMP2]], 1
+; CHECK-NEXT:    [[TMP3:%.*]] = inttoptr i64 [[HWASAN_SHADOW]] to ptr
+; CHECK-NEXT:    [[TMP4:%.*]] = ptrtoint ptr [[A]] to i64
+; CHECK-NEXT:    call void @__hwasan_loadN(i64 [[TMP4]], i64 5)
+; CHECK-NEXT:    [[B:%.*]] = load i40, ptr [[A]], align 4
+; CHECK-NEXT:    ret i40 [[B]]
+;
+; NOFASTPATH-LABEL: define i40 @test_load40
+; NOFASTPATH-SAME: (ptr [[A:%.*]]) #[[ATTR0]] {
+; NOFASTPATH-NEXT:  entry:
+; NOFASTPATH-NEXT:    [[TMP0:%.*]] = load i64, ptr @__hwasan_tls, align 8
+; NOFASTPATH-NEXT:    [[TMP1:%.*]] = and i64 [[TMP0]], 72057594037927935
+; NOFASTPATH-NEXT:    [[TMP2:%.*]] = or i64 [[TMP1]], 4294967295
+; NOFASTPATH-NEXT:    [[HWASAN_SHADOW:%.*]] = add i64 [[TMP2]], 1
+; NOFASTPATH-NEXT:    [[TMP3:%.*]] = inttoptr i64 [[HWASAN_SHADOW]] to ptr
+; NOFASTPATH-NEXT:    [[TMP4:%.*]] = ptrtoint ptr [[A]] to i64
+; NOFASTPATH-NEXT:    call void @__hwasan_loadN(i64 [[TMP4]], i64 5)
+; NOFASTPATH-NEXT:    [[B:%.*]] = load i40, ptr [[A]], align 4
+; NOFASTPATH-NEXT:    ret i40 [[B]]
+;
+; FASTPATH-LABEL: define i40 @test_load40
+; FASTPATH-SAME: (ptr [[A:%.*]]) #[[ATTR0]] {
+; FASTPATH-NEXT:  entry:
+; FASTPATH-NEXT:    [[TMP0:%.*]] = load i64, ptr @__hwasan_tls, align 8
+; FASTPATH-NEXT:    [[TMP1:%.*]] = and i64 [[TMP0]], 72057594037927935
+; FASTPATH-NEXT:    [[TMP2:%.*]] = or i64 [[TMP1]], 4294967295
+; FASTPATH-NEXT:    [[HWASAN_SHADOW:%.*]] = add i64 [[TMP2]], 1
+; FASTPATH-NEXT:    [[TMP3:%.*]] = inttoptr i64 [[HWASAN_SHADOW]] to ptr
+; FASTPATH-NEXT:    [[TMP4:%.*]] = ptrtoint ptr [[A]] to i64
+; FASTPATH-NEXT:    call void @__hwasan_loadN(i64 [[TMP4]], i64 5)
+; FASTPATH-NEXT:    [[B:%.*]] = load i40, ptr [[A]], align 4
+; FASTPATH-NEXT:    ret i40 [[B]]
+;
 ; ABORT-DYNAMIC-SHADOW-LABEL: define i40 @test_load40
 ; ABORT-DYNAMIC-SHADOW-SAME: (ptr [[A:%.*]]) #[[ATTR0]] {
 ; ABORT-DYNAMIC-SHADOW-NEXT:  entry:
@@ -535,6 +757,42 @@
 }
 
 define void @test_store8(ptr %a, i8 %b) sanitize_hwaddress {
+; CHECK-LABEL: define void @test_store8
+; CHECK-SAME: (ptr [[A:%.*]], i8 [[B:%.*]]) #[[ATTR0]] {
+; CHECK-NEXT:  entry:
+; CHECK-NEXT:    [[TMP0:%.*]] = load i64, ptr @__hwasan_tls, align 8
+; CHECK-NEXT:    [[TMP1:%.*]] = and i64 [[TMP0]], 72057594037927935
+; CHECK-NEXT:    [[TMP2:%.*]] = or i64 [[TMP1]], 4294967295
+; CHECK-NEXT:    [[HWASAN_SHADOW:%.*]] = add i64 [[TMP2]], 1
+; CHECK-NEXT:    [[TMP3:%.*]] = inttoptr i64 [[HWASAN_SHADOW]] to ptr
+; CHECK-NEXT:    call void @llvm.hwasan.check.memaccess.shortgranules(ptr [[TMP3]], ptr [[A]], i32 16)
+; CHECK-NEXT:    store i8 [[B]], ptr [[A]], align 4
+; CHECK-NEXT:    ret void
+;
+; NOFASTPATH-LABEL: define void @test_store8
+; NOFASTPATH-SAME: (ptr [[A:%.*]], i8 [[B:%.*]]) #[[ATTR0]] {
+; NOFASTPATH-NEXT:  entry:
+; NOFASTPATH-NEXT:    [[TMP0:%.*]] = load i64, ptr @__hwasan_tls, align 8
+; NOFASTPATH-NEXT:    [[TMP1:%.*]] = and i64 [[TMP0]], 72057594037927935
+; NOFASTPATH-NEXT:    [[TMP2:%.*]] = or i64 [[TMP1]], 4294967295
+; NOFASTPATH-NEXT:    [[HWASAN_SHADOW:%.*]] = add i64 [[TMP2]], 1
+; NOFASTPATH-NEXT:    [[TMP3:%.*]] = inttoptr i64 [[HWASAN_SHADOW]] to ptr
+; NOFASTPATH-NEXT:    call void @llvm.hwasan.check.memaccess.shortgranules(ptr [[TMP3]], ptr [[A]], i32 16)
+; NOFASTPATH-NEXT:    store i8 [[B]], ptr [[A]], align 4
+; NOFASTPATH-NEXT:    ret void
+;
+; FASTPATH-LABEL: define void @test_store8
+; FASTPATH-SAME: (ptr [[A:%.*]], i8 [[B:%.*]]) #[[ATTR0]] {
+; FASTPATH-NEXT:  entry:
+; FASTPATH-NEXT:    [[TMP0:%.*]] = load i64, ptr @__hwasan_tls, align 8
+; FASTPATH-NEXT:    [[TMP1:%.*]] = and i64 [[TMP0]], 72057594037927935
+; FASTPATH-NEXT:    [[TMP2:%.*]] = or i64 [[TMP1]], 4294967295
+; FASTPATH-NEXT:    [[HWASAN_SHADOW:%.*]] = add i64 [[TMP2]], 1
+; FASTPATH-NEXT:    [[TMP3:%.*]] = inttoptr i64 [[HWASAN_SHADOW]] to ptr
+; FASTPATH-NEXT:    call void @llvm.hwasan.check.memaccess.shortgranules(ptr [[TMP3]], ptr [[A]], i32 16)
+; FASTPATH-NEXT:    store i8 [[B]], ptr [[A]], align 4
+; FASTPATH-NEXT:    ret void
+;
 ; ABORT-DYNAMIC-SHADOW-LABEL: define void @test_store8
 ; ABORT-DYNAMIC-SHADOW-SAME: (ptr [[A:%.*]], i8 [[B:%.*]]) #[[ATTR0]] {
 ; ABORT-DYNAMIC-SHADOW-NEXT:  entry:
@@ -631,6 +889,42 @@
 }
 
 define void @test_store16(ptr %a, i16 %b) sanitize_hwaddress {
+; CHECK-LABEL: define void @test_store16
+; CHECK-SAME: (ptr [[A:%.*]], i16 [[B:%.*]]) #[[ATTR0]] {
+; CHECK-NEXT:  entry:
+; CHECK-NEXT:    [[TMP0:%.*]] = load i64, ptr @__hwasan_tls, align 8
+; CHECK-NEXT:    [[TMP1:%.*]] = and i64 [[TMP0]], 72057594037927935
+; CHECK-NEXT:    [[TMP2:%.*]] = or i64 [[TMP1]], 4294967295
+; CHECK-NEXT:    [[HWASAN_SHADOW:%.*]] = add i64 [[TMP2]], 1
+; CHECK-NEXT:    [[TMP3:%.*]] = inttoptr i64 [[HWASAN_SHADOW]] to ptr
+; CHECK-NEXT:    call void @llvm.hwasan.check.memaccess.shortgranules(ptr [[TMP3]], ptr [[A]], i32 17)
+; CHECK-NEXT:    store i16 [[B]], ptr [[A]], align 4
+; CHECK-NEXT:    ret void
+;
+; NOFASTPATH-LABEL: define void @test_store16
+; NOFASTPATH-SAME: (ptr [[A:%.*]], i16 [[B:%.*]]) #[[ATTR0]] {
+; NOFASTPATH-NEXT:  entry:
+; NOFASTPATH-NEXT:    [[TMP0:%.*]] = load i64, ptr @__hwasan_tls, align 8
+; NOFASTPATH-NEXT:    [[TMP1:%.*]] = and i64 [[TMP0]], 72057594037927935
+; NOFASTPATH-NEXT:    [[TMP2:%.*]] = or i64 [[TMP1]], 4294967295
+; NOFASTPATH-NEXT:    [[HWASAN_SHADOW:%.*]] = add i64 [[TMP2]], 1
+; NOFASTPATH-NEXT:    [[TMP3:%.*]] = inttoptr i64 [[HWASAN_SHADOW]] to ptr
+; NOFASTPATH-NEXT:    call void @llvm.hwasan.check.memaccess.shortgranules(ptr [[TMP3]], ptr [[A]], i32 17)
+; NOFASTPATH-NEXT:    store i16 [[B]], ptr [[A]], align 4
+; NOFASTPATH-NEXT:    ret void
+;
+; FASTPATH-LABEL: define void @test_store16
+; FASTPATH-SAME: (ptr [[A:%.*]], i16 [[B:%.*]]) #[[ATTR0]] {
+; FASTPATH-NEXT:  entry:
+; FASTPATH-NEXT:    [[TMP0:%.*]] = load i64, ptr @__hwasan_tls, align 8
+; FASTPATH-NEXT:    [[TMP1:%.*]] = and i64 [[TMP0]], 72057594037927935
+; FASTPATH-NEXT:    [[TMP2:%.*]] = or i64 [[TMP1]], 4294967295
+; FASTPATH-NEXT:    [[HWASAN_SHADOW:%.*]] = add i64 [[TMP2]], 1
+; FASTPATH-NEXT:    [[TMP3:%.*]] = inttoptr i64 [[HWASAN_SHADOW]] to ptr
+; FASTPATH-NEXT:    call void @llvm.hwasan.check.memaccess.shortgranules(ptr [[TMP3]], ptr [[A]], i32 17)
+; FASTPATH-NEXT:    store i16 [[B]], ptr [[A]], align 4
+; FASTPATH-NEXT:    ret void
+;
 ; ABORT-DYNAMIC-SHADOW-LABEL: define void @test_store16
 ; ABORT-DYNAMIC-SHADOW-SAME: (ptr [[A:%.*]], i16 [[B:%.*]]) #[[ATTR0]] {
 ; ABORT-DYNAMIC-SHADOW-NEXT:  entry:
@@ -727,6 +1021,42 @@
 }
 
