AtomicExpand: Fix expanding atomics into unconstrained FP in strictfp functions

Ideally the normal fadd/fmin/fmax this was creating would fail the verifier.
It's probably also necessary to force off FP exception handlers in the cmpxchg
loop but we don't have a generic way to do that now.

Note strictfp builder is broken in the minnum/maxnum case

https://reviews.llvm.org/D154993

6 files changed, 432 insertions, 189 deletions

diff --git a/llvm/lib/CodeGen/AtomicExpandPass.cpp b/llvm/lib/CodeGen/AtomicExpandPass.cpp
index 8f71ec2..80a0bb9 100644
--- a/llvm/lib/CodeGen/AtomicExpandPass.cpp
+++ b/llvm/lib/CodeGen/AtomicExpandPass.cpp
@@ -1576,6 +1576,11 @@ bool AtomicExpand::tryExpandAtomicCmpXchg(AtomicCmpXchgInst *CI) {
 bool llvm::expandAtomicRMWToCmpXchg(AtomicRMWInst *AI,
                                     CreateCmpXchgInstFun CreateCmpXchg) {
   ReplacementIRBuilder Builder(AI, AI->getModule()->getDataLayout());
+  Builder.setIsFPConstrained(
+      AI->getFunction()->hasFnAttribute(Attribute::StrictFP));
+
+  // FIXME: If FP exceptions are observable, we should force them off for the
+  // loop for the FP atomics.
   Value *Loaded = AtomicExpand::insertRMWCmpXchgLoop(
       Builder, AI->getType(), AI->getPointerOperand(), AI->getAlign(),
       AI->getOrdering(), AI->getSyncScopeID(),
diff --git a/llvm/lib/Transforms/Utils/LowerAtomic.cpp b/llvm/lib/Transforms/Utils/LowerAtomic.cpp
index 5c2d53c..b203970 100644
--- a/llvm/lib/Transforms/Utils/LowerAtomic.cpp
+++ b/llvm/lib/Transforms/Utils/LowerAtomic.cpp
@@ -101,6 +101,9 @@ Value *llvm::buildAtomicRMWValue(AtomicRMWInst::BinOp Op,
 
 bool llvm::lowerAtomicRMWInst(AtomicRMWInst *RMWI) {
   IRBuilder<> Builder(RMWI);
+  Builder.setIsFPConstrained(
+      RMWI->getFunction()->hasFnAttribute(Attribute::StrictFP));
+
   Value *Ptr = RMWI->getPointerOperand();
   Value *Val = RMWI->getValOperand();
 
diff --git a/llvm/test/Transforms/AtomicExpand/AMDGPU/expand-atomic-rmw-fadd.ll b/llvm/test/Transforms/AtomicExpand/AMDGPU/expand-atomic-rmw-fadd.ll
index 6287b5f7..6a29572 100644
--- a/llvm/test/Transforms/AtomicExpand/AMDGPU/expand-atomic-rmw-fadd.ll
+++ b/llvm/test/Transforms/AtomicExpand/AMDGPU/expand-atomic-rmw-fadd.ll
@@ -1901,4 +1901,176 @@ define float @test_atomicrmw_fadd_f32_global_one_as(ptr addrspace(1) %ptr, float
   ret float %res
 }
 
+define void @test_atomicrmw_fadd_f32_global_no_use_unsafe_structfp(ptr addrspace(1) %ptr, float %value) #1 {
+; CI-LABEL: @test_atomicrmw_fadd_f32_global_no_use_unsafe_structfp(
+; CI-NEXT:    [[TMP1:%.*]] = load float, ptr addrspace(1) [[PTR:%.*]], align 4
+; CI-NEXT:    br label [[ATOMICRMW_START:%.*]]
+; CI:       atomicrmw.start:
+; CI-NEXT:    [[LOADED:%.*]] = phi float [ [[TMP1]], [[TMP0:%.*]] ], [ [[TMP5:%.*]], [[ATOMICRMW_START]] ]
+; CI-NEXT:    [[NEW:%.*]] = call float @llvm.experimental.constrained.fadd.f32(float [[LOADED]], float [[VALUE:%.*]], metadata !"round.dynamic", metadata !"fpexcept.strict") #[[ATTR6:[0-9]+]]
+; CI-NEXT:    [[TMP2:%.*]] = bitcast float [[NEW]] to i32
+; CI-NEXT:    [[TMP3:%.*]] = bitcast float [[LOADED]] to i32
+; CI-NEXT:    [[TMP4:%.*]] = cmpxchg ptr addrspace(1) [[PTR]], i32 [[TMP3]], i32 [[TMP2]] syncscope("wavefront") monotonic monotonic, align 4
+; CI-NEXT:    [[SUCCESS:%.*]] = extractvalue { i32, i1 } [[TMP4]], 1
+; CI-NEXT:    [[NEWLOADED:%.*]] = extractvalue { i32, i1 } [[TMP4]], 0
+; CI-NEXT:    [[TMP5]] = bitcast i32 [[NEWLOADED]] to float
+; CI-NEXT:    br i1 [[SUCCESS]], label [[ATOMICRMW_END:%.*]], label [[ATOMICRMW_START]]
+; CI:       atomicrmw.end:
+; CI-NEXT:    ret void
+;
+; GFX9-LABEL: @test_atomicrmw_fadd_f32_global_no_use_unsafe_structfp(
+; GFX9-NEXT:    [[TMP1:%.*]] = load float, ptr addrspace(1) [[PTR:%.*]], align 4
+; GFX9-NEXT:    br label [[ATOMICRMW_START:%.*]]
+; GFX9:       atomicrmw.start:
+; GFX9-NEXT:    [[LOADED:%.*]] = phi float [ [[TMP1]], [[TMP0:%.*]] ], [ [[TMP5:%.*]], [[ATOMICRMW_START]] ]
+; GFX9-NEXT:    [[NEW:%.*]] = call float @llvm.experimental.constrained.fadd.f32(float [[LOADED]], float [[VALUE:%.*]], metadata !"round.dynamic", metadata !"fpexcept.strict") #[[ATTR6:[0-9]+]]
+; GFX9-NEXT:    [[TMP2:%.*]] = bitcast float [[NEW]] to i32
+; GFX9-NEXT:    [[TMP3:%.*]] = bitcast float [[LOADED]] to i32
+; GFX9-NEXT:    [[TMP4:%.*]] = cmpxchg ptr addrspace(1) [[PTR]], i32 [[TMP3]], i32 [[TMP2]] syncscope("wavefront") monotonic monotonic, align 4
+; GFX9-NEXT:    [[SUCCESS:%.*]] = extractvalue { i32, i1 } [[TMP4]], 1
+; GFX9-NEXT:    [[NEWLOADED:%.*]] = extractvalue { i32, i1 } [[TMP4]], 0
+; GFX9-NEXT:    [[TMP5]] = bitcast i32 [[NEWLOADED]] to float
+; GFX9-NEXT:    br i1 [[SUCCESS]], label [[ATOMICRMW_END:%.*]], label [[ATOMICRMW_START]]
+; GFX9:       atomicrmw.end:
+; GFX9-NEXT:    ret void
+;
+; GFX908-LABEL: @test_atomicrmw_fadd_f32_global_no_use_unsafe_structfp(
+; GFX908-NEXT:    [[RES:%.*]] = atomicrmw fadd ptr addrspace(1) [[PTR:%.*]], float [[VALUE:%.*]] syncscope("wavefront") monotonic, align 4
+; GFX908-NEXT:    ret void
+;
+; GFX90A-LABEL: @test_atomicrmw_fadd_f32_global_no_use_unsafe_structfp(
+; GFX90A-NEXT:    [[RES:%.*]] = atomicrmw fadd ptr addrspace(1) [[PTR:%.*]], float [[VALUE:%.*]] syncscope("wavefront") monotonic, align 4
+; GFX90A-NEXT:    ret void
+;
+; GFX940-LABEL: @test_atomicrmw_fadd_f32_global_no_use_unsafe_structfp(
+; GFX940-NEXT:    [[RES:%.*]] = atomicrmw fadd ptr addrspace(1) [[PTR:%.*]], float [[VALUE:%.*]] syncscope("wavefront") monotonic, align 4
+; GFX940-NEXT:    ret void
+;
+; GFX11-LABEL: @test_atomicrmw_fadd_f32_global_no_use_unsafe_structfp(
+; GFX11-NEXT:    [[RES:%.*]] = atomicrmw fadd ptr addrspace(1) [[PTR:%.*]], float [[VALUE:%.*]] syncscope("wavefront") monotonic, align 4
+; GFX11-NEXT:    ret void
+;
+  %res = atomicrmw fadd ptr addrspace(1) %ptr, float %value syncscope("wavefront") monotonic
+  ret void
+}
+
+define double @test_atomicrmw_fadd_f64_global_unsafe_strictfp(ptr addrspace(1) %ptr, double %value) #1 {
+; CI-LABEL: @test_atomicrmw_fadd_f64_global_unsafe_strictfp(
+; CI-NEXT:    [[TMP1:%.*]] = load double, ptr addrspace(1) [[PTR:%.*]], align 8
+; CI-NEXT:    br label [[ATOMICRMW_START:%.*]]
+; CI:       atomicrmw.start:
+; CI-NEXT:    [[LOADED:%.*]] = phi double [ [[TMP1]], [[TMP0:%.*]] ], [ [[TMP5:%.*]], [[ATOMICRMW_START]] ]
+; CI-NEXT:    [[NEW:%.*]] = call double @llvm.experimental.constrained.fadd.f64(double [[LOADED]], double [[VALUE:%.*]], metadata !"round.dynamic", metadata !"fpexcept.strict") #[[ATTR6]]
+; CI-NEXT:    [[TMP2:%.*]] = bitcast double [[NEW]] to i64
+; CI-NEXT:    [[TMP3:%.*]] = bitcast double [[LOADED]] to i64
+; CI-NEXT:    [[TMP4:%.*]] = cmpxchg ptr addrspace(1) [[PTR]], i64 [[TMP3]], i64 [[TMP2]] syncscope("wavefront") monotonic monotonic, align 8
+; CI-NEXT:    [[SUCCESS:%.*]] = extractvalue { i64, i1 } [[TMP4]], 1
+; CI-NEXT:    [[NEWLOADED:%.*]] = extractvalue { i64, i1 } [[TMP4]], 0
+; CI-NEXT:    [[TMP5]] = bitcast i64 [[NEWLOADED]] to double
+; CI-NEXT:    br i1 [[SUCCESS]], label [[ATOMICRMW_END:%.*]], label [[ATOMICRMW_START]]
+; CI:       atomicrmw.end:
+; CI-NEXT:    ret double [[TMP5]]
+;
+; GFX9-LABEL: @test_atomicrmw_fadd_f64_global_unsafe_strictfp(
+; GFX9-NEXT:    [[TMP1:%.*]] = load double, ptr addrspace(1) [[PTR:%.*]], align 8
+; GFX9-NEXT:    br label [[ATOMICRMW_START:%.*]]
+; GFX9:       atomicrmw.start:
+; GFX9-NEXT:    [[LOADED:%.*]] = phi double [ [[TMP1]], [[TMP0:%.*]] ], [ [[TMP5:%.*]], [[ATOMICRMW_START]] ]
+; GFX9-NEXT:    [[NEW:%.*]] = call double @llvm.experimental.constrained.fadd.f64(double [[LOADED]], double [[VALUE:%.*]], metadata !"round.dynamic", metadata !"fpexcept.strict") #[[ATTR6]]
+; GFX9-NEXT:    [[TMP2:%.*]] = bitcast double [[NEW]] to i64
+; GFX9-NEXT:    [[TMP3:%.*]] = bitcast double [[LOADED]] to i64
+; GFX9-NEXT:    [[TMP4:%.*]] = cmpxchg ptr addrspace(1) [[PTR]], i64 [[TMP3]], i64 [[TMP2]] syncscope("wavefront") monotonic monotonic, align 8
+; GFX9-NEXT:    [[SUCCESS:%.*]] = extractvalue { i64, i1 } [[TMP4]], 1
