AArch64: remove barriers from AArch64 atomic operations.

I've managed to convince myself that AArch64's acquire/release
instructions are sufficient to guarantee C++11's required semantics,
even in the sequentially-consistent case.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@179005 91177308-0d34-0410-b5e6-96231b3b80d8

1 files changed, 203 insertions, 178 deletions

diff --git a/test/CodeGen/AArch64/atomic-ops.ll b/test/CodeGen/AArch64/atomic-ops.ll
index f3c16171cc..5e87f21a21 100644
--- a/test/CodeGen/AArch64/atomic-ops.ll
+++ b/test/CodeGen/AArch64/atomic-ops.ll
@@ -8,18 +8,18 @@
 define i8 @test_atomic_load_add_i8(i8 %offset) nounwind {
 ; CHECK: test_atomic_load_add_i8:
    %old = atomicrmw add i8* @var8, i8 %offset seq_cst
-; CHECK: dmb ish
+; CHECK-NOT: dmb
 ; CHECK: adrp [[TMPADDR:x[0-9]+]], var8
 ; CHECK: add x[[ADDR:[0-9]+]], [[TMPADDR]], #:lo12:var8
 
 ; CHECK: .LBB{{[0-9]+}}_1:
-; CHECK-NEXT: ldxrb w[[OLD:[0-9]+]], [x[[ADDR]]]
+; CHECK-NEXT: ldaxrb w[[OLD:[0-9]+]], [x[[ADDR]]]
   ; w0 below is a reasonable guess but could change: it certainly comes into the
   ;  function there.
 ; CHECK-NEXT: add [[NEW:w[0-9]+]], w[[OLD]], w0
-; CHECK-NEXT: stxrb [[STATUS:w[0-9]+]], [[NEW]], [x[[ADDR]]]
+; CHECK-NEXT: stlxrb [[STATUS:w[0-9]+]], [[NEW]], [x[[ADDR]]]
 ; CHECK-NEXT: cbnz [[STATUS]], .LBB{{[0-9]+}}_1
-; CHECK: dmb ish
+; CHECK-NOT: dmb
 
 ; CHECK: mov x0, x[[OLD]]
    ret i8 %old
@@ -27,19 +27,19 @@ define i8 @test_atomic_load_add_i8(i8 %offset) nounwind {
 
 define i16 @test_atomic_load_add_i16(i16 %offset) nounwind {
 ; CHECK: test_atomic_load_add_i16:
-   %old = atomicrmw add i16* @var16, i16 %offset seq_cst
-; CHECK: dmb ish
+   %old = atomicrmw add i16* @var16, i16 %offset acquire
+; CHECK-NOT: dmb
 ; CHECK: adrp [[TMPADDR:x[0-9]+]], var16
 ; CHECK: add x[[ADDR:[0-9]+]], [[TMPADDR]], #:lo12:var16
 
 ; CHECK: .LBB{{[0-9]+}}_1:
-; CHECK-NEXT: ldxrh w[[OLD:[0-9]+]], [x[[ADDR]]]
+; CHECK-NEXT: ldaxrh w[[OLD:[0-9]+]], [x[[ADDR]]]
   ; w0 below is a reasonable guess but could change: it certainly comes into the
   ;  function there.
 ; CHECK-NEXT: add [[NEW:w[0-9]+]], w[[OLD]], w0
 ; CHECK-NEXT: stxrh [[STATUS:w[0-9]+]], [[NEW]], [x[[ADDR]]]
 ; CHECK-NEXT: cbnz [[STATUS]], .LBB{{[0-9]+}}_1
-; CHECK: dmb ish
+; CHECK-NOT: dmb
 
 ; CHECK: mov x0, x[[OLD]]
    ret i16 %old
@@ -47,8 +47,8 @@ define i16 @test_atomic_load_add_i16(i16 %offset) nounwind {
 
 define i32 @test_atomic_load_add_i32(i32 %offset) nounwind {
 ; CHECK: test_atomic_load_add_i32:
-   %old = atomicrmw add i32* @var32, i32 %offset seq_cst
-; CHECK: dmb ish
+   %old = atomicrmw add i32* @var32, i32 %offset release
+; CHECK-NOT: dmb
 ; CHECK: adrp [[TMPADDR:x[0-9]+]], var32
 ; CHECK: add x[[ADDR:[0-9]+]], [[TMPADDR]], #:lo12:var32
 
@@ -57,9 +57,9 @@ define i32 @test_atomic_load_add_i32(i32 %offset) nounwind {
   ; w0 below is a reasonable guess but could change: it certainly comes into the
   ;  function there.
 ; CHECK-NEXT: add [[NEW:w[0-9]+]], w[[OLD]], w0
-; CHECK-NEXT: stxr [[STATUS:w[0-9]+]], [[NEW]], [x[[ADDR]]]
+; CHECK-NEXT: stlxr [[STATUS:w[0-9]+]], [[NEW]], [x[[ADDR]]]
 ; CHECK-NEXT: cbnz [[STATUS]], .LBB{{[0-9]+}}_1
-; CHECK: dmb ish
+; CHECK-NOT: dmb
 
 ; CHECK: mov x0, x[[OLD]]
    ret i32 %old
@@ -67,8 +67,8 @@ define i32 @test_atomic_load_add_i32(i32 %offset) nounwind {
 
 define i64 @test_atomic_load_add_i64(i64 %offset) nounwind {
 ; CHECK: test_atomic_load_add_i64:
-   %old = atomicrmw add i64* @var64, i64 %offset seq_cst
-; CHECK: dmb ish
+   %old = atomicrmw add i64* @var64, i64 %offset monotonic
+; CHECK-NOT: dmb
 ; CHECK: adrp [[TMPADDR:x[0-9]+]], var64
 ; CHECK: add x[[ADDR:[0-9]+]], [[TMPADDR]], #:lo12:var64
 
@@ -79,7 +79,7 @@ define i64 @test_atomic_load_add_i64(i64 %offset) nounwind {
 ; CHECK-NEXT: add [[NEW:x[0-9]+]], x[[OLD]], x0
 ; CHECK-NEXT: stxr [[STATUS:w[0-9]+]], [[NEW]], [x[[ADDR]]]
 ; CHECK-NEXT: cbnz [[STATUS]], .LBB{{[0-9]+}}_1
-; CHECK: dmb ish
+; CHECK-NOT: dmb
 
 ; CHECK: mov x0, x[[OLD]]
    ret i64 %old
@@ -87,8 +87,8 @@ define i64 @test_atomic_load_add_i64(i64 %offset) nounwind {
 
 define i8 @test_atomic_load_sub_i8(i8 %offset) nounwind {
 ; CHECK: test_atomic_load_sub_i8:
-   %old = atomicrmw sub i8* @var8, i8 %offset seq_cst
-; CHECK: dmb ish
+   %old = atomicrmw sub i8* @var8, i8 %offset monotonic
+; CHECK-NOT: dmb
 ; CHECK: adrp [[TMPADDR:x[0-9]+]], var8
 ; CHECK: add x[[ADDR:[0-9]+]], [[TMPADDR]], #:lo12:var8
 
@@ -99,7 +99,7 @@ define i8 @test_atomic_load_sub_i8(i8 %offset) nounwind {
 ; CHECK-NEXT: sub [[NEW:w[0-9]+]], w[[OLD]], w0
 ; CHECK-NEXT: stxrb [[STATUS:w[0-9]+]], [[NEW]], [x[[ADDR]]]
 ; CHECK-NEXT: cbnz [[STATUS]], .LBB{{[0-9]+}}_1
-; CHECK: dmb ish
+; CHECK-NOT: dmb
 
 ; CHECK: mov x0, x[[OLD]]
    ret i8 %old
@@ -107,8 +107,8 @@ define i8 @test_atomic_load_sub_i8(i8 %offset) nounwind {
 
 define i16 @test_atomic_load_sub_i16(i16 %offset) nounwind {
 ; CHECK: test_atomic_load_sub_i16:
-   %old = atomicrmw sub i16* @var16, i16 %offset seq_cst
-; CHECK: dmb ish
+   %old = atomicrmw sub i16* @var16, i16 %offset release
+; CHECK-NOT: dmb
 ; CHECK: adrp [[TMPADDR:x[0-9]+]], var16
 ; CHECK: add x[[ADDR:[0-9]+]], [[TMPADDR]], #:lo12:var16
 
@@ -117,9 +117,9 @@ define i16 @test_atomic_load_sub_i16(i16 %offset) nounwind {
   ; w0 below is a reasonable guess but could change: it certainly comes into the
   ;  function there.
 ; CHECK-NEXT: sub [[NEW:w[0-9]+]], w[[OLD]], w0
-; CHECK-NEXT: stxrh [[STATUS:w[0-9]+]], [[NEW]], [x[[ADDR]]]
+; CHECK-NEXT: stlxrh [[STATUS:w[0-9]+]], [[NEW]], [x[[ADDR]]]
 ; CHECK-NEXT: cbnz [[STATUS]], .LBB{{[0-9]+}}_1
-; CHECK: dmb ish
+; CHECK-NOT: dmb
 