 define void @test_store32(ptr %a, i32 %b) sanitize_hwaddress {
+; CHECK-LABEL: define void @test_store32
+; CHECK-SAME: (ptr [[A:%.*]], i32 [[B:%.*]]) #[[ATTR0]] {
+; CHECK-NEXT:  entry:
+; CHECK-NEXT:    [[TMP0:%.*]] = load i64, ptr @__hwasan_tls, align 8
+; CHECK-NEXT:    [[TMP1:%.*]] = and i64 [[TMP0]], 72057594037927935
+; CHECK-NEXT:    [[TMP2:%.*]] = or i64 [[TMP1]], 4294967295
+; CHECK-NEXT:    [[HWASAN_SHADOW:%.*]] = add i64 [[TMP2]], 1
+; CHECK-NEXT:    [[TMP3:%.*]] = inttoptr i64 [[HWASAN_SHADOW]] to ptr
+; CHECK-NEXT:    call void @llvm.hwasan.check.memaccess.shortgranules(ptr [[TMP3]], ptr [[A]], i32 18)
+; CHECK-NEXT:    store i32 [[B]], ptr [[A]], align 4
+; CHECK-NEXT:    ret void
+;
+; NOFASTPATH-LABEL: define void @test_store32
+; NOFASTPATH-SAME: (ptr [[A:%.*]], i32 [[B:%.*]]) #[[ATTR0]] {
+; NOFASTPATH-NEXT:  entry:
+; NOFASTPATH-NEXT:    [[TMP0:%.*]] = load i64, ptr @__hwasan_tls, align 8
+; NOFASTPATH-NEXT:    [[TMP1:%.*]] = and i64 [[TMP0]], 72057594037927935
+; NOFASTPATH-NEXT:    [[TMP2:%.*]] = or i64 [[TMP1]], 4294967295
+; NOFASTPATH-NEXT:    [[HWASAN_SHADOW:%.*]] = add i64 [[TMP2]], 1
+; NOFASTPATH-NEXT:    [[TMP3:%.*]] = inttoptr i64 [[HWASAN_SHADOW]] to ptr
+; NOFASTPATH-NEXT:    call void @llvm.hwasan.check.memaccess.shortgranules(ptr [[TMP3]], ptr [[A]], i32 18)
+; NOFASTPATH-NEXT:    store i32 [[B]], ptr [[A]], align 4
+; NOFASTPATH-NEXT:    ret void
+;
+; FASTPATH-LABEL: define void @test_store32
+; FASTPATH-SAME: (ptr [[A:%.*]], i32 [[B:%.*]]) #[[ATTR0]] {
+; FASTPATH-NEXT:  entry:
+; FASTPATH-NEXT:    [[TMP0:%.*]] = load i64, ptr @__hwasan_tls, align 8
+; FASTPATH-NEXT:    [[TMP1:%.*]] = and i64 [[TMP0]], 72057594037927935
+; FASTPATH-NEXT:    [[TMP2:%.*]] = or i64 [[TMP1]], 4294967295
+; FASTPATH-NEXT:    [[HWASAN_SHADOW:%.*]] = add i64 [[TMP2]], 1
+; FASTPATH-NEXT:    [[TMP3:%.*]] = inttoptr i64 [[HWASAN_SHADOW]] to ptr
+; FASTPATH-NEXT:    call void @llvm.hwasan.check.memaccess.shortgranules(ptr [[TMP3]], ptr [[A]], i32 18)
+; FASTPATH-NEXT:    store i32 [[B]], ptr [[A]], align 4
+; FASTPATH-NEXT:    ret void
+;
 ; ABORT-DYNAMIC-SHADOW-LABEL: define void @test_store32
 ; ABORT-DYNAMIC-SHADOW-SAME: (ptr [[A:%.*]], i32 [[B:%.*]]) #[[ATTR0]] {
 ; ABORT-DYNAMIC-SHADOW-NEXT:  entry:
@@ -823,6 +1153,42 @@
 }
 
 define void @test_store64(ptr %a, i64 %b) sanitize_hwaddress {
+; CHECK-LABEL: define void @test_store64
+; CHECK-SAME: (ptr [[A:%.*]], i64 [[B:%.*]]) #[[ATTR0]] {
+; CHECK-NEXT:  entry:
+; CHECK-NEXT:    [[TMP0:%.*]] = load i64, ptr @__hwasan_tls, align 8
+; CHECK-NEXT:    [[TMP1:%.*]] = and i64 [[TMP0]], 72057594037927935
+; CHECK-NEXT:    [[TMP2:%.*]] = or i64 [[TMP1]], 4294967295
+; CHECK-NEXT:    [[HWASAN_SHADOW:%.*]] = add i64 [[TMP2]], 1
+; CHECK-NEXT:    [[TMP3:%.*]] = inttoptr i64 [[HWASAN_SHADOW]] to ptr
+; CHECK-NEXT:    call void @llvm.hwasan.check.memaccess.shortgranules(ptr [[TMP3]], ptr [[A]], i32 19)
+; CHECK-NEXT:    store i64 [[B]], ptr [[A]], align 8
+; CHECK-NEXT:    ret void
+;
+; NOFASTPATH-LABEL: define void @test_store64
+; NOFASTPATH-SAME: (ptr [[A:%.*]], i64 [[B:%.*]]) #[[ATTR0]] {
+; NOFASTPATH-NEXT:  entry:
+; NOFASTPATH-NEXT:    [[TMP0:%.*]] = load i64, ptr @__hwasan_tls, align 8
+; NOFASTPATH-NEXT:    [[TMP1:%.*]] = and i64 [[TMP0]], 72057594037927935
+; NOFASTPATH-NEXT:    [[TMP2:%.*]] = or i64 [[TMP1]], 4294967295
+; NOFASTPATH-NEXT:    [[HWASAN_SHADOW:%.*]] = add i64 [[TMP2]], 1
+; NOFASTPATH-NEXT:    [[TMP3:%.*]] = inttoptr i64 [[HWASAN_SHADOW]] to ptr
+; NOFASTPATH-NEXT:    call void @llvm.hwasan.check.memaccess.shortgranules(ptr [[TMP3]], ptr [[A]], i32 19)
+; NOFASTPATH-NEXT:    store i64 [[B]], ptr [[A]], align 8
+; NOFASTPATH-NEXT:    ret void
+;
+; FASTPATH-LABEL: define void @test_store64
+; FASTPATH-SAME: (ptr [[A:%.*]], i64 [[B:%.*]]) #[[ATTR0]] {
+; FASTPATH-NEXT:  entry:
+; FASTPATH-NEXT:    [[TMP0:%.*]] = load i64, ptr @__hwasan_tls, align 8
+; FASTPATH-NEXT:    [[TMP1:%.*]] = and i64 [[TMP0]], 72057594037927935
+; FASTPATH-NEXT:    [[TMP2:%.*]] = or i64 [[TMP1]], 4294967295
+; FASTPATH-NEXT:    [[HWASAN_SHADOW:%.*]] = add i64 [[TMP2]], 1
+; FASTPATH-NEXT:    [[TMP3:%.*]] = inttoptr i64 [[HWASAN_SHADOW]] to ptr
+; FASTPATH-NEXT:    call void @llvm.hwasan.check.memaccess.shortgranules(ptr [[TMP3]], ptr [[A]], i32 19)
+; FASTPATH-NEXT:    store i64 [[B]], ptr [[A]], align 8
+; FASTPATH-NEXT:    ret void
+;
 ; ABORT-DYNAMIC-SHADOW-LABEL: define void @test_store64
 ; ABORT-DYNAMIC-SHADOW-SAME: (ptr [[A:%.*]], i64 [[B:%.*]]) #[[ATTR0]] {
 ; ABORT-DYNAMIC-SHADOW-NEXT:  entry:
@@ -919,6 +1285,42 @@
 }
 
 define void @test_store128(ptr %a, i128 %b) sanitize_hwaddress {
+; CHECK-LABEL: define void @test_store128
+; CHECK-SAME: (ptr [[A:%.*]], i128 [[B:%.*]]) #[[ATTR0]] {
+; CHECK-NEXT:  entry:
+; CHECK-NEXT:    [[TMP0:%.*]] = load i64, ptr @__hwasan_tls, align 8
+; CHECK-NEXT:    [[TMP1:%.*]] = and i64 [[TMP0]], 72057594037927935
+; CHECK-NEXT:    [[TMP2:%.*]] = or i64 [[TMP1]], 4294967295
+; CHECK-NEXT:    [[HWASAN_SHADOW:%.*]] = add i64 [[TMP2]], 1
+; CHECK-NEXT:    [[TMP3:%.*]] = inttoptr i64 [[HWASAN_SHADOW]] to ptr
+; CHECK-NEXT:    call void @llvm.hwasan.check.memaccess.shortgranules(ptr [[TMP3]], ptr [[A]], i32 20)
+; CHECK-NEXT:    store i128 [[B]], ptr [[A]], align 16
+; CHECK-NEXT:    ret void
+;
+; NOFASTPATH-LABEL: define void @test_store128
+; NOFASTPATH-SAME: (ptr [[A:%.*]], i128 [[B:%.*]]) #[[ATTR0]] {
+; NOFASTPATH-NEXT:  entry:
+; NOFASTPATH-NEXT:    [[TMP0:%.*]] = load i64, ptr @__hwasan_tls, align 8
+; NOFASTPATH-NEXT:    [[TMP1:%.*]] = and i64 [[TMP0]], 72057594037927935
+; NOFASTPATH-NEXT:    [[TMP2:%.*]] = or i64 [[TMP1]], 4294967295
+; NOFASTPATH-NEXT:    [[HWASAN_SHADOW:%.*]] = add i64 [[TMP2]], 1
+; NOFASTPATH-NEXT:    [[TMP3:%.*]] = inttoptr i64 [[HWASAN_SHADOW]] to ptr
+; NOFASTPATH-NEXT:    call void @llvm.hwasan.check.memaccess.shortgranules(ptr [[TMP3]], ptr [[A]], i32 20)
+; NOFASTPATH-NEXT:    store i128 [[B]], ptr [[A]], align 16
+; NOFASTPATH-NEXT:    ret void
+;
+; FASTPATH-LABEL: define void @test_store128
+; FASTPATH-SAME: (ptr [[A:%.*]], i128 [[B:%.*]]) #[[ATTR0]] {
+; FASTPATH-NEXT:  entry:
+; FASTPATH-NEXT:    [[TMP0:%.*]] = load i64, ptr @__hwasan_tls, align 8
+; FASTPATH-NEXT:    [[TMP1:%.*]] = and i64 [[TMP0]], 72057594037927935
+; FASTPATH-NEXT:    [[TMP2:%.*]] = or i64 [[TMP1]], 4294967295
+; FASTPATH-NEXT:    [[HWASAN_SHADOW:%.*]] = add i64 [[TMP2]], 1
+; FASTPATH-NEXT:    [[TMP3:%.*]] = inttoptr i64 [[HWASAN_SHADOW]] to ptr
+; FASTPATH-NEXT:    call void @llvm.hwasan.check.memaccess.shortgranules(ptr [[TMP3]], ptr [[A]], i32 20)
+; FASTPATH-NEXT:    store i128 [[B]], ptr [[A]], align 16
+; FASTPATH-NEXT:    ret void
+;
 ; ABORT-DYNAMIC-SHADOW-LABEL: define void @test_store128
 ; ABORT-DYNAMIC-SHADOW-SAME: (ptr [[A:%.*]], i128 [[B:%.*]]) #[[ATTR0]] {
 ; ABORT-DYNAMIC-SHADOW-NEXT:  entry:
@@ -1015,6 +1417,45 @@
 }
 