+; GFX9-NEXT:    [[NEWLOADED:%.*]] = extractvalue { i64, i1 } [[TMP4]], 0
+; GFX9-NEXT:    [[TMP5]] = bitcast i64 [[NEWLOADED]] to double
+; GFX9-NEXT:    br i1 [[SUCCESS]], label [[ATOMICRMW_END:%.*]], label [[ATOMICRMW_START]]
+; GFX9:       atomicrmw.end:
+; GFX9-NEXT:    ret double [[TMP5]]
+;
+; GFX908-LABEL: @test_atomicrmw_fadd_f64_global_unsafe_strictfp(
+; GFX908-NEXT:    [[TMP1:%.*]] = load double, ptr addrspace(1) [[PTR:%.*]], align 8
+; GFX908-NEXT:    br label [[ATOMICRMW_START:%.*]]
+; GFX908:       atomicrmw.start:
+; GFX908-NEXT:    [[LOADED:%.*]] = phi double [ [[TMP1]], [[TMP0:%.*]] ], [ [[TMP5:%.*]], [[ATOMICRMW_START]] ]
+; GFX908-NEXT:    [[NEW:%.*]] = call double @llvm.experimental.constrained.fadd.f64(double [[LOADED]], double [[VALUE:%.*]], metadata !"round.dynamic", metadata !"fpexcept.strict") #[[ATTR6:[0-9]+]]
+; GFX908-NEXT:    [[TMP2:%.*]] = bitcast double [[NEW]] to i64
+; GFX908-NEXT:    [[TMP3:%.*]] = bitcast double [[LOADED]] to i64
+; GFX908-NEXT:    [[TMP4:%.*]] = cmpxchg ptr addrspace(1) [[PTR]], i64 [[TMP3]], i64 [[TMP2]] syncscope("wavefront") monotonic monotonic, align 8
+; GFX908-NEXT:    [[SUCCESS:%.*]] = extractvalue { i64, i1 } [[TMP4]], 1
+; GFX908-NEXT:    [[NEWLOADED:%.*]] = extractvalue { i64, i1 } [[TMP4]], 0
+; GFX908-NEXT:    [[TMP5]] = bitcast i64 [[NEWLOADED]] to double
+; GFX908-NEXT:    br i1 [[SUCCESS]], label [[ATOMICRMW_END:%.*]], label [[ATOMICRMW_START]]
+; GFX908:       atomicrmw.end:
+; GFX908-NEXT:    ret double [[TMP5]]
+;
+; GFX90A-LABEL: @test_atomicrmw_fadd_f64_global_unsafe_strictfp(
+; GFX90A-NEXT:    [[RES:%.*]] = atomicrmw fadd ptr addrspace(1) [[PTR:%.*]], double [[VALUE:%.*]] syncscope("wavefront") monotonic, align 8
+; GFX90A-NEXT:    ret double [[RES]]
+;
+; GFX940-LABEL: @test_atomicrmw_fadd_f64_global_unsafe_strictfp(
+; GFX940-NEXT:    [[RES:%.*]] = atomicrmw fadd ptr addrspace(1) [[PTR:%.*]], double [[VALUE:%.*]] syncscope("wavefront") monotonic, align 8
+; GFX940-NEXT:    ret double [[RES]]
+;
+; GFX11-LABEL: @test_atomicrmw_fadd_f64_global_unsafe_strictfp(
+; GFX11-NEXT:    [[TMP1:%.*]] = load double, ptr addrspace(1) [[PTR:%.*]], align 8
+; GFX11-NEXT:    br label [[ATOMICRMW_START:%.*]]
+; GFX11:       atomicrmw.start:
+; GFX11-NEXT:    [[LOADED:%.*]] = phi double [ [[TMP1]], [[TMP0:%.*]] ], [ [[TMP5:%.*]], [[ATOMICRMW_START]] ]
+; GFX11-NEXT:    [[NEW:%.*]] = call double @llvm.experimental.constrained.fadd.f64(double [[LOADED]], double [[VALUE:%.*]], metadata !"round.dynamic", metadata !"fpexcept.strict") #[[ATTR6:[0-9]+]]
+; GFX11-NEXT:    [[TMP2:%.*]] = bitcast double [[NEW]] to i64
+; GFX11-NEXT:    [[TMP3:%.*]] = bitcast double [[LOADED]] to i64
+; GFX11-NEXT:    [[TMP4:%.*]] = cmpxchg ptr addrspace(1) [[PTR]], i64 [[TMP3]], i64 [[TMP2]] syncscope("wavefront") monotonic monotonic, align 8
+; GFX11-NEXT:    [[SUCCESS:%.*]] = extractvalue { i64, i1 } [[TMP4]], 1
+; GFX11-NEXT:    [[NEWLOADED:%.*]] = extractvalue { i64, i1 } [[TMP4]], 0
+; GFX11-NEXT:    [[TMP5]] = bitcast i64 [[NEWLOADED]] to double
+; GFX11-NEXT:    br i1 [[SUCCESS]], label [[ATOMICRMW_END:%.*]], label [[ATOMICRMW_START]]
+; GFX11:       atomicrmw.end:
+; GFX11-NEXT:    ret double [[TMP5]]
+;
+  %res = atomicrmw fadd ptr addrspace(1) %ptr, double %value syncscope("wavefront") monotonic
+  ret double %res
+}
+
+define float @test_atomicrmw_fadd_f32_local_strictfp(ptr addrspace(3) %ptr, float %value) #2 {
+; CI-LABEL: @test_atomicrmw_fadd_f32_local_strictfp(
+; CI-NEXT:    [[TMP1:%.*]] = load float, ptr addrspace(3) [[PTR:%.*]], align 4
+; CI-NEXT:    br label [[ATOMICRMW_START:%.*]]
+; CI:       atomicrmw.start:
+; CI-NEXT:    [[LOADED:%.*]] = phi float [ [[TMP1]], [[TMP0:%.*]] ], [ [[TMP5:%.*]], [[ATOMICRMW_START]] ]
+; CI-NEXT:    [[NEW:%.*]] = call float @llvm.experimental.constrained.fadd.f32(float [[LOADED]], float [[VALUE:%.*]], metadata !"round.dynamic", metadata !"fpexcept.strict") #[[ATTR6]]
+; CI-NEXT:    [[TMP2:%.*]] = bitcast float [[NEW]] to i32
+; CI-NEXT:    [[TMP3:%.*]] = bitcast float [[LOADED]] to i32
+; CI-NEXT:    [[TMP4:%.*]] = cmpxchg ptr addrspace(3) [[PTR]], i32 [[TMP3]], i32 [[TMP2]] seq_cst seq_cst, align 4
+; CI-NEXT:    [[SUCCESS:%.*]] = extractvalue { i32, i1 } [[TMP4]], 1
+; CI-NEXT:    [[NEWLOADED:%.*]] = extractvalue { i32, i1 } [[TMP4]], 0
+; CI-NEXT:    [[TMP5]] = bitcast i32 [[NEWLOADED]] to float
+; CI-NEXT:    br i1 [[SUCCESS]], label [[ATOMICRMW_END:%.*]], label [[ATOMICRMW_START]]
+; CI:       atomicrmw.end:
+; CI-NEXT:    ret float [[TMP5]]
+;
+; GFX9-LABEL: @test_atomicrmw_fadd_f32_local_strictfp(
+; GFX9-NEXT:    [[RES:%.*]] = atomicrmw fadd ptr addrspace(3) [[PTR:%.*]], float [[VALUE:%.*]] seq_cst, align 4
+; GFX9-NEXT:    ret float [[RES]]
+;
+; GFX908-LABEL: @test_atomicrmw_fadd_f32_local_strictfp(
+; GFX908-NEXT:    [[RES:%.*]] = atomicrmw fadd ptr addrspace(3) [[PTR:%.*]], float [[VALUE:%.*]] seq_cst, align 4
+; GFX908-NEXT:    ret float [[RES]]
+;
+; GFX90A-LABEL: @test_atomicrmw_fadd_f32_local_strictfp(
+; GFX90A-NEXT:    [[RES:%.*]] = atomicrmw fadd ptr addrspace(3) [[PTR:%.*]], float [[VALUE:%.*]] seq_cst, align 4
+; GFX90A-NEXT:    ret float [[RES]]
+;
+; GFX940-LABEL: @test_atomicrmw_fadd_f32_local_strictfp(
+; GFX940-NEXT:    [[RES:%.*]] = atomicrmw fadd ptr addrspace(3) [[PTR:%.*]], float [[VALUE:%.*]] seq_cst, align 4
+; GFX940-NEXT:    ret float [[RES]]
+;
+; GFX11-LABEL: @test_atomicrmw_fadd_f32_local_strictfp(
+; GFX11-NEXT:    [[RES:%.*]] = atomicrmw fadd ptr addrspace(3) [[PTR:%.*]], float [[VALUE:%.*]] seq_cst, align 4
+; GFX11-NEXT:    ret float [[RES]]
+;
+  %res = atomicrmw fadd ptr addrspace(3) %ptr, float %value seq_cst
+  ret float %res
+}
 attributes #0 = { "denormal-fp-math-f32"="preserve-sign,preserve-sign" "amdgpu-unsafe-fp-atomics"="true" }
+attributes #1 = { strictfp "denormal-fp-math-f32"="preserve-sign,preserve-sign" "amdgpu-unsafe-fp-atomics"="true" }
+attributes #2 = { strictfp}
diff --git a/llvm/test/Transforms/AtomicExpand/AMDGPU/expand-atomic-rmw-fmax.ll b/llvm/test/Transforms/AtomicExpand/AMDGPU/expand-atomic-rmw-fmax.ll
index 87e39d6..12441d3 100644
--- a/llvm/test/Transforms/AtomicExpand/AMDGPU/expand-atomic-rmw-fmax.ll
+++ b/llvm/test/Transforms/AtomicExpand/AMDGPU/expand-atomic-rmw-fmax.ll
@@ -7,17 +7,17 @@ define float @test_atomicrmw_fmax_f32_flat(ptr %ptr, float %value) {
 ; GCN-NEXT:    [[TMP1:%.*]] = load float, ptr [[PTR:%.*]], align 4
 ; GCN-NEXT:    br label [[ATOMICRMW_START:%.*]]
 ; GCN:       atomicrmw.start:
-; GCN-NEXT:    [[LOADED:%.*]] = phi float [ [[TMP1]], [[TMP0:%.*]] ], [ [[TMP7:%.*]], [[ATOMICRMW_START]] ]
+; GCN-NEXT:    [[LOADED:%.*]] = phi float [ [[TMP1]], [[TMP0:%.*]] ], [ [[TMP6:%.*]], [[ATOMICRMW_START]] ]
 ; GCN-NEXT:    [[TMP2:%.*]] = call float @llvm.maxnum.f32(float [[LOADED]], float [[VALUE:%.*]])
-; GCN-NEXT:    [[TMP4:%.*]] = bitcast float [[TMP2]] to i32
-; GCN-NEXT:    [[TMP5:%.*]] = bitcast float [[LOADED]] to i32
-; GCN-NEXT:    [[TMP6:%.*]] = cmpxchg ptr [[PTR]], i32 [[TMP5]], i32 [[TMP4]] seq_cst seq_cst, align 4
-; GCN-NEXT:    [[SUCCESS:%.*]] = extractvalue { i32, i1 } [[TMP6]], 1
-; GCN-NEXT:    [[NEWLOADED:%.*]] = extractvalue { i32, i1 } [[TMP6]], 0
-; GCN-NEXT:    [[TMP7]] = bitcast i32 [[NEWLOADED]] to float
+; GCN-NEXT:    [[TMP3:%.*]] = bitcast float [[TMP2]] to i32
+; GCN-NEXT:    [[TMP4:%.*]] = bitcast float [[LOADED]] to i32
+; GCN-NEXT:    [[TMP5:%.*]] = cmpxchg ptr [[PTR]], i32 [[TMP4]], i32 [[TMP3]] seq_cst seq_cst, align 4
+; GCN-NEXT:    [[SUCCESS:%.*]] = extractvalue { i32, i1 } [[TMP5]], 1
+; GCN-NEXT:    [[NEWLOADED:%.*]] = extractvalue { i32, i1 } [[TMP5]], 0
+; GCN-NEXT:    [[TMP6]] = bitcast i32 [[NEWLOADED]] to float
 ; GCN-NEXT:    br i1 [[SUCCESS]], label [[ATOMICRMW_END:%.*]], label [[ATOMICRMW_START]]
 ; GCN:       atomicrmw.end:
-; GCN-NEXT:    ret float [[TMP7]]
+; GCN-NEXT:    ret float [[TMP6]]
 ;
   %res = atomicrmw fmax ptr %ptr, float %value seq_cst
   ret float %res