 ; CHECK: mov x0, x[[OLD]]
    ret i16 %old
@@ -127,19 +127,19 @@ define i16 @test_atomic_load_sub_i16(i16 %offset) nounwind {
 
 define i32 @test_atomic_load_sub_i32(i32 %offset) nounwind {
 ; CHECK: test_atomic_load_sub_i32:
-   %old = atomicrmw sub i32* @var32, i32 %offset seq_cst
-; CHECK: dmb ish
+   %old = atomicrmw sub i32* @var32, i32 %offset acquire
+; CHECK-NOT: dmb
 ; CHECK: adrp [[TMPADDR:x[0-9]+]], var32
 ; CHECK: add x[[ADDR:[0-9]+]], [[TMPADDR]], #:lo12:var32
 
 ; CHECK: .LBB{{[0-9]+}}_1:
-; CHECK-NEXT: ldxr w[[OLD:[0-9]+]], [x[[ADDR]]]
+; CHECK-NEXT: ldaxr w[[OLD:[0-9]+]], [x[[ADDR]]]
   ; w0 below is a reasonable guess but could change: it certainly comes into the
   ;  function there.
 ; CHECK-NEXT: sub [[NEW:w[0-9]+]], w[[OLD]], w0
 ; CHECK-NEXT: stxr [[STATUS:w[0-9]+]], [[NEW]], [x[[ADDR]]]
 ; CHECK-NEXT: cbnz [[STATUS]], .LBB{{[0-9]+}}_1
-; CHECK: dmb ish
+; CHECK-NOT: dmb
 
 ; CHECK: mov x0, x[[OLD]]
    ret i32 %old
@@ -148,18 +148,18 @@ define i32 @test_atomic_load_sub_i32(i32 %offset) nounwind {
 define i64 @test_atomic_load_sub_i64(i64 %offset) nounwind {
 ; CHECK: test_atomic_load_sub_i64:
    %old = atomicrmw sub i64* @var64, i64 %offset seq_cst
-; CHECK: dmb ish
+; CHECK-NOT: dmb
 ; CHECK: adrp [[TMPADDR:x[0-9]+]], var64
 ; CHECK: add x[[ADDR:[0-9]+]], [[TMPADDR]], #:lo12:var64
 
 ; CHECK: .LBB{{[0-9]+}}_1:
-; CHECK-NEXT: ldxr x[[OLD:[0-9]+]], [x[[ADDR]]]
+; CHECK-NEXT: ldaxr x[[OLD:[0-9]+]], [x[[ADDR]]]
   ; x0 below is a reasonable guess but could change: it certainly comes into the
   ; function there.
 ; CHECK-NEXT: sub [[NEW:x[0-9]+]], x[[OLD]], x0
-; CHECK-NEXT: stxr [[STATUS:w[0-9]+]], [[NEW]], [x[[ADDR]]]
+; CHECK-NEXT: stlxr [[STATUS:w[0-9]+]], [[NEW]], [x[[ADDR]]]
 ; CHECK-NEXT: cbnz [[STATUS]], .LBB{{[0-9]+}}_1
-; CHECK: dmb ish
+; CHECK-NOT: dmb
 
 ; CHECK: mov x0, x[[OLD]]
    ret i64 %old
@@ -167,8 +167,8 @@ define i64 @test_atomic_load_sub_i64(i64 %offset) nounwind {
 
 define i8 @test_atomic_load_and_i8(i8 %offset) nounwind {
 ; CHECK: test_atomic_load_and_i8:
-   %old = atomicrmw and i8* @var8, i8 %offset seq_cst
-; CHECK: dmb ish
+   %old = atomicrmw and i8* @var8, i8 %offset release
+; CHECK-NOT: dmb
 ; CHECK: adrp [[TMPADDR:x[0-9]+]], var8
 ; CHECK: add x[[ADDR:[0-9]+]], [[TMPADDR]], #:lo12:var8
 
@@ -177,9 +177,9 @@ define i8 @test_atomic_load_and_i8(i8 %offset) nounwind {
   ; w0 below is a reasonable guess but could change: it certainly comes into the
   ;  function there.
 ; CHECK-NEXT: and [[NEW:w[0-9]+]], w[[OLD]], w0
-; CHECK-NEXT: stxrb [[STATUS:w[0-9]+]], [[NEW]], [x[[ADDR]]]
+; CHECK-NEXT: stlxrb [[STATUS:w[0-9]+]], [[NEW]], [x[[ADDR]]]
 ; CHECK-NEXT: cbnz [[STATUS]], .LBB{{[0-9]+}}_1
-; CHECK: dmb ish
+; CHECK-NOT: dmb
 
 ; CHECK: mov x0, x[[OLD]]
    ret i8 %old
@@ -187,8 +187,8 @@ define i8 @test_atomic_load_and_i8(i8 %offset) nounwind {
 
 define i16 @test_atomic_load_and_i16(i16 %offset) nounwind {
 ; CHECK: test_atomic_load_and_i16:
-   %old = atomicrmw and i16* @var16, i16 %offset seq_cst
-; CHECK: dmb ish
+   %old = atomicrmw and i16* @var16, i16 %offset monotonic
+; CHECK-NOT: dmb
 ; CHECK: adrp [[TMPADDR:x[0-9]+]], var16
 ; CHECK: add x[[ADDR:[0-9]+]], [[TMPADDR]], #:lo12:var16
 
@@ -199,7 +199,7 @@ define i16 @test_atomic_load_and_i16(i16 %offset) nounwind {
 ; CHECK-NEXT: and [[NEW:w[0-9]+]], w[[OLD]], w0
 ; CHECK-NEXT: stxrh [[STATUS:w[0-9]+]], [[NEW]], [x[[ADDR]]]
 ; CHECK-NEXT: cbnz [[STATUS]], .LBB{{[0-9]+}}_1
-; CHECK: dmb ish
+; CHECK-NOT: dmb
 
 ; CHECK: mov x0, x[[OLD]]
    ret i16 %old
@@ -208,18 +208,18 @@ define i16 @test_atomic_load_and_i16(i16 %offset) nounwind {
 define i32 @test_atomic_load_and_i32(i32 %offset) nounwind {
 ; CHECK: test_atomic_load_and_i32:
    %old = atomicrmw and i32* @var32, i32 %offset seq_cst
-; CHECK: dmb ish
+; CHECK-NOT: dmb
 ; CHECK: adrp [[TMPADDR:x[0-9]+]], var32
 ; CHECK: add x[[ADDR:[0-9]+]], [[TMPADDR]], #:lo12:var32
 
 ; CHECK: .LBB{{[0-9]+}}_1:
-; CHECK-NEXT: ldxr w[[OLD:[0-9]+]], [x[[ADDR]]]
+; CHECK-NEXT: ldaxr w[[OLD:[0-9]+]], [x[[ADDR]]]
   ; w0 below is a reasonable guess but could change: it certainly comes into the
   ;  function there.
 ; CHECK-NEXT: and [[NEW:w[0-9]+]], w[[OLD]], w0
-; CHECK-NEXT: stxr [[STATUS:w[0-9]+]], [[NEW]], [x[[ADDR]]]
+; CHECK-NEXT: stlxr [[STATUS:w[0-9]+]], [[NEW]], [x[[ADDR]]]
 ; CHECK-NEXT: cbnz [[STATUS]], .LBB{{[0-9]+}}_1
-; CHECK: dmb ish
+; CHECK-NOT: dmb
 
 ; CHECK: mov x0, x[[OLD]]
    ret i32 %old
@@ -227,19 +227,19 @@ define i32 @test_atomic_load_and_i32(i32 %offset) nounwind {
 
 define i64 @test_atomic_load_and_i64(i64 %offset) nounwind {
 ; CHECK: test_atomic_load_and_i64:
-   %old = atomicrmw and i64* @var64, i64 %offset seq_cst
-; CHECK: dmb ish
+   %old = atomicrmw and i64* @var64, i64 %offset acquire
+; CHECK-NOT: dmb
 ; CHECK: adrp [[TMPADDR:x[0-9]+]], var64
 ; CHECK: add x[[ADDR:[0-9]+]], [[TMPADDR]], #:lo12:var64
 
 ; CHECK: .LBB{{[0-9]+}}_1:
-; CHECK-NEXT: ldxr x[[OLD:[0-9]+]], [x[[ADDR]]]
+; CHECK-NEXT: ldaxr x[[OLD:[0-9]+]], [x[[ADDR]]]
   ; x0 below is a reasonable guess but could change: it certainly comes into the
   ; function there.
 ; CHECK-NEXT: and [[NEW:x[0-9]+]], x[[OLD]], x0
 ; CHECK-NEXT: stxr [[STATUS:w[0-9]+]], [[NEW]], [x[[ADDR]]]
 ; CHECK-NEXT: cbnz [[STATUS]], .LBB{{[0-9]+}}_1
-; CHECK: dmb ish
+; CHECK-NOT: dmb
 