 define void @test_store40(ptr %a, i40 %b) sanitize_hwaddress {
+; CHECK-LABEL: define void @test_store40
+; CHECK-SAME: (ptr [[A:%.*]], i40 [[B:%.*]]) #[[ATTR0]] {
+; CHECK-NEXT:  entry:
+; CHECK-NEXT:    [[TMP0:%.*]] = load i64, ptr @__hwasan_tls, align 8
+; CHECK-NEXT:    [[TMP1:%.*]] = and i64 [[TMP0]], 72057594037927935
+; CHECK-NEXT:    [[TMP2:%.*]] = or i64 [[TMP1]], 4294967295
+; CHECK-NEXT:    [[HWASAN_SHADOW:%.*]] = add i64 [[TMP2]], 1
+; CHECK-NEXT:    [[TMP3:%.*]] = inttoptr i64 [[HWASAN_SHADOW]] to ptr
+; CHECK-NEXT:    [[TMP4:%.*]] = ptrtoint ptr [[A]] to i64
+; CHECK-NEXT:    call void @__hwasan_storeN(i64 [[TMP4]], i64 5)
+; CHECK-NEXT:    store i40 [[B]], ptr [[A]], align 4
+; CHECK-NEXT:    ret void
+;
+; NOFASTPATH-LABEL: define void @test_store40
+; NOFASTPATH-SAME: (ptr [[A:%.*]], i40 [[B:%.*]]) #[[ATTR0]] {
+; NOFASTPATH-NEXT:  entry:
+; NOFASTPATH-NEXT:    [[TMP0:%.*]] = load i64, ptr @__hwasan_tls, align 8
+; NOFASTPATH-NEXT:    [[TMP1:%.*]] = and i64 [[TMP0]], 72057594037927935
+; NOFASTPATH-NEXT:    [[TMP2:%.*]] = or i64 [[TMP1]], 4294967295
+; NOFASTPATH-NEXT:    [[HWASAN_SHADOW:%.*]] = add i64 [[TMP2]], 1
+; NOFASTPATH-NEXT:    [[TMP3:%.*]] = inttoptr i64 [[HWASAN_SHADOW]] to ptr
+; NOFASTPATH-NEXT:    [[TMP4:%.*]] = ptrtoint ptr [[A]] to i64
+; NOFASTPATH-NEXT:    call void @__hwasan_storeN(i64 [[TMP4]], i64 5)
+; NOFASTPATH-NEXT:    store i40 [[B]], ptr [[A]], align 4
+; NOFASTPATH-NEXT:    ret void
+;
+; FASTPATH-LABEL: define void @test_store40
+; FASTPATH-SAME: (ptr [[A:%.*]], i40 [[B:%.*]]) #[[ATTR0]] {
+; FASTPATH-NEXT:  entry:
+; FASTPATH-NEXT:    [[TMP0:%.*]] = load i64, ptr @__hwasan_tls, align 8
+; FASTPATH-NEXT:    [[TMP1:%.*]] = and i64 [[TMP0]], 72057594037927935
+; FASTPATH-NEXT:    [[TMP2:%.*]] = or i64 [[TMP1]], 4294967295
+; FASTPATH-NEXT:    [[HWASAN_SHADOW:%.*]] = add i64 [[TMP2]], 1
+; FASTPATH-NEXT:    [[TMP3:%.*]] = inttoptr i64 [[HWASAN_SHADOW]] to ptr
+; FASTPATH-NEXT:    [[TMP4:%.*]] = ptrtoint ptr [[A]] to i64
+; FASTPATH-NEXT:    call void @__hwasan_storeN(i64 [[TMP4]], i64 5)
+; FASTPATH-NEXT:    store i40 [[B]], ptr [[A]], align 4
+; FASTPATH-NEXT:    ret void
+;
 ; ABORT-DYNAMIC-SHADOW-LABEL: define void @test_store40
 ; ABORT-DYNAMIC-SHADOW-SAME: (ptr [[A:%.*]], i40 [[B:%.*]]) #[[ATTR0]] {
 ; ABORT-DYNAMIC-SHADOW-NEXT:  entry:
@@ -1057,6 +1498,45 @@
 }
 
 define void @test_store_unaligned(ptr %a, i64 %b) sanitize_hwaddress {
+; CHECK-LABEL: define void @test_store_unaligned
+; CHECK-SAME: (ptr [[A:%.*]], i64 [[B:%.*]]) #[[ATTR0]] {
+; CHECK-NEXT:  entry:
+; CHECK-NEXT:    [[TMP0:%.*]] = load i64, ptr @__hwasan_tls, align 8
+; CHECK-NEXT:    [[TMP1:%.*]] = and i64 [[TMP0]], 72057594037927935
+; CHECK-NEXT:    [[TMP2:%.*]] = or i64 [[TMP1]], 4294967295
+; CHECK-NEXT:    [[HWASAN_SHADOW:%.*]] = add i64 [[TMP2]], 1
+; CHECK-NEXT:    [[TMP3:%.*]] = inttoptr i64 [[HWASAN_SHADOW]] to ptr
+; CHECK-NEXT:    [[TMP4:%.*]] = ptrtoint ptr [[A]] to i64
+; CHECK-NEXT:    call void @__hwasan_storeN(i64 [[TMP4]], i64 8)
+; CHECK-NEXT:    store i64 [[B]], ptr [[A]], align 4
+; CHECK-NEXT:    ret void
+;
+; NOFASTPATH-LABEL: define void @test_store_unaligned
+; NOFASTPATH-SAME: (ptr [[A:%.*]], i64 [[B:%.*]]) #[[ATTR0]] {
+; NOFASTPATH-NEXT:  entry:
+; NOFASTPATH-NEXT:    [[TMP0:%.*]] = load i64, ptr @__hwasan_tls, align 8
+; NOFASTPATH-NEXT:    [[TMP1:%.*]] = and i64 [[TMP0]], 72057594037927935
+; NOFASTPATH-NEXT:    [[TMP2:%.*]] = or i64 [[TMP1]], 4294967295
+; NOFASTPATH-NEXT:    [[HWASAN_SHADOW:%.*]] = add i64 [[TMP2]], 1
+; NOFASTPATH-NEXT:    [[TMP3:%.*]] = inttoptr i64 [[HWASAN_SHADOW]] to ptr
+; NOFASTPATH-NEXT:    [[TMP4:%.*]] = ptrtoint ptr [[A]] to i64
+; NOFASTPATH-NEXT:    call void @__hwasan_storeN(i64 [[TMP4]], i64 8)
+; NOFASTPATH-NEXT:    store i64 [[B]], ptr [[A]], align 4
+; NOFASTPATH-NEXT:    ret void
+;
+; FASTPATH-LABEL: define void @test_store_unaligned
+; FASTPATH-SAME: (ptr [[A:%.*]], i64 [[B:%.*]]) #[[ATTR0]] {
+; FASTPATH-NEXT:  entry:
+; FASTPATH-NEXT:    [[TMP0:%.*]] = load i64, ptr @__hwasan_tls, align 8
+; FASTPATH-NEXT:    [[TMP1:%.*]] = and i64 [[TMP0]], 72057594037927935
+; FASTPATH-NEXT:    [[TMP2:%.*]] = or i64 [[TMP1]], 4294967295
+; FASTPATH-NEXT:    [[HWASAN_SHADOW:%.*]] = add i64 [[TMP2]], 1
+; FASTPATH-NEXT:    [[TMP3:%.*]] = inttoptr i64 [[HWASAN_SHADOW]] to ptr
+; FASTPATH-NEXT:    [[TMP4:%.*]] = ptrtoint ptr [[A]] to i64
+; FASTPATH-NEXT:    call void @__hwasan_storeN(i64 [[TMP4]], i64 8)
+; FASTPATH-NEXT:    store i64 [[B]], ptr [[A]], align 4
+; FASTPATH-NEXT:    ret void
+;
 ; ABORT-DYNAMIC-SHADOW-LABEL: define void @test_store_unaligned
 ; ABORT-DYNAMIC-SHADOW-SAME: (ptr [[A:%.*]], i64 [[B:%.*]]) #[[ATTR0]] {
 ; ABORT-DYNAMIC-SHADOW-NEXT:  entry:
@@ -1105,6 +1585,18 @@
 ; CHECK-NEXT:    [[B:%.*]] = load i8, ptr [[A]], align 4
 ; CHECK-NEXT:    ret i8 [[B]]
 ;
+; NOFASTPATH-LABEL: define i8 @test_load_noattr
+; NOFASTPATH-SAME: (ptr [[A:%.*]]) {
+; NOFASTPATH-NEXT:  entry:
+; NOFASTPATH-NEXT:    [[B:%.*]] = load i8, ptr [[A]], align 4
+; NOFASTPATH-NEXT:    ret i8 [[B]]
+;
+; FASTPATH-LABEL: define i8 @test_load_noattr
+; FASTPATH-SAME: (ptr [[A:%.*]]) {
+; FASTPATH-NEXT:  entry:
+; FASTPATH-NEXT:    [[B:%.*]] = load i8, ptr [[A]], align 4
+; FASTPATH-NEXT:    ret i8 [[B]]
+;
 ; ABORT-DYNAMIC-SHADOW-LABEL: define i8 @test_load_noattr
 ; ABORT-DYNAMIC-SHADOW-SAME: (ptr [[A:%.*]]) {
 ; ABORT-DYNAMIC-SHADOW-NEXT:  entry:
@@ -1141,6 +1633,18 @@
 ; CHECK-NEXT:    [[B:%.*]] = load i8, ptr [[A]], align 4
 ; CHECK-NEXT:    ret i8 [[B]]
 ;
+; NOFASTPATH-LABEL: define i8 @test_load_notmyattr
+; NOFASTPATH-SAME: (ptr [[A:%.*]]) #[[ATTR1:[0-9]+]] {
+; NOFASTPATH-NEXT:  entry:
+; NOFASTPATH-NEXT:    [[B:%.*]] = load i8, ptr [[A]], align 4
+; NOFASTPATH-NEXT:    ret i8 [[B]]
+;
+; FASTPATH-LABEL: define i8 @test_load_notmyattr
+; FASTPATH-SAME: (ptr [[A:%.*]]) #[[ATTR1:[0-9]+]] {
+; FASTPATH-NEXT:  entry:
+; FASTPATH-NEXT:    [[B:%.*]] = load i8, ptr [[A]], align 4
+; FASTPATH-NEXT:    ret i8 [[B]]
+;
 ; ABORT-DYNAMIC-SHADOW-LABEL: define i8 @test_load_notmyattr
 ; ABORT-DYNAMIC-SHADOW-SAME: (ptr [[A:%.*]]) #[[ATTR1:[0-9]+]] {
 ; ABORT-DYNAMIC-SHADOW-NEXT:  entry:
@@ -1172,11 +1676,23 @@
 
