[NVPTX][FIX] Expand atomics we cannot handle natively in the ISA

[jdoerfert]

Summary

NOTE: This is lacking a test and more of a request for feedback (I'm not an NVPTX person).

See also PR4219.

[arsenm]

Needs tests. The AMDGPU ones in test/Transforms/AtomicExpand can probably be copied as-is (plus another codegen one to make sure AtomicExpand is actually running)

[merge_guards_bot]

Build result: pass - 60562 tests passed, 0 failed and 726 were skipped.

Log files: console-log.txt, CMakeCache.txt

[tra]

+1 for tests.

llvm/lib/Target/NVPTX/NVPTXISelLowering.cpp

360

Typo: it's *4*4219: https://bugs.llvm.org/show_bug.cgi?id=44219

[jdoerfert]

Add tests and run pass in the pipeline

[jdoerfert]

Fixed the typo, copied test/Transforms/AtomicExpand/AMDGPU into NVPTX and changed the run lines accordingly. Then I run the update_test_checks. The result is different than before (some expansion happens), and close to the AMDGPU result, but I haven't verified everything.

I also added the pass explicitly to the NVPTX required passes. Are there existing tests to check the target specific pipeline?

[merge_guards_bot]

Build result: fail - 60568 tests passed, 3 failed and 726 were skipped.

failed: LLVM.CodeGen/NVPTX/atomics-sm60.ll
failed: LLVM.CodeGen/NVPTX/atomics.ll
failed: LLVM.CodeGen/NVPTX/load-store.ll

Log files: console-log.txt, CMakeCache.txt

[tra]

> Command Output (stderr):
> --
> /mnt/disks/ssd0/agent/workspace/amd64_debian_testing_clang8/llvm/test/CodeGen/NVPTX/atomics-sm60.ll:6:10: error: CHECK: expected string not found in input
> ; CHECK: atom.add.f64
>          ^
> <stdin>:1:1: note: scanning from here
> //
> ^
> <stdin>:34:2: note: possible intended match here
>  atom.cas.b64 %rd3, [%r1], %rd2, %rd1;
>  ^

This appears to be a regression. We do have fp32/fp64 atomic adds in NVPTX. Replacing them with add+CAS is suboptimal.

[jdoerfert]

In D71128#1773447, @tra wrote:

> Command Output (stderr):
> --
> /mnt/disks/ssd0/agent/workspace/amd64_debian_testing_clang8/llvm/test/CodeGen/NVPTX/atomics-sm60.ll:6:10: error: CHECK: expected string not found in input
> ; CHECK: atom.add.f64
>          ^
> <stdin>:1:1: note: scanning from here
> //
> ^
> <stdin>:34:2: note: possible intended match here
>  atom.cas.b64 %rd3, [%r1], %rd2, %rd1;
>  ^

This appears to be a regression. We do have fp32/fp64 atomic adds in NVPTX. Replacing them with add+CAS is suboptimal.

I'm already working on it.

[jdoerfert]

Fix test cases by exposing more TLI hooks

[merge_guards_bot]

Build result: FAILURE -
Log files: console-log.txt, CMakeCache.txt

[jdoerfert]

In D71128#1773635, @merge_guards_bot wrote:

Build result: FAILURE -
Log files: console-log.txt, CMakeCache.txt

I have the feeling that wasn't my fault.

[tra]

llvm/test/Transforms/AtomicExpand/NVPTX/expand-atomic-rmw-fadd.ll

17

Don't we want to preserve atomicrmw fadd in this case and lower it to atom.add.f32 ? Why do we want to expand here?

132

Ditto here and below. We do have atom.add.f64

llvm/test/Transforms/AtomicExpand/NVPTX/expand-atomic-rmw-fsub.ll

6

Functilon name fadd does not seem to match the instruction fsub.

16

I must be missing something -- I would think that we do *not* want to expand atomicrmw variants which we can lower to an existing instruction, but a lot of the tests show the opposite and expand atomics that have direct support in hardware. The patch subject seems to agree with my assumptions, but the tests appear to contradict it. Is that intentional? If so, what is it that I'm missing?

llvm/test/Transforms/AtomicExpand/NVPTX/unaligned-atomic.ll

3

Nit: no need for -check-prefix as you only using CHECK in the test.

[jdoerfert]

I will try to look into the problematic lowerings @tra pointed out (thanks btw!). Any hints to why they are expanded are appreciated :)

llvm/test/Transforms/AtomicExpand/NVPTX/expand-atomic-rmw-fadd.ll

17

Same as below.

132

To be honest, I don't even know why we do not match it. All I (tried) to do is add the limits wrt. size and alignment. Somehow that had more effect than I wanted. The new hooks already remove some of the weirdness we saw but it seems something is missing here (maybe during the instruction "registration").

llvm/test/Transforms/AtomicExpand/NVPTX/expand-atomic-rmw-fsub.ll

6

Good catch, copy & paste from the AMD tests ;) (@arsenm

16

It is not intentional to pesimise anything, as mentioned above. The problem is I am neither a backend nor NVPTX person and my changes do seem to have unwanted effects I cannot even categorize.

llvm/test/Transforms/AtomicExpand/NVPTX/unaligned-atomic.ll

3

Fair, I think I copied this ;)

[arsenm]

llvm/test/Transforms/AtomicExpand/NVPTX/expand-atomic-rmw-fsub.ll

16

For the purpose of this change, that this isn't optimal doesn't matter. These aren't implemented correct, but doing so is a separate change and those changes will show up in the same tests here

[tra]

llvm/test/Transforms/AtomicExpand/NVPTX/expand-atomic-rmw-fsub.ll

16

OK. Looks like atomicrmw fsub currently fails to lower on NVPTX, so expanding it is an improvement. However, expanding atomicrmw fadd is a substantial regression and is likely to be a showstopper. Atomic FP32 addition is a commonly used instruction in various reduction kernels so anything that prevents mapping it to atom.add.f32 instruction will be very noticeable.

I realize that there are many moving parts involved in getting this to work properly. If proper fix needs multiple patches, please try to commit them atomically to avoid the performance regression in between those changes.

Also, if there are dependent patches, it would be great to arrange all of them as such in phabricator, so it's easier to see the big picture.

[arsenm]

llvm/include/llvm/CodeGen/TargetLowering.h

1856–1871 ↗

(On Diff #232662)

Do we really need 4 of these when just the one for Instruction will work

[jdoerfert]

llvm/include/llvm/CodeGen/TargetLowering.h

1856–1871 ↗

(On Diff #232662)

Alternatively we can overload the instruction one and check for the kind to decide what to do. I don't remember how I ended up like this, I'll address this once I get around to this patch again...

llvm/lib/Target/NVPTX/NVPTXISelLowering.cpp

NVPTXTargetLowering::NVPTXTargetLowering(const NVPTXTargetMachine &TM,
 
   setBooleanContents(ZeroOrNegativeOneBooleanContent);
   setBooleanVectorContents(ZeroOrNegativeOneBooleanContent);
 
   // Jump is Expensive. Don't create extra control flow for 'and', 'or'
   // condition branches.
   setJumpIsExpensive(true);
 
+  // Force atomics to be expanded if the ISA doesn't support them: PR44219
+  setMinCmpXchgSizeInBits(32);
+  setMaxAtomicSizeInBitsSupported(64);
+  setSupportsUnalignedAtomics(false);
+
   // Wide divides are _very_ slow. Try to reduce the width of the divide if
   // possible.
   addBypassSlowDiv(64, 32);
 
   // By default, use the Source scheduling
   if (sched4reg)
     setSchedulingPreference(Sched::RegPressure);
   else

llvm/lib/Target/NVPTX/NVPTXTargetMachine.cpp

void NVPTXPassConfig::addIRPasses() {
   // call addEarlyAsPossiblePasses.
   const NVPTXSubtarget &ST = *getTM<NVPTXTargetMachine>().getSubtargetImpl();
   addPass(createNVVMReflectPass(ST.getSmVersion()));
 
   if (getOptLevel() != CodeGenOpt::None)
     addPass(createNVPTXImageOptimizerPass());
   addPass(createNVPTXAssignValidGlobalNamesPass());
   addPass(createGenericToNVVMPass());
+  addPass(createAtomicExpandPass());
 
   // NVPTXLowerArgs is required for correctness and should be run right
   // before the address space inference passes.
   addPass(createNVPTXLowerArgsPass(&getNVPTXTargetMachine()));
   if (getOptLevel() != CodeGenOpt::None) {
     addAddressSpaceInferencePasses();
     if (!DisableLoadStoreVectorizer)
       addPass(createLoadStoreVectorizerPass());