 ; CHECK: mov x0, x[[OLD]]
    ret i64 %old
@@ -248,18 +248,18 @@ define i64 @test_atomic_load_and_i64(i64 %offset) nounwind {
 define i8 @test_atomic_load_or_i8(i8 %offset) nounwind {
 ; CHECK: test_atomic_load_or_i8:
    %old = atomicrmw or i8* @var8, i8 %offset seq_cst
-; CHECK: dmb ish
+; CHECK-NOT: dmb
 ; CHECK: adrp [[TMPADDR:x[0-9]+]], var8
 ; CHECK: add x[[ADDR:[0-9]+]], [[TMPADDR]], #:lo12:var8
 
 ; CHECK: .LBB{{[0-9]+}}_1:
-; CHECK-NEXT: ldxrb w[[OLD:[0-9]+]], [x[[ADDR]]]
+; CHECK-NEXT: ldaxrb w[[OLD:[0-9]+]], [x[[ADDR]]]
   ; w0 below is a reasonable guess but could change: it certainly comes into the
   ;  function there.
 ; CHECK-NEXT: orr [[NEW:w[0-9]+]], w[[OLD]], w0
-; CHECK-NEXT: stxrb [[STATUS:w[0-9]+]], [[NEW]], [x[[ADDR]]]
+; CHECK-NEXT: stlxrb [[STATUS:w[0-9]+]], [[NEW]], [x[[ADDR]]]
 ; CHECK-NEXT: cbnz [[STATUS]], .LBB{{[0-9]+}}_1
-; CHECK: dmb ish
+; CHECK-NOT: dmb
 
 ; CHECK: mov x0, x[[OLD]]
    ret i8 %old
@@ -267,8 +267,8 @@ define i8 @test_atomic_load_or_i8(i8 %offset) nounwind {
 
 define i16 @test_atomic_load_or_i16(i16 %offset) nounwind {
 ; CHECK: test_atomic_load_or_i16:
-   %old = atomicrmw or i16* @var16, i16 %offset seq_cst
-; CHECK: dmb ish
+   %old = atomicrmw or i16* @var16, i16 %offset monotonic
+; CHECK-NOT: dmb
 ; CHECK: adrp [[TMPADDR:x[0-9]+]], var16
 ; CHECK: add x[[ADDR:[0-9]+]], [[TMPADDR]], #:lo12:var16
 
@@ -279,7 +279,7 @@ define i16 @test_atomic_load_or_i16(i16 %offset) nounwind {
 ; CHECK-NEXT: orr [[NEW:w[0-9]+]], w[[OLD]], w0
 ; CHECK-NEXT: stxrh [[STATUS:w[0-9]+]], [[NEW]], [x[[ADDR]]]
 ; CHECK-NEXT: cbnz [[STATUS]], .LBB{{[0-9]+}}_1
-; CHECK: dmb ish
+; CHECK-NOT: dmb
 
 ; CHECK: mov x0, x[[OLD]]
    ret i16 %old
@@ -287,19 +287,19 @@ define i16 @test_atomic_load_or_i16(i16 %offset) nounwind {
 
 define i32 @test_atomic_load_or_i32(i32 %offset) nounwind {
 ; CHECK: test_atomic_load_or_i32:
-   %old = atomicrmw or i32* @var32, i32 %offset seq_cst
-; CHECK: dmb ish
+   %old = atomicrmw or i32* @var32, i32 %offset acquire
+; CHECK-NOT: dmb
 ; CHECK: adrp [[TMPADDR:x[0-9]+]], var32
 ; CHECK: add x[[ADDR:[0-9]+]], [[TMPADDR]], #:lo12:var32
 
 ; CHECK: .LBB{{[0-9]+}}_1:
-; CHECK-NEXT: ldxr w[[OLD:[0-9]+]], [x[[ADDR]]]
+; CHECK-NEXT: ldaxr w[[OLD:[0-9]+]], [x[[ADDR]]]
   ; w0 below is a reasonable guess but could change: it certainly comes into the
   ;  function there.
 ; CHECK-NEXT: orr [[NEW:w[0-9]+]], w[[OLD]], w0
 ; CHECK-NEXT: stxr [[STATUS:w[0-9]+]], [[NEW]], [x[[ADDR]]]
 ; CHECK-NEXT: cbnz [[STATUS]], .LBB{{[0-9]+}}_1
-; CHECK: dmb ish
+; CHECK-NOT: dmb
 
 ; CHECK: mov x0, x[[OLD]]
    ret i32 %old
@@ -307,8 +307,8 @@ define i32 @test_atomic_load_or_i32(i32 %offset) nounwind {
 
 define i64 @test_atomic_load_or_i64(i64 %offset) nounwind {
 ; CHECK: test_atomic_load_or_i64:
-   %old = atomicrmw or i64* @var64, i64 %offset seq_cst
-; CHECK: dmb ish
+   %old = atomicrmw or i64* @var64, i64 %offset release
+; CHECK-NOT: dmb
 ; CHECK: adrp [[TMPADDR:x[0-9]+]], var64
 ; CHECK: add x[[ADDR:[0-9]+]], [[TMPADDR]], #:lo12:var64
 
@@ -317,9 +317,9 @@ define i64 @test_atomic_load_or_i64(i64 %offset) nounwind {
   ; x0 below is a reasonable guess but could change: it certainly comes into the
   ; function there.
 ; CHECK-NEXT: orr [[NEW:x[0-9]+]], x[[OLD]], x0
-; CHECK-NEXT: stxr [[STATUS:w[0-9]+]], [[NEW]], [x[[ADDR]]]
+; CHECK-NEXT: stlxr [[STATUS:w[0-9]+]], [[NEW]], [x[[ADDR]]]
 ; CHECK-NEXT: cbnz [[STATUS]], .LBB{{[0-9]+}}_1
-; CHECK: dmb ish
+; CHECK-NOT: dmb
 
 ; CHECK: mov x0, x[[OLD]]
    ret i64 %old
@@ -327,19 +327,19 @@ define i64 @test_atomic_load_or_i64(i64 %offset) nounwind {
 
 define i8 @test_atomic_load_xor_i8(i8 %offset) nounwind {
 ; CHECK: test_atomic_load_xor_i8:
-   %old = atomicrmw xor i8* @var8, i8 %offset seq_cst
-; CHECK: dmb ish
+   %old = atomicrmw xor i8* @var8, i8 %offset acquire
+; CHECK-NOT: dmb
 ; CHECK: adrp [[TMPADDR:x[0-9]+]], var8
 ; CHECK: add x[[ADDR:[0-9]+]], [[TMPADDR]], #:lo12:var8
 
 ; CHECK: .LBB{{[0-9]+}}_1:
-; CHECK-NEXT: ldxrb w[[OLD:[0-9]+]], [x[[ADDR]]]
+; CHECK-NEXT: ldaxrb w[[OLD:[0-9]+]], [x[[ADDR]]]
   ; w0 below is a reasonable guess but could change: it certainly comes into the
   ;  function there.
 ; CHECK-NEXT: eor [[NEW:w[0-9]+]], w[[OLD]], w0
 ; CHECK-NEXT: stxrb [[STATUS:w[0-9]+]], [[NEW]], [x[[ADDR]]]
 ; CHECK-NEXT: cbnz [[STATUS]], .LBB{{[0-9]+}}_1
-; CHECK: dmb ish
+; CHECK-NOT: dmb
 
 ; CHECK: mov x0, x[[OLD]]
    ret i8 %old
@@ -347,8 +347,8 @@ define i8 @test_atomic_load_xor_i8(i8 %offset) nounwind {
 
 define i16 @test_atomic_load_xor_i16(i16 %offset) nounwind {
 ; CHECK: test_atomic_load_xor_i16:
-   %old = atomicrmw xor i16* @var16, i16 %offset seq_cst
-; CHECK: dmb ish
+   %old = atomicrmw xor i16* @var16, i16 %offset release
+; CHECK-NOT: dmb
 ; CHECK: adrp [[TMPADDR:x[0-9]+]], var16
 ; CHECK: add x[[ADDR:[0-9]+]], [[TMPADDR]], #:lo12:var16
 
@@ -357,9 +357,9 @@ define i16 @test_atomic_load_xor_i16(i16 %offset) nounwind {
   ; w0 below is a reasonable guess but could change: it certainly comes into the
   ;  function there.
 ; CHECK-NEXT: eor [[NEW:w[0-9]+]], w[[OLD]], w0
-; CHECK-NEXT: stxrh [[STATUS:w[0-9]+]], [[NEW]], [x[[ADDR]]]
+; CHECK-NEXT: stlxrh [[STATUS:w[0-9]+]], [[NEW]], [x[[ADDR]]]
 ; CHECK-NEXT: cbnz [[STATUS]], .LBB{{[0-9]+}}_1
-; CHECK: dmb ish
+; CHECK-NOT: dmb
 