 define i8 @test_load_addrspace(ptr addrspace(256) %a) sanitize_hwaddress {
 ; CHECK-LABEL: define i8 @test_load_addrspace
-; CHECK-SAME: (ptr addrspace(256) [[A:%.*]]) #[[ATTR0:[0-9]+]] {
+; CHECK-SAME: (ptr addrspace(256) [[A:%.*]]) #[[ATTR0]] {
 ; CHECK-NEXT:  entry:
 ; CHECK-NEXT:    [[B:%.*]] = load i8, ptr addrspace(256) [[A]], align 4
 ; CHECK-NEXT:    ret i8 [[B]]
 ;
+; NOFASTPATH-LABEL: define i8 @test_load_addrspace
+; NOFASTPATH-SAME: (ptr addrspace(256) [[A:%.*]]) #[[ATTR0]] {
+; NOFASTPATH-NEXT:  entry:
+; NOFASTPATH-NEXT:    [[B:%.*]] = load i8, ptr addrspace(256) [[A]], align 4
+; NOFASTPATH-NEXT:    ret i8 [[B]]
+;
+; FASTPATH-LABEL: define i8 @test_load_addrspace
+; FASTPATH-SAME: (ptr addrspace(256) [[A:%.*]]) #[[ATTR0]] {
+; FASTPATH-NEXT:  entry:
+; FASTPATH-NEXT:    [[B:%.*]] = load i8, ptr addrspace(256) [[A]], align 4
+; FASTPATH-NEXT:    ret i8 [[B]]
+;
 ; ABORT-DYNAMIC-SHADOW-LABEL: define i8 @test_load_addrspace
 ; ABORT-DYNAMIC-SHADOW-SAME: (ptr addrspace(256) [[A:%.*]]) #[[ATTR0]] {
 ; ABORT-DYNAMIC-SHADOW-NEXT:  entry:
diff --git a/llvm/test/Instrumentation/HWAddressSanitizer/X86/basic.ll b/llvm/test/Instrumentation/HWAddressSanitizer/X86/basic.ll
--- a/llvm/test/Instrumentation/HWAddressSanitizer/X86/basic.ll
+++ b/llvm/test/Instrumentation/HWAddressSanitizer/X86/basic.ll
@@ -3,6 +3,9 @@
 ; Generic code is covered by ../basic.ll, only the x86_64 specific code is
 ; tested here.
 ;
+; RUN: opt < %s -passes=hwasan -S | FileCheck %s
+; RUN: opt < %s -passes=hwasan -hwasan-inline-fast-path-checks=0 -S | FileCheck %s --check-prefixes=NOFASTPATH
+; RUN: opt < %s -passes=hwasan -hwasan-inline-fast-path-checks=1 -S | FileCheck %s --check-prefixes=FASTPATH
 ; RUN: opt < %s -passes=hwasan -hwasan-recover=0 -S | FileCheck %s  --check-prefixes=ABORT
 ; RUN: opt < %s -passes=hwasan -hwasan-recover=1 -S | FileCheck %s  --check-prefixes=RECOVER
 ; RUN: opt < %s -passes=hwasan -hwasan-recover=0 -hwasan-instrument-with-calls=0 -S | FileCheck %s  --check-prefixes=ABORT-INLINE
@@ -12,6 +15,33 @@
 target triple = "x86_64-unknown-linux-gnu"
 
 define i8 @test_load8(ptr %a) sanitize_hwaddress {
+; CHECK-LABEL: define i8 @test_load8
+; CHECK-SAME: (ptr [[A:%.*]]) #[[ATTR0:[0-9]+]] {
+; CHECK-NEXT:  entry:
+; CHECK-NEXT:    [[DOTHWASAN_SHADOW:%.*]] = call ptr asm "", "=r,0"(ptr null)
+; CHECK-NEXT:    [[TMP0:%.*]] = ptrtoint ptr [[A]] to i64
+; CHECK-NEXT:    call void @__hwasan_load1(i64 [[TMP0]])
+; CHECK-NEXT:    [[B:%.*]] = load i8, ptr [[A]], align 4
+; CHECK-NEXT:    ret i8 [[B]]
+;
+; NOFASTPATH-LABEL: define i8 @test_load8
+; NOFASTPATH-SAME: (ptr [[A:%.*]]) #[[ATTR0:[0-9]+]] {
+; NOFASTPATH-NEXT:  entry:
+; NOFASTPATH-NEXT:    [[DOTHWASAN_SHADOW:%.*]] = call ptr asm "", "=r,0"(ptr null)
+; NOFASTPATH-NEXT:    [[TMP0:%.*]] = ptrtoint ptr [[A]] to i64
+; NOFASTPATH-NEXT:    call void @__hwasan_load1(i64 [[TMP0]])
+; NOFASTPATH-NEXT:    [[B:%.*]] = load i8, ptr [[A]], align 4
+; NOFASTPATH-NEXT:    ret i8 [[B]]
+;
+; FASTPATH-LABEL: define i8 @test_load8
+; FASTPATH-SAME: (ptr [[A:%.*]]) #[[ATTR0:[0-9]+]] {
+; FASTPATH-NEXT:  entry:
+; FASTPATH-NEXT:    [[DOTHWASAN_SHADOW:%.*]] = call ptr asm "", "=r,0"(ptr null)
+; FASTPATH-NEXT:    [[TMP0:%.*]] = ptrtoint ptr [[A]] to i64
+; FASTPATH-NEXT:    call void @__hwasan_load1(i64 [[TMP0]])
+; FASTPATH-NEXT:    [[B:%.*]] = load i8, ptr [[A]], align 4
+; FASTPATH-NEXT:    ret i8 [[B]]
+;
 ; ABORT-LABEL: define i8 @test_load8
 ; ABORT-SAME: (ptr [[A:%.*]]) #[[ATTR0:[0-9]+]] {
 ; ABORT-NEXT:  entry:
@@ -118,6 +148,33 @@
 }
 
 define i40 @test_load40(ptr %a) sanitize_hwaddress {
+; CHECK-LABEL: define i40 @test_load40
+; CHECK-SAME: (ptr [[A:%.*]]) #[[ATTR0]] {
+; CHECK-NEXT:  entry:
+; CHECK-NEXT:    [[DOTHWASAN_SHADOW:%.*]] = call ptr asm "", "=r,0"(ptr null)
+; CHECK-NEXT:    [[TMP0:%.*]] = ptrtoint ptr [[A]] to i64
+; CHECK-NEXT:    call void @__hwasan_loadN(i64 [[TMP0]], i64 5)
+; CHECK-NEXT:    [[B:%.*]] = load i40, ptr [[A]], align 4
+; CHECK-NEXT:    ret i40 [[B]]
+;
+; NOFASTPATH-LABEL: define i40 @test_load40
+; NOFASTPATH-SAME: (ptr [[A:%.*]]) #[[ATTR0]] {
+; NOFASTPATH-NEXT:  entry:
+; NOFASTPATH-NEXT:    [[DOTHWASAN_SHADOW:%.*]] = call ptr asm "", "=r,0"(ptr null)
+; NOFASTPATH-NEXT:    [[TMP0:%.*]] = ptrtoint ptr [[A]] to i64
+; NOFASTPATH-NEXT:    call void @__hwasan_loadN(i64 [[TMP0]], i64 5)
+; NOFASTPATH-NEXT:    [[B:%.*]] = load i40, ptr [[A]], align 4
+; NOFASTPATH-NEXT:    ret i40 [[B]]
+;
+; FASTPATH-LABEL: define i40 @test_load40
+; FASTPATH-SAME: (ptr [[A:%.*]]) #[[ATTR0]] {
+; FASTPATH-NEXT:  entry:
+; FASTPATH-NEXT:    [[DOTHWASAN_SHADOW:%.*]] = call ptr asm "", "=r,0"(ptr null)
+; FASTPATH-NEXT:    [[TMP0:%.*]] = ptrtoint ptr [[A]] to i64
+; FASTPATH-NEXT:    call void @__hwasan_loadN(i64 [[TMP0]], i64 5)
+; FASTPATH-NEXT:    [[B:%.*]] = load i40, ptr [[A]], align 4
+; FASTPATH-NEXT:    ret i40 [[B]]
+;
 ; ABORT-LABEL: define i40 @test_load40
 ; ABORT-SAME: (ptr [[A:%.*]]) #[[ATTR0]] {
 ; ABORT-NEXT:  entry:
@@ -168,6 +225,33 @@
 }
 
 define void @test_store8(ptr %a, i8 %b) sanitize_hwaddress {
+; CHECK-LABEL: define void @test_store8
+; CHECK-SAME: (ptr [[A:%.*]], i8 [[B:%.*]]) #[[ATTR0]] {
+; CHECK-NEXT:  entry:
+; CHECK-NEXT:    [[DOTHWASAN_SHADOW:%.*]] = call ptr asm "", "=r,0"(ptr null)
+; CHECK-NEXT:    [[TMP0:%.*]] = ptrtoint ptr [[A]] to i64
+; CHECK-NEXT:    call void @__hwasan_store1(i64 [[TMP0]])
+; CHECK-NEXT:    store i8 [[B]], ptr [[A]], align 4
+; CHECK-NEXT:    ret void
+;
+; NOFASTPATH-LABEL: define void @test_store8
+; NOFASTPATH-SAME: (ptr [[A:%.*]], i8 [[B:%.*]]) #[[ATTR0]] {
+; NOFASTPATH-NEXT:  entry:
+; NOFASTPATH-NEXT:    [[DOTHWASAN_SHADOW:%.*]] = call ptr asm "", "=r,0"(ptr null)
+; NOFASTPATH-NEXT:    [[TMP0:%.*]] = ptrtoint ptr [[A]] to i64
+; NOFASTPATH-NEXT:    call void @__hwasan_store1(i64 [[TMP0]])
+; NOFASTPATH-NEXT:    store i8 [[B]], ptr [[A]], align 4
+; NOFASTPATH-NEXT:    ret void
+;
+; FASTPATH-LABEL: define void @test_store8
+; FASTPATH-SAME: (ptr [[A:%.*]], i8 [[B:%.*]]) #[[ATTR0]] {
+; FASTPATH-NEXT:  entry:
+; FASTPATH-NEXT:    [[DOTHWASAN_SHADOW:%.*]] = call ptr asm "", "=r,0"(ptr null)
+; FASTPATH-NEXT:    [[TMP0:%.*]] = ptrtoint ptr [[A]] to i64
+; FASTPATH-NEXT:    call void @__hwasan_store1(i64 [[TMP0]])
+; FASTPATH-NEXT:    store i8 [[B]], ptr [[A]], align 4
+; FASTPATH-NEXT:    ret void
+;
 ; ABORT-LABEL: define void @test_store8
 ; ABORT-SAME: (ptr [[A:%.*]], i8 [[B:%.*]]) #[[ATTR0]] {
 ; ABORT-NEXT:  entry:
@@ -268,9 +352,39 @@
 ; RECOVER-INLINE-NEXT:    store i8 [[B]], ptr [[A]], align 4
 ; RECOVER-INLINE-NEXT:    ret void
 ;
+entry:
+  store i8 %b, ptr %a, align 4
+  ret void
 }
 