@@ -28,17 +28,17 @@ define float @test_atomicrmw_fmax_f32_global(ptr addrspace(1) %ptr, float %value
 ; GCN-NEXT:    [[TMP1:%.*]] = load float, ptr addrspace(1) [[PTR:%.*]], align 4
 ; GCN-NEXT:    br label [[ATOMICRMW_START:%.*]]
 ; GCN:       atomicrmw.start:
-; GCN-NEXT:    [[LOADED:%.*]] = phi float [ [[TMP1]], [[TMP0:%.*]] ], [ [[TMP7:%.*]], [[ATOMICRMW_START]] ]
+; GCN-NEXT:    [[LOADED:%.*]] = phi float [ [[TMP1]], [[TMP0:%.*]] ], [ [[TMP6:%.*]], [[ATOMICRMW_START]] ]
 ; GCN-NEXT:    [[TMP2:%.*]] = call float @llvm.maxnum.f32(float [[LOADED]], float [[VALUE:%.*]])
-; GCN-NEXT:    [[TMP4:%.*]] = bitcast float [[TMP2]] to i32
-; GCN-NEXT:    [[TMP5:%.*]] = bitcast float [[LOADED]] to i32
-; GCN-NEXT:    [[TMP6:%.*]] = cmpxchg ptr addrspace(1) [[PTR]], i32 [[TMP5]], i32 [[TMP4]] seq_cst seq_cst, align 4
-; GCN-NEXT:    [[SUCCESS:%.*]] = extractvalue { i32, i1 } [[TMP6]], 1
-; GCN-NEXT:    [[NEWLOADED:%.*]] = extractvalue { i32, i1 } [[TMP6]], 0
-; GCN-NEXT:    [[TMP7]] = bitcast i32 [[NEWLOADED]] to float
+; GCN-NEXT:    [[TMP3:%.*]] = bitcast float [[TMP2]] to i32
+; GCN-NEXT:    [[TMP4:%.*]] = bitcast float [[LOADED]] to i32
+; GCN-NEXT:    [[TMP5:%.*]] = cmpxchg ptr addrspace(1) [[PTR]], i32 [[TMP4]], i32 [[TMP3]] seq_cst seq_cst, align 4
+; GCN-NEXT:    [[SUCCESS:%.*]] = extractvalue { i32, i1 } [[TMP5]], 1
+; GCN-NEXT:    [[NEWLOADED:%.*]] = extractvalue { i32, i1 } [[TMP5]], 0
+; GCN-NEXT:    [[TMP6]] = bitcast i32 [[NEWLOADED]] to float
 ; GCN-NEXT:    br i1 [[SUCCESS]], label [[ATOMICRMW_END:%.*]], label [[ATOMICRMW_START]]
 ; GCN:       atomicrmw.end:
-; GCN-NEXT:    ret float [[TMP7]]
+; GCN-NEXT:    ret float [[TMP6]]
 ;
   %res = atomicrmw fmax ptr addrspace(1) %ptr, float %value seq_cst
   ret float %res
@@ -49,17 +49,17 @@ define float @test_atomicrmw_fmax_f32_local(ptr addrspace(3) %ptr, float %value)
 ; GCN-NEXT:    [[TMP1:%.*]] = load float, ptr addrspace(3) [[PTR:%.*]], align 4
 ; GCN-NEXT:    br label [[ATOMICRMW_START:%.*]]
 ; GCN:       atomicrmw.start:
-; GCN-NEXT:    [[LOADED:%.*]] = phi float [ [[TMP1]], [[TMP0:%.*]] ], [ [[TMP7:%.*]], [[ATOMICRMW_START]] ]
+; GCN-NEXT:    [[LOADED:%.*]] = phi float [ [[TMP1]], [[TMP0:%.*]] ], [ [[TMP6:%.*]], [[ATOMICRMW_START]] ]
 ; GCN-NEXT:    [[TMP2:%.*]] = call float @llvm.maxnum.f32(float [[LOADED]], float [[VALUE:%.*]])
-; GCN-NEXT:    [[TMP4:%.*]] = bitcast float [[TMP2]] to i32
-; GCN-NEXT:    [[TMP5:%.*]] = bitcast float [[LOADED]] to i32
-; GCN-NEXT:    [[TMP6:%.*]] = cmpxchg ptr addrspace(3) [[PTR]], i32 [[TMP5]], i32 [[TMP4]] seq_cst seq_cst, align 4
-; GCN-NEXT:    [[SUCCESS:%.*]] = extractvalue { i32, i1 } [[TMP6]], 1
-; GCN-NEXT:    [[NEWLOADED:%.*]] = extractvalue { i32, i1 } [[TMP6]], 0
-; GCN-NEXT:    [[TMP7]] = bitcast i32 [[NEWLOADED]] to float
+; GCN-NEXT:    [[TMP3:%.*]] = bitcast float [[TMP2]] to i32
+; GCN-NEXT:    [[TMP4:%.*]] = bitcast float [[LOADED]] to i32
+; GCN-NEXT:    [[TMP5:%.*]] = cmpxchg ptr addrspace(3) [[PTR]], i32 [[TMP4]], i32 [[TMP3]] seq_cst seq_cst, align 4
+; GCN-NEXT:    [[SUCCESS:%.*]] = extractvalue { i32, i1 } [[TMP5]], 1
+; GCN-NEXT:    [[NEWLOADED:%.*]] = extractvalue { i32, i1 } [[TMP5]], 0
+; GCN-NEXT:    [[TMP6]] = bitcast i32 [[NEWLOADED]] to float
 ; GCN-NEXT:    br i1 [[SUCCESS]], label [[ATOMICRMW_END:%.*]], label [[ATOMICRMW_START]]
 ; GCN:       atomicrmw.end:
-; GCN-NEXT:    ret float [[TMP7]]
+; GCN-NEXT:    ret float [[TMP6]]
 ;
   %res = atomicrmw fmax ptr addrspace(3) %ptr, float %value seq_cst
   ret float %res
@@ -84,17 +84,17 @@ define half @test_atomicrmw_fmax_f16_flat(ptr %ptr, half %value) {
 ; GCN-NEXT:    [[TMP5:%.*]] = call half @llvm.maxnum.f16(half [[TMP4]], half [[VALUE:%.*]])
 ; GCN-NEXT:    [[TMP6:%.*]] = bitcast half [[TMP5]] to i16
 ; GCN-NEXT:    [[EXTENDED:%.*]] = zext i16 [[TMP6]] to i32
-; GCN-NEXT:    [[SHIFTED2:%.*]] = shl nuw i32 [[EXTENDED]], [[SHIFTAMT]]
+; GCN-NEXT:    [[SHIFTED1:%.*]] = shl nuw i32 [[EXTENDED]], [[SHIFTAMT]]
 ; GCN-NEXT:    [[UNMASKED:%.*]] = and i32 [[LOADED]], [[INV_MASK]]
-; GCN-NEXT:    [[INSERTED:%.*]] = or i32 [[UNMASKED]], [[SHIFTED2]]
+; GCN-NEXT:    [[INSERTED:%.*]] = or i32 [[UNMASKED]], [[SHIFTED1]]
 ; GCN-NEXT:    [[TMP7:%.*]] = cmpxchg ptr [[ALIGNEDADDR]], i32 [[LOADED]], i32 [[INSERTED]] seq_cst seq_cst, align 4
 ; GCN-NEXT:    [[SUCCESS:%.*]] = extractvalue { i32, i1 } [[TMP7]], 1
 ; GCN-NEXT:    [[NEWLOADED]] = extractvalue { i32, i1 } [[TMP7]], 0
 ; GCN-NEXT:    br i1 [[SUCCESS]], label [[ATOMICRMW_END:%.*]], label [[ATOMICRMW_START]]
 ; GCN:       atomicrmw.end:
-; GCN-NEXT:    [[SHIFTED3:%.*]] = lshr i32 [[NEWLOADED]], [[SHIFTAMT]]
-; GCN-NEXT:    [[EXTRACTED4:%.*]] = trunc i32 [[SHIFTED3]] to i16
-; GCN-NEXT:    [[TMP8:%.*]] = bitcast i16 [[EXTRACTED4]] to half
+; GCN-NEXT:    [[SHIFTED2:%.*]] = lshr i32 [[NEWLOADED]], [[SHIFTAMT]]
+; GCN-NEXT:    [[EXTRACTED3:%.*]] = trunc i32 [[SHIFTED2]] to i16
+; GCN-NEXT:    [[TMP8:%.*]] = bitcast i16 [[EXTRACTED3]] to half
 ; GCN-NEXT:    ret half [[TMP8]]
 ;
   %res = atomicrmw fmax ptr %ptr, half %value seq_cst
@@ -120,17 +120,17 @@ define half @test_atomicrmw_fmax_f16_global(ptr addrspace(1) %ptr, half %value)
 ; GCN-NEXT:    [[TMP5:%.*]] = call half @llvm.maxnum.f16(half [[TMP4]], half [[VALUE:%.*]])
 ; GCN-NEXT:    [[TMP6:%.*]] = bitcast half [[TMP5]] to i16
 ; GCN-NEXT:    [[EXTENDED:%.*]] = zext i16 [[TMP6]] to i32
-; GCN-NEXT:    [[SHIFTED2:%.*]] = shl nuw i32 [[EXTENDED]], [[SHIFTAMT]]
+; GCN-NEXT:    [[SHIFTED1:%.*]] = shl nuw i32 [[EXTENDED]], [[SHIFTAMT]]
 ; GCN-NEXT:    [[UNMASKED:%.*]] = and i32 [[LOADED]], [[INV_MASK]]
-; GCN-NEXT:    [[INSERTED:%.*]] = or i32 [[UNMASKED]], [[SHIFTED2]]
+; GCN-NEXT:    [[INSERTED:%.*]] = or i32 [[UNMASKED]], [[SHIFTED1]]
 ; GCN-NEXT:    [[TMP7:%.*]] = cmpxchg ptr addrspace(1) [[ALIGNEDADDR]], i32 [[LOADED]], i32 [[INSERTED]] seq_cst seq_cst, align 4
 ; GCN-NEXT:    [[SUCCESS:%.*]] = extractvalue { i32, i1 } [[TMP7]], 1
 ; GCN-NEXT:    [[NEWLOADED]] = extractvalue { i32, i1 } [[TMP7]], 0
 ; GCN-NEXT:    br i1 [[SUCCESS]], label [[ATOMICRMW_END:%.*]], label [[ATOMICRMW_START]]
 ; GCN:       atomicrmw.end:
-; GCN-NEXT:    [[SHIFTED3:%.*]] = lshr i32 [[NEWLOADED]], [[SHIFTAMT]]
-; GCN-NEXT:    [[EXTRACTED4:%.*]] = trunc i32 [[SHIFTED3]] to i16
-; GCN-NEXT:    [[TMP8:%.*]] = bitcast i16 [[EXTRACTED4]] to half
+; GCN-NEXT:    [[SHIFTED2:%.*]] = lshr i32 [[NEWLOADED]], [[SHIFTAMT]]
+; GCN-NEXT:    [[EXTRACTED3:%.*]] = trunc i32 [[SHIFTED2]] to i16
+; GCN-NEXT:    [[TMP8:%.*]] = bitcast i16 [[EXTRACTED3]] to half
 ; GCN-NEXT:    ret half [[TMP8]]
 ;
   %res = atomicrmw fmax ptr addrspace(1) %ptr, half %value seq_cst
@@ -182,17 +182,17 @@ define half @test_atomicrmw_fmax_f16_local(ptr addrspace(3) %ptr, half %value) {
 ; GCN-NEXT:    [[TMP5:%.*]] = call half @llvm.maxnum.f16(half [[TMP4]], half [[VALUE:%.*]])
 ; GCN-NEXT:    [[TMP6:%.*]] = bitcast half [[TMP5]] to i16
 ; GCN-NEXT:    [[EXTENDED:%.*]] = zext i16 [[TMP6]] to i32
-; GCN-NEXT:    [[SHIFTED2:%.*]] = shl nuw i32 [[EXTENDED]], [[SHIFTAMT]]
+; GCN-NEXT:    [[SHIFTED1:%.*]] = shl nuw i32 [[EXTENDED]], [[SHIFTAMT]]
 ; GCN-NEXT:    [[UNMASKED:%.*]] = and i32 [[LOADED]], [[INV_MASK]]
-; GCN-NEXT:    [[INSERTED:%.*]] = or i32 [[UNMASKED]], [[SHIFTED2]]
+; GCN-NEXT:    [[INSERTED:%.*]] = or i32 [[UNMASKED]], [[SHIFTED1]]
 ; GCN-NEXT:    [[TMP7:%.*]] = cmpxchg ptr addrspace(3) [[ALIGNEDADDR]], i32 [[LOADED]], i32 [[INSERTED]] seq_cst seq_cst, align 4
 ; GCN-NEXT:    [[SUCCESS:%.*]] = extractvalue { i32, i1 } [[TMP7]], 1
 ; GCN-NEXT:    [[NEWLOADED]] = extractvalue { i32, i1 } [[TMP7]], 0
 ; GCN-NEXT:    br i1 [[SUCCESS]], label [[ATOMICRMW_END:%.*]], label [[ATOMICRMW_START]]
 ; GCN:       atomicrmw.end:
-; GCN-NEXT:    [[SHIFTED3:%.*]] = lshr i32 [[NEWLOADED]], [[SHIFTAMT]]
-; GCN-NEXT:    [[EXTRACTED4:%.*]] = trunc i32 [[SHIFTED3]] to i16
-; GCN-NEXT:    [[TMP8:%.*]] = bitcast i16 [[EXTRACTED4]] to half
+; GCN-NEXT:    [[SHIFTED2:%.*]] = lshr i32 [[NEWLOADED]], [[SHIFTAMT]]
+; GCN-NEXT:    [[EXTRACTED3:%.*]] = trunc i32 [[SHIFTED2]] to i16