 ; CHECK: mov x0, x[[OLD]]
    ret i16 %old
@@ -368,18 +368,18 @@ define i16 @test_atomic_load_xor_i16(i16 %offset) nounwind {
 define i32 @test_atomic_load_xor_i32(i32 %offset) nounwind {
 ; CHECK: test_atomic_load_xor_i32:
    %old = atomicrmw xor i32* @var32, i32 %offset seq_cst
-; CHECK: dmb ish
+; CHECK-NOT: dmb
 ; CHECK: adrp [[TMPADDR:x[0-9]+]], var32
 ; CHECK: add x[[ADDR:[0-9]+]], [[TMPADDR]], #:lo12:var32
 
 ; CHECK: .LBB{{[0-9]+}}_1:
-; CHECK-NEXT: ldxr w[[OLD:[0-9]+]], [x[[ADDR]]]
+; CHECK-NEXT: ldaxr w[[OLD:[0-9]+]], [x[[ADDR]]]
   ; w0 below is a reasonable guess but could change: it certainly comes into the
   ;  function there.
 ; CHECK-NEXT: eor [[NEW:w[0-9]+]], w[[OLD]], w0
-; CHECK-NEXT: stxr [[STATUS:w[0-9]+]], [[NEW]], [x[[ADDR]]]
+; CHECK-NEXT: stlxr [[STATUS:w[0-9]+]], [[NEW]], [x[[ADDR]]]
 ; CHECK-NEXT: cbnz [[STATUS]], .LBB{{[0-9]+}}_1
-; CHECK: dmb ish
+; CHECK-NOT: dmb
 
 ; CHECK: mov x0, x[[OLD]]
    ret i32 %old
@@ -387,8 +387,8 @@ define i32 @test_atomic_load_xor_i32(i32 %offset) nounwind {
 
 define i64 @test_atomic_load_xor_i64(i64 %offset) nounwind {
 ; CHECK: test_atomic_load_xor_i64:
-   %old = atomicrmw xor i64* @var64, i64 %offset seq_cst
-; CHECK: dmb ish
+   %old = atomicrmw xor i64* @var64, i64 %offset monotonic
+; CHECK-NOT: dmb
 ; CHECK: adrp [[TMPADDR:x[0-9]+]], var64
 ; CHECK: add x[[ADDR:[0-9]+]], [[TMPADDR]], #:lo12:var64
 
@@ -399,7 +399,7 @@ define i64 @test_atomic_load_xor_i64(i64 %offset) nounwind {
 ; CHECK-NEXT: eor [[NEW:x[0-9]+]], x[[OLD]], x0
 ; CHECK-NEXT: stxr [[STATUS:w[0-9]+]], [[NEW]], [x[[ADDR]]]
 ; CHECK-NEXT: cbnz [[STATUS]], .LBB{{[0-9]+}}_1
-; CHECK: dmb ish
+; CHECK-NOT: dmb
 
 ; CHECK: mov x0, x[[OLD]]
    ret i64 %old
@@ -407,8 +407,8 @@ define i64 @test_atomic_load_xor_i64(i64 %offset) nounwind {
 
 define i8 @test_atomic_load_xchg_i8(i8 %offset) nounwind {
 ; CHECK: test_atomic_load_xchg_i8:
-   %old = atomicrmw xchg i8* @var8, i8 %offset seq_cst
-; CHECK: dmb ish
+   %old = atomicrmw xchg i8* @var8, i8 %offset monotonic
+; CHECK-NOT: dmb
 ; CHECK: adrp [[TMPADDR:x[0-9]+]], var8
 ; CHECK: add x[[ADDR:[0-9]+]], [[TMPADDR]], #:lo12:var8
 
@@ -418,7 +418,7 @@ define i8 @test_atomic_load_xchg_i8(i8 %offset) nounwind {
   ; function there.
 ; CHECK-NEXT: stxrb [[STATUS:w[0-9]+]], w0, [x[[ADDR]]]
 ; CHECK-NEXT: cbnz [[STATUS]], .LBB{{[0-9]+}}_1
-; CHECK: dmb ish
+; CHECK-NOT: dmb
 
 ; CHECK: mov x0, x[[OLD]]
    ret i8 %old
@@ -427,17 +427,17 @@ define i8 @test_atomic_load_xchg_i8(i8 %offset) nounwind {
 define i16 @test_atomic_load_xchg_i16(i16 %offset) nounwind {
 ; CHECK: test_atomic_load_xchg_i16:
    %old = atomicrmw xchg i16* @var16, i16 %offset seq_cst
-; CHECK: dmb ish
+; CHECK-NOT: dmb
 ; CHECK: adrp [[TMPADDR:x[0-9]+]], var16
 ; CHECK: add x[[ADDR:[0-9]+]], [[TMPADDR]], #:lo12:var16
 
 ; CHECK: .LBB{{[0-9]+}}_1:
-; CHECK-NEXT: ldxrh w[[OLD:[0-9]+]], [x[[ADDR]]]
+; CHECK-NEXT: ldaxrh w[[OLD:[0-9]+]], [x[[ADDR]]]
   ; w0 below is a reasonable guess but could change: it certainly comes into the
   ; function there.
-; CHECK-NEXT: stxrh [[STATUS:w[0-9]+]], w0, [x[[ADDR]]]
+; CHECK-NEXT: stlxrh [[STATUS:w[0-9]+]], w0, [x[[ADDR]]]
 ; CHECK-NEXT: cbnz [[STATUS]], .LBB{{[0-9]+}}_1
-; CHECK: dmb ish
+; CHECK-NOT: dmb
 
 ; CHECK: mov x0, x[[OLD]]
    ret i16 %old
@@ -445,8 +445,8 @@ define i16 @test_atomic_load_xchg_i16(i16 %offset) nounwind {
 
 define i32 @test_atomic_load_xchg_i32(i32 %offset) nounwind {
 ; CHECK: test_atomic_load_xchg_i32:
-   %old = atomicrmw xchg i32* @var32, i32 %offset seq_cst
-; CHECK: dmb ish
+   %old = atomicrmw xchg i32* @var32, i32 %offset release
+; CHECK-NOT: dmb
 ; CHECK: adrp [[TMPADDR:x[0-9]+]], var32
 ; CHECK: add x[[ADDR:[0-9]+]], [[TMPADDR]], #:lo12:var32
 
@@ -454,9 +454,9 @@ define i32 @test_atomic_load_xchg_i32(i32 %offset) nounwind {
 ; CHECK-NEXT: ldxr w[[OLD:[0-9]+]], [x[[ADDR]]]
   ; w0 below is a reasonable guess but could change: it certainly comes into the
   ;  function there.
-; CHECK-NEXT: stxr [[STATUS:w[0-9]+]], w0, [x[[ADDR]]]
+; CHECK-NEXT: stlxr [[STATUS:w[0-9]+]], w0, [x[[ADDR]]]
 ; CHECK-NEXT: cbnz [[STATUS]], .LBB{{[0-9]+}}_1
-; CHECK: dmb ish
+; CHECK-NOT: dmb
 
 ; CHECK: mov x0, x[[OLD]]
    ret i32 %old
@@ -464,18 +464,18 @@ define i32 @test_atomic_load_xchg_i32(i32 %offset) nounwind {
 
 define i64 @test_atomic_load_xchg_i64(i64 %offset) nounwind {
 ; CHECK: test_atomic_load_xchg_i64:
-   %old = atomicrmw xchg i64* @var64, i64 %offset seq_cst
-; CHECK: dmb ish
+   %old = atomicrmw xchg i64* @var64, i64 %offset acquire
+; CHECK-NOT: dmb
 ; CHECK: adrp [[TMPADDR:x[0-9]+]], var64
 ; CHECK: add x[[ADDR:[0-9]+]], [[TMPADDR]], #:lo12:var64
 
 ; CHECK: .LBB{{[0-9]+}}_1:
-; CHECK-NEXT: ldxr x[[OLD:[0-9]+]], [x[[ADDR]]]
+; CHECK-NEXT: ldaxr x[[OLD:[0-9]+]], [x[[ADDR]]]
   ; x0 below is a reasonable guess but could change: it certainly comes into the
   ; function there.
 ; CHECK-NEXT: stxr [[STATUS:w[0-9]+]], x0, [x[[ADDR]]]
 ; CHECK-NEXT: cbnz [[STATUS]], .LBB{{[0-9]+}}_1
-; CHECK: dmb ish
+; CHECK-NOT: dmb
 
 ; CHECK: mov x0, x[[OLD]]
    ret i64 %old
@@ -484,20 +484,20 @@ define i64 @test_atomic_load_xchg_i64(i64 %offset) nounwind {
 
 define i8 @test_atomic_load_min_i8(i8 %offset) nounwind {
 ; CHECK: test_atomic_load_min_i8:
-   %old = atomicrmw min i8* @var8, i8 %offset seq_cst
-; CHECK: dmb ish
+   %old = atomicrmw min i8* @var8, i8 %offset acquire
+; CHECK-NOT: dmb
 ; CHECK: adrp [[TMPADDR:x[0-9]+]], var8
 ; CHECK: add x[[ADDR:[0-9]+]], [[TMPADDR]], #:lo12:var8
 
 ; CHECK: .LBB{{[0-9]+}}_1:
-; CHECK-NEXT: ldxrb w[[OLD:[0-9]+]], [x[[ADDR]]]
+; CHECK-NEXT: ldaxrb w[[OLD:[0-9]+]], [x[[ADDR]]]
   ; w0 below is a reasonable guess but could change: it certainly comes into the
   ;  function there.
 ; CHECK-NEXT: cmp w0, w[[OLD]], sxtb
 ; CHECK-NEXT: csel [[NEW:w[0-9]+]], w[[OLD]], w0, gt
 ; CHECK-NEXT: stxrb [[STATUS:w[0-9]+]], [[NEW]], [x[[ADDR]]]
 ; CHECK-NEXT: cbnz [[STATUS]], .LBB{{[0-9]+}}_1
-; CHECK: dmb ish
+; CHECK-NOT: dmb
 
 ; CHECK: mov x0, x[[OLD]]
    ret i8 %old
@@ -505,8 +505,8 @@ define i8 @test_atomic_load_min_i8(i8 %offset) nounwind {
 
 define i16 @test_atomic_load_min_i16(i16 %offset) nounwind {
 ; CHECK: test_atomic_load_min_i16:
-   %old = atomicrmw min i16* @var16, i16 %offset seq_cst
-; CHECK: dmb ish
+   %old = atomicrmw min i16* @var16, i16 %offset release
+; CHECK-NOT: dmb
 ; CHECK: adrp [[TMPADDR:x[0-9]+]], var16
 ; CHECK: add x[[ADDR:[0-9]+]], [[TMPADDR]], #:lo12:var16
 
@@ -516,9 +516,9 @@ define i16 @test_atomic_load_min_i16(i16 %offset) nounwind {
   ;  function there.
 ; CHECK-NEXT: cmp w0, w[[OLD]], sxth
 ; CHECK-NEXT: csel [[NEW:w[0-9]+]], w[[OLD]], w0, gt
-; CHECK-NEXT: stxrh [[STATUS:w[0-9]+]], [[NEW]], [x[[ADDR]]]
+; CHECK-NEXT: stlxrh [[STATUS:w[0-9]+]], [[NEW]], [x[[ADDR]]]
 ; CHECK-NEXT: cbnz [[STATUS]], .LBB{{[0-9]+}}_1
-; CHECK: dmb ish
+; CHECK-NOT: dmb
 
 ; CHECK: mov x0, x[[OLD]]
    ret i16 %old
@@ -526,8 +526,8 @@ define i16 @test_atomic_load_min_i16(i16 %offset) nounwind {
 
 define i32 @test_atomic_load_min_i32(i32 %offset) nounwind {
 ; CHECK: test_atomic_load_min_i32:
-   %old = atomicrmw min i32* @var32, i32 %offset seq_cst
-; CHECK: dmb ish
+   %old = atomicrmw min i32* @var32, i32 %offset monotonic
+; CHECK-NOT: dmb
 ; CHECK: adrp [[TMPADDR:x[0-9]+]], var32
 ; CHECK: add x[[ADDR:[0-9]+]], [[TMPADDR]], #:lo12:var32
 
@@ -539,7 +539,7 @@ define i32 @test_atomic_load_min_i32(i32 %offset) nounwind {
 ; CHECK-NEXT: csel [[NEW:w[0-9]+]], w[[OLD]], w0, gt
 ; CHECK-NEXT: stxr [[STATUS:w[0-9]+]], [[NEW]], [x[[ADDR]]]
 ; CHECK-NEXT: cbnz [[STATUS]], .LBB{{[0-9]+}}_1
-; CHECK: dmb ish
+; CHECK-NOT: dmb
 
 ; CHECK: mov x0, x[[OLD]]
    ret i32 %old
@@ -548,19 +548,19 @@ define i32 @test_atomic_load_min_i32(i32 %offset) nounwind {
 define i64 @test_atomic_load_min_i64(i64 %offset) nounwind {
 ; CHECK: test_atomic_load_min_i64:
    %old = atomicrmw min i64* @var64, i64 %offset seq_cst
-; CHECK: dmb ish
+; CHECK-NOT: dmb
 ; CHECK: adrp [[TMPADDR:x[0-9]+]], var64
 ; CHECK: add x[[ADDR:[0-9]+]], [[TMPADDR]], #:lo12:var64
 
 ; CHECK: .LBB{{[0-9]+}}_1:
-; CHECK-NEXT: ldxr x[[OLD:[0-9]+]], [x[[ADDR]]]
+; CHECK-NEXT: ldaxr x[[OLD:[0-9]+]], [x[[ADDR]]]
   ; x0 below is a reasonable guess but could change: it certainly comes into the
   ; function there.
 ; CHECK-NEXT: cmp x0, x[[OLD]]
 ; CHECK-NEXT: csel [[NEW:x[0-9]+]], x[[OLD]], x0, gt
-; CHECK-NEXT: stxr [[STATUS:w[0-9]+]], [[NEW]], [x[[ADDR]]]
+; CHECK-NEXT: stlxr [[STATUS:w[0-9]+]], [[NEW]], [x[[ADDR]]]
 ; CHECK-NEXT: cbnz [[STATUS]], .LBB{{[0-9]+}}_1
-; CHECK: dmb ish
+; CHECK-NOT: dmb
 
 ; CHECK: mov x0, x[[OLD]]
    ret i64 %old
@@ -569,19 +569,19 @@ define i64 @test_atomic_load_min_i64(i64 %offset) nounwind {
 define i8 @test_atomic_load_max_i8(i8 %offset) nounwind {
 ; CHECK: test_atomic_load_max_i8:
    %old = atomicrmw max i8* @var8, i8 %offset seq_cst
-; CHECK: dmb ish
+; CHECK-NOT: dmb
 ; CHECK: adrp [[TMPADDR:x[0-9]+]], var8
 ; CHECK: add x[[ADDR:[0-9]+]], [[TMPADDR]], #:lo12:var8
 
 ; CHECK: .LBB{{[0-9]+}}_1:
-; CHECK-NEXT: ldxrb w[[OLD:[0-9]+]], [x[[ADDR]]]
+; CHECK-NEXT: ldaxrb w[[OLD:[0-9]+]], [x[[ADDR]]]
   ; w0 below is a reasonable guess but could change: it certainly comes into the
   ;  function there.
 ; CHECK-NEXT: cmp w0, w[[OLD]], sxtb
 ; CHECK-NEXT: csel [[NEW:w[0-9]+]], w[[OLD]], w0, lt
-; CHECK-NEXT: stxrb [[STATUS:w[0-9]+]], [[NEW]], [x[[ADDR]]]
+; CHECK-NEXT: stlxrb [[STATUS:w[0-9]+]], [[NEW]], [x[[ADDR]]]
 ; CHECK-NEXT: cbnz [[STATUS]], .LBB{{[0-9]+}}_1
-; CHECK: dmb ish
+; CHECK-NOT: dmb
 
 ; CHECK: mov x0, x[[OLD]]
    ret i8 %old
@@ -589,20 +589,20 @@ define i8 @test_atomic_load_max_i8(i8 %offset) nounwind {
 
 define i16 @test_atomic_load_max_i16(i16 %offset) nounwind {
 ; CHECK: test_atomic_load_max_i16:
-   %old = atomicrmw max i16* @var16, i16 %offset seq_cst
-; CHECK: dmb ish
+   %old = atomicrmw max i16* @var16, i16 %offset acquire
+; CHECK-NOT: dmb
 ; CHECK: adrp [[TMPADDR:x[0-9]+]], var16
 ; CHECK: add x[[ADDR:[0-9]+]], [[TMPADDR]], #:lo12:var16
 
 ; CHECK: .LBB{{[0-9]+}}_1:
-; CHECK-NEXT: ldxrh w[[OLD:[0-9]+]], [x[[ADDR]]]
+; CHECK-NEXT: ldaxrh w[[OLD:[0-9]+]], [x[[ADDR]]]
   ; w0 below is a reasonable guess but could change: it certainly comes into the
   ;  function there.
 ; CHECK-NEXT: cmp w0, w[[OLD]], sxth
 ; CHECK-NEXT: csel [[NEW:w[0-9]+]], w[[OLD]], w0, lt
 ; CHECK-NEXT: stxrh [[STATUS:w[0-9]+]], [[NEW]], [x[[ADDR]]]
 ; CHECK-NEXT: cbnz [[STATUS]], .LBB{{[0-9]+}}_1
-; CHECK: dmb ish
+; CHECK-NOT: dmb
 
 ; CHECK: mov x0, x[[OLD]]
    ret i16 %old
@@ -610,8 +610,8 @@ define i16 @test_atomic_load_max_i16(i16 %offset) nounwind {
 
 define i32 @test_atomic_load_max_i32(i32 %offset) nounwind {
 ; CHECK: test_atomic_load_max_i32:
-   %old = atomicrmw max i32* @var32, i32 %offset seq_cst
-; CHECK: dmb ish
+   %old = atomicrmw max i32* @var32, i32 %offset release
+; CHECK-NOT: dmb
 ; CHECK: adrp [[TMPADDR:x[0-9]+]], var32
 ; CHECK: add x[[ADDR:[0-9]+]], [[TMPADDR]], #:lo12:var32
 
@@ -621,9 +621,9 @@ define i32 @test_atomic_load_max_i32(i32 %offset) nounwind {
   ;  function there.
 ; CHECK-NEXT: cmp w0, w[[OLD]]
 ; CHECK-NEXT: csel [[NEW:w[0-9]+]], w[[OLD]], w0, lt
-; CHECK-NEXT: stxr [[STATUS:w[0-9]+]], [[NEW]], [x[[ADDR]]]
+; CHECK-NEXT: stlxr [[STATUS:w[0-9]+]], [[NEW]], [x[[ADDR]]]
 ; CHECK-NEXT: cbnz [[STATUS]], .LBB{{[0-9]+}}_1
-; CHECK: dmb ish
+; CHECK-NOT: dmb
 
 ; CHECK: mov x0, x[[OLD]]
    ret i32 %old
@@ -631,8 +631,8 @@ define i32 @test_atomic_load_max_i32(i32 %offset) nounwind {
 
 define i64 @test_atomic_load_max_i64(i64 %offset) nounwind {
 ; CHECK: test_atomic_load_max_i64:
-   %old = atomicrmw max i64* @var64, i64 %offset seq_cst
-; CHECK: dmb ish
+   %old = atomicrmw max i64* @var64, i64 %offset monotonic
+; CHECK-NOT: dmb
 ; CHECK: adrp [[TMPADDR:x[0-9]+]], var64
 ; CHECK: add x[[ADDR:[0-9]+]], [[TMPADDR]], #:lo12:var64
 
@@ -644,7 +644,7 @@ define i64 @test_atomic_load_max_i64(i64 %offset) nounwind {
 ; CHECK-NEXT: csel [[NEW:x[0-9]+]], x[[OLD]], x0, lt
 ; CHECK-NEXT: stxr [[STATUS:w[0-9]+]], [[NEW]], [x[[ADDR]]]
 ; CHECK-NEXT: cbnz [[STATUS]], .LBB{{[0-9]+}}_1
-; CHECK: dmb ish
+; CHECK-NOT: dmb
 
 ; CHECK: mov x0, x[[OLD]]
    ret i64 %old
@@ -652,8 +652,8 @@ define i64 @test_atomic_load_max_i64(i64 %offset) nounwind {
 
 define i8 @test_atomic_load_umin_i8(i8 %offset) nounwind {
 ; CHECK: test_atomic_load_umin_i8:
-   %old = atomicrmw umin i8* @var8, i8 %offset seq_cst
-; CHECK: dmb ish
+   %old = atomicrmw umin i8* @var8, i8 %offset monotonic
+; CHECK-NOT: dmb
 ; CHECK: adrp [[TMPADDR:x[0-9]+]], var8
 ; CHECK: add x[[ADDR:[0-9]+]], [[TMPADDR]], #:lo12:var8
 
@@ -665,7 +665,7 @@ define i8 @test_atomic_load_umin_i8(i8 %offset) nounwind {
 ; CHECK-NEXT: csel [[NEW:w[0-9]+]], w[[OLD]], w0, hi
 ; CHECK-NEXT: stxrb [[STATUS:w[0-9]+]], [[NEW]], [x[[ADDR]]]
 ; CHECK-NEXT: cbnz [[STATUS]], .LBB{{[0-9]+}}_1
-; CHECK: dmb ish
+; CHECK-NOT: dmb
 
 ; CHECK: mov x0, x[[OLD]]
    ret i8 %old
@@ -673,20 +673,20 @@ define i8 @test_atomic_load_umin_i8(i8 %offset) nounwind {
 
 define i16 @test_atomic_load_umin_i16(i16 %offset) nounwind {
 ; CHECK: test_atomic_load_umin_i16:
-   %old = atomicrmw umin i16* @var16, i16 %offset seq_cst
-; CHECK: dmb ish
+   %old = atomicrmw umin i16* @var16, i16 %offset acquire
+; CHECK-NOT: dmb
 ; CHECK: adrp [[TMPADDR:x[0-9]+]], var16
 ; CHECK: add x[[ADDR:[0-9]+]], [[TMPADDR]], #:lo12:var16
 
 ; CHECK: .LBB{{[0-9]+}}_1:
-; CHECK-NEXT: ldxrh w[[OLD:[0-9]+]], [x[[ADDR]]]
+; CHECK-NEXT: ldaxrh w[[OLD:[0-9]+]], [x[[ADDR]]]
   ; w0 below is a reasonable guess but could change: it certainly comes into the
   ;  function there.
 ; CHECK-NEXT: cmp w0, w[[OLD]], uxth
 ; CHECK-NEXT: csel [[NEW:w[0-9]+]], w[[OLD]], w0, hi
 ; CHECK-NEXT: stxrh [[STATUS:w[0-9]+]], [[NEW]], [x[[ADDR]]]
 ; CHECK-NEXT: cbnz [[STATUS]], .LBB{{[0-9]+}}_1
-; CHECK: dmb ish
+; CHECK-NOT: dmb
 
 ; CHECK: mov x0, x[[OLD]]
    ret i16 %old
@@ -695,19 +695,19 @@ define i16 @test_atomic_load_umin_i16(i16 %offset) nounwind {
 define i32 @test_atomic_load_umin_i32(i32 %offset) nounwind {
 ; CHECK: test_atomic_load_umin_i32:
    %old = atomicrmw umin i32* @var32, i32 %offset seq_cst
-; CHECK: dmb ish
+; CHECK-NOT: dmb
 ; CHECK: adrp [[TMPADDR:x[0-9]+]], var32
 ; CHECK: add x[[ADDR:[0-9]+]], [[TMPADDR]], #:lo12:var32
 
 ; CHECK: .LBB{{[0-9]+}}_1:
-; CHECK-NEXT: ldxr w[[OLD:[0-9]+]], [x[[ADDR]]]
+; CHECK-NEXT: ldaxr w[[OLD:[0-9]+]], [x[[ADDR]]]
   ; w0 below is a reasonable guess but could change: it certainly comes into the
   ;  function there.
 ; CHECK-NEXT: cmp w0, w[[OLD]]
 ; CHECK-NEXT: csel [[NEW:w[0-9]+]], w[[OLD]], w0, hi
-; CHECK-NEXT: stxr [[STATUS:w[0-9]+]], [[NEW]], [x[[ADDR]]]
+; CHECK-NEXT: stlxr [[STATUS:w[0-9]+]], [[NEW]], [x[[ADDR]]]
 ; CHECK-NEXT: cbnz [[STATUS]], .LBB{{[0-9]+}}_1
-; CHECK: dmb ish
+; CHECK-NOT: dmb
 
 ; CHECK: mov x0, x[[OLD]]
    ret i32 %old
@@ -715,20 +715,20 @@ define i32 @test_atomic_load_umin_i32(i32 %offset) nounwind {
 
 define i64 @test_atomic_load_umin_i64(i64 %offset) nounwind {
 ; CHECK: test_atomic_load_umin_i64:
-   %old = atomicrmw umin i64* @var64, i64 %offset seq_cst
-; CHECK: dmb ish
+   %old = atomicrmw umin i64* @var64, i64 %offset acq_rel
+; CHECK-NOT: dmb
 ; CHECK: adrp [[TMPADDR:x[0-9]+]], var64
 ; CHECK: add x[[ADDR:[0-9]+]], [[TMPADDR]], #:lo12:var64
 
 ; CHECK: .LBB{{[0-9]+}}_1:
-; CHECK-NEXT: ldxr x[[OLD:[0-9]+]], [x[[ADDR]]]
+; CHECK-NEXT: ldaxr x[[OLD:[0-9]+]], [x[[ADDR]]]
   ; x0 below is a reasonable guess but could change: it certainly comes into the
   ; function there.
 ; CHECK-NEXT: cmp x0, x[[OLD]]
 ; CHECK-NEXT: csel [[NEW:x[0-9]+]], x[[OLD]], x0, hi
-; CHECK-NEXT: stxr [[STATUS:w[0-9]+]], [[NEW]], [x[[ADDR]]]
+; CHECK-NEXT: stlxr [[STATUS:w[0-9]+]], [[NEW]], [x[[ADDR]]]
 ; CHECK-NEXT: cbnz [[STATUS]], .LBB{{[0-9]+}}_1
-; CHECK: dmb ish
+; CHECK-NOT: dmb
 
 ; CHECK: mov x0, x[[OLD]]
    ret i64 %old
@@ -736,20 +736,20 @@ define i64 @test_atomic_load_umin_i64(i64 %offset) nounwind {
 
 define i8 @test_atomic_load_umax_i8(i8 %offset) nounwind {
 ; CHECK: test_atomic_load_umax_i8:
-   %old = atomicrmw umax i8* @var8, i8 %offset seq_cst
-; CHECK: dmb ish
+   %old = atomicrmw umax i8* @var8, i8 %offset acq_rel
+; CHECK-NOT: dmb
 ; CHECK: adrp [[TMPADDR:x[0-9]+]], var8
 ; CHECK: add x[[ADDR:[0-9]+]], [[TMPADDR]], #:lo12:var8
 
 ; CHECK: .LBB{{[0-9]+}}_1:
-; CHECK-NEXT: ldxrb w[[OLD:[0-9]+]], [x[[ADDR]]]
+; CHECK-NEXT: ldaxrb w[[OLD:[0-9]+]], [x[[ADDR]]]
   ; w0 below is a reasonable guess but could change: it certainly comes into the
   ;  function there.
 ; CHECK-NEXT: cmp w0, w[[OLD]], uxtb
 ; CHECK-NEXT: csel [[NEW:w[0-9]+]], w[[OLD]], w0, lo
-; CHECK-NEXT: stxrb [[STATUS:w[0-9]+]], [[NEW]], [x[[ADDR]]]
+; CHECK-NEXT: stlxrb [[STATUS:w[0-9]+]], [[NEW]], [x[[ADDR]]]
 ; CHECK-NEXT: cbnz [[STATUS]], .LBB{{[0-9]+}}_1
-; CHECK: dmb ish
+; CHECK-NOT: dmb
 
 ; CHECK: mov x0, x[[OLD]]
    ret i8 %old
@@ -757,8 +757,8 @@ define i8 @test_atomic_load_umax_i8(i8 %offset) nounwind {
 
 define i16 @test_atomic_load_umax_i16(i16 %offset) nounwind {
 ; CHECK: test_atomic_load_umax_i16:
-   %old = atomicrmw umax i16* @var16, i16 %offset seq_cst
-; CHECK: dmb ish
+   %old = atomicrmw umax i16* @var16, i16 %offset monotonic
+; CHECK-NOT: dmb
 ; CHECK: adrp [[TMPADDR:x[0-9]+]], var16
 ; CHECK: add x[[ADDR:[0-9]+]], [[TMPADDR]], #:lo12:var16
 
@@ -770,7 +770,7 @@ define i16 @test_atomic_load_umax_i16(i16 %offset) nounwind {
 ; CHECK-NEXT: csel [[NEW:w[0-9]+]], w[[OLD]], w0, lo
 ; CHECK-NEXT: stxrh [[STATUS:w[0-9]+]], [[NEW]], [x[[ADDR]]]
 ; CHECK-NEXT: cbnz [[STATUS]], .LBB{{[0-9]+}}_1
-; CHECK: dmb ish
+; CHECK-NOT: dmb
 
 ; CHECK: mov x0, x[[OLD]]
    ret i16 %old
@@ -779,19 +779,19 @@ define i16 @test_atomic_load_umax_i16(i16 %offset) nounwind {
 define i32 @test_atomic_load_umax_i32(i32 %offset) nounwind {
 ; CHECK: test_atomic_load_umax_i32:
    %old = atomicrmw umax i32* @var32, i32 %offset seq_cst
-; CHECK: dmb ish
+; CHECK-NOT: dmb
 ; CHECK: adrp [[TMPADDR:x[0-9]+]], var32
 ; CHECK: add x[[ADDR:[0-9]+]], [[TMPADDR]], #:lo12:var32
 
 ; CHECK: .LBB{{[0-9]+}}_1:
-; CHECK-NEXT: ldxr w[[OLD:[0-9]+]], [x[[ADDR]]]
+; CHECK-NEXT: ldaxr w[[OLD:[0-9]+]], [x[[ADDR]]]
   ; w0 below is a reasonable guess but could change: it certainly comes into the
   ;  function there.
 ; CHECK-NEXT: cmp w0, w[[OLD]]
 ; CHECK-NEXT: csel [[NEW:w[0-9]+]], w[[OLD]], w0, lo
-; CHECK-NEXT: stxr [[STATUS:w[0-9]+]], [[NEW]], [x[[ADDR]]]
+; CHECK-NEXT: stlxr [[STATUS:w[0-9]+]], [[NEW]], [x[[ADDR]]]
 ; CHECK-NEXT: cbnz [[STATUS]], .LBB{{[0-9]+}}_1
-; CHECK: dmb ish
+; CHECK-NOT: dmb
 
 ; CHECK: mov x0, x[[OLD]]
    ret i32 %old
@@ -799,8 +799,8 @@ define i32 @test_atomic_load_umax_i32(i32 %offset) nounwind {
 
 define i64 @test_atomic_load_umax_i64(i64 %offset) nounwind {
 ; CHECK: test_atomic_load_umax_i64:
-   %old = atomicrmw umax i64* @var64, i64 %offset seq_cst
-; CHECK: dmb ish
+   %old = atomicrmw umax i64* @var64, i64 %offset release
+; CHECK-NOT: dmb
 ; CHECK: adrp [[TMPADDR:x[0-9]+]], var64
 ; CHECK: add x[[ADDR:[0-9]+]], [[TMPADDR]], #:lo12:var64
 
@@ -810,9 +810,9 @@ define i64 @test_atomic_load_umax_i64(i64 %offset) nounwind {
   ; function there.
 ; CHECK-NEXT: cmp x0, x[[OLD]]
 ; CHECK-NEXT: csel [[NEW:x[0-9]+]], x[[OLD]], x0, lo
-; CHECK-NEXT: stxr [[STATUS:w[0-9]+]], [[NEW]], [x[[ADDR]]]
+; CHECK-NEXT: stlxr [[STATUS:w[0-9]+]], [[NEW]], [x[[ADDR]]]
 ; CHECK-NEXT: cbnz [[STATUS]], .LBB{{[0-9]+}}_1
-; CHECK: dmb ish
+; CHECK-NOT: dmb
 
 ; CHECK: mov x0, x[[OLD]]
    ret i64 %old
@@ -820,13 +820,13 @@ define i64 @test_atomic_load_umax_i64(i64 %offset) nounwind {
 
 define i8 @test_atomic_cmpxchg_i8(i8 %wanted, i8 %new) nounwind {
 ; CHECK: test_atomic_cmpxchg_i8:
-   %old = cmpxchg i8* @var8, i8 %wanted, i8 %new seq_cst
-; CHECK: dmb ish
+   %old = cmpxchg i8* @var8, i8 %wanted, i8 %new acquire
+; CHECK-NOT: dmb
 ; CHECK: adrp [[TMPADDR:x[0-9]+]], var8
 ; CHECK: add x[[ADDR:[0-9]+]], [[TMPADDR]], #:lo12:var8
 
 ; CHECK: [[STARTAGAIN:.LBB[0-9]+_[0-9]+]]:
-; CHECK-NEXT: ldxrb w[[OLD:[0-9]+]], [x[[ADDR]]]
+; CHECK-NEXT: ldaxrb w[[OLD:[0-9]+]], [x[[ADDR]]]
   ; w0 below is a reasonable guess but could change: it certainly comes into the
   ;  function there.
 ; CHECK-NEXT: cmp w[[OLD]], w0
@@ -834,7 +834,7 @@ define i8 @test_atomic_cmpxchg_i8(i8 %wanted, i8 %new) nounwind {
   ; As above, w1 is a reasonable guess.
 ; CHECK: stxrb [[STATUS:w[0-9]+]], w1, [x[[ADDR]]]
 ; CHECK-NEXT: cbnz [[STATUS]], [[STARTAGAIN]]
-; CHECK: dmb ish
+; CHECK-NOT: dmb
 
 ; CHECK: mov x0, x[[OLD]]
    ret i8 %old
@@ -843,20 +843,20 @@ define i8 @test_atomic_cmpxchg_i8(i8 %wanted, i8 %new) nounwind {
 define i16 @test_atomic_cmpxchg_i16(i16 %wanted, i16 %new) nounwind {
 ; CHECK: test_atomic_cmpxchg_i16:
    %old = cmpxchg i16* @var16, i16 %wanted, i16 %new seq_cst
-; CHECK: dmb ish
+; CHECK-NOT: dmb
 ; CHECK: adrp [[TMPADDR:x[0-9]+]], var16
 ; CHECK: add x[[ADDR:[0-9]+]], [[TMPADDR]], #:lo12:var16
 
 ; CHECK: [[STARTAGAIN:.LBB[0-9]+_[0-9]+]]:
-; CHECK-NEXT: ldxrh w[[OLD:[0-9]+]], [x[[ADDR]]]
+; CHECK-NEXT: ldaxrh w[[OLD:[0-9]+]], [x[[ADDR]]]
   ; w0 below is a reasonable guess but could change: it certainly comes into the
   ;  function there.
 ; CHECK-NEXT: cmp w[[OLD]], w0
 ; CHECK-NEXT: b.ne [[GET_OUT:.LBB[0-9]+_[0-9]+]]
   ; As above, w1 is a reasonable guess.
-; CHECK: stxrh [[STATUS:w[0-9]+]], w1, [x[[ADDR]]]
+; CHECK: stlxrh [[STATUS:w[0-9]+]], w1, [x[[ADDR]]]
 ; CHECK-NEXT: cbnz [[STATUS]], [[STARTAGAIN]]
-; CHECK: dmb ish
+; CHECK-NOT: dmb
 