 define void @test_store40(ptr %a, i40 %b) sanitize_hwaddress {
+; CHECK-LABEL: define void @test_store40
+; CHECK-SAME: (ptr [[A:%.*]], i40 [[B:%.*]]) #[[ATTR0]] {
+; CHECK-NEXT:  entry:
+; CHECK-NEXT:    [[DOTHWASAN_SHADOW:%.*]] = call ptr asm "", "=r,0"(ptr null)
+; CHECK-NEXT:    [[TMP0:%.*]] = ptrtoint ptr [[A]] to i64
+; CHECK-NEXT:    call void @__hwasan_storeN(i64 [[TMP0]], i64 5)
+; CHECK-NEXT:    store i40 [[B]], ptr [[A]], align 4
+; CHECK-NEXT:    ret void
+;
+; NOFASTPATH-LABEL: define void @test_store40
+; NOFASTPATH-SAME: (ptr [[A:%.*]], i40 [[B:%.*]]) #[[ATTR0]] {
+; NOFASTPATH-NEXT:  entry:
+; NOFASTPATH-NEXT:    [[DOTHWASAN_SHADOW:%.*]] = call ptr asm "", "=r,0"(ptr null)
+; NOFASTPATH-NEXT:    [[TMP0:%.*]] = ptrtoint ptr [[A]] to i64
+; NOFASTPATH-NEXT:    call void @__hwasan_storeN(i64 [[TMP0]], i64 5)
+; NOFASTPATH-NEXT:    store i40 [[B]], ptr [[A]], align 4
+; NOFASTPATH-NEXT:    ret void
+;
+; FASTPATH-LABEL: define void @test_store40
+; FASTPATH-SAME: (ptr [[A:%.*]], i40 [[B:%.*]]) #[[ATTR0]] {
+; FASTPATH-NEXT:  entry:
+; FASTPATH-NEXT:    [[DOTHWASAN_SHADOW:%.*]] = call ptr asm "", "=r,0"(ptr null)
+; FASTPATH-NEXT:    [[TMP0:%.*]] = ptrtoint ptr [[A]] to i64
+; FASTPATH-NEXT:    call void @__hwasan_storeN(i64 [[TMP0]], i64 5)
+; FASTPATH-NEXT:    store i40 [[B]], ptr [[A]], align 4
+; FASTPATH-NEXT:    ret void
+;
 ; ABORT-LABEL: define void @test_store40
 ; ABORT-SAME: (ptr [[A:%.*]], i40 [[B:%.*]]) #[[ATTR0]] {
 ; ABORT-NEXT:  entry:
@@ -321,6 +435,33 @@
 }
 
 define void @test_store_unaligned(ptr %a, i64 %b) sanitize_hwaddress {
+; CHECK-LABEL: define void @test_store_unaligned
+; CHECK-SAME: (ptr [[A:%.*]], i64 [[B:%.*]]) #[[ATTR0]] {
+; CHECK-NEXT:  entry:
+; CHECK-NEXT:    [[DOTHWASAN_SHADOW:%.*]] = call ptr asm "", "=r,0"(ptr null)
+; CHECK-NEXT:    [[TMP0:%.*]] = ptrtoint ptr [[A]] to i64
+; CHECK-NEXT:    call void @__hwasan_storeN(i64 [[TMP0]], i64 8)
+; CHECK-NEXT:    store i64 [[B]], ptr [[A]], align 4
+; CHECK-NEXT:    ret void
+;
+; NOFASTPATH-LABEL: define void @test_store_unaligned
+; NOFASTPATH-SAME: (ptr [[A:%.*]], i64 [[B:%.*]]) #[[ATTR0]] {
+; NOFASTPATH-NEXT:  entry:
+; NOFASTPATH-NEXT:    [[DOTHWASAN_SHADOW:%.*]] = call ptr asm "", "=r,0"(ptr null)
+; NOFASTPATH-NEXT:    [[TMP0:%.*]] = ptrtoint ptr [[A]] to i64
+; NOFASTPATH-NEXT:    call void @__hwasan_storeN(i64 [[TMP0]], i64 8)
+; NOFASTPATH-NEXT:    store i64 [[B]], ptr [[A]], align 4
+; NOFASTPATH-NEXT:    ret void
+;
+; FASTPATH-LABEL: define void @test_store_unaligned
+; FASTPATH-SAME: (ptr [[A:%.*]], i64 [[B:%.*]]) #[[ATTR0]] {
+; FASTPATH-NEXT:  entry:
+; FASTPATH-NEXT:    [[DOTHWASAN_SHADOW:%.*]] = call ptr asm "", "=r,0"(ptr null)
+; FASTPATH-NEXT:    [[TMP0:%.*]] = ptrtoint ptr [[A]] to i64
+; FASTPATH-NEXT:    call void @__hwasan_storeN(i64 [[TMP0]], i64 8)
+; FASTPATH-NEXT:    store i64 [[B]], ptr [[A]], align 4
+; FASTPATH-NEXT:    ret void
+;
 ; ABORT-LABEL: define void @test_store_unaligned
 ; ABORT-SAME: (ptr [[A:%.*]], i64 [[B:%.*]]) #[[ATTR0]] {
 ; ABORT-NEXT:  entry:
diff --git a/llvm/test/Instrumentation/HWAddressSanitizer/basic.ll b/llvm/test/Instrumentation/HWAddressSanitizer/basic.ll
--- a/llvm/test/Instrumentation/HWAddressSanitizer/basic.ll
+++ b/llvm/test/Instrumentation/HWAddressSanitizer/basic.ll
@@ -1,6 +1,9 @@
 ; NOTE: Assertions have been autogenerated by utils/update_test_checks.py UTC_ARGS: --version 2
 ; Test basic address sanitizer instrumentation.
 ;
+; RUN: opt < %s -passes=hwasan -S | FileCheck %s
+; RUN: opt < %s -passes=hwasan -hwasan-inline-fast-path-checks=0 -S | FileCheck %s --check-prefixes=NOFASTPATH
+; RUN: opt < %s -passes=hwasan -hwasan-inline-fast-path-checks=1 -S | FileCheck %s --check-prefixes=FASTPATH
 ; RUN: opt < %s -passes=hwasan -hwasan-recover=0 -hwasan-with-ifunc=1 -hwasan-with-tls=0 -S | FileCheck %s --check-prefixes=ABORT-DYNAMIC-SHADOW
 ; RUN: opt < %s -passes=hwasan -hwasan-recover=1 -hwasan-with-ifunc=1 -hwasan-with-tls=0 -S | FileCheck %s --check-prefixes=RECOVER-DYNAMIC-SHADOW
 ; RUN: opt < %s -passes=hwasan -hwasan-recover=0 -hwasan-mapping-offset=0 -S | FileCheck %s --check-prefixes=ABORT-ZERO-BASED-SHADOW
@@ -13,6 +16,30 @@
 target triple = "aarch64--linux-android10000"
 
 define i8 @test_load8(ptr %a) sanitize_hwaddress {
+; CHECK-LABEL: define i8 @test_load8
+; CHECK-SAME: (ptr [[A:%.*]]) #[[ATTR0:[0-9]+]] {
+; CHECK-NEXT:  entry:
+; CHECK-NEXT:    [[DOTHWASAN_SHADOW:%.*]] = call ptr asm "", "=r,0"(ptr @__hwasan_shadow)
+; CHECK-NEXT:    call void @llvm.hwasan.check.memaccess.shortgranules(ptr [[DOTHWASAN_SHADOW]], ptr [[A]], i32 0)
+; CHECK-NEXT:    [[B:%.*]] = load i8, ptr [[A]], align 4
+; CHECK-NEXT:    ret i8 [[B]]
+;
+; NOFASTPATH-LABEL: define i8 @test_load8
+; NOFASTPATH-SAME: (ptr [[A:%.*]]) #[[ATTR0:[0-9]+]] {
+; NOFASTPATH-NEXT:  entry:
+; NOFASTPATH-NEXT:    [[DOTHWASAN_SHADOW:%.*]] = call ptr asm "", "=r,0"(ptr @__hwasan_shadow)
+; NOFASTPATH-NEXT:    call void @llvm.hwasan.check.memaccess.shortgranules(ptr [[DOTHWASAN_SHADOW]], ptr [[A]], i32 0)
+; NOFASTPATH-NEXT:    [[B:%.*]] = load i8, ptr [[A]], align 4
+; NOFASTPATH-NEXT:    ret i8 [[B]]
+;
+; FASTPATH-LABEL: define i8 @test_load8
+; FASTPATH-SAME: (ptr [[A:%.*]]) #[[ATTR0:[0-9]+]] {
+; FASTPATH-NEXT:  entry:
+; FASTPATH-NEXT:    [[DOTHWASAN_SHADOW:%.*]] = call ptr asm "", "=r,0"(ptr @__hwasan_shadow)
+; FASTPATH-NEXT:    call void @llvm.hwasan.check.memaccess.shortgranules(ptr [[DOTHWASAN_SHADOW]], ptr [[A]], i32 0)
+; FASTPATH-NEXT:    [[B:%.*]] = load i8, ptr [[A]], align 4
+; FASTPATH-NEXT:    ret i8 [[B]]
+;
 ; ABORT-DYNAMIC-SHADOW-LABEL: define i8 @test_load8
 ; ABORT-DYNAMIC-SHADOW-SAME: (ptr [[A:%.*]]) #[[ATTR0:[0-9]+]] {
 ; ABORT-DYNAMIC-SHADOW-NEXT:  entry:
@@ -109,6 +136,30 @@
 }
 
 define i16 @test_load16(ptr %a) sanitize_hwaddress {
+; CHECK-LABEL: define i16 @test_load16
+; CHECK-SAME: (ptr [[A:%.*]]) #[[ATTR0]] {
+; CHECK-NEXT:  entry:
+; CHECK-NEXT:    [[DOTHWASAN_SHADOW:%.*]] = call ptr asm "", "=r,0"(ptr @__hwasan_shadow)
+; CHECK-NEXT:    call void @llvm.hwasan.check.memaccess.shortgranules(ptr [[DOTHWASAN_SHADOW]], ptr [[A]], i32 1)
+; CHECK-NEXT:    [[B:%.*]] = load i16, ptr [[A]], align 4
+; CHECK-NEXT:    ret i16 [[B]]
+;
+; NOFASTPATH-LABEL: define i16 @test_load16
+; NOFASTPATH-SAME: (ptr [[A:%.*]]) #[[ATTR0]] {
+; NOFASTPATH-NEXT:  entry:
+; NOFASTPATH-NEXT:    [[DOTHWASAN_SHADOW:%.*]] = call ptr asm "", "=r,0"(ptr @__hwasan_shadow)
+; NOFASTPATH-NEXT:    call void @llvm.hwasan.check.memaccess.shortgranules(ptr [[DOTHWASAN_SHADOW]], ptr [[A]], i32 1)
+; NOFASTPATH-NEXT:    [[B:%.*]] = load i16, ptr [[A]], align 4
+; NOFASTPATH-NEXT:    ret i16 [[B]]
+;
+; FASTPATH-LABEL: define i16 @test_load16
+; FASTPATH-SAME: (ptr [[A:%.*]]) #[[ATTR0]] {
+; FASTPATH-NEXT:  entry:
+; FASTPATH-NEXT:    [[DOTHWASAN_SHADOW:%.*]] = call ptr asm "", "=r,0"(ptr @__hwasan_shadow)
+; FASTPATH-NEXT:    call void @llvm.hwasan.check.memaccess.shortgranules(ptr [[DOTHWASAN_SHADOW]], ptr [[A]], i32 1)
+; FASTPATH-NEXT:    [[B:%.*]] = load i16, ptr [[A]], align 4
+; FASTPATH-NEXT:    ret i16 [[B]]
+;
 ; ABORT-DYNAMIC-SHADOW-LABEL: define i16 @test_load16
 ; ABORT-DYNAMIC-SHADOW-SAME: (ptr [[A:%.*]]) #[[ATTR0]] {
 ; ABORT-DYNAMIC-SHADOW-NEXT:  entry:
@@ -206,6 +257,30 @@
 }
 