llvm/test/Transforms/AtomicExpand/NVPTX/expand-atomic-i16.ll

+; NOTE: Assertions have been autogenerated by utils/update_test_checks.py
+; RUN: opt -mtriple=nvptx-unknown-unknown -S -atomic-expand %s | FileCheck %s
+; RUN: opt -mtriple=nvptx64-unknown-unknown -S -atomic-expand %s | FileCheck %s
+
+define i16 @test_atomicrmw_xchg_i16_global(i16 addrspace(1)* %ptr, i16 %value) {
+; CHECK-LABEL: @test_atomicrmw_xchg_i16_global(
+; CHECK-NEXT:    [[RES:%.*]] = atomicrmw xchg i16 addrspace(1)* [[PTR:%.*]], i16 [[VALUE:%.*]] seq_cst
+; CHECK-NEXT:    ret i16 [[RES]]
+;
+  %res = atomicrmw xchg i16 addrspace(1)* %ptr, i16 %value seq_cst
+  ret i16 %res
+}
+
+define i16 @test_atomicrmw_add_i16_global(i16 addrspace(1)* %ptr, i16 %value) {
+; CHECK-LABEL: @test_atomicrmw_add_i16_global(
+; CHECK-NEXT:    [[RES:%.*]] = atomicrmw add i16 addrspace(1)* [[PTR:%.*]], i16 [[VALUE:%.*]] seq_cst
+; CHECK-NEXT:    ret i16 [[RES]]
+;
+  %res = atomicrmw add i16 addrspace(1)* %ptr, i16 %value seq_cst
+  ret i16 %res
+}
+
+define i16 @test_atomicrmw_sub_i16_global(i16 addrspace(1)* %ptr, i16 %value) {
+; CHECK-LABEL: @test_atomicrmw_sub_i16_global(
+; CHECK-NEXT:    [[RES:%.*]] = atomicrmw sub i16 addrspace(1)* [[PTR:%.*]], i16 [[VALUE:%.*]] seq_cst
+; CHECK-NEXT:    ret i16 [[RES]]
+;
+  %res = atomicrmw sub i16 addrspace(1)* %ptr, i16 %value seq_cst
+  ret i16 %res
+}
+
+define i16 @test_atomicrmw_and_i16_global(i16 addrspace(1)* %ptr, i16 %value) {
+; CHECK-LABEL: @test_atomicrmw_and_i16_global(
+; CHECK-NEXT:    [[TMP1:%.*]] = ptrtoint i16 addrspace(1)* [[PTR:%.*]] to i64
+; CHECK-NEXT:    [[TMP2:%.*]] = and i64 [[TMP1]], -4
+; CHECK-NEXT:    [[ALIGNEDADDR:%.*]] = inttoptr i64 [[TMP2]] to i32 addrspace(1)*
+; CHECK-NEXT:    [[PTRLSB:%.*]] = and i64 [[TMP1]], 3
+; CHECK-NEXT:    [[TMP3:%.*]] = shl i64 [[PTRLSB]], 3
+; CHECK-NEXT:    [[SHIFTAMT:%.*]] = trunc i64 [[TMP3]] to i32
+; CHECK-NEXT:    [[MASK:%.*]] = shl i32 65535, [[SHIFTAMT]]
+; CHECK-NEXT:    [[INV_MASK:%.*]] = xor i32 [[MASK]], -1
+; CHECK-NEXT:    [[TMP4:%.*]] = zext i16 [[VALUE:%.*]] to i32
+; CHECK-NEXT:    [[VALOPERAND_SHIFTED:%.*]] = shl i32 [[TMP4]], [[SHIFTAMT]]
+; CHECK-NEXT:    [[ANDOPERAND:%.*]] = or i32 [[INV_MASK]], [[VALOPERAND_SHIFTED]]
+; CHECK-NEXT:    [[TMP5:%.*]] = atomicrmw and i32 addrspace(1)* [[ALIGNEDADDR]], i32 [[ANDOPERAND]] seq_cst
+; CHECK-NEXT:    [[TMP6:%.*]] = lshr i32 [[TMP5]], [[SHIFTAMT]]
+; CHECK-NEXT:    [[TMP7:%.*]] = trunc i32 [[TMP6]] to i16
+; CHECK-NEXT:    ret i16 [[TMP7]]
+;
+  %res = atomicrmw and i16 addrspace(1)* %ptr, i16 %value seq_cst
+  ret i16 %res
+}
+
+define i16 @test_atomicrmw_nand_i16_global(i16 addrspace(1)* %ptr, i16 %value) {
+; CHECK-LABEL: @test_atomicrmw_nand_i16_global(
+; CHECK-NEXT:    [[RES:%.*]] = atomicrmw nand i16 addrspace(1)* [[PTR:%.*]], i16 [[VALUE:%.*]] seq_cst
+; CHECK-NEXT:    ret i16 [[RES]]
+;
+  %res = atomicrmw nand i16 addrspace(1)* %ptr, i16 %value seq_cst
+  ret i16 %res
+}
+
+define i16 @test_atomicrmw_or_i16_global(i16 addrspace(1)* %ptr, i16 %value) {
+; CHECK-LABEL: @test_atomicrmw_or_i16_global(
+; CHECK-NEXT:    [[TMP1:%.*]] = ptrtoint i16 addrspace(1)* [[PTR:%.*]] to i64
+; CHECK-NEXT:    [[TMP2:%.*]] = and i64 [[TMP1]], -4
+; CHECK-NEXT:    [[ALIGNEDADDR:%.*]] = inttoptr i64 [[TMP2]] to i32 addrspace(1)*
+; CHECK-NEXT:    [[PTRLSB:%.*]] = and i64 [[TMP1]], 3
+; CHECK-NEXT:    [[TMP3:%.*]] = shl i64 [[PTRLSB]], 3
+; CHECK-NEXT:    [[SHIFTAMT:%.*]] = trunc i64 [[TMP3]] to i32
+; CHECK-NEXT:    [[MASK:%.*]] = shl i32 65535, [[SHIFTAMT]]
+; CHECK-NEXT:    [[INV_MASK:%.*]] = xor i32 [[MASK]], -1
+; CHECK-NEXT:    [[TMP4:%.*]] = zext i16 [[VALUE:%.*]] to i32
+; CHECK-NEXT:    [[VALOPERAND_SHIFTED:%.*]] = shl i32 [[TMP4]], [[SHIFTAMT]]
+; CHECK-NEXT:    [[TMP5:%.*]] = atomicrmw or i32 addrspace(1)* [[ALIGNEDADDR]], i32 [[VALOPERAND_SHIFTED]] seq_cst
+; CHECK-NEXT:    [[TMP6:%.*]] = lshr i32 [[TMP5]], [[SHIFTAMT]]
+; CHECK-NEXT:    [[TMP7:%.*]] = trunc i32 [[TMP6]] to i16
+; CHECK-NEXT:    ret i16 [[TMP7]]
+;
+  %res = atomicrmw or i16 addrspace(1)* %ptr, i16 %value seq_cst
+  ret i16 %res
+}
+
+define i16 @test_atomicrmw_xor_i16_global(i16 addrspace(1)* %ptr, i16 %value) {
+; CHECK-LABEL: @test_atomicrmw_xor_i16_global(
+; CHECK-NEXT:    [[TMP1:%.*]] = ptrtoint i16 addrspace(1)* [[PTR:%.*]] to i64
+; CHECK-NEXT:    [[TMP2:%.*]] = and i64 [[TMP1]], -4
+; CHECK-NEXT:    [[ALIGNEDADDR:%.*]] = inttoptr i64 [[TMP2]] to i32 addrspace(1)*
+; CHECK-NEXT:    [[PTRLSB:%.*]] = and i64 [[TMP1]], 3
+; CHECK-NEXT:    [[TMP3:%.*]] = shl i64 [[PTRLSB]], 3
+; CHECK-NEXT:    [[SHIFTAMT:%.*]] = trunc i64 [[TMP3]] to i32
+; CHECK-NEXT:    [[MASK:%.*]] = shl i32 65535, [[SHIFTAMT]]
+; CHECK-NEXT:    [[INV_MASK:%.*]] = xor i32 [[MASK]], -1