 ; CHECK: mov x0, x[[OLD]]
    ret i16 %old
@@ -864,8 +864,8 @@ define i16 @test_atomic_cmpxchg_i16(i16 %wanted, i16 %new) nounwind {
 
 define i32 @test_atomic_cmpxchg_i32(i32 %wanted, i32 %new) nounwind {
 ; CHECK: test_atomic_cmpxchg_i32:
-   %old = cmpxchg i32* @var32, i32 %wanted, i32 %new seq_cst
-; CHECK: dmb ish
+   %old = cmpxchg i32* @var32, i32 %wanted, i32 %new release
+; CHECK-NOT: dmb
 ; CHECK: adrp [[TMPADDR:x[0-9]+]], var32
 ; CHECK: add x[[ADDR:[0-9]+]], [[TMPADDR]], #:lo12:var32
 
@@ -876,9 +876,9 @@ define i32 @test_atomic_cmpxchg_i32(i32 %wanted, i32 %new) nounwind {
 ; CHECK-NEXT: cmp w[[OLD]], w0
 ; CHECK-NEXT: b.ne [[GET_OUT:.LBB[0-9]+_[0-9]+]]
   ; As above, w1 is a reasonable guess.
-; CHECK: stxr [[STATUS:w[0-9]+]], w1, [x[[ADDR]]]
+; CHECK: stlxr [[STATUS:w[0-9]+]], w1, [x[[ADDR]]]
 ; CHECK-NEXT: cbnz [[STATUS]], [[STARTAGAIN]]
-; CHECK: dmb ish
+; CHECK-NOT: dmb
 
 ; CHECK: mov x0, x[[OLD]]
    ret i32 %old
@@ -886,8 +886,8 @@ define i32 @test_atomic_cmpxchg_i32(i32 %wanted, i32 %new) nounwind {
 
 define i64 @test_atomic_cmpxchg_i64(i64 %wanted, i64 %new) nounwind {
 ; CHECK: test_atomic_cmpxchg_i64:
-   %old = cmpxchg i64* @var64, i64 %wanted, i64 %new seq_cst
-; CHECK: dmb ish
+   %old = cmpxchg i64* @var64, i64 %wanted, i64 %new monotonic
+; CHECK-NOT: dmb
 ; CHECK: adrp [[TMPADDR:x[0-9]+]], var64
 ; CHECK: add x[[ADDR:[0-9]+]], [[TMPADDR]], #:lo12:var64
 
@@ -900,7 +900,7 @@ define i64 @test_atomic_cmpxchg_i64(i64 %wanted, i64 %new) nounwind {
   ; As above, w1 is a reasonable guess.
 ; CHECK: stxr [[STATUS:w[0-9]+]], x1, [x[[ADDR]]]
 ; CHECK-NEXT: cbnz [[STATUS]], [[STARTAGAIN]]
-; CHECK: dmb ish
+; CHECK-NOT: dmb
 
 ; CHECK: mov x0, x[[OLD]]
    ret i64 %old
@@ -933,19 +933,26 @@ define i8 @test_atomic_load_monotonic_regoff_i8(i64 %base, i64 %off) nounwind {
 define i8 @test_atomic_load_acquire_i8() nounwind {
 ; CHECK: test_atomic_load_acquire_i8:
   %val = load atomic i8* @var8 acquire, align 1
+; CHECK-NOT: dmb
 ; CHECK: adrp [[TMPADDR:x[0-9]+]], var8
+; CHECK-NOT: dmb
 ; CHECK: add x[[ADDR:[0-9]+]], [[TMPADDR]], #:lo12:var8
-
+; CHECK-NOT: dmb
 ; CHECK: ldarb w0, [x[[ADDR]]]
+; CHECK-NOT: dmb
   ret i8 %val
 }
 
 define i8 @test_atomic_load_seq_cst_i8() nounwind {
 ; CHECK: test_atomic_load_seq_cst_i8:
   %val = load atomic i8* @var8 seq_cst, align 1
-; CHECK: adrp x[[HIADDR:[0-9]+]], var8
-; CHECK: ldrb w0, [x[[HIADDR]], #:lo12:var8]
-; CHECK: dmb ish
+; CHECK-NOT: dmb
+; CHECK: adrp [[HIADDR:x[0-9]+]], var8
+; CHECK-NOT: dmb
+; CHECK: add x[[ADDR:[0-9]+]], [[HIADDR]], #:lo12:var8
+; CHECK-NOT: dmb
+; CHECK: ldarb w0, [x[[ADDR]]]
+; CHECK-NOT: dmb
   ret i8 %val
 }
 
@@ -954,6 +961,7 @@ define i16 @test_atomic_load_monotonic_i16() nounwind {
   %val = load atomic i16* @var16 monotonic, align 2
 ; CHECK-NOT: dmb
 ; CHECK: adrp x[[HIADDR:[0-9]+]], var16
+; CHECK-NOT: dmb
 ; CHECK: ldrh w0, [x[[HIADDR]], #:lo12:var16]
 ; CHECK-NOT: dmb
 
@@ -976,9 +984,13 @@ define i32 @test_atomic_load_monotonic_regoff_i32(i64 %base, i64 %off) nounwind
 define i64 @test_atomic_load_seq_cst_i64() nounwind {
 ; CHECK: test_atomic_load_seq_cst_i64:
   %val = load atomic i64* @var64 seq_cst, align 8
-; CHECK: adrp x[[HIADDR:[0-9]+]], var64
-; CHECK: ldr x0, [x[[HIADDR]], #:lo12:var64]
-; CHECK: dmb ish
+; CHECK-NOT: dmb
+; CHECK: adrp [[HIADDR:x[0-9]+]], var64
+; CHECK-NOT: dmb
+; CHECK: add x[[ADDR:[0-9]+]], [[HIADDR]], #:lo12:var64
+; CHECK-NOT: dmb
+; CHECK: ldar x0, [x[[ADDR]]]
+; CHECK-NOT: dmb
   ret i64 %val
 }
 
@@ -1005,20 +1017,26 @@ define void @test_atomic_store_monotonic_regoff_i8(i64 %base, i64 %off, i8 %val)
 define void @test_atomic_store_release_i8(i8 %val) nounwind {
 ; CHECK: test_atomic_store_release_i8:
   store atomic i8 %val, i8* @var8 release, align 1
+; CHECK-NOT: dmb
 ; CHECK: adrp [[HIADDR:x[0-9]+]], var8
+; CHECK-NOT: dmb
 ; CHECK: add x[[ADDR:[0-9]+]], [[HIADDR]], #:lo12:var8
+; CHECK-NOT: dmb
 ; CHECK: stlrb w0, [x[[ADDR]]]
-
+; CHECK-NOT: dmb
   ret void
 }
 
 define void @test_atomic_store_seq_cst_i8(i8 %val) nounwind {
 ; CHECK: test_atomic_store_seq_cst_i8:
   store atomic i8 %val, i8* @var8 seq_cst, align 1
+; CHECK-NOT: dmb
 ; CHECK: adrp [[HIADDR:x[0-9]+]], var8
+; CHECK-NOT: dmb
 ; CHECK: add x[[ADDR:[0-9]+]], [[HIADDR]], #:lo12:var8
+; CHECK-NOT: dmb
 ; CHECK: stlrb w0, [x[[ADDR]]]
-; CHECK: dmb ish
+; CHECK-NOT: dmb
 
   ret void
 }
@@ -1026,9 +1044,11 @@ define void @test_atomic_store_seq_cst_i8(i8 %val) nounwind {
 define void @test_atomic_store_monotonic_i16(i16 %val) nounwind {
 ; CHECK: test_atomic_store_monotonic_i16:
   store atomic i16 %val, i16* @var16 monotonic, align 2
+; CHECK-NOT: dmb
 ; CHECK: adrp x[[HIADDR:[0-9]+]], var16
+; CHECK-NOT: dmb
 ; CHECK: strh w0, [x[[HIADDR]], #:lo12:var16]
-
+; CHECK-NOT: dmb
   ret void
 }
 
@@ -1039,7 +1059,9 @@ define void @test_atomic_store_monotonic_regoff_i32(i64 %base, i64 %off, i32 %va
   %addr = inttoptr i64 %addr_int to i32*
 
   store atomic i32 %val, i32* %addr monotonic, align 4
+; CHECK-NOT: dmb
 ; CHECK: str w2, [x0, x1]
+; CHECK-NOT: dmb
 
   ret void
 }
@@ -1047,9 +1069,12 @@ define void @test_atomic_store_monotonic_regoff_i32(i64 %base, i64 %off, i32 %va
 define void @test_atomic_store_release_i64(i64 %val) nounwind {
 ; CHECK: test_atomic_store_release_i64:
   store atomic i64 %val, i64* @var64 release, align 8
+; CHECK-NOT: dmb
 ; CHECK: adrp [[HIADDR:x[0-9]+]], var64
+; CHECK-NOT: dmb
 ; CHECK: add x[[ADDR:[0-9]+]], [[HIADDR]], #:lo12:var64
+; CHECK-NOT: dmb
 ; CHECK: stlr x0, [x[[ADDR]]]
-
+; CHECK-NOT: dmb
   ret void
 }