 define i32 @test_load32(ptr %a) sanitize_hwaddress {
+; CHECK-LABEL: define i32 @test_load32
+; CHECK-SAME: (ptr [[A:%.*]]) #[[ATTR0]] {
+; CHECK-NEXT:  entry:
+; CHECK-NEXT:    [[DOTHWASAN_SHADOW:%.*]] = call ptr asm "", "=r,0"(ptr @__hwasan_shadow)
+; CHECK-NEXT:    call void @llvm.hwasan.check.memaccess.shortgranules(ptr [[DOTHWASAN_SHADOW]], ptr [[A]], i32 2)
+; CHECK-NEXT:    [[B:%.*]] = load i32, ptr [[A]], align 4
+; CHECK-NEXT:    ret i32 [[B]]
+;
+; NOFASTPATH-LABEL: define i32 @test_load32
+; NOFASTPATH-SAME: (ptr [[A:%.*]]) #[[ATTR0]] {
+; NOFASTPATH-NEXT:  entry:
+; NOFASTPATH-NEXT:    [[DOTHWASAN_SHADOW:%.*]] = call ptr asm "", "=r,0"(ptr @__hwasan_shadow)
+; NOFASTPATH-NEXT:    call void @llvm.hwasan.check.memaccess.shortgranules(ptr [[DOTHWASAN_SHADOW]], ptr [[A]], i32 2)
+; NOFASTPATH-NEXT:    [[B:%.*]] = load i32, ptr [[A]], align 4
+; NOFASTPATH-NEXT:    ret i32 [[B]]
+;
+; FASTPATH-LABEL: define i32 @test_load32
+; FASTPATH-SAME: (ptr [[A:%.*]]) #[[ATTR0]] {
+; FASTPATH-NEXT:  entry:
+; FASTPATH-NEXT:    [[DOTHWASAN_SHADOW:%.*]] = call ptr asm "", "=r,0"(ptr @__hwasan_shadow)
+; FASTPATH-NEXT:    call void @llvm.hwasan.check.memaccess.shortgranules(ptr [[DOTHWASAN_SHADOW]], ptr [[A]], i32 2)
+; FASTPATH-NEXT:    [[B:%.*]] = load i32, ptr [[A]], align 4
+; FASTPATH-NEXT:    ret i32 [[B]]
+;
 ; ABORT-DYNAMIC-SHADOW-LABEL: define i32 @test_load32
 ; ABORT-DYNAMIC-SHADOW-SAME: (ptr [[A:%.*]]) #[[ATTR0]] {
 ; ABORT-DYNAMIC-SHADOW-NEXT:  entry:
@@ -303,6 +378,30 @@
 }
 
 define i64 @test_load64(ptr %a) sanitize_hwaddress {
+; CHECK-LABEL: define i64 @test_load64
+; CHECK-SAME: (ptr [[A:%.*]]) #[[ATTR0]] {
+; CHECK-NEXT:  entry:
+; CHECK-NEXT:    [[DOTHWASAN_SHADOW:%.*]] = call ptr asm "", "=r,0"(ptr @__hwasan_shadow)
+; CHECK-NEXT:    call void @llvm.hwasan.check.memaccess.shortgranules(ptr [[DOTHWASAN_SHADOW]], ptr [[A]], i32 3)
+; CHECK-NEXT:    [[B:%.*]] = load i64, ptr [[A]], align 8
+; CHECK-NEXT:    ret i64 [[B]]
+;
+; NOFASTPATH-LABEL: define i64 @test_load64
+; NOFASTPATH-SAME: (ptr [[A:%.*]]) #[[ATTR0]] {
+; NOFASTPATH-NEXT:  entry:
+; NOFASTPATH-NEXT:    [[DOTHWASAN_SHADOW:%.*]] = call ptr asm "", "=r,0"(ptr @__hwasan_shadow)
+; NOFASTPATH-NEXT:    call void @llvm.hwasan.check.memaccess.shortgranules(ptr [[DOTHWASAN_SHADOW]], ptr [[A]], i32 3)
+; NOFASTPATH-NEXT:    [[B:%.*]] = load i64, ptr [[A]], align 8
+; NOFASTPATH-NEXT:    ret i64 [[B]]
+;
+; FASTPATH-LABEL: define i64 @test_load64
+; FASTPATH-SAME: (ptr [[A:%.*]]) #[[ATTR0]] {
+; FASTPATH-NEXT:  entry:
+; FASTPATH-NEXT:    [[DOTHWASAN_SHADOW:%.*]] = call ptr asm "", "=r,0"(ptr @__hwasan_shadow)
+; FASTPATH-NEXT:    call void @llvm.hwasan.check.memaccess.shortgranules(ptr [[DOTHWASAN_SHADOW]], ptr [[A]], i32 3)
+; FASTPATH-NEXT:    [[B:%.*]] = load i64, ptr [[A]], align 8
+; FASTPATH-NEXT:    ret i64 [[B]]
+;
 ; ABORT-DYNAMIC-SHADOW-LABEL: define i64 @test_load64
 ; ABORT-DYNAMIC-SHADOW-SAME: (ptr [[A:%.*]]) #[[ATTR0]] {
 ; ABORT-DYNAMIC-SHADOW-NEXT:  entry:
@@ -400,6 +499,30 @@
 }
 
 define i128 @test_load128(ptr %a) sanitize_hwaddress {
+; CHECK-LABEL: define i128 @test_load128
+; CHECK-SAME: (ptr [[A:%.*]]) #[[ATTR0]] {
+; CHECK-NEXT:  entry:
+; CHECK-NEXT:    [[DOTHWASAN_SHADOW:%.*]] = call ptr asm "", "=r,0"(ptr @__hwasan_shadow)
+; CHECK-NEXT:    call void @llvm.hwasan.check.memaccess.shortgranules(ptr [[DOTHWASAN_SHADOW]], ptr [[A]], i32 4)
+; CHECK-NEXT:    [[B:%.*]] = load i128, ptr [[A]], align 16
+; CHECK-NEXT:    ret i128 [[B]]
+;
+; NOFASTPATH-LABEL: define i128 @test_load128
+; NOFASTPATH-SAME: (ptr [[A:%.*]]) #[[ATTR0]] {
+; NOFASTPATH-NEXT:  entry:
+; NOFASTPATH-NEXT:    [[DOTHWASAN_SHADOW:%.*]] = call ptr asm "", "=r,0"(ptr @__hwasan_shadow)
+; NOFASTPATH-NEXT:    call void @llvm.hwasan.check.memaccess.shortgranules(ptr [[DOTHWASAN_SHADOW]], ptr [[A]], i32 4)
+; NOFASTPATH-NEXT:    [[B:%.*]] = load i128, ptr [[A]], align 16
+; NOFASTPATH-NEXT:    ret i128 [[B]]
+;
+; FASTPATH-LABEL: define i128 @test_load128
+; FASTPATH-SAME: (ptr [[A:%.*]]) #[[ATTR0]] {
+; FASTPATH-NEXT:  entry:
+; FASTPATH-NEXT:    [[DOTHWASAN_SHADOW:%.*]] = call ptr asm "", "=r,0"(ptr @__hwasan_shadow)
+; FASTPATH-NEXT:    call void @llvm.hwasan.check.memaccess.shortgranules(ptr [[DOTHWASAN_SHADOW]], ptr [[A]], i32 4)
+; FASTPATH-NEXT:    [[B:%.*]] = load i128, ptr [[A]], align 16
+; FASTPATH-NEXT:    ret i128 [[B]]
+;
 ; ABORT-DYNAMIC-SHADOW-LABEL: define i128 @test_load128
 ; ABORT-DYNAMIC-SHADOW-SAME: (ptr [[A:%.*]]) #[[ATTR0]] {
 ; ABORT-DYNAMIC-SHADOW-NEXT:  entry:
@@ -497,6 +620,33 @@
 }
 
 define i40 @test_load40(ptr %a) sanitize_hwaddress {
+; CHECK-LABEL: define i40 @test_load40
+; CHECK-SAME: (ptr [[A:%.*]]) #[[ATTR0]] {
+; CHECK-NEXT:  entry:
+; CHECK-NEXT:    [[DOTHWASAN_SHADOW:%.*]] = call ptr asm "", "=r,0"(ptr @__hwasan_shadow)
+; CHECK-NEXT:    [[TMP0:%.*]] = ptrtoint ptr [[A]] to i64
+; CHECK-NEXT:    call void @__hwasan_loadN(i64 [[TMP0]], i64 5)
+; CHECK-NEXT:    [[B:%.*]] = load i40, ptr [[A]], align 4
+; CHECK-NEXT:    ret i40 [[B]]
+;
+; NOFASTPATH-LABEL: define i40 @test_load40
+; NOFASTPATH-SAME: (ptr [[A:%.*]]) #[[ATTR0]] {
+; NOFASTPATH-NEXT:  entry:
+; NOFASTPATH-NEXT:    [[DOTHWASAN_SHADOW:%.*]] = call ptr asm "", "=r,0"(ptr @__hwasan_shadow)
+; NOFASTPATH-NEXT:    [[TMP0:%.*]] = ptrtoint ptr [[A]] to i64
+; NOFASTPATH-NEXT:    call void @__hwasan_loadN(i64 [[TMP0]], i64 5)
+; NOFASTPATH-NEXT:    [[B:%.*]] = load i40, ptr [[A]], align 4
+; NOFASTPATH-NEXT:    ret i40 [[B]]
+;
+; FASTPATH-LABEL: define i40 @test_load40
+; FASTPATH-SAME: (ptr [[A:%.*]]) #[[ATTR0]] {
+; FASTPATH-NEXT:  entry:
+; FASTPATH-NEXT:    [[DOTHWASAN_SHADOW:%.*]] = call ptr asm "", "=r,0"(ptr @__hwasan_shadow)
+; FASTPATH-NEXT:    [[TMP0:%.*]] = ptrtoint ptr [[A]] to i64
+; FASTPATH-NEXT:    call void @__hwasan_loadN(i64 [[TMP0]], i64 5)
+; FASTPATH-NEXT:    [[B:%.*]] = load i40, ptr [[A]], align 4
+; FASTPATH-NEXT:    ret i40 [[B]]
+;
 ; ABORT-DYNAMIC-SHADOW-LABEL: define i40 @test_load40
 ; ABORT-DYNAMIC-SHADOW-SAME: (ptr [[A:%.*]]) #[[ATTR0]] {
 ; ABORT-DYNAMIC-SHADOW-NEXT:  entry:
@@ -539,6 +689,30 @@
 }
 