+; CHECK-NEXT:    [[TMP4:%.*]] = zext i16 [[VALUE:%.*]] to i32
+; CHECK-NEXT:    [[VALOPERAND_SHIFTED:%.*]] = shl i32 [[TMP4]], [[SHIFTAMT]]
+; CHECK-NEXT:    [[TMP5:%.*]] = atomicrmw xor i32 addrspace(1)* [[ALIGNEDADDR]], i32 [[VALOPERAND_SHIFTED]] seq_cst
+; CHECK-NEXT:    [[TMP6:%.*]] = lshr i32 [[TMP5]], [[SHIFTAMT]]
+; CHECK-NEXT:    [[TMP7:%.*]] = trunc i32 [[TMP6]] to i16
+; CHECK-NEXT:    ret i16 [[TMP7]]
+;
+  %res = atomicrmw xor i16 addrspace(1)* %ptr, i16 %value seq_cst
+  ret i16 %res
+}
+
+define i16 @test_atomicrmw_max_i16_global(i16 addrspace(1)* %ptr, i16 %value) {
+; CHECK-LABEL: @test_atomicrmw_max_i16_global(
+; CHECK-NEXT:    [[RES:%.*]] = atomicrmw max i16 addrspace(1)* [[PTR:%.*]], i16 [[VALUE:%.*]] seq_cst
+; CHECK-NEXT:    ret i16 [[RES]]
+;
+  %res = atomicrmw max i16 addrspace(1)* %ptr, i16 %value seq_cst
+  ret i16 %res
+}
+
+define i16 @test_atomicrmw_min_i16_global(i16 addrspace(1)* %ptr, i16 %value) {
+; CHECK-LABEL: @test_atomicrmw_min_i16_global(
+; CHECK-NEXT:    [[RES:%.*]] = atomicrmw min i16 addrspace(1)* [[PTR:%.*]], i16 [[VALUE:%.*]] seq_cst
+; CHECK-NEXT:    ret i16 [[RES]]
+;
+  %res = atomicrmw min i16 addrspace(1)* %ptr, i16 %value seq_cst
+  ret i16 %res
+}
+
+define i16 @test_atomicrmw_umax_i16_global(i16 addrspace(1)* %ptr, i16 %value) {
+; CHECK-LABEL: @test_atomicrmw_umax_i16_global(
+; CHECK-NEXT:    [[RES:%.*]] = atomicrmw umax i16 addrspace(1)* [[PTR:%.*]], i16 [[VALUE:%.*]] seq_cst
+; CHECK-NEXT:    ret i16 [[RES]]
+;
+  %res = atomicrmw umax i16 addrspace(1)* %ptr, i16 %value seq_cst
+  ret i16 %res
+}
+
+define i16 @test_atomicrmw_umin_i16_global(i16 addrspace(1)* %ptr, i16 %value) {
+; CHECK-LABEL: @test_atomicrmw_umin_i16_global(
+; CHECK-NEXT:    [[RES:%.*]] = atomicrmw umin i16 addrspace(1)* [[PTR:%.*]], i16 [[VALUE:%.*]] seq_cst
+; CHECK-NEXT:    ret i16 [[RES]]
+;
+  %res = atomicrmw umin i16 addrspace(1)* %ptr, i16 %value seq_cst
+  ret i16 %res
+}
+
+define i16 @test_cmpxchg_i16_global(i16 addrspace(1)* %out, i16 %in, i16 %old) {
+; CHECK-LABEL: @test_cmpxchg_i16_global(
+; CHECK-NEXT:    [[GEP:%.*]] = getelementptr i16, i16 addrspace(1)* [[OUT:%.*]], i64 4
+; CHECK-NEXT:    [[TMP1:%.*]] = ptrtoint i16 addrspace(1)* [[GEP]] to i64
+; CHECK-NEXT:    [[TMP2:%.*]] = and i64 [[TMP1]], -4
+; CHECK-NEXT:    [[ALIGNEDADDR:%.*]] = inttoptr i64 [[TMP2]] to i32 addrspace(1)*
+; CHECK-NEXT:    [[PTRLSB:%.*]] = and i64 [[TMP1]], 3
+; CHECK-NEXT:    [[TMP3:%.*]] = shl i64 [[PTRLSB]], 3
+; CHECK-NEXT:    [[SHIFTAMT:%.*]] = trunc i64 [[TMP3]] to i32
+; CHECK-NEXT:    [[MASK:%.*]] = shl i32 65535, [[SHIFTAMT]]
+; CHECK-NEXT:    [[INV_MASK:%.*]] = xor i32 [[MASK]], -1
+; CHECK-NEXT:    [[TMP4:%.*]] = zext i16 [[IN:%.*]] to i32
+; CHECK-NEXT:    [[TMP5:%.*]] = shl i32 [[TMP4]], [[SHIFTAMT]]
+; CHECK-NEXT:    [[TMP6:%.*]] = zext i16 [[OLD:%.*]] to i32
+; CHECK-NEXT:    [[TMP7:%.*]] = shl i32 [[TMP6]], [[SHIFTAMT]]
+; CHECK-NEXT:    [[TMP8:%.*]] = load i32, i32 addrspace(1)* [[ALIGNEDADDR]]
+; CHECK-NEXT:    [[TMP9:%.*]] = and i32 [[TMP8]], [[INV_MASK]]
+; CHECK-NEXT:    br label [[PARTWORD_CMPXCHG_LOOP:%.*]]
+; CHECK:       partword.cmpxchg.loop:
+; CHECK-NEXT:    [[TMP10:%.*]] = phi i32 [ [[TMP9]], [[TMP0:%.*]] ], [ [[TMP16:%.*]], [[PARTWORD_CMPXCHG_FAILURE:%.*]] ]
+; CHECK-NEXT:    [[TMP11:%.*]] = or i32 [[TMP10]], [[TMP5]]
+; CHECK-NEXT:    [[TMP12:%.*]] = or i32 [[TMP10]], [[TMP7]]
+; CHECK-NEXT:    [[TMP13:%.*]] = cmpxchg i32 addrspace(1)* [[ALIGNEDADDR]], i32 [[TMP12]], i32 [[TMP11]] seq_cst seq_cst
+; CHECK-NEXT:    [[TMP14:%.*]] = extractvalue { i32, i1 } [[TMP13]], 0
+; CHECK-NEXT:    [[TMP15:%.*]] = extractvalue { i32, i1 } [[TMP13]], 1
+; CHECK-NEXT:    br i1 [[TMP15]], label [[PARTWORD_CMPXCHG_END:%.*]], label [[PARTWORD_CMPXCHG_FAILURE]]
+; CHECK:       partword.cmpxchg.failure:
+; CHECK-NEXT:    [[TMP16]] = and i32 [[TMP14]], [[INV_MASK]]
+; CHECK-NEXT:    [[TMP17:%.*]] = icmp ne i32 [[TMP10]], [[TMP16]]
+; CHECK-NEXT:    br i1 [[TMP17]], label [[PARTWORD_CMPXCHG_LOOP]], label [[PARTWORD_CMPXCHG_END]]
+; CHECK:       partword.cmpxchg.end:
+; CHECK-NEXT:    [[TMP18:%.*]] = lshr i32 [[TMP14]], [[SHIFTAMT]]
+; CHECK-NEXT:    [[TMP19:%.*]] = trunc i32 [[TMP18]] to i16
+; CHECK-NEXT:    [[TMP20:%.*]] = insertvalue { i16, i1 } undef, i16 [[TMP19]], 0
+; CHECK-NEXT:    [[TMP21:%.*]] = insertvalue { i16, i1 } [[TMP20]], i1 [[TMP15]], 1
+; CHECK-NEXT:    [[EXTRACT:%.*]] = extractvalue { i16, i1 } [[TMP21]], 0
+; CHECK-NEXT:    ret i16 [[EXTRACT]]
+;
+  %gep = getelementptr i16, i16 addrspace(1)* %out, i64 4
+  %res = cmpxchg i16 addrspace(1)* %gep, i16 %old, i16 %in seq_cst seq_cst
+  %extract = extractvalue {i16, i1} %res, 0
+  ret i16 %extract
+}