+; GCN-NEXT:    [[TMP8:%.*]] = bitcast i16 [[EXTRACTED3]] to half
 ; GCN-NEXT:    ret half [[TMP8]]
 ;
   %res = atomicrmw fmax ptr addrspace(3) %ptr, half %value seq_cst
@@ -204,17 +204,17 @@ define double @test_atomicrmw_fmax_f64_flat(ptr %ptr, double %value) {
 ; GCN-NEXT:    [[TMP1:%.*]] = load double, ptr [[PTR:%.*]], align 8
 ; GCN-NEXT:    br label [[ATOMICRMW_START:%.*]]
 ; GCN:       atomicrmw.start:
-; GCN-NEXT:    [[LOADED:%.*]] = phi double [ [[TMP1]], [[TMP0:%.*]] ], [ [[TMP7:%.*]], [[ATOMICRMW_START]] ]
+; GCN-NEXT:    [[LOADED:%.*]] = phi double [ [[TMP1]], [[TMP0:%.*]] ], [ [[TMP6:%.*]], [[ATOMICRMW_START]] ]
 ; GCN-NEXT:    [[TMP2:%.*]] = call double @llvm.maxnum.f64(double [[LOADED]], double [[VALUE:%.*]])
-; GCN-NEXT:    [[TMP4:%.*]] = bitcast double [[TMP2]] to i64
-; GCN-NEXT:    [[TMP5:%.*]] = bitcast double [[LOADED]] to i64
-; GCN-NEXT:    [[TMP6:%.*]] = cmpxchg ptr [[PTR]], i64 [[TMP5]], i64 [[TMP4]] seq_cst seq_cst, align 8
-; GCN-NEXT:    [[SUCCESS:%.*]] = extractvalue { i64, i1 } [[TMP6]], 1
-; GCN-NEXT:    [[NEWLOADED:%.*]] = extractvalue { i64, i1 } [[TMP6]], 0
-; GCN-NEXT:    [[TMP7]] = bitcast i64 [[NEWLOADED]] to double
+; GCN-NEXT:    [[TMP3:%.*]] = bitcast double [[TMP2]] to i64
+; GCN-NEXT:    [[TMP4:%.*]] = bitcast double [[LOADED]] to i64
+; GCN-NEXT:    [[TMP5:%.*]] = cmpxchg ptr [[PTR]], i64 [[TMP4]], i64 [[TMP3]] seq_cst seq_cst, align 8
+; GCN-NEXT:    [[SUCCESS:%.*]] = extractvalue { i64, i1 } [[TMP5]], 1
+; GCN-NEXT:    [[NEWLOADED:%.*]] = extractvalue { i64, i1 } [[TMP5]], 0
+; GCN-NEXT:    [[TMP6]] = bitcast i64 [[NEWLOADED]] to double
 ; GCN-NEXT:    br i1 [[SUCCESS]], label [[ATOMICRMW_END:%.*]], label [[ATOMICRMW_START]]
 ; GCN:       atomicrmw.end:
-; GCN-NEXT:    ret double [[TMP7]]
+; GCN-NEXT:    ret double [[TMP6]]
 ;
   %res = atomicrmw fmax ptr %ptr, double %value seq_cst
   ret double %res
@@ -225,17 +225,17 @@ define double @test_atomicrmw_fmax_f64_global(ptr addrspace(1) %ptr, double %val
 ; GCN-NEXT:    [[TMP1:%.*]] = load double, ptr addrspace(1) [[PTR:%.*]], align 8
 ; GCN-NEXT:    br label [[ATOMICRMW_START:%.*]]
 ; GCN:       atomicrmw.start:
-; GCN-NEXT:    [[LOADED:%.*]] = phi double [ [[TMP1]], [[TMP0:%.*]] ], [ [[TMP7:%.*]], [[ATOMICRMW_START]] ]
+; GCN-NEXT:    [[LOADED:%.*]] = phi double [ [[TMP1]], [[TMP0:%.*]] ], [ [[TMP6:%.*]], [[ATOMICRMW_START]] ]
 ; GCN-NEXT:    [[TMP2:%.*]] = call double @llvm.maxnum.f64(double [[LOADED]], double [[VALUE:%.*]])
-; GCN-NEXT:    [[TMP4:%.*]] = bitcast double [[TMP2]] to i64
-; GCN-NEXT:    [[TMP5:%.*]] = bitcast double [[LOADED]] to i64
-; GCN-NEXT:    [[TMP6:%.*]] = cmpxchg ptr addrspace(1) [[PTR]], i64 [[TMP5]], i64 [[TMP4]] seq_cst seq_cst, align 8
-; GCN-NEXT:    [[SUCCESS:%.*]] = extractvalue { i64, i1 } [[TMP6]], 1
-; GCN-NEXT:    [[NEWLOADED:%.*]] = extractvalue { i64, i1 } [[TMP6]], 0
-; GCN-NEXT:    [[TMP7]] = bitcast i64 [[NEWLOADED]] to double
+; GCN-NEXT:    [[TMP3:%.*]] = bitcast double [[TMP2]] to i64
+; GCN-NEXT:    [[TMP4:%.*]] = bitcast double [[LOADED]] to i64
+; GCN-NEXT:    [[TMP5:%.*]] = cmpxchg ptr addrspace(1) [[PTR]], i64 [[TMP4]], i64 [[TMP3]] seq_cst seq_cst, align 8
+; GCN-NEXT:    [[SUCCESS:%.*]] = extractvalue { i64, i1 } [[TMP5]], 1
+; GCN-NEXT:    [[NEWLOADED:%.*]] = extractvalue { i64, i1 } [[TMP5]], 0
+; GCN-NEXT:    [[TMP6]] = bitcast i64 [[NEWLOADED]] to double
 ; GCN-NEXT:    br i1 [[SUCCESS]], label [[ATOMICRMW_END:%.*]], label [[ATOMICRMW_START]]
 ; GCN:       atomicrmw.end:
-; GCN-NEXT:    ret double [[TMP7]]
+; GCN-NEXT:    ret double [[TMP6]]
 ;
   %res = atomicrmw fmax ptr addrspace(1) %ptr, double %value seq_cst
   ret double %res
@@ -246,18 +246,39 @@ define double @test_atomicrmw_fmax_f64_local(ptr addrspace(3) %ptr, double %valu
 ; GCN-NEXT:    [[TMP1:%.*]] = load double, ptr addrspace(3) [[PTR:%.*]], align 8
 ; GCN-NEXT:    br label [[ATOMICRMW_START:%.*]]
 ; GCN:       atomicrmw.start:
-; GCN-NEXT:    [[LOADED:%.*]] = phi double [ [[TMP1]], [[TMP0:%.*]] ], [ [[TMP7:%.*]], [[ATOMICRMW_START]] ]
+; GCN-NEXT:    [[LOADED:%.*]] = phi double [ [[TMP1]], [[TMP0:%.*]] ], [ [[TMP6:%.*]], [[ATOMICRMW_START]] ]
 ; GCN-NEXT:    [[TMP2:%.*]] = call double @llvm.maxnum.f64(double [[LOADED]], double [[VALUE:%.*]])
-; GCN-NEXT:    [[TMP4:%.*]] = bitcast double [[TMP2]] to i64
-; GCN-NEXT:    [[TMP5:%.*]] = bitcast double [[LOADED]] to i64
-; GCN-NEXT:    [[TMP6:%.*]] = cmpxchg ptr addrspace(3) [[PTR]], i64 [[TMP5]], i64 [[TMP4]] seq_cst seq_cst, align 8
-; GCN-NEXT:    [[SUCCESS:%.*]] = extractvalue { i64, i1 } [[TMP6]], 1
-; GCN-NEXT:    [[NEWLOADED:%.*]] = extractvalue { i64, i1 } [[TMP6]], 0
-; GCN-NEXT:    [[TMP7]] = bitcast i64 [[NEWLOADED]] to double
+; GCN-NEXT:    [[TMP3:%.*]] = bitcast double [[TMP2]] to i64
+; GCN-NEXT:    [[TMP4:%.*]] = bitcast double [[LOADED]] to i64
+; GCN-NEXT:    [[TMP5:%.*]] = cmpxchg ptr addrspace(3) [[PTR]], i64 [[TMP4]], i64 [[TMP3]] seq_cst seq_cst, align 8
+; GCN-NEXT:    [[SUCCESS:%.*]] = extractvalue { i64, i1 } [[TMP5]], 1
+; GCN-NEXT:    [[NEWLOADED:%.*]] = extractvalue { i64, i1 } [[TMP5]], 0
+; GCN-NEXT:    [[TMP6]] = bitcast i64 [[NEWLOADED]] to double
 ; GCN-NEXT:    br i1 [[SUCCESS]], label [[ATOMICRMW_END:%.*]], label [[ATOMICRMW_START]]
 ; GCN:       atomicrmw.end:
-; GCN-NEXT:    ret double [[TMP7]]
+; GCN-NEXT:    ret double [[TMP6]]
 ;
   %res = atomicrmw fmax ptr addrspace(3) %ptr, double %value seq_cst
   ret double %res
 }
+
+define double @test_atomicrmw_fmax_f64_global_strictfp(ptr addrspace(1) %ptr, double %value) strictfp {
+; GCN-LABEL: @test_atomicrmw_fmax_f64_global_strictfp(
+; GCN-NEXT:    [[TMP1:%.*]] = load double, ptr addrspace(1) [[PTR:%.*]], align 8
+; GCN-NEXT:    br label [[ATOMICRMW_START:%.*]]
+; GCN:       atomicrmw.start:
+; GCN-NEXT:    [[LOADED:%.*]] = phi double [ [[TMP1]], [[TMP0:%.*]] ], [ [[TMP6:%.*]], [[ATOMICRMW_START]] ]
+; GCN-NEXT:    [[TMP2:%.*]] = call double @llvm.maxnum.f64(double [[LOADED]], double [[VALUE:%.*]]) #[[ATTR3:[0-9]+]]
+; GCN-NEXT:    [[TMP3:%.*]] = bitcast double [[TMP2]] to i64
+; GCN-NEXT:    [[TMP4:%.*]] = bitcast double [[LOADED]] to i64
+; GCN-NEXT:    [[TMP5:%.*]] = cmpxchg ptr addrspace(1) [[PTR]], i64 [[TMP4]], i64 [[TMP3]] seq_cst seq_cst, align 8
+; GCN-NEXT:    [[SUCCESS:%.*]] = extractvalue { i64, i1 } [[TMP5]], 1
+; GCN-NEXT:    [[NEWLOADED:%.*]] = extractvalue { i64, i1 } [[TMP5]], 0
+; GCN-NEXT:    [[TMP6]] = bitcast i64 [[NEWLOADED]] to double
+; GCN-NEXT:    br i1 [[SUCCESS]], label [[ATOMICRMW_END:%.*]], label [[ATOMICRMW_START]]
+; GCN:       atomicrmw.end:
+; GCN-NEXT:    ret double [[TMP6]]
+;
+  %res = atomicrmw fmax ptr addrspace(1) %ptr, double %value seq_cst
+  ret double %res
+}
diff --git a/llvm/test/Transforms/AtomicExpand/AMDGPU/expand-atomic-rmw-fmin.ll b/llvm/test/Transforms/AtomicExpand/AMDGPU/expand-atomic-rmw-fmin.ll
index c1ed6b8..3eda559 100644
--- a/llvm/test/Transforms/AtomicExpand/AMDGPU/expand-atomic-rmw-fmin.ll
+++ b/llvm/test/Transforms/AtomicExpand/AMDGPU/expand-atomic-rmw-fmin.ll
@@ -7,17 +7,17 @@ define float @test_atomicrmw_fmin_f32_flat(ptr %ptr, float %value) {
 ; GCN-NEXT:    [[TMP1:%.*]] = load float, ptr [[PTR:%.*]], align 4
 ; GCN-NEXT:    br label [[ATOMICRMW_START:%.*]]
 ; GCN:       atomicrmw.start:
-; GCN-NEXT:    [[LOADED:%.*]] = phi float [ [[TMP1]], [[TMP0:%.*]] ], [ [[TMP7:%.*]], [[ATOMICRMW_START]] ]
+; GCN-NEXT:    [[LOADED:%.*]] = phi float [ [[TMP1]], [[TMP0:%.*]] ], [ [[TMP6:%.*]], [[ATOMICRMW_START]] ]
 ; GCN-NEXT:    [[TMP2:%.*]] = call float @llvm.minnum.f32(float [[LOADED]], float [[VALUE:%.*]])
-; GCN-NEXT:    [[TMP4:%.*]] = bitcast float [[TMP2]] to i32
-; GCN-NEXT:    [[TMP5:%.*]] = bitcast float [[LOADED]] to i32
-; GCN-NEXT:    [[TMP6:%.*]] = cmpxchg ptr [[PTR]], i32 [[TMP5]], i32 [[TMP4]] seq_cst seq_cst, align 4
-; GCN-NEXT:    [[SUCCESS:%.*]] = extractvalue { i32, i1 } [[TMP6]], 1
-; GCN-NEXT:    [[NEWLOADED:%.*]] = extractvalue { i32, i1 } [[TMP6]], 0
-; GCN-NEXT:    [[TMP7]] = bitcast i32 [[NEWLOADED]] to float
+; GCN-NEXT:    [[TMP3:%.*]] = bitcast float [[TMP2]] to i32
+; GCN-NEXT:    [[TMP4:%.*]] = bitcast float [[LOADED]] to i32
+; GCN-NEXT:    [[TMP5:%.*]] = cmpxchg ptr [[PTR]], i32 [[TMP4]], i32 [[TMP3]] seq_cst seq_cst, align 4
+; GCN-NEXT:    [[SUCCESS:%.*]] = extractvalue { i32, i1 } [[TMP5]], 1
+; GCN-NEXT:    [[NEWLOADED:%.*]] = extractvalue { i32, i1 } [[TMP5]], 0
+; GCN-NEXT:    [[TMP6]] = bitcast i32 [[NEWLOADED]] to float
 ; GCN-NEXT:    br i1 [[SUCCESS]], label [[ATOMICRMW_END:%.*]], label [[ATOMICRMW_START]]
 ; GCN:       atomicrmw.end:
-; GCN-NEXT:    ret float [[TMP7]]
+; GCN-NEXT:    ret float [[TMP6]]
 ;
   %res = atomicrmw fmin ptr %ptr, float %value seq_cst
   ret float %res
@@ -28,17 +28,17 @@ define float @test_atomicrmw_fmin_f32_global(ptr addrspace(1) %ptr, float %value
 ; GCN-NEXT:    [[TMP1:%.*]] = load float, ptr addrspace(1) [[PTR:%.*]], align 4
 ; GCN-NEXT:    br label [[ATOMICRMW_START:%.*]]
 ; GCN:       atomicrmw.start:
-; GCN-NEXT:    [[LOADED:%.*]] = phi float [ [[TMP1]], [[TMP0:%.*]] ], [ [[TMP7:%.*]], [[ATOMICRMW_START]] ]
+; GCN-NEXT:    [[LOADED:%.*]] = phi float [ [[TMP1]], [[TMP0:%.*]] ], [ [[TMP6:%.*]], [[ATOMICRMW_START]] ]
 ; GCN-NEXT:    [[TMP2:%.*]] = call float @llvm.minnum.f32(float [[LOADED]], float [[VALUE:%.*]])
-; GCN-NEXT:    [[TMP4:%.*]] = bitcast float [[TMP2]] to i32
-; GCN-NEXT:    [[TMP5:%.*]] = bitcast float [[LOADED]] to i32
-; GCN-NEXT:    [[TMP6:%.*]] = cmpxchg ptr addrspace(1) [[PTR]], i32 [[TMP5]], i32 [[TMP4]] seq_cst seq_cst, align 4
-; GCN-NEXT:    [[SUCCESS:%.*]] = extractvalue { i32, i1 } [[TMP6]], 1
-; GCN-NEXT:    [[NEWLOADED:%.*]] = extractvalue { i32, i1 } [[TMP6]], 0
-; GCN-NEXT:    [[TMP7]] = bitcast i32 [[NEWLOADED]] to float
+; GCN-NEXT:    [[TMP3:%.*]] = bitcast float [[TMP2]] to i32
+; GCN-NEXT:    [[TMP4:%.*]] = bitcast float [[LOADED]] to i32
+; GCN-NEXT:    [[TMP5:%.*]] = cmpxchg ptr addrspace(1) [[PTR]], i32 [[TMP4]], i32 [[TMP3]] seq_cst seq_cst, align 4
+; GCN-NEXT:    [[SUCCESS:%.*]] = extractvalue { i32, i1 } [[TMP5]], 1
+; GCN-NEXT:    [[NEWLOADED:%.*]] = extractvalue { i32, i1 } [[TMP5]], 0
+; GCN-NEXT:    [[TMP6]] = bitcast i32 [[NEWLOADED]] to float
 ; GCN-NEXT:    br i1 [[SUCCESS]], label [[ATOMICRMW_END:%.*]], label [[ATOMICRMW_START]]
 ; GCN:       atomicrmw.end:
-; GCN-NEXT:    ret float [[TMP7]]
+; GCN-NEXT:    ret float [[TMP6]]
 ;
   %res = atomicrmw fmin ptr addrspace(1) %ptr, float %value seq_cst
   ret float %res
@@ -49,17 +49,17 @@ define float @test_atomicrmw_fmin_f32_local(ptr addrspace(3) %ptr, float %value)
 ; GCN-NEXT:    [[TMP1:%.*]] = load float, ptr addrspace(3) [[PTR:%.*]], align 4
 ; GCN-NEXT:    br label [[ATOMICRMW_START:%.*]]
 ; GCN:       atomicrmw.start:
-; GCN-NEXT:    [[LOADED:%.*]] = phi float [ [[TMP1]], [[TMP0:%.*]] ], [ [[TMP7:%.*]], [[ATOMICRMW_START]] ]
+; GCN-NEXT:    [[LOADED:%.*]] = phi float [ [[TMP1]], [[TMP0:%.*]] ], [ [[TMP6:%.*]], [[ATOMICRMW_START]] ]
 ; GCN-NEXT:    [[TMP2:%.*]] = call float @llvm.minnum.f32(float [[LOADED]], float [[VALUE:%.*]])
-; GCN-NEXT:    [[TMP4:%.*]] = bitcast float [[TMP2]] to i32
-; GCN-NEXT:    [[TMP5:%.*]] = bitcast float [[LOADED]] to i32
-; GCN-NEXT:    [[TMP6:%.*]] = cmpxchg ptr addrspace(3) [[PTR]], i32 [[TMP5]], i32 [[TMP4]] seq_cst seq_cst, align 4
-; GCN-NEXT:    [[SUCCESS:%.*]] = extractvalue { i32, i1 } [[TMP6]], 1
-; GCN-NEXT:    [[NEWLOADED:%.*]] = extractvalue { i32, i1 } [[TMP6]], 0
-; GCN-NEXT:    [[TMP7]] = bitcast i32 [[NEWLOADED]] to float
+; GCN-NEXT:    [[TMP3:%.*]] = bitcast float [[TMP2]] to i32
+; GCN-NEXT:    [[TMP4:%.*]] = bitcast float [[LOADED]] to i32
+; GCN-NEXT:    [[TMP5:%.*]] = cmpxchg ptr addrspace(3) [[PTR]], i32 [[TMP4]], i32 [[TMP3]] seq_cst seq_cst, align 4
+; GCN-NEXT:    [[SUCCESS:%.*]] = extractvalue { i32, i1 } [[TMP5]], 1
+; GCN-NEXT:    [[NEWLOADED:%.*]] = extractvalue { i32, i1 } [[TMP5]], 0
+; GCN-NEXT:    [[TMP6]] = bitcast i32 [[NEWLOADED]] to float
 ; GCN-NEXT:    br i1 [[SUCCESS]], label [[ATOMICRMW_END:%.*]], label [[ATOMICRMW_START]]
 ; GCN:       atomicrmw.end:
-; GCN-NEXT:    ret float [[TMP7]]
+; GCN-NEXT:    ret float [[TMP6]]
 ;
   %res = atomicrmw fmin ptr addrspace(3) %ptr, float %value seq_cst
   ret float %res
@@ -84,17 +84,17 @@ define half @test_atomicrmw_fmin_f16_flat(ptr %ptr, half %value) {
 ; GCN-NEXT:    [[TMP5:%.*]] = call half @llvm.minnum.f16(half [[TMP4]], half [[VALUE:%.*]])
 ; GCN-NEXT:    [[TMP6:%.*]] = bitcast half [[TMP5]] to i16
 ; GCN-NEXT:    [[EXTENDED:%.*]] = zext i16 [[TMP6]] to i32
-; GCN-NEXT:    [[SHIFTED2:%.*]] = shl nuw i32 [[EXTENDED]], [[SHIFTAMT]]
+; GCN-NEXT:    [[SHIFTED1:%.*]] = shl nuw i32 [[EXTENDED]], [[SHIFTAMT]]
 ; GCN-NEXT:    [[UNMASKED:%.*]] = and i32 [[LOADED]], [[INV_MASK]]
-; GCN-NEXT:    [[INSERTED:%.*]] = or i32 [[UNMASKED]], [[SHIFTED2]]
+; GCN-NEXT:    [[INSERTED:%.*]] = or i32 [[UNMASKED]], [[SHIFTED1]]
 ; GCN-NEXT:    [[TMP7:%.*]] = cmpxchg ptr [[ALIGNEDADDR]], i32 [[LOADED]], i32 [[INSERTED]] seq_cst seq_cst, align 4
 ; GCN-NEXT:    [[SUCCESS:%.*]] = extractvalue { i32, i1 } [[TMP7]], 1
 ; GCN-NEXT:    [[NEWLOADED]] = extractvalue { i32, i1 } [[TMP7]], 0
 ; GCN-NEXT:    br i1 [[SUCCESS]], label [[ATOMICRMW_END:%.*]], label [[ATOMICRMW_START]]
 ; GCN:       atomicrmw.end:
-; GCN-NEXT:    [[SHIFTED3:%.*]] = lshr i32 [[NEWLOADED]], [[SHIFTAMT]]
-; GCN-NEXT:    [[EXTRACTED4:%.*]] = trunc i32 [[SHIFTED3]] to i16
-; GCN-NEXT:    [[TMP8:%.*]] = bitcast i16 [[EXTRACTED4]] to half
+; GCN-NEXT:    [[SHIFTED2:%.*]] = lshr i32 [[NEWLOADED]], [[SHIFTAMT]]
+; GCN-NEXT:    [[EXTRACTED3:%.*]] = trunc i32 [[SHIFTED2]] to i16
+; GCN-NEXT:    [[TMP8:%.*]] = bitcast i16 [[EXTRACTED3]] to half
 ; GCN-NEXT:    ret half [[TMP8]]
 ;
   %res = atomicrmw fmin ptr %ptr, half %value seq_cst
@@ -120,17 +120,17 @@ define half @test_atomicrmw_fmin_f16_global(ptr addrspace(1) %ptr, half %value)
 ; GCN-NEXT:    [[TMP5:%.*]] = call half @llvm.minnum.f16(half [[TMP4]], half [[VALUE:%.*]])
 ; GCN-NEXT:    [[TMP6:%.*]] = bitcast half [[TMP5]] to i16
 ; GCN-NEXT:    [[EXTENDED:%.*]] = zext i16 [[TMP6]] to i32
-; GCN-NEXT:    [[SHIFTED2:%.*]] = shl nuw i32 [[EXTENDED]], [[SHIFTAMT]]
+; GCN-NEXT:    [[SHIFTED1:%.*]] = shl nuw i32 [[EXTENDED]], [[SHIFTAMT]]
 ; GCN-NEXT:    [[UNMASKED:%.*]] = and i32 [[LOADED]], [[INV_MASK]]
-; GCN-NEXT:    [[INSERTED:%.*]] = or i32 [[UNMASKED]], [[SHIFTED2]]
+; GCN-NEXT:    [[INSERTED:%.*]] = or i32 [[UNMASKED]], [[SHIFTED1]]
 ; GCN-NEXT:    [[TMP7:%.*]] = cmpxchg ptr addrspace(1) [[ALIGNEDADDR]], i32 [[LOADED]], i32 [[INSERTED]] seq_cst seq_cst, align 4
 ; GCN-NEXT:    [[SUCCESS:%.*]] = extractvalue { i32, i1 } [[TMP7]], 1
 ; GCN-NEXT:    [[NEWLOADED]] = extractvalue { i32, i1 } [[TMP7]], 0
 ; GCN-NEXT:    br i1 [[SUCCESS]], label [[ATOMICRMW_END:%.*]], label [[ATOMICRMW_START]]
 ; GCN:       atomicrmw.end:
-; GCN-NEXT:    [[SHIFTED3:%.*]] = lshr i32 [[NEWLOADED]], [[SHIFTAMT]]
-; GCN-NEXT:    [[EXTRACTED4:%.*]] = trunc i32 [[SHIFTED3]] to i16
-; GCN-NEXT:    [[TMP8:%.*]] = bitcast i16 [[EXTRACTED4]] to half
+; GCN-NEXT:    [[SHIFTED2:%.*]] = lshr i32 [[NEWLOADED]], [[SHIFTAMT]]
+; GCN-NEXT:    [[EXTRACTED3:%.*]] = trunc i32 [[SHIFTED2]] to i16
+; GCN-NEXT:    [[TMP8:%.*]] = bitcast i16 [[EXTRACTED3]] to half
 ; GCN-NEXT:    ret half [[TMP8]]
 ;
   %res = atomicrmw fmin ptr addrspace(1) %ptr, half %value seq_cst
@@ -182,17 +182,17 @@ define half @test_atomicrmw_fmin_f16_local(ptr addrspace(3) %ptr, half %value) {
 ; GCN-NEXT:    [[TMP5:%.*]] = call half @llvm.minnum.f16(half [[TMP4]], half [[VALUE:%.*]])
 ; GCN-NEXT:    [[TMP6:%.*]] = bitcast half [[TMP5]] to i16
 ; GCN-NEXT:    [[EXTENDED:%.*]] = zext i16 [[TMP6]] to i32
-; GCN-NEXT:    [[SHIFTED2:%.*]] = shl nuw i32 [[EXTENDED]], [[SHIFTAMT]]
+; GCN-NEXT:    [[SHIFTED1:%.*]] = shl nuw i32 [[EXTENDED]], [[SHIFTAMT]]
 ; GCN-NEXT:    [[UNMASKED:%.*]] = and i32 [[LOADED]], [[INV_MASK]]
-; GCN-NEXT:    [[INSERTED:%.*]] = or i32 [[UNMASKED]], [[SHIFTED2]]
+; GCN-NEXT:    [[INSERTED:%.*]] = or i32 [[UNMASKED]], [[SHIFTED1]]
 ; GCN-NEXT:    [[TMP7:%.*]] = cmpxchg ptr addrspace(3) [[ALIGNEDADDR]], i32 [[LOADED]], i32 [[INSERTED]] seq_cst seq_cst, align 4
 ; GCN-NEXT:    [[SUCCESS:%.*]] = extractvalue { i32, i1 } [[TMP7]], 1
 ; GCN-NEXT:    [[NEWLOADED]] = extractvalue { i32, i1 } [[TMP7]], 0
 ; GCN-NEXT:    br i1 [[SUCCESS]], label [[ATOMICRMW_END:%.*]], label [[ATOMICRMW_START]]
 ; GCN:       atomicrmw.end:
-; GCN-NEXT:    [[SHIFTED3:%.*]] = lshr i32 [[NEWLOADED]], [[SHIFTAMT]]
-; GCN-NEXT:    [[EXTRACTED4:%.*]] = trunc i32 [[SHIFTED3]] to i16
-; GCN-NEXT:    [[TMP8:%.*]] = bitcast i16 [[EXTRACTED4]] to half
+; GCN-NEXT:    [[SHIFTED2:%.*]] = lshr i32 [[NEWLOADED]], [[SHIFTAMT]]
+; GCN-NEXT:    [[EXTRACTED3:%.*]] = trunc i32 [[SHIFTED2]] to i16
+; GCN-NEXT:    [[TMP8:%.*]] = bitcast i16 [[EXTRACTED3]] to half
 ; GCN-NEXT:    ret half [[TMP8]]
 ;
   %res = atomicrmw fmin ptr addrspace(3) %ptr, half %value seq_cst
@@ -204,17 +204,17 @@ define double @test_atomicrmw_fmin_f64_flat(ptr %ptr, double %value) {
 ; GCN-NEXT:    [[TMP1:%.*]] = load double, ptr [[PTR:%.*]], align 8
 ; GCN-NEXT:    br label [[ATOMICRMW_START:%.*]]
 ; GCN:       atomicrmw.start:
-; GCN-NEXT:    [[LOADED:%.*]] = phi double [ [[TMP1]], [[TMP0:%.*]] ], [ [[TMP7:%.*]], [[ATOMICRMW_START]] ]
+; GCN-NEXT:    [[LOADED:%.*]] = phi double [ [[TMP1]], [[TMP0:%.*]] ], [ [[TMP6:%.*]], [[ATOMICRMW_START]] ]
 ; GCN-NEXT:    [[TMP2:%.*]] = call double @llvm.minnum.f64(double [[LOADED]], double [[VALUE:%.*]])
-; GCN-NEXT:    [[TMP4:%.*]] = bitcast double [[TMP2]] to i64
-; GCN-NEXT:    [[TMP5:%.*]] = bitcast double [[LOADED]] to i64
-; GCN-NEXT:    [[TMP6:%.*]] = cmpxchg ptr [[PTR]], i64 [[TMP5]], i64 [[TMP4]] seq_cst seq_cst, align 8
-; GCN-NEXT:    [[SUCCESS:%.*]] = extractvalue { i64, i1 } [[TMP6]], 1
-; GCN-NEXT:    [[NEWLOADED:%.*]] = extractvalue { i64, i1 } [[TMP6]], 0
-; GCN-NEXT:    [[TMP7]] = bitcast i64 [[NEWLOADED]] to double
+; GCN-NEXT:    [[TMP3:%.*]] = bitcast double [[TMP2]] to i64
+; GCN-NEXT:    [[TMP4:%.*]] = bitcast double [[LOADED]] to i64
+; GCN-NEXT:    [[TMP5:%.*]] = cmpxchg ptr [[PTR]], i64 [[TMP4]], i64 [[TMP3]] seq_cst seq_cst, align 8
+; GCN-NEXT:    [[SUCCESS:%.*]] = extractvalue { i64, i1 } [[TMP5]], 1
+; GCN-NEXT:    [[NEWLOADED:%.*]] = extractvalue { i64, i1 } [[TMP5]], 0
+; GCN-NEXT:    [[TMP6]] = bitcast i64 [[NEWLOADED]] to double
 ; GCN-NEXT:    br i1 [[SUCCESS]], label [[ATOMICRMW_END:%.*]], label [[ATOMICRMW_START]]
 ; GCN:       atomicrmw.end:
-; GCN-NEXT:    ret double [[TMP7]]
+; GCN-NEXT:    ret double [[TMP6]]
 ;
   %res = atomicrmw fmin ptr %ptr, double %value seq_cst
   ret double %res
@@ -225,17 +225,17 @@ define double @test_atomicrmw_fmin_f64_global(ptr addrspace(1) %ptr, double %val
 ; GCN-NEXT:    [[TMP1:%.*]] = load double, ptr addrspace(1) [[PTR:%.*]], align 8
 ; GCN-NEXT:    br label [[ATOMICRMW_START:%.*]]
 ; GCN:       atomicrmw.start:
-; GCN-NEXT:    [[LOADED:%.*]] = phi double [ [[TMP1]], [[TMP0:%.*]] ], [ [[TMP7:%.*]], [[ATOMICRMW_START]] ]
+; GCN-NEXT:    [[LOADED:%.*]] = phi double [ [[TMP1]], [[TMP0:%.*]] ], [ [[TMP6:%.*]], [[ATOMICRMW_START]] ]
 ; GCN-NEXT:    [[TMP2:%.*]] = call double @llvm.minnum.f64(double [[LOADED]], double [[VALUE:%.*]])
-; GCN-NEXT:    [[TMP4:%.*]] = bitcast double [[TMP2]] to i64
-; GCN-NEXT:    [[TMP5:%.*]] = bitcast double [[LOADED]] to i64
-; GCN-NEXT:    [[TMP6:%.*]] = cmpxchg ptr addrspace(1) [[PTR]], i64 [[TMP5]], i64 [[TMP4]] seq_cst seq_cst, align 8
-; GCN-NEXT:    [[SUCCESS:%.*]] = extractvalue { i64, i1 } [[TMP6]], 1
-; GCN-NEXT:    [[NEWLOADED:%.*]] = extractvalue { i64, i1 } [[TMP6]], 0
-; GCN-NEXT:    [[TMP7]] = bitcast i64 [[NEWLOADED]] to double
+; GCN-NEXT:    [[TMP3:%.*]] = bitcast double [[TMP2]] to i64
+; GCN-NEXT:    [[TMP4:%.*]] = bitcast double [[LOADED]] to i64
+; GCN-NEXT:    [[TMP5:%.*]] = cmpxchg ptr addrspace(1) [[PTR]], i64 [[TMP4]], i64 [[TMP3]] seq_cst seq_cst, align 8
+; GCN-NEXT:    [[SUCCESS:%.*]] = extractvalue { i64, i1 } [[TMP5]], 1
+; GCN-NEXT:    [[NEWLOADED:%.*]] = extractvalue { i64, i1 } [[TMP5]], 0
+; GCN-NEXT:    [[TMP6]] = bitcast i64 [[NEWLOADED]] to double
 ; GCN-NEXT:    br i1 [[SUCCESS]], label [[ATOMICRMW_END:%.*]], label [[ATOMICRMW_START]]
 ; GCN:       atomicrmw.end:
-; GCN-NEXT:    ret double [[TMP7]]
+; GCN-NEXT:    ret double [[TMP6]]
 ;
   %res = atomicrmw fmin ptr addrspace(1) %ptr, double %value seq_cst
   ret double %res
@@ -246,18 +246,39 @@ define double @test_atomicrmw_fmin_f64_local(ptr addrspace(3) %ptr, double %valu
 ; GCN-NEXT:    [[TMP1:%.*]] = load double, ptr addrspace(3) [[PTR:%.*]], align 8
 ; GCN-NEXT:    br label [[ATOMICRMW_START:%.*]]
 ; GCN:       atomicrmw.start:
-; GCN-NEXT:    [[LOADED:%.*]] = phi double [ [[TMP1]], [[TMP0:%.*]] ], [ [[TMP7:%.*]], [[ATOMICRMW_START]] ]
+; GCN-NEXT:    [[LOADED:%.*]] = phi double [ [[TMP1]], [[TMP0:%.*]] ], [ [[TMP6:%.*]], [[ATOMICRMW_START]] ]
 ; GCN-NEXT:    [[TMP2:%.*]] = call double @llvm.minnum.f64(double [[LOADED]], double [[VALUE:%.*]])
-; GCN-NEXT:    [[TMP4:%.*]] = bitcast double [[TMP2]] to i64
-; GCN-NEXT:    [[TMP5:%.*]] = bitcast double [[LOADED]] to i64
-; GCN-NEXT:    [[TMP6:%.*]] = cmpxchg ptr addrspace(3) [[PTR]], i64 [[TMP5]], i64 [[TMP4]] seq_cst seq_cst, align 8
-; GCN-NEXT:    [[SUCCESS:%.*]] = extractvalue { i64, i1 } [[TMP6]], 1
-; GCN-NEXT:    [[NEWLOADED:%.*]] = extractvalue { i64, i1 } [[TMP6]], 0
-; GCN-NEXT:    [[TMP7]] = bitcast i64 [[NEWLOADED]] to double
+; GCN-NEXT:    [[TMP3:%.*]] = bitcast double [[TMP2]] to i64
+; GCN-NEXT:    [[TMP4:%.*]] = bitcast double [[LOADED]] to i64
+; GCN-NEXT:    [[TMP5:%.*]] = cmpxchg ptr addrspace(3) [[PTR]], i64 [[TMP4]], i64 [[TMP3]] seq_cst seq_cst, align 8
+; GCN-NEXT:    [[SUCCESS:%.*]] = extractvalue { i64, i1 } [[TMP5]], 1
+; GCN-NEXT:    [[NEWLOADED:%.*]] = extractvalue { i64, i1 } [[TMP5]], 0
+; GCN-NEXT:    [[TMP6]] = bitcast i64 [[NEWLOADED]] to double
 ; GCN-NEXT:    br i1 [[SUCCESS]], label [[ATOMICRMW_END:%.*]], label [[ATOMICRMW_START]]
 ; GCN:       atomicrmw.end:
-; GCN-NEXT:    ret double [[TMP7]]
+; GCN-NEXT:    ret double [[TMP6]]
 ;
   %res = atomicrmw fmin ptr addrspace(3) %ptr, double %value seq_cst
   ret double %res
 }
+
+define double @test_atomicrmw_fmin_f64_global_strictfp(ptr addrspace(1) %ptr, double %value) strictfp {
+; GCN-LABEL: @test_atomicrmw_fmin_f64_global_strictfp(
+; GCN-NEXT:    [[TMP1:%.*]] = load double, ptr addrspace(1) [[PTR:%.*]], align 8
+; GCN-NEXT:    br label [[ATOMICRMW_START:%.*]]
+; GCN:       atomicrmw.start:
+; GCN-NEXT:    [[LOADED:%.*]] = phi double [ [[TMP1]], [[TMP0:%.*]] ], [ [[TMP6:%.*]], [[ATOMICRMW_START]] ]
+; GCN-NEXT:    [[TMP2:%.*]] = call double @llvm.minnum.f64(double [[LOADED]], double [[VALUE:%.*]])
+; GCN-NEXT:    [[TMP3:%.*]] = bitcast double [[TMP2]] to i64
+; GCN-NEXT:    [[TMP4:%.*]] = bitcast double [[LOADED]] to i64
+; GCN-NEXT:    [[TMP5:%.*]] = cmpxchg ptr addrspace(1) [[PTR]], i64 [[TMP4]], i64 [[TMP3]] seq_cst seq_cst, align 8
+; GCN-NEXT:    [[SUCCESS:%.*]] = extractvalue { i64, i1 } [[TMP5]], 1
+; GCN-NEXT:    [[NEWLOADED:%.*]] = extractvalue { i64, i1 } [[TMP5]], 0
+; GCN-NEXT:    [[TMP6]] = bitcast i64 [[NEWLOADED]] to double
+; GCN-NEXT:    br i1 [[SUCCESS]], label [[ATOMICRMW_END:%.*]], label [[ATOMICRMW_START]]
+; GCN:       atomicrmw.end:
+; GCN-NEXT:    ret double [[TMP6]]
+;
+  %res = atomicrmw fmin ptr addrspace(1) %ptr, double %value seq_cst
+  ret double %res
+}
diff --git a/llvm/test/Transforms/AtomicExpand/AMDGPU/expand-atomic-rmw-fsub.ll b/llvm/test/Transforms/AtomicExpand/AMDGPU/expand-atomic-rmw-fsub.ll
index 44e2446..d8ab058 100644
--- a/llvm/test/Transforms/AtomicExpand/AMDGPU/expand-atomic-rmw-fsub.ll
+++ b/llvm/test/Transforms/AtomicExpand/AMDGPU/expand-atomic-rmw-fsub.ll
@@ -7,17 +7,17 @@ define float @test_atomicrmw_fsub_f32_flat(ptr %ptr, float %value) {
 ; GCN-NEXT:    [[TMP1:%.*]] = load float, ptr [[PTR:%.*]], align 4
 ; GCN-NEXT:    br label [[ATOMICRMW_START:%.*]]
 ; GCN:       atomicrmw.start:
-; GCN-NEXT:    [[LOADED:%.*]] = phi float [ [[TMP1]], [[TMP0:%.*]] ], [ [[TMP6:%.*]], [[ATOMICRMW_START]] ]
+; GCN-NEXT:    [[LOADED:%.*]] = phi float [ [[TMP1]], [[TMP0:%.*]] ], [ [[TMP5:%.*]], [[ATOMICRMW_START]] ]
 ; GCN-NEXT:    [[NEW:%.*]] = fsub float [[LOADED]], [[VALUE:%.*]]
-; GCN-NEXT:    [[TMP3:%.*]] = bitcast float [[NEW]] to i32
-; GCN-NEXT:    [[TMP4:%.*]] = bitcast float [[LOADED]] to i32
-; GCN-NEXT:    [[TMP5:%.*]] = cmpxchg ptr [[PTR]], i32 [[TMP4]], i32 [[TMP3]] seq_cst seq_cst, align 4
-; GCN-NEXT:    [[SUCCESS:%.*]] = extractvalue { i32, i1 } [[TMP5]], 1
-; GCN-NEXT:    [[NEWLOADED:%.*]] = extractvalue { i32, i1 } [[TMP5]], 0
-; GCN-NEXT:    [[TMP6]] = bitcast i32 [[NEWLOADED]] to float
+; GCN-NEXT:    [[TMP2:%.*]] = bitcast float [[NEW]] to i32
+; GCN-NEXT:    [[TMP3:%.*]] = bitcast float [[LOADED]] to i32
+; GCN-NEXT:    [[TMP4:%.*]] = cmpxchg ptr [[PTR]], i32 [[TMP3]], i32 [[TMP2]] seq_cst seq_cst, align 4
+; GCN-NEXT:    [[SUCCESS:%.*]] = extractvalue { i32, i1 } [[TMP4]], 1
+; GCN-NEXT:    [[NEWLOADED:%.*]] = extractvalue { i32, i1 } [[TMP4]], 0
+; GCN-NEXT:    [[TMP5]] = bitcast i32 [[NEWLOADED]] to float
 ; GCN-NEXT:    br i1 [[SUCCESS]], label [[ATOMICRMW_END:%.*]], label [[ATOMICRMW_START]]
 ; GCN:       atomicrmw.end:
-; GCN-NEXT:    ret float [[TMP6]]
+; GCN-NEXT:    ret float [[TMP5]]
 ;
   %res = atomicrmw fsub ptr %ptr, float %value seq_cst
   ret float %res
@@ -28,17 +28,17 @@ define float @test_atomicrmw_fsub_f32_global(ptr addrspace(1) %ptr, float %value
 ; GCN-NEXT:    [[TMP1:%.*]] = load float, ptr addrspace(1) [[PTR:%.*]], align 4
 ; GCN-NEXT:    br label [[ATOMICRMW_START:%.*]]
 ; GCN:       atomicrmw.start:
-; GCN-NEXT:    [[LOADED:%.*]] = phi float [ [[TMP1]], [[TMP0:%.*]] ], [ [[TMP6:%.*]], [[ATOMICRMW_START]] ]
+; GCN-NEXT:    [[LOADED:%.*]] = phi float [ [[TMP1]], [[TMP0:%.*]] ], [ [[TMP5:%.*]], [[ATOMICRMW_START]] ]
 ; GCN-NEXT:    [[NEW:%.*]] = fsub float [[LOADED]], [[VALUE:%.*]]
-; GCN-NEXT:    [[TMP3:%.*]] = bitcast float [[NEW]] to i32
-; GCN-NEXT:    [[TMP4:%.*]] = bitcast float [[LOADED]] to i32
-; GCN-NEXT:    [[TMP5:%.*]] = cmpxchg ptr addrspace(1) [[PTR]], i32 [[TMP4]], i32 [[TMP3]] seq_cst seq_cst, align 4
-; GCN-NEXT:    [[SUCCESS:%.*]] = extractvalue { i32, i1 } [[TMP5]], 1
-; GCN-NEXT:    [[NEWLOADED:%.*]] = extractvalue { i32, i1 } [[TMP5]], 0
-; GCN-NEXT:    [[TMP6]] = bitcast i32 [[NEWLOADED]] to float
+; GCN-NEXT:    [[TMP2:%.*]] = bitcast float [[NEW]] to i32
+; GCN-NEXT:    [[TMP3:%.*]] = bitcast float [[LOADED]] to i32
+; GCN-NEXT:    [[TMP4:%.*]] = cmpxchg ptr addrspace(1) [[PTR]], i32 [[TMP3]], i32 [[TMP2]] seq_cst seq_cst, align 4
+; GCN-NEXT:    [[SUCCESS:%.*]] = extractvalue { i32, i1 } [[TMP4]], 1
+; GCN-NEXT:    [[NEWLOADED:%.*]] = extractvalue { i32, i1 } [[TMP4]], 0
+; GCN-NEXT:    [[TMP5]] = bitcast i32 [[NEWLOADED]] to float
 ; GCN-NEXT:    br i1 [[SUCCESS]], label [[ATOMICRMW_END:%.*]], label [[ATOMICRMW_START]]
 ; GCN:       atomicrmw.end:
-; GCN-NEXT:    ret float [[TMP6]]
+; GCN-NEXT:    ret float [[TMP5]]
 ;
   %res = atomicrmw fsub ptr addrspace(1) %ptr, float %value seq_cst
   ret float %res
@@ -49,17 +49,17 @@ define float @test_atomicrmw_fsub_f32_local(ptr addrspace(3) %ptr, float %value)
 ; GCN-NEXT:    [[TMP1:%.*]] = load float, ptr addrspace(3) [[PTR:%.*]], align 4
 ; GCN-NEXT:    br label [[ATOMICRMW_START:%.*]]
 ; GCN:       atomicrmw.start:
-; GCN-NEXT:    [[LOADED:%.*]] = phi float [ [[TMP1]], [[TMP0:%.*]] ], [ [[TMP6:%.*]], [[ATOMICRMW_START]] ]
+; GCN-NEXT:    [[LOADED:%.*]] = phi float [ [[TMP1]], [[TMP0:%.*]] ], [ [[TMP5:%.*]], [[ATOMICRMW_START]] ]
 ; GCN-NEXT:    [[NEW:%.*]] = fsub float [[LOADED]], [[VALUE:%.*]]
-; GCN-NEXT:    [[TMP3:%.*]] = bitcast float [[NEW]] to i32
-; GCN-NEXT:    [[TMP4:%.*]] = bitcast float [[LOADED]] to i32
-; GCN-NEXT:    [[TMP5:%.*]] = cmpxchg ptr addrspace(3) [[PTR]], i32 [[TMP4]], i32 [[TMP3]] seq_cst seq_cst, align 4
-; GCN-NEXT:    [[SUCCESS:%.*]] = extractvalue { i32, i1 } [[TMP5]], 1
-; GCN-NEXT:    [[NEWLOADED:%.*]] = extractvalue { i32, i1 } [[TMP5]], 0
-; GCN-NEXT:    [[TMP6]] = bitcast i32 [[NEWLOADED]] to float
+; GCN-NEXT:    [[TMP2:%.*]] = bitcast float [[NEW]] to i32
+; GCN-NEXT:    [[TMP3:%.*]] = bitcast float [[LOADED]] to i32
+; GCN-NEXT:    [[TMP4:%.*]] = cmpxchg ptr addrspace(3) [[PTR]], i32 [[TMP3]], i32 [[TMP2]] seq_cst seq_cst, align 4
+; GCN-NEXT:    [[SUCCESS:%.*]] = extractvalue { i32, i1 } [[TMP4]], 1
+; GCN-NEXT:    [[NEWLOADED:%.*]] = extractvalue { i32, i1 } [[TMP4]], 0
+; GCN-NEXT:    [[TMP5]] = bitcast i32 [[NEWLOADED]] to float
 ; GCN-NEXT:    br i1 [[SUCCESS]], label [[ATOMICRMW_END:%.*]], label [[ATOMICRMW_START]]
 ; GCN:       atomicrmw.end:
-; GCN-NEXT:    ret float [[TMP6]]
+; GCN-NEXT:    ret float [[TMP5]]
 ;
   %res = atomicrmw fsub ptr addrspace(3) %ptr, float %value seq_cst
   ret float %res
@@ -84,17 +84,17 @@ define half @test_atomicrmw_fsub_f16_flat(ptr %ptr, half %value) {
 ; GCN-NEXT:    [[NEW:%.*]] = fsub half [[TMP4]], [[VALUE:%.*]]
 ; GCN-NEXT:    [[TMP5:%.*]] = bitcast half [[NEW]] to i16
 ; GCN-NEXT:    [[EXTENDED:%.*]] = zext i16 [[TMP5]] to i32
-; GCN-NEXT:    [[SHIFTED2:%.*]] = shl nuw i32 [[EXTENDED]], [[SHIFTAMT]]
+; GCN-NEXT:    [[SHIFTED1:%.*]] = shl nuw i32 [[EXTENDED]], [[SHIFTAMT]]
 ; GCN-NEXT:    [[UNMASKED:%.*]] = and i32 [[LOADED]], [[INV_MASK]]
-; GCN-NEXT:    [[INSERTED:%.*]] = or i32 [[UNMASKED]], [[SHIFTED2]]
+; GCN-NEXT:    [[INSERTED:%.*]] = or i32 [[UNMASKED]], [[SHIFTED1]]
 ; GCN-NEXT:    [[TMP6:%.*]] = cmpxchg ptr [[ALIGNEDADDR]], i32 [[LOADED]], i32 [[INSERTED]] seq_cst seq_cst, align 4
 ; GCN-NEXT:    [[SUCCESS:%.*]] = extractvalue { i32, i1 } [[TMP6]], 1
 ; GCN-NEXT:    [[NEWLOADED]] = extractvalue { i32, i1 } [[TMP6]], 0
 ; GCN-NEXT:    br i1 [[SUCCESS]], label [[ATOMICRMW_END:%.*]], label [[ATOMICRMW_START]]
 ; GCN:       atomicrmw.end:
-; GCN-NEXT:    [[SHIFTED3:%.*]] = lshr i32 [[NEWLOADED]], [[SHIFTAMT]]
-; GCN-NEXT:    [[EXTRACTED4:%.*]] = trunc i32 [[SHIFTED3]] to i16
-; GCN-NEXT:    [[TMP7:%.*]] = bitcast i16 [[EXTRACTED4]] to half
+; GCN-NEXT:    [[SHIFTED2:%.*]] = lshr i32 [[NEWLOADED]], [[SHIFTAMT]]
+; GCN-NEXT:    [[EXTRACTED3:%.*]] = trunc i32 [[SHIFTED2]] to i16
+; GCN-NEXT:    [[TMP7:%.*]] = bitcast i16 [[EXTRACTED3]] to half
 ; GCN-NEXT:    ret half [[TMP7]]
 ;
   %res = atomicrmw fsub ptr %ptr, half %value seq_cst
@@ -120,17 +120,17 @@ define half @test_atomicrmw_fsub_f16_global(ptr addrspace(1) %ptr, half %value)
 ; GCN-NEXT:    [[NEW:%.*]] = fsub half [[TMP4]], [[VALUE:%.*]]
 ; GCN-NEXT:    [[TMP5:%.*]] = bitcast half [[NEW]] to i16
 ; GCN-NEXT:    [[EXTENDED:%.*]] = zext i16 [[TMP5]] to i32
-; GCN-NEXT:    [[SHIFTED2:%.*]] = shl nuw i32 [[EXTENDED]], [[SHIFTAMT]]
+; GCN-NEXT:    [[SHIFTED1:%.*]] = shl nuw i32 [[EXTENDED]], [[SHIFTAMT]]
 ; GCN-NEXT:    [[UNMASKED:%.*]] = and i32 [[LOADED]], [[INV_MASK]]
-; GCN-NEXT:    [[INSERTED:%.*]] = or i32 [[UNMASKED]], [[SHIFTED2]]
+; GCN-NEXT:    [[INSERTED:%.*]] = or i32 [[UNMASKED]], [[SHIFTED1]]
 ; GCN-NEXT:    [[TMP6:%.*]] = cmpxchg ptr addrspace(1) [[ALIGNEDADDR]], i32 [[LOADED]], i32 [[INSERTED]] seq_cst seq_cst, align 4
 ; GCN-NEXT:    [[SUCCESS:%.*]] = extractvalue { i32, i1 } [[TMP6]], 1
 ; GCN-NEXT:    [[NEWLOADED]] = extractvalue { i32, i1 } [[TMP6]], 0
 ; GCN-NEXT:    br i1 [[SUCCESS]], label [[ATOMICRMW_END:%.*]], label [[ATOMICRMW_START]]
 ; GCN:       atomicrmw.end:
-; GCN-NEXT:    [[SHIFTED3:%.*]] = lshr i32 [[NEWLOADED]], [[SHIFTAMT]]
-; GCN-NEXT:    [[EXTRACTED4:%.*]] = trunc i32 [[SHIFTED3]] to i16
-; GCN-NEXT:    [[TMP7:%.*]] = bitcast i16 [[EXTRACTED4]] to half
+; GCN-NEXT:    [[SHIFTED2:%.*]] = lshr i32 [[NEWLOADED]], [[SHIFTAMT]]
+; GCN-NEXT:    [[EXTRACTED3:%.*]] = trunc i32 [[SHIFTED2]] to i16
+; GCN-NEXT:    [[TMP7:%.*]] = bitcast i16 [[EXTRACTED3]] to half
 ; GCN-NEXT:    ret half [[TMP7]]
 ;
   %res = atomicrmw fsub ptr addrspace(1) %ptr, half %value seq_cst
@@ -182,17 +182,17 @@ define half @test_atomicrmw_fsub_f16_local(ptr addrspace(3) %ptr, half %value) {
 ; GCN-NEXT:    [[NEW:%.*]] = fsub half [[TMP4]], [[VALUE:%.*]]
 ; GCN-NEXT:    [[TMP5:%.*]] = bitcast half [[NEW]] to i16
 ; GCN-NEXT:    [[EXTENDED:%.*]] = zext i16 [[TMP5]] to i32
-; GCN-NEXT:    [[SHIFTED2:%.*]] = shl nuw i32 [[EXTENDED]], [[SHIFTAMT]]
+; GCN-NEXT:    [[SHIFTED1:%.*]] = shl nuw i32 [[EXTENDED]], [[SHIFTAMT]]
 ; GCN-NEXT:    [[UNMASKED:%.*]] = and i32 [[LOADED]], [[INV_MASK]]
-; GCN-NEXT:    [[INSERTED:%.*]] = or i32 [[UNMASKED]], [[SHIFTED2]]
+; GCN-NEXT:    [[INSERTED:%.*]] = or i32 [[UNMASKED]], [[SHIFTED1]]
 ; GCN-NEXT:    [[TMP6:%.*]] = cmpxchg ptr addrspace(3) [[ALIGNEDADDR]], i32 [[LOADED]], i32 [[INSERTED]] seq_cst seq_cst, align 4
 ; GCN-NEXT:    [[SUCCESS:%.*]] = extractvalue { i32, i1 } [[TMP6]], 1
 ; GCN-NEXT:    [[NEWLOADED]] = extractvalue { i32, i1 } [[TMP6]], 0
 ; GCN-NEXT:    br i1 [[SUCCESS]], label [[ATOMICRMW_END:%.*]], label [[ATOMICRMW_START]]
 ; GCN:       atomicrmw.end:
-; GCN-NEXT:    [[SHIFTED3:%.*]] = lshr i32 [[NEWLOADED]], [[SHIFTAMT]]
-; GCN-NEXT:    [[EXTRACTED4:%.*]] = trunc i32 [[SHIFTED3]] to i16
-; GCN-NEXT:    [[TMP7:%.*]] = bitcast i16 [[EXTRACTED4]] to half
+; GCN-NEXT:    [[SHIFTED2:%.*]] = lshr i32 [[NEWLOADED]], [[SHIFTAMT]]
+; GCN-NEXT:    [[EXTRACTED3:%.*]] = trunc i32 [[SHIFTED2]] to i16
+; GCN-NEXT:    [[TMP7:%.*]] = bitcast i16 [[EXTRACTED3]] to half
 ; GCN-NEXT:    ret half [[TMP7]]
 ;
   %res = atomicrmw fsub ptr addrspace(3) %ptr, half %value seq_cst
@@ -204,17 +204,17 @@ define double @test_atomicrmw_fsub_f64_flat(ptr %ptr, double %value) {
 ; GCN-NEXT:    [[TMP1:%.*]] = load double, ptr [[PTR:%.*]], align 8
 ; GCN-NEXT:    br label [[ATOMICRMW_START:%.*]]
 ; GCN:       atomicrmw.start:
-; GCN-NEXT:    [[LOADED:%.*]] = phi double [ [[TMP1]], [[TMP0:%.*]] ], [ [[TMP6:%.*]], [[ATOMICRMW_START]] ]
+; GCN-NEXT:    [[LOADED:%.*]] = phi double [ [[TMP1]], [[TMP0:%.*]] ], [ [[TMP5:%.*]], [[ATOMICRMW_START]] ]
 ; GCN-NEXT:    [[NEW:%.*]] = fsub double [[LOADED]], [[VALUE:%.*]]
-; GCN-NEXT:    [[TMP3:%.*]] = bitcast double [[NEW]] to i64
-; GCN-NEXT:    [[TMP4:%.*]] = bitcast double [[LOADED]] to i64
-; GCN-NEXT:    [[TMP5:%.*]] = cmpxchg ptr [[PTR]], i64 [[TMP4]], i64 [[TMP3]] seq_cst seq_cst, align 8
-; GCN-NEXT:    [[SUCCESS:%.*]] = extractvalue { i64, i1 } [[TMP5]], 1
-; GCN-NEXT:    [[NEWLOADED:%.*]] = extractvalue { i64, i1 } [[TMP5]], 0
-; GCN-NEXT:    [[TMP6]] = bitcast i64 [[NEWLOADED]] to double
+; GCN-NEXT:    [[TMP2:%.*]] = bitcast double [[NEW]] to i64
+; GCN-NEXT:    [[TMP3:%.*]] = bitcast double [[LOADED]] to i64
+; GCN-NEXT:    [[TMP4:%.*]] = cmpxchg ptr [[PTR]], i64 [[TMP3]], i64 [[TMP2]] seq_cst seq_cst, align 8
+; GCN-NEXT:    [[SUCCESS:%.*]] = extractvalue { i64, i1 } [[TMP4]], 1
+; GCN-NEXT:    [[NEWLOADED:%.*]] = extractvalue { i64, i1 } [[TMP4]], 0
+; GCN-NEXT:    [[TMP5]] = bitcast i64 [[NEWLOADED]] to double
 ; GCN-NEXT:    br i1 [[SUCCESS]], label [[ATOMICRMW_END:%.*]], label [[ATOMICRMW_START]]
 ; GCN:       atomicrmw.end:
-; GCN-NEXT:    ret double [[TMP6]]
+; GCN-NEXT:    ret double [[TMP5]]
 ;
   %res = atomicrmw fsub ptr %ptr, double %value seq_cst
   ret double %res
@@ -225,17 +225,17 @@ define double @test_atomicrmw_fsub_f64_global(ptr addrspace(1) %ptr, double %val
 ; GCN-NEXT:    [[TMP1:%.*]] = load double, ptr addrspace(1) [[PTR:%.*]], align 8
 ; GCN-NEXT:    br label [[ATOMICRMW_START:%.*]]
 ; GCN:       atomicrmw.start:
-; GCN-NEXT:    [[LOADED:%.*]] = phi double [ [[TMP1]], [[TMP0:%.*]] ], [ [[TMP6:%.*]], [[ATOMICRMW_START]] ]
+; GCN-NEXT:    [[LOADED:%.*]] = phi double [ [[TMP1]], [[TMP0:%.*]] ], [ [[TMP5:%.*]], [[ATOMICRMW_START]] ]
 ; GCN-NEXT:    [[NEW:%.*]] = fsub double [[LOADED]], [[VALUE:%.*]]
-; GCN-NEXT:    [[TMP3:%.*]] = bitcast double [[NEW]] to i64
-; GCN-NEXT:    [[TMP4:%.*]] = bitcast double [[LOADED]] to i64
-; GCN-NEXT:    [[TMP5:%.*]] = cmpxchg ptr addrspace(1) [[PTR]], i64 [[TMP4]], i64 [[TMP3]] seq_cst seq_cst, align 8
-; GCN-NEXT:    [[SUCCESS:%.*]] = extractvalue { i64, i1 } [[TMP5]], 1
-; GCN-NEXT:    [[NEWLOADED:%.*]] = extractvalue { i64, i1 } [[TMP5]], 0
-; GCN-NEXT:    [[TMP6]] = bitcast i64 [[NEWLOADED]] to double
+; GCN-NEXT:    [[TMP2:%.*]] = bitcast double [[NEW]] to i64
+; GCN-NEXT:    [[TMP3:%.*]] = bitcast double [[LOADED]] to i64
+; GCN-NEXT:    [[TMP4:%.*]] = cmpxchg ptr addrspace(1) [[PTR]], i64 [[TMP3]], i64 [[TMP2]] seq_cst seq_cst, align 8
+; GCN-NEXT:    [[SUCCESS:%.*]] = extractvalue { i64, i1 } [[TMP4]], 1
+; GCN-NEXT:    [[NEWLOADED:%.*]] = extractvalue { i64, i1 } [[TMP4]], 0
+; GCN-NEXT:    [[TMP5]] = bitcast i64 [[NEWLOADED]] to double
 ; GCN-NEXT:    br i1 [[SUCCESS]], label [[ATOMICRMW_END:%.*]], label [[ATOMICRMW_START]]
 ; GCN:       atomicrmw.end:
-; GCN-NEXT:    ret double [[TMP6]]
+; GCN-NEXT:    ret double [[TMP5]]
 ;
   %res = atomicrmw fsub ptr addrspace(1) %ptr, double %value seq_cst
   ret double %res
@@ -246,18 +246,39 @@ define double @test_atomicrmw_fsub_f64_local(ptr addrspace(3) %ptr, double %valu
 ; GCN-NEXT:    [[TMP1:%.*]] = load double, ptr addrspace(3) [[PTR:%.*]], align 8
 ; GCN-NEXT:    br label [[ATOMICRMW_START:%.*]]
 ; GCN:       atomicrmw.start:
-; GCN-NEXT:    [[LOADED:%.*]] = phi double [ [[TMP1]], [[TMP0:%.*]] ], [ [[TMP6:%.*]], [[ATOMICRMW_START]] ]
+; GCN-NEXT:    [[LOADED:%.*]] = phi double [ [[TMP1]], [[TMP0:%.*]] ], [ [[TMP5:%.*]], [[ATOMICRMW_START]] ]
 ; GCN-NEXT:    [[NEW:%.*]] = fsub double [[LOADED]], [[VALUE:%.*]]
-; GCN-NEXT:    [[TMP3:%.*]] = bitcast double [[NEW]] to i64
-; GCN-NEXT:    [[TMP4:%.*]] = bitcast double [[LOADED]] to i64
-; GCN-NEXT:    [[TMP5:%.*]] = cmpxchg ptr addrspace(3) [[PTR]], i64 [[TMP4]], i64 [[TMP3]] seq_cst seq_cst, align 8
-; GCN-NEXT:    [[SUCCESS:%.*]] = extractvalue { i64, i1 } [[TMP5]], 1
-; GCN-NEXT:    [[NEWLOADED:%.*]] = extractvalue { i64, i1 } [[TMP5]], 0
-; GCN-NEXT:    [[TMP6]] = bitcast i64 [[NEWLOADED]] to double
+; GCN-NEXT:    [[TMP2:%.*]] = bitcast double [[NEW]] to i64
+; GCN-NEXT:    [[TMP3:%.*]] = bitcast double [[LOADED]] to i64
+; GCN-NEXT:    [[TMP4:%.*]] = cmpxchg ptr addrspace(3) [[PTR]], i64 [[TMP3]], i64 [[TMP2]] seq_cst seq_cst, align 8
+; GCN-NEXT:    [[SUCCESS:%.*]] = extractvalue { i64, i1 } [[TMP4]], 1
+; GCN-NEXT:    [[NEWLOADED:%.*]] = extractvalue { i64, i1 } [[TMP4]], 0
+; GCN-NEXT:    [[TMP5]] = bitcast i64 [[NEWLOADED]] to double
 ; GCN-NEXT:    br i1 [[SUCCESS]], label [[ATOMICRMW_END:%.*]], label [[ATOMICRMW_START]]
 ; GCN:       atomicrmw.end:
-; GCN-NEXT:    ret double [[TMP6]]
+; GCN-NEXT:    ret double [[TMP5]]
 ;
   %res = atomicrmw fsub ptr addrspace(3) %ptr, double %value seq_cst
   ret double %res
 }
+
+define float @test_atomicrmw_fsub_f32_global_strictfp(ptr addrspace(1) %ptr, float %value) strictfp {
+; GCN-LABEL: @test_atomicrmw_fsub_f32_global_strictfp(
+; GCN-NEXT:    [[TMP1:%.*]] = load float, ptr addrspace(1) [[PTR:%.*]], align 4
+; GCN-NEXT:    br label [[ATOMICRMW_START:%.*]]
+; GCN:       atomicrmw.start:
+; GCN-NEXT:    [[LOADED:%.*]] = phi float [ [[TMP1]], [[TMP0:%.*]] ], [ [[TMP5:%.*]], [[ATOMICRMW_START]] ]
+; GCN-NEXT:    [[NEW:%.*]] = call float @llvm.experimental.constrained.fsub.f32(float [[LOADED]], float [[VALUE:%.*]], metadata !"round.dynamic", metadata !"fpexcept.strict") #[[ATTR4:[0-9]+]]
+; GCN-NEXT:    [[TMP2:%.*]] = bitcast float [[NEW]] to i32
+; GCN-NEXT:    [[TMP3:%.*]] = bitcast float [[LOADED]] to i32
+; GCN-NEXT:    [[TMP4:%.*]] = cmpxchg ptr addrspace(1) [[PTR]], i32 [[TMP3]], i32 [[TMP2]] seq_cst seq_cst, align 4
+; GCN-NEXT:    [[SUCCESS:%.*]] = extractvalue { i32, i1 } [[TMP4]], 1
+; GCN-NEXT:    [[NEWLOADED:%.*]] = extractvalue { i32, i1 } [[TMP4]], 0
+; GCN-NEXT:    [[TMP5]] = bitcast i32 [[NEWLOADED]] to float
+; GCN-NEXT:    br i1 [[SUCCESS]], label [[ATOMICRMW_END:%.*]], label [[ATOMICRMW_START]]
+; GCN:       atomicrmw.end:
+; GCN-NEXT:    ret float [[TMP5]]
+;
+  %res = atomicrmw fsub ptr addrspace(1) %ptr, float %value seq_cst
+  ret float %res
+}