 define void @test_store8(ptr %a, i8 %b) sanitize_hwaddress {
+; CHECK-LABEL: define void @test_store8
+; CHECK-SAME: (ptr [[A:%.*]], i8 [[B:%.*]]) #[[ATTR0]] {
+; CHECK-NEXT:  entry:
+; CHECK-NEXT:    [[DOTHWASAN_SHADOW:%.*]] = call ptr asm "", "=r,0"(ptr @__hwasan_shadow)
+; CHECK-NEXT:    call void @llvm.hwasan.check.memaccess.shortgranules(ptr [[DOTHWASAN_SHADOW]], ptr [[A]], i32 16)
+; CHECK-NEXT:    store i8 [[B]], ptr [[A]], align 4
+; CHECK-NEXT:    ret void
+;
+; NOFASTPATH-LABEL: define void @test_store8
+; NOFASTPATH-SAME: (ptr [[A:%.*]], i8 [[B:%.*]]) #[[ATTR0]] {
+; NOFASTPATH-NEXT:  entry:
+; NOFASTPATH-NEXT:    [[DOTHWASAN_SHADOW:%.*]] = call ptr asm "", "=r,0"(ptr @__hwasan_shadow)
+; NOFASTPATH-NEXT:    call void @llvm.hwasan.check.memaccess.shortgranules(ptr [[DOTHWASAN_SHADOW]], ptr [[A]], i32 16)
+; NOFASTPATH-NEXT:    store i8 [[B]], ptr [[A]], align 4
+; NOFASTPATH-NEXT:    ret void
+;
+; FASTPATH-LABEL: define void @test_store8
+; FASTPATH-SAME: (ptr [[A:%.*]], i8 [[B:%.*]]) #[[ATTR0]] {
+; FASTPATH-NEXT:  entry:
+; FASTPATH-NEXT:    [[DOTHWASAN_SHADOW:%.*]] = call ptr asm "", "=r,0"(ptr @__hwasan_shadow)
+; FASTPATH-NEXT:    call void @llvm.hwasan.check.memaccess.shortgranules(ptr [[DOTHWASAN_SHADOW]], ptr [[A]], i32 16)
+; FASTPATH-NEXT:    store i8 [[B]], ptr [[A]], align 4
+; FASTPATH-NEXT:    ret void
+;
 ; ABORT-DYNAMIC-SHADOW-LABEL: define void @test_store8
 ; ABORT-DYNAMIC-SHADOW-SAME: (ptr [[A:%.*]], i8 [[B:%.*]]) #[[ATTR0]] {
 ; ABORT-DYNAMIC-SHADOW-NEXT:  entry:
@@ -636,6 +810,30 @@
 }
 
 define void @test_store16(ptr %a, i16 %b) sanitize_hwaddress {
+; CHECK-LABEL: define void @test_store16
+; CHECK-SAME: (ptr [[A:%.*]], i16 [[B:%.*]]) #[[ATTR0]] {
+; CHECK-NEXT:  entry:
+; CHECK-NEXT:    [[DOTHWASAN_SHADOW:%.*]] = call ptr asm "", "=r,0"(ptr @__hwasan_shadow)
+; CHECK-NEXT:    call void @llvm.hwasan.check.memaccess.shortgranules(ptr [[DOTHWASAN_SHADOW]], ptr [[A]], i32 17)
+; CHECK-NEXT:    store i16 [[B]], ptr [[A]], align 4
+; CHECK-NEXT:    ret void
+;
+; NOFASTPATH-LABEL: define void @test_store16
+; NOFASTPATH-SAME: (ptr [[A:%.*]], i16 [[B:%.*]]) #[[ATTR0]] {
+; NOFASTPATH-NEXT:  entry:
+; NOFASTPATH-NEXT:    [[DOTHWASAN_SHADOW:%.*]] = call ptr asm "", "=r,0"(ptr @__hwasan_shadow)
+; NOFASTPATH-NEXT:    call void @llvm.hwasan.check.memaccess.shortgranules(ptr [[DOTHWASAN_SHADOW]], ptr [[A]], i32 17)
+; NOFASTPATH-NEXT:    store i16 [[B]], ptr [[A]], align 4
+; NOFASTPATH-NEXT:    ret void
+;
+; FASTPATH-LABEL: define void @test_store16
+; FASTPATH-SAME: (ptr [[A:%.*]], i16 [[B:%.*]]) #[[ATTR0]] {
+; FASTPATH-NEXT:  entry:
+; FASTPATH-NEXT:    [[DOTHWASAN_SHADOW:%.*]] = call ptr asm "", "=r,0"(ptr @__hwasan_shadow)
+; FASTPATH-NEXT:    call void @llvm.hwasan.check.memaccess.shortgranules(ptr [[DOTHWASAN_SHADOW]], ptr [[A]], i32 17)
+; FASTPATH-NEXT:    store i16 [[B]], ptr [[A]], align 4
+; FASTPATH-NEXT:    ret void
+;
 ; ABORT-DYNAMIC-SHADOW-LABEL: define void @test_store16
 ; ABORT-DYNAMIC-SHADOW-SAME: (ptr [[A:%.*]], i16 [[B:%.*]]) #[[ATTR0]] {
 ; ABORT-DYNAMIC-SHADOW-NEXT:  entry:
@@ -733,6 +931,30 @@
 }
 
 define void @test_store32(ptr %a, i32 %b) sanitize_hwaddress {
+; CHECK-LABEL: define void @test_store32
+; CHECK-SAME: (ptr [[A:%.*]], i32 [[B:%.*]]) #[[ATTR0]] {
+; CHECK-NEXT:  entry:
+; CHECK-NEXT:    [[DOTHWASAN_SHADOW:%.*]] = call ptr asm "", "=r,0"(ptr @__hwasan_shadow)
+; CHECK-NEXT:    call void @llvm.hwasan.check.memaccess.shortgranules(ptr [[DOTHWASAN_SHADOW]], ptr [[A]], i32 18)
+; CHECK-NEXT:    store i32 [[B]], ptr [[A]], align 4
+; CHECK-NEXT:    ret void
+;
+; NOFASTPATH-LABEL: define void @test_store32
+; NOFASTPATH-SAME: (ptr [[A:%.*]], i32 [[B:%.*]]) #[[ATTR0]] {
+; NOFASTPATH-NEXT:  entry:
+; NOFASTPATH-NEXT:    [[DOTHWASAN_SHADOW:%.*]] = call ptr asm "", "=r,0"(ptr @__hwasan_shadow)
+; NOFASTPATH-NEXT:    call void @llvm.hwasan.check.memaccess.shortgranules(ptr [[DOTHWASAN_SHADOW]], ptr [[A]], i32 18)
+; NOFASTPATH-NEXT:    store i32 [[B]], ptr [[A]], align 4
+; NOFASTPATH-NEXT:    ret void
+;
+; FASTPATH-LABEL: define void @test_store32
+; FASTPATH-SAME: (ptr [[A:%.*]], i32 [[B:%.*]]) #[[ATTR0]] {
+; FASTPATH-NEXT:  entry:
+; FASTPATH-NEXT:    [[DOTHWASAN_SHADOW:%.*]] = call ptr asm "", "=r,0"(ptr @__hwasan_shadow)
+; FASTPATH-NEXT:    call void @llvm.hwasan.check.memaccess.shortgranules(ptr [[DOTHWASAN_SHADOW]], ptr [[A]], i32 18)
+; FASTPATH-NEXT:    store i32 [[B]], ptr [[A]], align 4
+; FASTPATH-NEXT:    ret void
+;
 ; ABORT-DYNAMIC-SHADOW-LABEL: define void @test_store32
 ; ABORT-DYNAMIC-SHADOW-SAME: (ptr [[A:%.*]], i32 [[B:%.*]]) #[[ATTR0]] {
 ; ABORT-DYNAMIC-SHADOW-NEXT:  entry:
@@ -830,6 +1052,30 @@
 }
 
 define void @test_store64(ptr %a, i64 %b) sanitize_hwaddress {
+; CHECK-LABEL: define void @test_store64
+; CHECK-SAME: (ptr [[A:%.*]], i64 [[B:%.*]]) #[[ATTR0]] {
+; CHECK-NEXT:  entry:
+; CHECK-NEXT:    [[DOTHWASAN_SHADOW:%.*]] = call ptr asm "", "=r,0"(ptr @__hwasan_shadow)
+; CHECK-NEXT:    call void @llvm.hwasan.check.memaccess.shortgranules(ptr [[DOTHWASAN_SHADOW]], ptr [[A]], i32 19)
+; CHECK-NEXT:    store i64 [[B]], ptr [[A]], align 8
+; CHECK-NEXT:    ret void
+;
+; NOFASTPATH-LABEL: define void @test_store64
+; NOFASTPATH-SAME: (ptr [[A:%.*]], i64 [[B:%.*]]) #[[ATTR0]] {
+; NOFASTPATH-NEXT:  entry:
+; NOFASTPATH-NEXT:    [[DOTHWASAN_SHADOW:%.*]] = call ptr asm "", "=r,0"(ptr @__hwasan_shadow)
+; NOFASTPATH-NEXT:    call void @llvm.hwasan.check.memaccess.shortgranules(ptr [[DOTHWASAN_SHADOW]], ptr [[A]], i32 19)
+; NOFASTPATH-NEXT:    store i64 [[B]], ptr [[A]], align 8
+; NOFASTPATH-NEXT:    ret void
+;
+; FASTPATH-LABEL: define void @test_store64
+; FASTPATH-SAME: (ptr [[A:%.*]], i64 [[B:%.*]]) #[[ATTR0]] {
+; FASTPATH-NEXT:  entry:
+; FASTPATH-NEXT:    [[DOTHWASAN_SHADOW:%.*]] = call ptr asm "", "=r,0"(ptr @__hwasan_shadow)
+; FASTPATH-NEXT:    call void @llvm.hwasan.check.memaccess.shortgranules(ptr [[DOTHWASAN_SHADOW]], ptr [[A]], i32 19)
+; FASTPATH-NEXT:    store i64 [[B]], ptr [[A]], align 8
+; FASTPATH-NEXT:    ret void
+;
 ; ABORT-DYNAMIC-SHADOW-LABEL: define void @test_store64
 ; ABORT-DYNAMIC-SHADOW-SAME: (ptr [[A:%.*]], i64 [[B:%.*]]) #[[ATTR0]] {
 ; ABORT-DYNAMIC-SHADOW-NEXT:  entry:
@@ -927,6 +1173,30 @@
 }
 