llvm/test/Transforms/AtomicExpand/NVPTX/expand-atomic-i8.ll

+; NOTE: Assertions have been autogenerated by utils/update_test_checks.py
+; RUN: opt -mtriple=nvptx-unknown-unknown -S -atomic-expand %s | FileCheck %s
+; RUN: opt -mtriple=nvptx64-unknown-unknown -S -atomic-expand %s | FileCheck %s
+
+define i8 @test_atomicrmw_xchg_i8_global(i8 addrspace(1)* %ptr, i8 %value) {
+; CHECK-LABEL: @test_atomicrmw_xchg_i8_global(
+; CHECK-NEXT:    [[RES:%.*]] = atomicrmw xchg i8 addrspace(1)* [[PTR:%.*]], i8 [[VALUE:%.*]] seq_cst
+; CHECK-NEXT:    ret i8 [[RES]]
+;
+  %res = atomicrmw xchg i8 addrspace(1)* %ptr, i8 %value seq_cst
+  ret i8 %res
+}
+
+define i8 @test_atomicrmw_add_i8_global(i8 addrspace(1)* %ptr, i8 %value) {
+; CHECK-LABEL: @test_atomicrmw_add_i8_global(
+; CHECK-NEXT:    [[RES:%.*]] = atomicrmw add i8 addrspace(1)* [[PTR:%.*]], i8 [[VALUE:%.*]] seq_cst
+; CHECK-NEXT:    ret i8 [[RES]]
+;
+  %res = atomicrmw add i8 addrspace(1)* %ptr, i8 %value seq_cst
+  ret i8 %res
+}
+
+define i8 @test_atomicrmw_sub_i8_global(i8 addrspace(1)* %ptr, i8 %value) {
+; CHECK-LABEL: @test_atomicrmw_sub_i8_global(
+; CHECK-NEXT:    [[RES:%.*]] = atomicrmw sub i8 addrspace(1)* [[PTR:%.*]], i8 [[VALUE:%.*]] seq_cst
+; CHECK-NEXT:    ret i8 [[RES]]
+;
+  %res = atomicrmw sub i8 addrspace(1)* %ptr, i8 %value seq_cst
+  ret i8 %res
+}
+
+define i8 @test_atomicrmw_and_i8_global(i8 addrspace(1)* %ptr, i8 %value) {
+; CHECK-LABEL: @test_atomicrmw_and_i8_global(
+; CHECK-NEXT:    [[TMP1:%.*]] = ptrtoint i8 addrspace(1)* [[PTR:%.*]] to i64
+; CHECK-NEXT:    [[TMP2:%.*]] = and i64 [[TMP1]], -4
+; CHECK-NEXT:    [[ALIGNEDADDR:%.*]] = inttoptr i64 [[TMP2]] to i32 addrspace(1)*
+; CHECK-NEXT:    [[PTRLSB:%.*]] = and i64 [[TMP1]], 3
+; CHECK-NEXT:    [[TMP3:%.*]] = shl i64 [[PTRLSB]], 3
+; CHECK-NEXT:    [[SHIFTAMT:%.*]] = trunc i64 [[TMP3]] to i32
+; CHECK-NEXT:    [[MASK:%.*]] = shl i32 255, [[SHIFTAMT]]
+; CHECK-NEXT:    [[INV_MASK:%.*]] = xor i32 [[MASK]], -1
+; CHECK-NEXT:    [[TMP4:%.*]] = zext i8 [[VALUE:%.*]] to i32
+; CHECK-NEXT:    [[VALOPERAND_SHIFTED:%.*]] = shl i32 [[TMP4]], [[SHIFTAMT]]
+; CHECK-NEXT:    [[ANDOPERAND:%.*]] = or i32 [[INV_MASK]], [[VALOPERAND_SHIFTED]]
+; CHECK-NEXT:    [[TMP5:%.*]] = atomicrmw and i32 addrspace(1)* [[ALIGNEDADDR]], i32 [[ANDOPERAND]] seq_cst
+; CHECK-NEXT:    [[TMP6:%.*]] = lshr i32 [[TMP5]], [[SHIFTAMT]]
+; CHECK-NEXT:    [[TMP7:%.*]] = trunc i32 [[TMP6]] to i8
+; CHECK-NEXT:    ret i8 [[TMP7]]
+;
+  %res = atomicrmw and i8 addrspace(1)* %ptr, i8 %value seq_cst
+  ret i8 %res
+}
+
+define i8 @test_atomicrmw_nand_i8_global(i8 addrspace(1)* %ptr, i8 %value) {
+; CHECK-LABEL: @test_atomicrmw_nand_i8_global(
+; CHECK-NEXT:    [[RES:%.*]] = atomicrmw nand i8 addrspace(1)* [[PTR:%.*]], i8 [[VALUE:%.*]] seq_cst
+; CHECK-NEXT:    ret i8 [[RES]]
+;
+  %res = atomicrmw nand i8 addrspace(1)* %ptr, i8 %value seq_cst
+  ret i8 %res
+}
+
+define i8 @test_atomicrmw_or_i8_global(i8 addrspace(1)* %ptr, i8 %value) {
+; CHECK-LABEL: @test_atomicrmw_or_i8_global(
+; CHECK-NEXT:    [[TMP1:%.*]] = ptrtoint i8 addrspace(1)* [[PTR:%.*]] to i64
+; CHECK-NEXT:    [[TMP2:%.*]] = and i64 [[TMP1]], -4
+; CHECK-NEXT:    [[ALIGNEDADDR:%.*]] = inttoptr i64 [[TMP2]] to i32 addrspace(1)*
+; CHECK-NEXT:    [[PTRLSB:%.*]] = and i64 [[TMP1]], 3
+; CHECK-NEXT:    [[TMP3:%.*]] = shl i64 [[PTRLSB]], 3
+; CHECK-NEXT:    [[SHIFTAMT:%.*]] = trunc i64 [[TMP3]] to i32
+; CHECK-NEXT:    [[MASK:%.*]] = shl i32 255, [[SHIFTAMT]]
+; CHECK-NEXT:    [[INV_MASK:%.*]] = xor i32 [[MASK]], -1
+; CHECK-NEXT:    [[TMP4:%.*]] = zext i8 [[VALUE:%.*]] to i32
+; CHECK-NEXT:    [[VALOPERAND_SHIFTED:%.*]] = shl i32 [[TMP4]], [[SHIFTAMT]]
+; CHECK-NEXT:    [[TMP5:%.*]] = atomicrmw or i32 addrspace(1)* [[ALIGNEDADDR]], i32 [[VALOPERAND_SHIFTED]] seq_cst
+; CHECK-NEXT:    [[TMP6:%.*]] = lshr i32 [[TMP5]], [[SHIFTAMT]]
+; CHECK-NEXT:    [[TMP7:%.*]] = trunc i32 [[TMP6]] to i8
+; CHECK-NEXT:    ret i8 [[TMP7]]
+;
+  %res = atomicrmw or i8 addrspace(1)* %ptr, i8 %value seq_cst
+  ret i8 %res
+}
+
+define i8 @test_atomicrmw_xor_i8_global(i8 addrspace(1)* %ptr, i8 %value) {
+; CHECK-LABEL: @test_atomicrmw_xor_i8_global(
+; CHECK-NEXT:    [[TMP1:%.*]] = ptrtoint i8 addrspace(1)* [[PTR:%.*]] to i64
+; CHECK-NEXT:    [[TMP2:%.*]] = and i64 [[TMP1]], -4
+; CHECK-NEXT:    [[ALIGNEDADDR:%.*]] = inttoptr i64 [[TMP2]] to i32 addrspace(1)*
+; CHECK-NEXT:    [[PTRLSB:%.*]] = and i64 [[TMP1]], 3
+; CHECK-NEXT:    [[TMP3:%.*]] = shl i64 [[PTRLSB]], 3
+; CHECK-NEXT:    [[SHIFTAMT:%.*]] = trunc i64 [[TMP3]] to i32
+; CHECK-NEXT:    [[MASK:%.*]] = shl i32 255, [[SHIFTAMT]]
+; CHECK-NEXT:    [[INV_MASK:%.*]] = xor i32 [[MASK]], -1
+; CHECK-NEXT:    [[TMP4:%.*]] = zext i8 [[VALUE:%.*]] to i32
+; CHECK-NEXT:    [[VALOPERAND_SHIFTED:%.*]] = shl i32 [[TMP4]], [[SHIFTAMT]]
+; CHECK-NEXT:    [[TMP5:%.*]] = atomicrmw xor i32 addrspace(1)* [[ALIGNEDADDR]], i32 [[VALOPERAND_SHIFTED]] seq_cst
+; CHECK-NEXT:    [[TMP6:%.*]] = lshr i32 [[TMP5]], [[SHIFTAMT]]
+; CHECK-NEXT:    [[TMP7:%.*]] = trunc i32 [[TMP6]] to i8
+; CHECK-NEXT:    ret i8 [[TMP7]]
+;
+  %res = atomicrmw xor i8 addrspace(1)* %ptr, i8 %value seq_cst
+  ret i8 %res
+}
+
+define i8 @test_atomicrmw_max_i8_global(i8 addrspace(1)* %ptr, i8 %value) {
+; CHECK-LABEL: @test_atomicrmw_max_i8_global(
+; CHECK-NEXT:    [[RES:%.*]] = atomicrmw max i8 addrspace(1)* [[PTR:%.*]], i8 [[VALUE:%.*]] seq_cst
+; CHECK-NEXT:    ret i8 [[RES]]
+;
+  %res = atomicrmw max i8 addrspace(1)* %ptr, i8 %value seq_cst
+  ret i8 %res
+}
+
+define i8 @test_atomicrmw_min_i8_global(i8 addrspace(1)* %ptr, i8 %value) {
+; CHECK-LABEL: @test_atomicrmw_min_i8_global(
+; CHECK-NEXT:    [[RES:%.*]] = atomicrmw min i8 addrspace(1)* [[PTR:%.*]], i8 [[VALUE:%.*]] seq_cst
+; CHECK-NEXT:    ret i8 [[RES]]
+;
+  %res = atomicrmw min i8 addrspace(1)* %ptr, i8 %value seq_cst
+  ret i8 %res
+}
+
+define i8 @test_atomicrmw_umax_i8_global(i8 addrspace(1)* %ptr, i8 %value) {
+; CHECK-LABEL: @test_atomicrmw_umax_i8_global(
+; CHECK-NEXT:    [[RES:%.*]] = atomicrmw umax i8 addrspace(1)* [[PTR:%.*]], i8 [[VALUE:%.*]] seq_cst
+; CHECK-NEXT:    ret i8 [[RES]]
+;
+  %res = atomicrmw umax i8 addrspace(1)* %ptr, i8 %value seq_cst
+  ret i8 %res
+}
+
+define i8 @test_atomicrmw_umin_i8_global(i8 addrspace(1)* %ptr, i8 %value) {
+; CHECK-LABEL: @test_atomicrmw_umin_i8_global(
+; CHECK-NEXT:    [[RES:%.*]] = atomicrmw umin i8 addrspace(1)* [[PTR:%.*]], i8 [[VALUE:%.*]] seq_cst
+; CHECK-NEXT:    ret i8 [[RES]]
+;
+  %res = atomicrmw umin i8 addrspace(1)* %ptr, i8 %value seq_cst
+  ret i8 %res
+}
+
+define i8 @test_cmpxchg_i8_global(i8 addrspace(1)* %out, i8 %in, i8 %old) {
+; CHECK-LABEL: @test_cmpxchg_i8_global(
+; CHECK-NEXT:    [[GEP:%.*]] = getelementptr i8, i8 addrspace(1)* [[OUT:%.*]], i64 4
+; CHECK-NEXT:    [[TMP1:%.*]] = ptrtoint i8 addrspace(1)* [[GEP]] to i64
+; CHECK-NEXT:    [[TMP2:%.*]] = and i64 [[TMP1]], -4
+; CHECK-NEXT:    [[ALIGNEDADDR:%.*]] = inttoptr i64 [[TMP2]] to i32 addrspace(1)*
+; CHECK-NEXT:    [[PTRLSB:%.*]] = and i64 [[TMP1]], 3
+; CHECK-NEXT:    [[TMP3:%.*]] = shl i64 [[PTRLSB]], 3
+; CHECK-NEXT:    [[SHIFTAMT:%.*]] = trunc i64 [[TMP3]] to i32
+; CHECK-NEXT:    [[MASK:%.*]] = shl i32 255, [[SHIFTAMT]]
+; CHECK-NEXT:    [[INV_MASK:%.*]] = xor i32 [[MASK]], -1
+; CHECK-NEXT:    [[TMP4:%.*]] = zext i8 [[IN:%.*]] to i32
+; CHECK-NEXT:    [[TMP5:%.*]] = shl i32 [[TMP4]], [[SHIFTAMT]]
+; CHECK-NEXT:    [[TMP6:%.*]] = zext i8 [[OLD:%.*]] to i32
+; CHECK-NEXT:    [[TMP7:%.*]] = shl i32 [[TMP6]], [[SHIFTAMT]]
+; CHECK-NEXT:    [[TMP8:%.*]] = load i32, i32 addrspace(1)* [[ALIGNEDADDR]]
+; CHECK-NEXT:    [[TMP9:%.*]] = and i32 [[TMP8]], [[INV_MASK]]
+; CHECK-NEXT:    br label [[PARTWORD_CMPXCHG_LOOP:%.*]]
+; CHECK:       partword.cmpxchg.loop:
+; CHECK-NEXT:    [[TMP10:%.*]] = phi i32 [ [[TMP9]], [[TMP0:%.*]] ], [ [[TMP16:%.*]], [[PARTWORD_CMPXCHG_FAILURE:%.*]] ]
+; CHECK-NEXT:    [[TMP11:%.*]] = or i32 [[TMP10]], [[TMP5]]
+; CHECK-NEXT:    [[TMP12:%.*]] = or i32 [[TMP10]], [[TMP7]]
+; CHECK-NEXT:    [[TMP13:%.*]] = cmpxchg i32 addrspace(1)* [[ALIGNEDADDR]], i32 [[TMP12]], i32 [[TMP11]] seq_cst seq_cst
+; CHECK-NEXT:    [[TMP14:%.*]] = extractvalue { i32, i1 } [[TMP13]], 0
+; CHECK-NEXT:    [[TMP15:%.*]] = extractvalue { i32, i1 } [[TMP13]], 1
+; CHECK-NEXT:    br i1 [[TMP15]], label [[PARTWORD_CMPXCHG_END:%.*]], label [[PARTWORD_CMPXCHG_FAILURE]]
+; CHECK:       partword.cmpxchg.failure:
+; CHECK-NEXT:    [[TMP16]] = and i32 [[TMP14]], [[INV_MASK]]
+; CHECK-NEXT:    [[TMP17:%.*]] = icmp ne i32 [[TMP10]], [[TMP16]]
+; CHECK-NEXT:    br i1 [[TMP17]], label [[PARTWORD_CMPXCHG_LOOP]], label [[PARTWORD_CMPXCHG_END]]
+; CHECK:       partword.cmpxchg.end:
+; CHECK-NEXT:    [[TMP18:%.*]] = lshr i32 [[TMP14]], [[SHIFTAMT]]
+; CHECK-NEXT:    [[TMP19:%.*]] = trunc i32 [[TMP18]] to i8
+; CHECK-NEXT:    [[TMP20:%.*]] = insertvalue { i8, i1 } undef, i8 [[TMP19]], 0
+; CHECK-NEXT:    [[TMP21:%.*]] = insertvalue { i8, i1 } [[TMP20]], i1 [[TMP15]], 1
+; CHECK-NEXT:    [[EXTRACT:%.*]] = extractvalue { i8, i1 } [[TMP21]], 0
+; CHECK-NEXT:    ret i8 [[EXTRACT]]
+;
+  %gep = getelementptr i8, i8 addrspace(1)* %out, i64 4
+  %res = cmpxchg i8 addrspace(1)* %gep, i8 %old, i8 %in seq_cst seq_cst
+  %extract = extractvalue {i8, i1} %res, 0
+  ret i8 %extract
+}

llvm/test/Transforms/AtomicExpand/NVPTX/expand-atomic-rmw-fadd.ll

+; NOTE: Assertions have been autogenerated by utils/update_test_checks.py
+; RUN: opt -S -mtriple=nvptx-unknown-unknown -mcpu=sm_30 -atomic-expand %s | FileCheck %s
+; RUN: opt -S -mtriple=nvptx-unknown-unknown -mcpu=sm_60 -atomic-expand %s | FileCheck %s
+; RUN: opt -S -mtriple=nvptx-unknown-unknown -mcpu=sm_75 -atomic-expand %s | FileCheck %s
+
+define float @test_atomicrmw_fadd_f32_flat(float* %ptr, float %value) {
+; CHECK-LABEL: @test_atomicrmw_fadd_f32_flat(
+; CHECK-NEXT:    [[TMP1:%.*]] = load float, float* [[PTR:%.*]], align 4
+; CHECK-NEXT:    br label [[ATOMICRMW_START:%.*]]
+; CHECK:       atomicrmw.start:
+; CHECK-NEXT:    [[LOADED:%.*]] = phi float [ [[TMP1]], [[TMP0:%.*]] ], [ [[TMP6:%.*]], [[ATOMICRMW_START]] ]
+; CHECK-NEXT:    [[NEW:%.*]] = fadd float [[LOADED]], [[VALUE:%.*]]
+; CHECK-NEXT:    [[TMP2:%.*]] = bitcast float* [[PTR]] to i32*
+; CHECK-NEXT:    [[TMP3:%.*]] = bitcast float [[NEW]] to i32
+; CHECK-NEXT:    [[TMP4:%.*]] = bitcast float [[LOADED]] to i32
+; CHECK-NEXT:    [[TMP5:%.*]] = cmpxchg i32* [[TMP2]], i32 [[TMP4]], i32 [[TMP3]] seq_cst seq_cst
+; CHECK-NEXT:    [[SUCCESS:%.*]] = extractvalue { i32, i1 } [[TMP5]], 1
+; CHECK-NEXT:    [[NEWLOADED:%.*]] = extractvalue { i32, i1 } [[TMP5]], 0
+; CHECK-NEXT:    [[TMP6]] = bitcast i32 [[NEWLOADED]] to float
+; CHECK-NEXT:    br i1 [[SUCCESS]], label [[ATOMICRMW_END:%.*]], label [[ATOMICRMW_START]]
+; CHECK:       atomicrmw.end:
+; CHECK-NEXT:    ret float [[TMP6]]
+;
+  %res = atomicrmw fadd float* %ptr, float %value seq_cst
+  ret float %res
+}
+
+define float @test_atomicrmw_fadd_f32_global(float addrspace(1)* %ptr, float %value) {
+; CHECK-LABEL: @test_atomicrmw_fadd_f32_global(
+; CHECK-NEXT:    [[TMP1:%.*]] = load float, float addrspace(1)* [[PTR:%.*]], align 4
+; CHECK-NEXT:    br label [[ATOMICRMW_START:%.*]]
+; CHECK:       atomicrmw.start:
+; CHECK-NEXT:    [[LOADED:%.*]] = phi float [ [[TMP1]], [[TMP0:%.*]] ], [ [[TMP6:%.*]], [[ATOMICRMW_START]] ]
+; CHECK-NEXT:    [[NEW:%.*]] = fadd float [[LOADED]], [[VALUE:%.*]]
+; CHECK-NEXT:    [[TMP2:%.*]] = bitcast float addrspace(1)* [[PTR]] to i32 addrspace(1)*
+; CHECK-NEXT:    [[TMP3:%.*]] = bitcast float [[NEW]] to i32
+; CHECK-NEXT:    [[TMP4:%.*]] = bitcast float [[LOADED]] to i32
+; CHECK-NEXT:    [[TMP5:%.*]] = cmpxchg i32 addrspace(1)* [[TMP2]], i32 [[TMP4]], i32 [[TMP3]] seq_cst seq_cst
+; CHECK-NEXT:    [[SUCCESS:%.*]] = extractvalue { i32, i1 } [[TMP5]], 1
+; CHECK-NEXT:    [[NEWLOADED:%.*]] = extractvalue { i32, i1 } [[TMP5]], 0
+; CHECK-NEXT:    [[TMP6]] = bitcast i32 [[NEWLOADED]] to float
+; CHECK-NEXT:    br i1 [[SUCCESS]], label [[ATOMICRMW_END:%.*]], label [[ATOMICRMW_START]]
+; CHECK:       atomicrmw.end:
+; CHECK-NEXT:    ret float [[TMP6]]
+;
+  %res = atomicrmw fadd float addrspace(1)* %ptr, float %value seq_cst
+  ret float %res
+}
+
+define void @test_atomicrmw_fadd_f32_global_no_use(float addrspace(1)* %ptr, float %value) {
+; CHECK-LABEL: @test_atomicrmw_fadd_f32_global_no_use(
+; CHECK-NEXT:    [[TMP1:%.*]] = load float, float addrspace(1)* [[PTR:%.*]], align 4
+; CHECK-NEXT:    br label [[ATOMICRMW_START:%.*]]
+; CHECK:       atomicrmw.start:
+; CHECK-NEXT:    [[LOADED:%.*]] = phi float [ [[TMP1]], [[TMP0:%.*]] ], [ [[TMP6:%.*]], [[ATOMICRMW_START]] ]
+; CHECK-NEXT:    [[NEW:%.*]] = fadd float [[LOADED]], [[VALUE:%.*]]
+; CHECK-NEXT:    [[TMP2:%.*]] = bitcast float addrspace(1)* [[PTR]] to i32 addrspace(1)*
+; CHECK-NEXT:    [[TMP3:%.*]] = bitcast float [[NEW]] to i32
+; CHECK-NEXT:    [[TMP4:%.*]] = bitcast float [[LOADED]] to i32
+; CHECK-NEXT:    [[TMP5:%.*]] = cmpxchg i32 addrspace(1)* [[TMP2]], i32 [[TMP4]], i32 [[TMP3]] seq_cst seq_cst
+; CHECK-NEXT:    [[SUCCESS:%.*]] = extractvalue { i32, i1 } [[TMP5]], 1
+; CHECK-NEXT:    [[NEWLOADED:%.*]] = extractvalue { i32, i1 } [[TMP5]], 0
+; CHECK-NEXT:    [[TMP6]] = bitcast i32 [[NEWLOADED]] to float
+; CHECK-NEXT:    br i1 [[SUCCESS]], label [[ATOMICRMW_END:%.*]], label [[ATOMICRMW_START]]
+; CHECK:       atomicrmw.end:
+; CHECK-NEXT:    ret void
+;
+  %res = atomicrmw fadd float addrspace(1)* %ptr, float %value seq_cst
+  ret void
+}
+
+define float @test_atomicrmw_fadd_f32_local(float addrspace(3)* %ptr, float %value) {
+; CHECK-LABEL: @test_atomicrmw_fadd_f32_local(
+; CHECK-NEXT:    [[TMP1:%.*]] = load float, float addrspace(3)* [[PTR:%.*]], align 4
+; CHECK-NEXT:    br label [[ATOMICRMW_START:%.*]]
+; CHECK:       atomicrmw.start:
+; CHECK-NEXT:    [[LOADED:%.*]] = phi float [ [[TMP1]], [[TMP0:%.*]] ], [ [[TMP6:%.*]], [[ATOMICRMW_START]] ]
+; CHECK-NEXT:    [[NEW:%.*]] = fadd float [[LOADED]], [[VALUE:%.*]]
+; CHECK-NEXT:    [[TMP2:%.*]] = bitcast float addrspace(3)* [[PTR]] to i32 addrspace(3)*
+; CHECK-NEXT:    [[TMP3:%.*]] = bitcast float [[NEW]] to i32
+; CHECK-NEXT:    [[TMP4:%.*]] = bitcast float [[LOADED]] to i32
+; CHECK-NEXT:    [[TMP5:%.*]] = cmpxchg i32 addrspace(3)* [[TMP2]], i32 [[TMP4]], i32 [[TMP3]] seq_cst seq_cst
+; CHECK-NEXT:    [[SUCCESS:%.*]] = extractvalue { i32, i1 } [[TMP5]], 1
+; CHECK-NEXT:    [[NEWLOADED:%.*]] = extractvalue { i32, i1 } [[TMP5]], 0
+; CHECK-NEXT:    [[TMP6]] = bitcast i32 [[NEWLOADED]] to float
+; CHECK-NEXT:    br i1 [[SUCCESS]], label [[ATOMICRMW_END:%.*]], label [[ATOMICRMW_START]]
+; CHECK:       atomicrmw.end:
+; CHECK-NEXT:    ret float [[TMP6]]
+;
+  %res = atomicrmw fadd float addrspace(3)* %ptr, float %value seq_cst
+  ret float %res
+}
+
+define half @test_atomicrmw_fadd_f16_flat(half* %ptr, half %value) {
+; CHECK-LABEL: @test_atomicrmw_fadd_f16_flat(
+; CHECK-NEXT:    [[RES:%.*]] = atomicrmw fadd half* [[PTR:%.*]], half [[VALUE:%.*]] seq_cst
+; CHECK-NEXT:    ret half [[RES]]
+;
+  %res = atomicrmw fadd half* %ptr, half %value seq_cst
+  ret half %res
+}
+
+define half @test_atomicrmw_fadd_f16_global(half addrspace(1)* %ptr, half %value) {
+; CHECK-LABEL: @test_atomicrmw_fadd_f16_global(
+; CHECK-NEXT:    [[RES:%.*]] = atomicrmw fadd half addrspace(1)* [[PTR:%.*]], half [[VALUE:%.*]] seq_cst
+; CHECK-NEXT:    ret half [[RES]]
+;
+  %res = atomicrmw fadd half addrspace(1)* %ptr, half %value seq_cst
+  ret half %res
+}
+
+define half @test_atomicrmw_fadd_f16_local(half addrspace(3)* %ptr, half %value) {
+; CHECK-LABEL: @test_atomicrmw_fadd_f16_local(
+; CHECK-NEXT:    [[RES:%.*]] = atomicrmw fadd half addrspace(3)* [[PTR:%.*]], half [[VALUE:%.*]] seq_cst
+; CHECK-NEXT:    ret half [[RES]]
+;
+  %res = atomicrmw fadd half addrspace(3)* %ptr, half %value seq_cst
+  ret half %res
+}
+
+define double @test_atomicrmw_fadd_f64_flat(double* %ptr, double %value) {
+; CHECK-LABEL: @test_atomicrmw_fadd_f64_flat(
+; CHECK-NEXT:    [[TMP1:%.*]] = load double, double* [[PTR:%.*]], align 8
+; CHECK-NEXT:    br label [[ATOMICRMW_START:%.*]]
+; CHECK:       atomicrmw.start:
+; CHECK-NEXT:    [[LOADED:%.*]] = phi double [ [[TMP1]], [[TMP0:%.*]] ], [ [[TMP6:%.*]], [[ATOMICRMW_START]] ]
+; CHECK-NEXT:    [[NEW:%.*]] = fadd double [[LOADED]], [[VALUE:%.*]]
+; CHECK-NEXT:    [[TMP2:%.*]] = bitcast double* [[PTR]] to i64*
+; CHECK-NEXT:    [[TMP3:%.*]] = bitcast double [[NEW]] to i64
+; CHECK-NEXT:    [[TMP4:%.*]] = bitcast double [[LOADED]] to i64
+; CHECK-NEXT:    [[TMP5:%.*]] = cmpxchg i64* [[TMP2]], i64 [[TMP4]], i64 [[TMP3]] seq_cst seq_cst
+; CHECK-NEXT:    [[SUCCESS:%.*]] = extractvalue { i64, i1 } [[TMP5]], 1
+; CHECK-NEXT:    [[NEWLOADED:%.*]] = extractvalue { i64, i1 } [[TMP5]], 0
+; CHECK-NEXT:    [[TMP6]] = bitcast i64 [[NEWLOADED]] to double
+; CHECK-NEXT:    br i1 [[SUCCESS]], label [[ATOMICRMW_END:%.*]], label [[ATOMICRMW_START]]
+; CHECK:       atomicrmw.end:
+; CHECK-NEXT:    ret double [[TMP6]]
+;
+  %res = atomicrmw fadd double* %ptr, double %value seq_cst
+  ret double %res
+}
+
+define double @test_atomicrmw_fadd_f64_global(double addrspace(1)* %ptr, double %value) {
+; CHECK-LABEL: @test_atomicrmw_fadd_f64_global(
+; CHECK-NEXT:    [[TMP1:%.*]] = load double, double addrspace(1)* [[PTR:%.*]], align 8
+; CHECK-NEXT:    br label [[ATOMICRMW_START:%.*]]
+; CHECK:       atomicrmw.start:
+; CHECK-NEXT:    [[LOADED:%.*]] = phi double [ [[TMP1]], [[TMP0:%.*]] ], [ [[TMP6:%.*]], [[ATOMICRMW_START]] ]
+; CHECK-NEXT:    [[NEW:%.*]] = fadd double [[LOADED]], [[VALUE:%.*]]
+; CHECK-NEXT:    [[TMP2:%.*]] = bitcast double addrspace(1)* [[PTR]] to i64 addrspace(1)*
+; CHECK-NEXT:    [[TMP3:%.*]] = bitcast double [[NEW]] to i64
+; CHECK-NEXT:    [[TMP4:%.*]] = bitcast double [[LOADED]] to i64
+; CHECK-NEXT:    [[TMP5:%.*]] = cmpxchg i64 addrspace(1)* [[TMP2]], i64 [[TMP4]], i64 [[TMP3]] seq_cst seq_cst
+; CHECK-NEXT:    [[SUCCESS:%.*]] = extractvalue { i64, i1 } [[TMP5]], 1
+; CHECK-NEXT:    [[NEWLOADED:%.*]] = extractvalue { i64, i1 } [[TMP5]], 0
+; CHECK-NEXT:    [[TMP6]] = bitcast i64 [[NEWLOADED]] to double
+; CHECK-NEXT:    br i1 [[SUCCESS]], label [[ATOMICRMW_END:%.*]], label [[ATOMICRMW_START]]
+; CHECK:       atomicrmw.end:
+; CHECK-NEXT:    ret double [[TMP6]]
+;
+  %res = atomicrmw fadd double addrspace(1)* %ptr, double %value seq_cst
+  ret double %res
+}
+
+define double @test_atomicrmw_fadd_f64_local(double addrspace(3)* %ptr, double %value) {
+; CHECK-LABEL: @test_atomicrmw_fadd_f64_local(
+; CHECK-NEXT:    [[TMP1:%.*]] = load double, double addrspace(3)* [[PTR:%.*]], align 8
+; CHECK-NEXT:    br label [[ATOMICRMW_START:%.*]]
+; CHECK:       atomicrmw.start:
+; CHECK-NEXT:    [[LOADED:%.*]] = phi double [ [[TMP1]], [[TMP0:%.*]] ], [ [[TMP6:%.*]], [[ATOMICRMW_START]] ]
+; CHECK-NEXT:    [[NEW:%.*]] = fadd double [[LOADED]], [[VALUE:%.*]]
+; CHECK-NEXT:    [[TMP2:%.*]] = bitcast double addrspace(3)* [[PTR]] to i64 addrspace(3)*
+; CHECK-NEXT:    [[TMP3:%.*]] = bitcast double [[NEW]] to i64
+; CHECK-NEXT:    [[TMP4:%.*]] = bitcast double [[LOADED]] to i64
+; CHECK-NEXT:    [[TMP5:%.*]] = cmpxchg i64 addrspace(3)* [[TMP2]], i64 [[TMP4]], i64 [[TMP3]] seq_cst seq_cst
+; CHECK-NEXT:    [[SUCCESS:%.*]] = extractvalue { i64, i1 } [[TMP5]], 1
+; CHECK-NEXT:    [[NEWLOADED:%.*]] = extractvalue { i64, i1 } [[TMP5]], 0
+; CHECK-NEXT:    [[TMP6]] = bitcast i64 [[NEWLOADED]] to double
+; CHECK-NEXT:    br i1 [[SUCCESS]], label [[ATOMICRMW_END:%.*]], label [[ATOMICRMW_START]]
+; CHECK:       atomicrmw.end:
+; CHECK-NEXT:    ret double [[TMP6]]
+;
+  %res = atomicrmw fadd double addrspace(3)* %ptr, double %value seq_cst
+  ret double %res
+}
+

llvm/test/Transforms/AtomicExpand/NVPTX/expand-atomic-rmw-fsub.ll

+; NOTE: Assertions have been autogenerated by utils/update_test_checks.py
+; RUN: opt -S -mtriple=nvptx64-unknown-unknown -mcpu=sm_30 -atomic-expand %s | FileCheck %s
+; RUN: opt -S -mtriple=nvptx64-unknown-unknown -mcpu=sm_75 -atomic-expand %s | FileCheck %s
+
+define float @test_atomicrmw_fadd_f32_flat(float* %ptr, float %value) {
+; CHECK-LABEL: @test_atomicrmw_fadd_f32_flat(
+; CHECK-NEXT:    [[TMP1:%.*]] = load float, float* [[PTR:%.*]], align 4
+; CHECK-NEXT:    br label [[ATOMICRMW_START:%.*]]
+; CHECK:       atomicrmw.start:
+; CHECK-NEXT:    [[LOADED:%.*]] = phi float [ [[TMP1]], [[TMP0:%.*]] ], [ [[TMP6:%.*]], [[ATOMICRMW_START]] ]
+; CHECK-NEXT:    [[NEW:%.*]] = fsub float [[LOADED]], [[VALUE:%.*]]
+; CHECK-NEXT:    [[TMP2:%.*]] = bitcast float* [[PTR]] to i32*
+; CHECK-NEXT:    [[TMP3:%.*]] = bitcast float [[NEW]] to i32
+; CHECK-NEXT:    [[TMP4:%.*]] = bitcast float [[LOADED]] to i32
+; CHECK-NEXT:    [[TMP5:%.*]] = cmpxchg i32* [[TMP2]], i32 [[TMP4]], i32 [[TMP3]] seq_cst seq_cst
+; CHECK-NEXT:    [[SUCCESS:%.*]] = extractvalue { i32, i1 } [[TMP5]], 1
+; CHECK-NEXT:    [[NEWLOADED:%.*]] = extractvalue { i32, i1 } [[TMP5]], 0
+; CHECK-NEXT:    [[TMP6]] = bitcast i32 [[NEWLOADED]] to float
+; CHECK-NEXT:    br i1 [[SUCCESS]], label [[ATOMICRMW_END:%.*]], label [[ATOMICRMW_START]]
+; CHECK:       atomicrmw.end:
+; CHECK-NEXT:    ret float [[TMP6]]
+;
+  %res = atomicrmw fsub float* %ptr, float %value seq_cst
+  ret float %res
+}
+
+define float @test_atomicrmw_fsub_f32_global(float addrspace(1)* %ptr, float %value) {
+; CHECK-LABEL: @test_atomicrmw_fsub_f32_global(
+; CHECK-NEXT:    [[TMP1:%.*]] = load float, float addrspace(1)* [[PTR:%.*]], align 4
+; CHECK-NEXT:    br label [[ATOMICRMW_START:%.*]]
+; CHECK:       atomicrmw.start:
+; CHECK-NEXT:    [[LOADED:%.*]] = phi float [ [[TMP1]], [[TMP0:%.*]] ], [ [[TMP6:%.*]], [[ATOMICRMW_START]] ]
+; CHECK-NEXT:    [[NEW:%.*]] = fsub float [[LOADED]], [[VALUE:%.*]]
+; CHECK-NEXT:    [[TMP2:%.*]] = bitcast float addrspace(1)* [[PTR]] to i32 addrspace(1)*
+; CHECK-NEXT:    [[TMP3:%.*]] = bitcast float [[NEW]] to i32
+; CHECK-NEXT:    [[TMP4:%.*]] = bitcast float [[LOADED]] to i32
+; CHECK-NEXT:    [[TMP5:%.*]] = cmpxchg i32 addrspace(1)* [[TMP2]], i32 [[TMP4]], i32 [[TMP3]] seq_cst seq_cst
+; CHECK-NEXT:    [[SUCCESS:%.*]] = extractvalue { i32, i1 } [[TMP5]], 1
+; CHECK-NEXT:    [[NEWLOADED:%.*]] = extractvalue { i32, i1 } [[TMP5]], 0
+; CHECK-NEXT:    [[TMP6]] = bitcast i32 [[NEWLOADED]] to float
+; CHECK-NEXT:    br i1 [[SUCCESS]], label [[ATOMICRMW_END:%.*]], label [[ATOMICRMW_START]]
+; CHECK:       atomicrmw.end:
+; CHECK-NEXT:    ret float [[TMP6]]
+;
+  %res = atomicrmw fsub float addrspace(1)* %ptr, float %value seq_cst
+  ret float %res
+}
+
+define float @test_atomicrmw_fsub_f32_local(float addrspace(3)* %ptr, float %value) {
+; CHECK-LABEL: @test_atomicrmw_fsub_f32_local(
+; CHECK-NEXT:    [[TMP1:%.*]] = load float, float addrspace(3)* [[PTR:%.*]], align 4
+; CHECK-NEXT:    br label [[ATOMICRMW_START:%.*]]
+; CHECK:       atomicrmw.start:
+; CHECK-NEXT:    [[LOADED:%.*]] = phi float [ [[TMP1]], [[TMP0:%.*]] ], [ [[TMP6:%.*]], [[ATOMICRMW_START]] ]
+; CHECK-NEXT:    [[NEW:%.*]] = fsub float [[LOADED]], [[VALUE:%.*]]
+; CHECK-NEXT:    [[TMP2:%.*]] = bitcast float addrspace(3)* [[PTR]] to i32 addrspace(3)*
+; CHECK-NEXT:    [[TMP3:%.*]] = bitcast float [[NEW]] to i32
+; CHECK-NEXT:    [[TMP4:%.*]] = bitcast float [[LOADED]] to i32
+; CHECK-NEXT:    [[TMP5:%.*]] = cmpxchg i32 addrspace(3)* [[TMP2]], i32 [[TMP4]], i32 [[TMP3]] seq_cst seq_cst
+; CHECK-NEXT:    [[SUCCESS:%.*]] = extractvalue { i32, i1 } [[TMP5]], 1
+; CHECK-NEXT:    [[NEWLOADED:%.*]] = extractvalue { i32, i1 } [[TMP5]], 0
+; CHECK-NEXT:    [[TMP6]] = bitcast i32 [[NEWLOADED]] to float
+; CHECK-NEXT:    br i1 [[SUCCESS]], label [[ATOMICRMW_END:%.*]], label [[ATOMICRMW_START]]
+; CHECK:       atomicrmw.end:
+; CHECK-NEXT:    ret float [[TMP6]]
+;
+  %res = atomicrmw fsub float addrspace(3)* %ptr, float %value seq_cst
+  ret float %res
+}
+
+define half @test_atomicrmw_fsub_f16_flat(half* %ptr, half %value) {
+; CHECK-LABEL: @test_atomicrmw_fsub_f16_flat(
+; CHECK-NEXT:    [[RES:%.*]] = atomicrmw fsub half* [[PTR:%.*]], half [[VALUE:%.*]] seq_cst
+; CHECK-NEXT:    ret half [[RES]]
+;
+  %res = atomicrmw fsub half* %ptr, half %value seq_cst
+  ret half %res
+}
+
+define half @test_atomicrmw_fsub_f16_global(half addrspace(1)* %ptr, half %value) {
+; CHECK-LABEL: @test_atomicrmw_fsub_f16_global(
+; CHECK-NEXT:    [[RES:%.*]] = atomicrmw fsub half addrspace(1)* [[PTR:%.*]], half [[VALUE:%.*]] seq_cst
+; CHECK-NEXT:    ret half [[RES]]
+;
+  %res = atomicrmw fsub half addrspace(1)* %ptr, half %value seq_cst
+  ret half %res
+}
+
+define half @test_atomicrmw_fsub_f16_local(half addrspace(3)* %ptr, half %value) {
+; CHECK-LABEL: @test_atomicrmw_fsub_f16_local(
+; CHECK-NEXT:    [[RES:%.*]] = atomicrmw fsub half addrspace(3)* [[PTR:%.*]], half [[VALUE:%.*]] seq_cst
+; CHECK-NEXT:    ret half [[RES]]
+;
+  %res = atomicrmw fsub half addrspace(3)* %ptr, half %value seq_cst
+  ret half %res
+}
+
+define double @test_atomicrmw_fsub_f64_flat(double* %ptr, double %value) {
+; CHECK-LABEL: @test_atomicrmw_fsub_f64_flat(
+; CHECK-NEXT:    [[TMP1:%.*]] = load double, double* [[PTR:%.*]], align 8
+; CHECK-NEXT:    br label [[ATOMICRMW_START:%.*]]
+; CHECK:       atomicrmw.start:
+; CHECK-NEXT:    [[LOADED:%.*]] = phi double [ [[TMP1]], [[TMP0:%.*]] ], [ [[TMP6:%.*]], [[ATOMICRMW_START]] ]
+; CHECK-NEXT:    [[NEW:%.*]] = fsub double [[LOADED]], [[VALUE:%.*]]
+; CHECK-NEXT:    [[TMP2:%.*]] = bitcast double* [[PTR]] to i64*
+; CHECK-NEXT:    [[TMP3:%.*]] = bitcast double [[NEW]] to i64
+; CHECK-NEXT:    [[TMP4:%.*]] = bitcast double [[LOADED]] to i64
+; CHECK-NEXT:    [[TMP5:%.*]] = cmpxchg i64* [[TMP2]], i64 [[TMP4]], i64 [[TMP3]] seq_cst seq_cst
+; CHECK-NEXT:    [[SUCCESS:%.*]] = extractvalue { i64, i1 } [[TMP5]], 1
+; CHECK-NEXT:    [[NEWLOADED:%.*]] = extractvalue { i64, i1 } [[TMP5]], 0
+; CHECK-NEXT:    [[TMP6]] = bitcast i64 [[NEWLOADED]] to double
+; CHECK-NEXT:    br i1 [[SUCCESS]], label [[ATOMICRMW_END:%.*]], label [[ATOMICRMW_START]]
+; CHECK:       atomicrmw.end:
+; CHECK-NEXT:    ret double [[TMP6]]
+;
+  %res = atomicrmw fsub double* %ptr, double %value seq_cst
+  ret double %res
+}
+
+define double @test_atomicrmw_fsub_f64_global(double addrspace(1)* %ptr, double %value) {
+; CHECK-LABEL: @test_atomicrmw_fsub_f64_global(
+; CHECK-NEXT:    [[TMP1:%.*]] = load double, double addrspace(1)* [[PTR:%.*]], align 8
+; CHECK-NEXT:    br label [[ATOMICRMW_START:%.*]]
+; CHECK:       atomicrmw.start:
+; CHECK-NEXT:    [[LOADED:%.*]] = phi double [ [[TMP1]], [[TMP0:%.*]] ], [ [[TMP6:%.*]], [[ATOMICRMW_START]] ]
+; CHECK-NEXT:    [[NEW:%.*]] = fsub double [[LOADED]], [[VALUE:%.*]]
+; CHECK-NEXT:    [[TMP2:%.*]] = bitcast double addrspace(1)* [[PTR]] to i64 addrspace(1)*
+; CHECK-NEXT:    [[TMP3:%.*]] = bitcast double [[NEW]] to i64
+; CHECK-NEXT:    [[TMP4:%.*]] = bitcast double [[LOADED]] to i64
+; CHECK-NEXT:    [[TMP5:%.*]] = cmpxchg i64 addrspace(1)* [[TMP2]], i64 [[TMP4]], i64 [[TMP3]] seq_cst seq_cst
+; CHECK-NEXT:    [[SUCCESS:%.*]] = extractvalue { i64, i1 } [[TMP5]], 1
+; CHECK-NEXT:    [[NEWLOADED:%.*]] = extractvalue { i64, i1 } [[TMP5]], 0
+; CHECK-NEXT:    [[TMP6]] = bitcast i64 [[NEWLOADED]] to double
+; CHECK-NEXT:    br i1 [[SUCCESS]], label [[ATOMICRMW_END:%.*]], label [[ATOMICRMW_START]]
+; CHECK:       atomicrmw.end:
+; CHECK-NEXT:    ret double [[TMP6]]
+;
+  %res = atomicrmw fsub double addrspace(1)* %ptr, double %value seq_cst
+  ret double %res
+}
+
+define double @test_atomicrmw_fsub_f64_local(double addrspace(3)* %ptr, double %value) {
+; CHECK-LABEL: @test_atomicrmw_fsub_f64_local(
+; CHECK-NEXT:    [[TMP1:%.*]] = load double, double addrspace(3)* [[PTR:%.*]], align 8
+; CHECK-NEXT:    br label [[ATOMICRMW_START:%.*]]
+; CHECK:       atomicrmw.start:
+; CHECK-NEXT:    [[LOADED:%.*]] = phi double [ [[TMP1]], [[TMP0:%.*]] ], [ [[TMP6:%.*]], [[ATOMICRMW_START]] ]
+; CHECK-NEXT:    [[NEW:%.*]] = fsub double [[LOADED]], [[VALUE:%.*]]
+; CHECK-NEXT:    [[TMP2:%.*]] = bitcast double addrspace(3)* [[PTR]] to i64 addrspace(3)*
+; CHECK-NEXT:    [[TMP3:%.*]] = bitcast double [[NEW]] to i64
+; CHECK-NEXT:    [[TMP4:%.*]] = bitcast double [[LOADED]] to i64
+; CHECK-NEXT:    [[TMP5:%.*]] = cmpxchg i64 addrspace(3)* [[TMP2]], i64 [[TMP4]], i64 [[TMP3]] seq_cst seq_cst
+; CHECK-NEXT:    [[SUCCESS:%.*]] = extractvalue { i64, i1 } [[TMP5]], 1
+; CHECK-NEXT:    [[NEWLOADED:%.*]] = extractvalue { i64, i1 } [[TMP5]], 0
+; CHECK-NEXT:    [[TMP6]] = bitcast i64 [[NEWLOADED]] to double
+; CHECK-NEXT:    br i1 [[SUCCESS]], label [[ATOMICRMW_END:%.*]], label [[ATOMICRMW_START]]
+; CHECK:       atomicrmw.end:
+; CHECK-NEXT:    ret double [[TMP6]]
+;
+  %res = atomicrmw fsub double addrspace(3)* %ptr, double %value seq_cst
+  ret double %res
+}

llvm/test/Transforms/AtomicExpand/NVPTX/expand-atomic-rmw-nand.ll

+; NOTE: Assertions have been autogenerated by utils/update_test_checks.py
+; RUN: opt -mtriple=nvptx64-unknown-unknown -S -atomic-expand %s | FileCheck %s
+; RUN: opt -mtriple=nvptx64-unknown-unknown -S -atomic-expand %s | FileCheck %s
+
+define i32 @test_atomicrmw_nand_i32_flat(i32* %ptr, i32 %value) {
+; CHECK-LABEL: @test_atomicrmw_nand_i32_flat(
+; CHECK-NEXT:    [[RES:%.*]] = atomicrmw nand i32* [[PTR:%.*]], i32 [[VALUE:%.*]] seq_cst
+; CHECK-NEXT:    ret i32 [[RES]]
+;
+  %res = atomicrmw nand i32* %ptr, i32 %value seq_cst
+  ret i32 %res
+}
+
+define i32 @test_atomicrmw_nand_i32_global(i32 addrspace(1)* %ptr, i32 %value) {
+; CHECK-LABEL: @test_atomicrmw_nand_i32_global(
+; CHECK-NEXT:    [[RES:%.*]] = atomicrmw nand i32 addrspace(1)* [[PTR:%.*]], i32 [[VALUE:%.*]] seq_cst
+; CHECK-NEXT:    ret i32 [[RES]]
+;
+  %res = atomicrmw nand i32 addrspace(1)* %ptr, i32 %value seq_cst
+  ret i32 %res
+}
+
+define i32 @test_atomicrmw_nand_i32_local(i32 addrspace(3)* %ptr, i32 %value) {
+; CHECK-LABEL: @test_atomicrmw_nand_i32_local(
+; CHECK-NEXT:    [[RES:%.*]] = atomicrmw nand i32 addrspace(3)* [[PTR:%.*]], i32 [[VALUE:%.*]] seq_cst
+; CHECK-NEXT:    ret i32 [[RES]]
+;
+  %res = atomicrmw nand i32 addrspace(3)* %ptr, i32 %value seq_cst
+  ret i32 %res
+}

llvm/test/Transforms/AtomicExpand/NVPTX/lit.local.cfg

+if not 'NVPTX' in config.root.targets:
+  config.unsupported = True

llvm/test/Transforms/AtomicExpand/NVPTX/unaligned-atomic.ll

+; NOTE: Assertions have been autogenerated by utils/update_test_checks.py
+; RUN: opt -S -mtriple=nvptx64-unknown-unknown -atomic-expand %s | FileCheck -check-prefix=CHECK %s
+
+define i32 @atomic_load_global_align1(i32 addrspace(1)* %ptr) {
+; CHECK-LABEL: @atomic_load_global_align1(
+; CHECK-NEXT:    [[TMP1:%.*]] = bitcast i32 addrspace(1)* [[PTR:%.*]] to i8 addrspace(1)*
+; CHECK-NEXT:    [[TMP2:%.*]] = addrspacecast i8 addrspace(1)* [[TMP1]] to i8*
+; CHECK-NEXT:    [[TMP3:%.*]] = alloca i32, align 4
+; CHECK-NEXT:    [[TMP4:%.*]] = bitcast i32* [[TMP3]] to i8*
+; CHECK-NEXT:    call void @llvm.lifetime.start.p0i8(i64 4, i8* [[TMP4]])
+; CHECK-NEXT:    call void @__atomic_load(i64 4, i8* [[TMP2]], i8* [[TMP4]], i32 5)
+; CHECK-NEXT:    [[TMP5:%.*]] = load i32, i32* [[TMP3]], align 4
+; CHECK-NEXT:    call void @llvm.lifetime.end.p0i8(i64 4, i8* [[TMP4]])
+; CHECK-NEXT:    ret i32 [[TMP5]]
+;
+  %val = load atomic i32, i32 addrspace(1)* %ptr  seq_cst, align 1
+  ret i32 %val
+}
+
+define void @atomic_store_global_align1(i32 addrspace(1)* %ptr, i32 %val) {
+; CHECK-LABEL: @atomic_store_global_align1(
+; CHECK-NEXT:    [[TMP1:%.*]] = bitcast i32 addrspace(1)* [[PTR:%.*]] to i8 addrspace(1)*
+; CHECK-NEXT:    [[TMP2:%.*]] = addrspacecast i8 addrspace(1)* [[TMP1]] to i8*
+; CHECK-NEXT:    [[TMP3:%.*]] = alloca i32, align 4
+; CHECK-NEXT:    [[TMP4:%.*]] = bitcast i32* [[TMP3]] to i8*
+; CHECK-NEXT:    call void @llvm.lifetime.start.p0i8(i64 4, i8* [[TMP4]])
+; CHECK-NEXT:    store i32 [[VAL:%.*]], i32* [[TMP3]], align 4
+; CHECK-NEXT:    call void @__atomic_store(i64 4, i8* [[TMP2]], i8* [[TMP4]], i32 0)
+; CHECK-NEXT:    call void @llvm.lifetime.end.p0i8(i64 4, i8* [[TMP4]])
+; CHECK-NEXT:    ret void
+;
+  store atomic i32 %val, i32 addrspace(1)* %ptr monotonic, align 1
+  ret void
+}