 define void @test_store128(ptr %a, i128 %b) sanitize_hwaddress {
+; CHECK-LABEL: define void @test_store128
+; CHECK-SAME: (ptr [[A:%.*]], i128 [[B:%.*]]) #[[ATTR0]] {
+; CHECK-NEXT:  entry:
+; CHECK-NEXT:    [[DOTHWASAN_SHADOW:%.*]] = call ptr asm "", "=r,0"(ptr @__hwasan_shadow)
+; CHECK-NEXT:    call void @llvm.hwasan.check.memaccess.shortgranules(ptr [[DOTHWASAN_SHADOW]], ptr [[A]], i32 20)
+; CHECK-NEXT:    store i128 [[B]], ptr [[A]], align 16
+; CHECK-NEXT:    ret void
+;
+; NOFASTPATH-LABEL: define void @test_store128
+; NOFASTPATH-SAME: (ptr [[A:%.*]], i128 [[B:%.*]]) #[[ATTR0]] {
+; NOFASTPATH-NEXT:  entry:
+; NOFASTPATH-NEXT:    [[DOTHWASAN_SHADOW:%.*]] = call ptr asm "", "=r,0"(ptr @__hwasan_shadow)
+; NOFASTPATH-NEXT:    call void @llvm.hwasan.check.memaccess.shortgranules(ptr [[DOTHWASAN_SHADOW]], ptr [[A]], i32 20)
+; NOFASTPATH-NEXT:    store i128 [[B]], ptr [[A]], align 16
+; NOFASTPATH-NEXT:    ret void
+;
+; FASTPATH-LABEL: define void @test_store128
+; FASTPATH-SAME: (ptr [[A:%.*]], i128 [[B:%.*]]) #[[ATTR0]] {
+; FASTPATH-NEXT:  entry:
+; FASTPATH-NEXT:    [[DOTHWASAN_SHADOW:%.*]] = call ptr asm "", "=r,0"(ptr @__hwasan_shadow)
+; FASTPATH-NEXT:    call void @llvm.hwasan.check.memaccess.shortgranules(ptr [[DOTHWASAN_SHADOW]], ptr [[A]], i32 20)
+; FASTPATH-NEXT:    store i128 [[B]], ptr [[A]], align 16
+; FASTPATH-NEXT:    ret void
+;
 ; ABORT-DYNAMIC-SHADOW-LABEL: define void @test_store128
 ; ABORT-DYNAMIC-SHADOW-SAME: (ptr [[A:%.*]], i128 [[B:%.*]]) #[[ATTR0]] {
 ; ABORT-DYNAMIC-SHADOW-NEXT:  entry:
@@ -1024,6 +1294,33 @@
 }
 
 define void @test_store40(ptr %a, i40 %b) sanitize_hwaddress {
+; CHECK-LABEL: define void @test_store40
+; CHECK-SAME: (ptr [[A:%.*]], i40 [[B:%.*]]) #[[ATTR0]] {
+; CHECK-NEXT:  entry:
+; CHECK-NEXT:    [[DOTHWASAN_SHADOW:%.*]] = call ptr asm "", "=r,0"(ptr @__hwasan_shadow)
+; CHECK-NEXT:    [[TMP0:%.*]] = ptrtoint ptr [[A]] to i64
+; CHECK-NEXT:    call void @__hwasan_storeN(i64 [[TMP0]], i64 5)
+; CHECK-NEXT:    store i40 [[B]], ptr [[A]], align 4
+; CHECK-NEXT:    ret void
+;
+; NOFASTPATH-LABEL: define void @test_store40
+; NOFASTPATH-SAME: (ptr [[A:%.*]], i40 [[B:%.*]]) #[[ATTR0]] {
+; NOFASTPATH-NEXT:  entry:
+; NOFASTPATH-NEXT:    [[DOTHWASAN_SHADOW:%.*]] = call ptr asm "", "=r,0"(ptr @__hwasan_shadow)
+; NOFASTPATH-NEXT:    [[TMP0:%.*]] = ptrtoint ptr [[A]] to i64
+; NOFASTPATH-NEXT:    call void @__hwasan_storeN(i64 [[TMP0]], i64 5)
+; NOFASTPATH-NEXT:    store i40 [[B]], ptr [[A]], align 4
+; NOFASTPATH-NEXT:    ret void
+;
+; FASTPATH-LABEL: define void @test_store40
+; FASTPATH-SAME: (ptr [[A:%.*]], i40 [[B:%.*]]) #[[ATTR0]] {
+; FASTPATH-NEXT:  entry:
+; FASTPATH-NEXT:    [[DOTHWASAN_SHADOW:%.*]] = call ptr asm "", "=r,0"(ptr @__hwasan_shadow)
+; FASTPATH-NEXT:    [[TMP0:%.*]] = ptrtoint ptr [[A]] to i64
+; FASTPATH-NEXT:    call void @__hwasan_storeN(i64 [[TMP0]], i64 5)
+; FASTPATH-NEXT:    store i40 [[B]], ptr [[A]], align 4
+; FASTPATH-NEXT:    ret void
+;
 ; ABORT-DYNAMIC-SHADOW-LABEL: define void @test_store40
 ; ABORT-DYNAMIC-SHADOW-SAME: (ptr [[A:%.*]], i40 [[B:%.*]]) #[[ATTR0]] {
 ; ABORT-DYNAMIC-SHADOW-NEXT:  entry:
@@ -1066,6 +1363,33 @@
 }
 
 define void @test_store_unaligned(ptr %a, i64 %b) sanitize_hwaddress {
+; CHECK-LABEL: define void @test_store_unaligned
+; CHECK-SAME: (ptr [[A:%.*]], i64 [[B:%.*]]) #[[ATTR0]] {
+; CHECK-NEXT:  entry:
+; CHECK-NEXT:    [[DOTHWASAN_SHADOW:%.*]] = call ptr asm "", "=r,0"(ptr @__hwasan_shadow)
+; CHECK-NEXT:    [[TMP0:%.*]] = ptrtoint ptr [[A]] to i64
+; CHECK-NEXT:    call void @__hwasan_storeN(i64 [[TMP0]], i64 8)
+; CHECK-NEXT:    store i64 [[B]], ptr [[A]], align 4
+; CHECK-NEXT:    ret void
+;
+; NOFASTPATH-LABEL: define void @test_store_unaligned
+; NOFASTPATH-SAME: (ptr [[A:%.*]], i64 [[B:%.*]]) #[[ATTR0]] {
+; NOFASTPATH-NEXT:  entry:
+; NOFASTPATH-NEXT:    [[DOTHWASAN_SHADOW:%.*]] = call ptr asm "", "=r,0"(ptr @__hwasan_shadow)
+; NOFASTPATH-NEXT:    [[TMP0:%.*]] = ptrtoint ptr [[A]] to i64
+; NOFASTPATH-NEXT:    call void @__hwasan_storeN(i64 [[TMP0]], i64 8)
+; NOFASTPATH-NEXT:    store i64 [[B]], ptr [[A]], align 4
+; NOFASTPATH-NEXT:    ret void
+;
+; FASTPATH-LABEL: define void @test_store_unaligned
+; FASTPATH-SAME: (ptr [[A:%.*]], i64 [[B:%.*]]) #[[ATTR0]] {
+; FASTPATH-NEXT:  entry:
+; FASTPATH-NEXT:    [[DOTHWASAN_SHADOW:%.*]] = call ptr asm "", "=r,0"(ptr @__hwasan_shadow)
+; FASTPATH-NEXT:    [[TMP0:%.*]] = ptrtoint ptr [[A]] to i64
+; FASTPATH-NEXT:    call void @__hwasan_storeN(i64 [[TMP0]], i64 8)
+; FASTPATH-NEXT:    store i64 [[B]], ptr [[A]], align 4
+; FASTPATH-NEXT:    ret void
+;
 ; ABORT-DYNAMIC-SHADOW-LABEL: define void @test_store_unaligned
 ; ABORT-DYNAMIC-SHADOW-SAME: (ptr [[A:%.*]], i64 [[B:%.*]]) #[[ATTR0]] {
 ; ABORT-DYNAMIC-SHADOW-NEXT:  entry:
@@ -1114,6 +1438,18 @@
 ; CHECK-NEXT:    [[B:%.*]] = load i8, ptr [[A]], align 4
 ; CHECK-NEXT:    ret i8 [[B]]
 ;
+; NOFASTPATH-LABEL: define i8 @test_load_noattr
+; NOFASTPATH-SAME: (ptr [[A:%.*]]) {
+; NOFASTPATH-NEXT:  entry:
+; NOFASTPATH-NEXT:    [[B:%.*]] = load i8, ptr [[A]], align 4
+; NOFASTPATH-NEXT:    ret i8 [[B]]
+;
+; FASTPATH-LABEL: define i8 @test_load_noattr
+; FASTPATH-SAME: (ptr [[A:%.*]]) {
+; FASTPATH-NEXT:  entry:
+; FASTPATH-NEXT:    [[B:%.*]] = load i8, ptr [[A]], align 4
+; FASTPATH-NEXT:    ret i8 [[B]]
+;
 ; ABORT-DYNAMIC-SHADOW-LABEL: define i8 @test_load_noattr
 ; ABORT-DYNAMIC-SHADOW-SAME: (ptr [[A:%.*]]) {
 ; ABORT-DYNAMIC-SHADOW-NEXT:  entry:
@@ -1150,6 +1486,18 @@
 ; CHECK-NEXT:    [[B:%.*]] = load i8, ptr [[A]], align 4
 ; CHECK-NEXT:    ret i8 [[B]]
 ;
+; NOFASTPATH-LABEL: define i8 @test_load_notmyattr
+; NOFASTPATH-SAME: (ptr [[A:%.*]]) #[[ATTR1:[0-9]+]] {
+; NOFASTPATH-NEXT:  entry:
+; NOFASTPATH-NEXT:    [[B:%.*]] = load i8, ptr [[A]], align 4
+; NOFASTPATH-NEXT:    ret i8 [[B]]
+;
+; FASTPATH-LABEL: define i8 @test_load_notmyattr
+; FASTPATH-SAME: (ptr [[A:%.*]]) #[[ATTR1:[0-9]+]] {
+; FASTPATH-NEXT:  entry:
+; FASTPATH-NEXT:    [[B:%.*]] = load i8, ptr [[A]], align 4
+; FASTPATH-NEXT:    ret i8 [[B]]
+;
 ; ABORT-DYNAMIC-SHADOW-LABEL: define i8 @test_load_notmyattr
 ; ABORT-DYNAMIC-SHADOW-SAME: (ptr [[A:%.*]]) #[[ATTR1:[0-9]+]] {
 ; ABORT-DYNAMIC-SHADOW-NEXT:  entry:
@@ -1181,11 +1529,23 @@
 
 define i8 @test_load_addrspace(ptr addrspace(256) %a) sanitize_hwaddress {
 ; CHECK-LABEL: define i8 @test_load_addrspace
-; CHECK-SAME: (ptr addrspace(256) [[A:%.*]]) #[[ATTR0:[0-9]+]] {
+; CHECK-SAME: (ptr addrspace(256) [[A:%.*]]) #[[ATTR0]] {
 ; CHECK-NEXT:  entry:
 ; CHECK-NEXT:    [[B:%.*]] = load i8, ptr addrspace(256) [[A]], align 4
 ; CHECK-NEXT:    ret i8 [[B]]
 ;
+; NOFASTPATH-LABEL: define i8 @test_load_addrspace
+; NOFASTPATH-SAME: (ptr addrspace(256) [[A:%.*]]) #[[ATTR0]] {
+; NOFASTPATH-NEXT:  entry:
+; NOFASTPATH-NEXT:    [[B:%.*]] = load i8, ptr addrspace(256) [[A]], align 4
+; NOFASTPATH-NEXT:    ret i8 [[B]]
+;
+; FASTPATH-LABEL: define i8 @test_load_addrspace
+; FASTPATH-SAME: (ptr addrspace(256) [[A:%.*]]) #[[ATTR0]] {
+; FASTPATH-NEXT:  entry:
+; FASTPATH-NEXT:    [[B:%.*]] = load i8, ptr addrspace(256) [[A]], align 4
+; FASTPATH-NEXT:    ret i8 [[B]]
+;
 ; ABORT-DYNAMIC-SHADOW-LABEL: define i8 @test_load_addrspace
 ; ABORT-DYNAMIC-SHADOW-SAME: (ptr addrspace(256) [[A:%.*]]) #[[ATTR0]] {
 ; ABORT-DYNAMIC-SHADOW-NEXT:  entry: