[llvm-branch-commits] [clang] [llvm] [HLSL] Add `InterlockedCompareStoreFloatBitwise` function and resource methods (PR #222166)

Joshua Batista via llvm-branch-commits llvm-branch-commits at lists.llvm.org
Thu Sep 10 11:07:46 PDT 2026


https://github.com/bob80905 updated https://github.com/llvm/llvm-project/pull/222166

>From aa979e48f933ee5da315bd84f849be863e2d54b3 Mon Sep 17 00:00:00 2001
From: Joshua Batista <jbatista at microsoft.com>
Date: Tue, 8 Sep 2026 13:36:59 -0700
Subject: [PATCH] First attempt implementing
 InterlockedCompareStoreFloatBitwise

---
 clang/include/clang/Basic/Builtins.td         |   7 ++
 clang/lib/CodeGen/CGHLSLBuiltins.cpp          |  14 ++-
 clang/lib/Sema/HLSLBuiltinTypeDeclBuilder.cpp |   5 +
 clang/lib/Sema/HLSLExternalSemaSource.cpp     |  17 ++-
 clang/lib/Sema/SemaHLSL.cpp                   |  23 ++--
 .../InterlockedCompareStoreFloatBitwise.hlsl  |  35 ++++++
 .../builtins/RWBuffer-Interlocked.hlsl        |   6 ++
 ...r-InterlockedCompareStoreFloatBitwise.hlsl |  27 +++++
 ...r-InterlockedCompareStoreFloatBitwise.hlsl |  21 ++++
 ...erlockedCompareStoreFloatBitwise-sm60.hlsl |  39 +++++++
 ...lockedCompareStoreFloatBitwise-errors.hlsl | 100 ++++++++++++++++++
 .../ResourceAtomicCompareStoreFloat.ll        |  30 ++++++
 12 files changed, 313 insertions(+), 11 deletions(-)
 create mode 100644 clang/test/CodeGenHLSL/builtins/InterlockedCompareStoreFloatBitwise.hlsl
 create mode 100644 clang/test/CodeGenHLSL/builtins/RWByteAddressBuffer-InterlockedCompareStoreFloatBitwise.hlsl
 create mode 100644 clang/test/CodeGenHLSL/builtins/RasterizerOrderedByteAddressBuffer-InterlockedCompareStoreFloatBitwise.hlsl
 create mode 100644 clang/test/SemaHLSL/BuiltIns/ByteAddressBuffer-InterlockedCompareStoreFloatBitwise-sm60.hlsl
 create mode 100644 clang/test/SemaHLSL/BuiltIns/InterlockedCompareStoreFloatBitwise-errors.hlsl
 create mode 100644 llvm/test/CodeGen/DirectX/ResourceAtomicCompareStoreFloat.ll

diff --git a/clang/include/clang/Basic/Builtins.td b/clang/include/clang/Basic/Builtins.td
index e34c2f0b1ea32..af9a743716e6e 100644
--- a/clang/include/clang/Basic/Builtins.td
+++ b/clang/include/clang/Basic/Builtins.td
@@ -5572,6 +5572,13 @@ def HLSLInterlockedCompareStore : LangBuiltin<"HLSL_LANG"> {
   let Prototype = "void (...)";
 }
 
+def HLSLInterlockedCompareStoreFloatBitwise : LangBuiltin<"HLSL_LANG"> {
+  let Spellings = ["__builtin_hlsl_interlocked_compare_store_float_bitwise"];
+  // Prevent inadvertent float -> double arg promotion.
+  let Attributes = [NoThrow, CustomTypeChecking];
+  let Prototype = "void (...)";
+}
+
 def HLSLInterlockedExchange : LangBuiltin<"HLSL_LANG"> {
   let Spellings = ["__builtin_hlsl_interlocked_exchange"];
   // Prevent inadvertent float -> double arg promotion.
diff --git a/clang/lib/CodeGen/CGHLSLBuiltins.cpp b/clang/lib/CodeGen/CGHLSLBuiltins.cpp
index 1af7af505b2e6..3af0b523df31a 100644
--- a/clang/lib/CodeGen/CGHLSLBuiltins.cpp
+++ b/clang/lib/CodeGen/CGHLSLBuiltins.cpp
@@ -360,6 +360,17 @@ static Value *handleInterlockedCompareOp(CodeGenFunction &CGF,
   Value *Compare = CGF.EmitScalarExpr(E->getArg(1));
   Value *Val = CGF.EmitScalarExpr(E->getArg(2));
 
+  // `cmpxchg` takes an integer or a pointer, so the float-bitwise operations
+  // work on the bit pattern of the float. This is what those operations mean,
+  // and DXIL and SPIR-V both need the integer form.
+  if (Compare->getType()->isFloatingPointTy()) {
+    llvm::Type *IntTy =
+        CGF.Builder.getIntNTy(Compare->getType()->getPrimitiveSizeInBits());
+    Compare = CGF.Builder.CreateBitCast(Compare, IntTy);
+    Val = CGF.Builder.CreateBitCast(Val, IntTy);
+    DestAddr = DestAddr.withElementType(IntTy);
+  }
+
   Value *Pair = CGF.Builder.CreateAtomicCmpXchg(
       DestAddr, Compare, Val, llvm::AtomicOrdering::Monotonic,
       llvm::AtomicOrdering::Monotonic, getHLSLAtomicScope(CGF, DestLV));
@@ -1518,7 +1529,8 @@ Value *CodeGenFunction::EmitHLSLBuiltinExpr(unsigned BuiltinID,
     return handleInterlockedOp(*this, E, llvm::AtomicRMWInst::And);
   }
   case Builtin::BI__builtin_hlsl_interlocked_compare_exchange:
-  case Builtin::BI__builtin_hlsl_interlocked_compare_store: {
+  case Builtin::BI__builtin_hlsl_interlocked_compare_store:
+  case Builtin::BI__builtin_hlsl_interlocked_compare_store_float_bitwise: {
     return handleInterlockedCompareOp(*this, E);
   }
   case Builtin::BI__builtin_hlsl_interlocked_exchange: {
diff --git a/clang/lib/Sema/HLSLBuiltinTypeDeclBuilder.cpp b/clang/lib/Sema/HLSLBuiltinTypeDeclBuilder.cpp
index 3c374f8649c55..9918a09338482 100644
--- a/clang/lib/Sema/HLSLBuiltinTypeDeclBuilder.cpp
+++ b/clang/lib/Sema/HLSLBuiltinTypeDeclBuilder.cpp
@@ -1760,6 +1760,11 @@ BuiltinTypeDeclBuilder::addByteAddressBufferInterlockedMethods() {
   addByteAddressBufferInterlockedCompareMethod(
       "InterlockedCompareStore", AST.UnsignedIntTy,
       "__builtin_hlsl_interlocked_compare_store");
+  // The float-bitwise compare reuses the 32-bit integer DXIL operation, so it
+  // needs no capability bits and works from SM 6.0.
+  addByteAddressBufferInterlockedCompareMethod(
+      "InterlockedCompareStoreFloatBitwise", AST.FloatTy,
+      "__builtin_hlsl_interlocked_compare_store_float_bitwise");
   addByteAddressBufferInterlockedMethod(
       "InterlockedExchange", AST.UnsignedIntTy,
       "__builtin_hlsl_interlocked_exchange", /*RequiresOriginalValue=*/true);
diff --git a/clang/lib/Sema/HLSLExternalSemaSource.cpp b/clang/lib/Sema/HLSLExternalSemaSource.cpp
index d3adae74db7bd..20efbf4cde861 100644
--- a/clang/lib/Sema/HLSLExternalSemaSource.cpp
+++ b/clang/lib/Sema/HLSLExternalSemaSource.cpp
@@ -925,14 +925,19 @@ static void defineHLSLInterlockedFunc(Sema &S, NamespaceDecl *NS,
 // Synthesize the compare-and-swap overload sets: {int, uint, int64_t,
 // uint64_t} x {groupshared, device}. Compare-store reports nothing and
 // compare-exchange always reports the previous value, so each has a single
-// arity.
+// arity. The float-bitwise operations carry their own names and take float
+// alone, so they get a float-only overload set.
 static void defineHLSLInterlockedCompareFunc(Sema &S, NamespaceDecl *NS,
                                              StringRef FuncName,
                                              StringRef BuiltinName,
-                                             AtomicOverloadShape Shape) {
+                                             AtomicOverloadShape Shape,
+                                             bool FloatOnly = false) {
   ASTContext &AST = S.getASTContext();
-  QualType Elems[] = {AST.IntTy, AST.UnsignedIntTy, AST.LongTy,
-                      AST.UnsignedLongTy};
+  QualType IntElems[] = {AST.IntTy, AST.UnsignedIntTy, AST.LongTy,
+                         AST.UnsignedLongTy};
+  QualType FloatElems[] = {AST.FloatTy};
+  ArrayRef<QualType> Elems =
+      FloatOnly ? ArrayRef<QualType>(FloatElems) : ArrayRef<QualType>(IntElems);
 
   for (QualType ElemTy : Elems)
     for (LangAS AS : {LangAS::hlsl_groupshared, LangAS::hlsl_device})
@@ -952,6 +957,10 @@ void HLSLExternalSemaSource::defineHLSLAtomicIntrinsics() {
                                    "InterlockedCompareStore",
                                    "__builtin_hlsl_interlocked_compare_store",
                                    AtomicOverloadShape::CompareStore);
+  defineHLSLInterlockedCompareFunc(
+      *SemaPtr, HLSLNamespace, "InterlockedCompareStoreFloatBitwise",
+      "__builtin_hlsl_interlocked_compare_store_float_bitwise",
+      AtomicOverloadShape::CompareStore, /*FloatOnly=*/true);
   defineHLSLInterlockedFunc(*SemaPtr, HLSLNamespace, "InterlockedExchange",
                             "__builtin_hlsl_interlocked_exchange",
                             /*RequiresOriginalValue=*/true,
diff --git a/clang/lib/Sema/SemaHLSL.cpp b/clang/lib/Sema/SemaHLSL.cpp
index 8a2dea4b0f803..244f0442ccc11 100644
--- a/clang/lib/Sema/SemaHLSL.cpp
+++ b/clang/lib/Sema/SemaHLSL.cpp
@@ -4223,7 +4223,7 @@ static bool CheckSamplingBuiltin(Sema &S, CallExpr *TheCall, SampleKind Kind) {
 /// The types an interlocked operation accepts for `dest`. A float operation
 /// works on the value's bit pattern, and DXC accepts 32-bit float only, so
 /// half and double are always rejected.
-enum class InterlockedDest { Int, IntOrFloat };
+enum class InterlockedDest { Int, IntOrFloat, Float };
 
 /// Check a call to an HLSL interlocked builtin. The operation accepts between
 /// `MinArgs` and `MaxArgs` arguments, `dest` accepts `Dest`, and
@@ -4255,13 +4255,17 @@ static bool CheckInterlockedBuiltin(Sema &S, CallExpr *TheCall,
   }
 
   QualType DestTy = TheCall->getArg(0)->getType().getUnqualifiedType();
-  const bool AllowsFloat = Dest == InterlockedDest::IntOrFloat;
-  if (!DestTy->isIntegerType() &&
-      !(AllowsFloat && DestTy->isSpecificBuiltinType(BuiltinType::Float))) {
+  const bool DestIsOK =
+      DestTy->isSpecificBuiltinType(BuiltinType::Float)
+          ? Dest != InterlockedDest::Int
+          : Dest != InterlockedDest::Float && DestTy->isIntegerType();
+  if (!DestIsOK) {
     S.Diag(TheCall->getArg(0)->getBeginLoc(),
            diag::err_builtin_invalid_arg_type)
-        << /*ordinal=*/1 << /*scalar*/ 1 << /*integer*/ 1
-        << /*32 bit floating-point*/ (AllowsFloat ? 3 : 0) << DestTy;
+        << /*ordinal=*/1 << /*scalar*/ 1
+        << /*integer*/ (Dest == InterlockedDest::Float ? 0 : 1)
+        << /*32 bit floating-point*/ (Dest == InterlockedDest::Int ? 0 : 3)
+        << DestTy;
     return true;
   }
 
@@ -4819,6 +4823,13 @@ bool SemaHLSL::CheckBuiltinFunctionCall(unsigned BuiltinID, CallExpr *TheCall) {
                                 /*ReportsOriginalValue=*/false))
       return true;
     break;
+  case Builtin::BI__builtin_hlsl_interlocked_compare_store_float_bitwise:
+    // The same shape, but it compares bit patterns, so float alone.
+    if (CheckInterlockedBuiltin(SemaRef, TheCall, /*MinArgs=*/3, /*MaxArgs=*/3,
+                                InterlockedDest::Float,
+                                /*ReportsOriginalValue=*/false))
+      return true;
+    break;
   case Builtin::BI__builtin_hlsl_interlocked_compare_exchange:
     // `compare_value` comes before `value`, and `original_value` is required.
     if (CheckInterlockedBuiltin(SemaRef, TheCall, /*MinArgs=*/4, /*MaxArgs=*/4,
diff --git a/clang/test/CodeGenHLSL/builtins/InterlockedCompareStoreFloatBitwise.hlsl b/clang/test/CodeGenHLSL/builtins/InterlockedCompareStoreFloatBitwise.hlsl
new file mode 100644
index 0000000000000..c5fbb4854e85c
--- /dev/null
+++ b/clang/test/CodeGenHLSL/builtins/InterlockedCompareStoreFloatBitwise.hlsl
@@ -0,0 +1,35 @@
+// RUN: %clang_cc1 -std=hlsl2021 -finclude-default-header -triple \
+// RUN:   dxil-pc-shadermodel6.6-library %s -emit-llvm -disable-llvm-passes -o - | \
+// RUN:   FileCheck %s --check-prefixes=CHECK,DXCHECK
+
+// RUN: %clang_cc1 -std=hlsl2021 -finclude-default-header -triple \
+// RUN:   spirv-pc-vulkan-library %s -emit-llvm -disable-llvm-passes -o - | \
+// RUN:   FileCheck %s --check-prefixes=CHECK,SPVCHECK
+
+// Test basic lowering of HLSL InterlockedCompareStoreFloatBitwise to `cmpxchg
+// monotonic`. `cmpxchg` takes an integer, so the float arguments become their
+// bit patterns first. The operation reports nothing, so the `cmpxchg` result
+// stays unused.
+
+groupshared float gs_f32;
+
+// CHECK-LABEL: define {{.*}}void @{{.*}}test_float
+// CHECK: [[CMP:%.*]] = bitcast float %{{.*}} to i32
+// CHECK-NEXT: [[VAL:%.*]] = bitcast float %{{.*}} to i32
+// DXCHECK-NEXT:  cmpxchg ptr addrspace(3) {{.*}}@gs_f32{{.*}}, i32 [[CMP]], i32 [[VAL]] syncscope("workgroup") monotonic monotonic
+// SPVCHECK-NEXT: cmpxchg ptr addrspace(3) {{.*}}@gs_f32{{.*}}, i32 [[CMP]], i32 [[VAL]] syncscope("workgroup") monotonic monotonic
+export void test_float(float cmp, float v) {
+  InterlockedCompareStoreFloatBitwise(gs_f32, cmp, v);
+}
+
+// A device-address-space destination uses the "device" scope instead.
+RWBuffer<float> Buf : register(u0);
+
+// CHECK-LABEL: define {{.*}}void @{{.*}}test_device
+// CHECK: [[CMP:%.*]] = bitcast float %{{.*}} to i32
+// CHECK-NEXT: [[VAL:%.*]] = bitcast float %{{.*}} to i32
+// DXCHECK-NEXT:  cmpxchg ptr %{{.*}}, i32 [[CMP]], i32 [[VAL]] syncscope("device") monotonic monotonic
+// SPVCHECK-NEXT: cmpxchg ptr addrspace(11) %{{.*}}, i32 [[CMP]], i32 [[VAL]] syncscope("device") monotonic monotonic
+export void test_device(float cmp, float v) {
+  InterlockedCompareStoreFloatBitwise(Buf[0], cmp, v);
+}
diff --git a/clang/test/CodeGenHLSL/builtins/RWBuffer-Interlocked.hlsl b/clang/test/CodeGenHLSL/builtins/RWBuffer-Interlocked.hlsl
index 47cff714a7df7..b85e40a9b1093 100644
--- a/clang/test/CodeGenHLSL/builtins/RWBuffer-Interlocked.hlsl
+++ b/clang/test/CodeGenHLSL/builtins/RWBuffer-Interlocked.hlsl
@@ -19,6 +19,7 @@
 
 RWBuffer<int> Out : register(u0);
 RWBuffer<uint> UOut : register(u1);
+RWBuffer<float> FOut : register(u2);
 
 // CHECK-LABEL: define void @main
 // DXCHECK:  %[[PTR1:.*]] = call {{.*}} @llvm.dx.resource.getpointer.p0.tdx.TypedBuffer_i32_1_0_1t.i32(target("dx.TypedBuffer", i32, 1, 0, 1) %{{.*}}, i32 %{{.*}})
@@ -43,6 +44,8 @@ RWBuffer<uint> UOut : register(u1);
 // DXCHECK:  cmpxchg ptr %[[PTR10]], i32 1, i32 2 syncscope("device") monotonic monotonic
 // DXCHECK:  %[[PTR11:.*]] = call {{.*}} @llvm.dx.resource.getpointer.p0.tdx.TypedBuffer_i32_1_0_1t.i32(target("dx.TypedBuffer", i32, 1, 0, 1) %{{.*}}, i32 %{{.*}})
 // DXCHECK:  cmpxchg ptr %[[PTR11]], i32 1, i32 2 syncscope("device") monotonic monotonic
+// DXCHECK:  %[[PTR12:.*]] = call {{.*}} @llvm.dx.resource.getpointer.p0.tdx.TypedBuffer_f32_1_0_0t.i32(target("dx.TypedBuffer", float, 1, 0, 0) %{{.*}}, i32 %{{.*}})
+// DXCHECK:  cmpxchg ptr %[[PTR12]], i32 1065353216, i32 1073741824 syncscope("device") monotonic monotonic
 // SPVCHECK: %[[PTR1:.*]] = call {{.*}} @llvm.spv.resource.getpointer.{{.*}}(target("spirv.SignedImage", i32, {{.*}}) %{{.*}}, i32 %{{.*}})
 // SPVCHECK: atomicrmw add ptr addrspace(11) %[[PTR1]], i32 1 syncscope("device") monotonic
 // SPVCHECK: %[[PTR2:.*]] = call {{.*}} @llvm.spv.resource.getpointer.{{.*}}(target("spirv.SignedImage", i32, {{.*}}) %{{.*}}, i32 %{{.*}})
@@ -65,6 +68,8 @@ RWBuffer<uint> UOut : register(u1);
 // SPVCHECK: cmpxchg ptr addrspace(11) %[[PTR10]], i32 1, i32 2 syncscope("device") monotonic monotonic
 // SPVCHECK: %[[PTR11:.*]] = call {{.*}} @llvm.spv.resource.getpointer.{{.*}}(target("spirv.SignedImage", i32, {{.*}}) %{{.*}}, i32 %{{.*}})
 // SPVCHECK: cmpxchg ptr addrspace(11) %[[PTR11]], i32 1, i32 2 syncscope("device") monotonic monotonic
+// SPVCHECK: %[[PTR12:.*]] = call {{.*}} @llvm.spv.resource.getpointer.{{.*}}(target("spirv.Image", float, {{.*}}) %{{.*}}, i32 %{{.*}})
+// SPVCHECK: cmpxchg ptr addrspace(11) %[[PTR12]], i32 1065353216, i32 1073741824 syncscope("device") monotonic monotonic
 [shader("compute")]
 [numthreads(1,1,1)]
 void main(uint3 id : SV_DispatchThreadID) {
@@ -80,4 +85,5 @@ void main(uint3 id : SV_DispatchThreadID) {
   InterlockedExchange(Out[id.x], 1, orig);
   InterlockedCompareStore(Out[id.x], 1, 2);
   InterlockedCompareExchange(Out[id.x], 1, 2, orig);
+  InterlockedCompareStoreFloatBitwise(FOut[id.x], 1.0f, 2.0f);
 }
diff --git a/clang/test/CodeGenHLSL/builtins/RWByteAddressBuffer-InterlockedCompareStoreFloatBitwise.hlsl b/clang/test/CodeGenHLSL/builtins/RWByteAddressBuffer-InterlockedCompareStoreFloatBitwise.hlsl
new file mode 100644
index 0000000000000..ca2f2b02cdaa4
--- /dev/null
+++ b/clang/test/CodeGenHLSL/builtins/RWByteAddressBuffer-InterlockedCompareStoreFloatBitwise.hlsl
@@ -0,0 +1,27 @@
+// RUN: %clang_cc1 -std=hlsl202x -finclude-default-header -triple \
+// RUN:   dxil-pc-shadermodel6.6-library %s -emit-llvm -disable-llvm-passes -o - | \
+// RUN:   FileCheck %s --check-prefixes=CHECK,DXCHECK
+
+// RUN: %clang_cc1 -std=hlsl202x -finclude-default-header -triple \
+// RUN:   spirv-pc-vulkan1.3-library %s -emit-llvm -disable-llvm-passes -o - | \
+// RUN:   FileCheck %s --check-prefixes=CHECK,SPVCHECK
+
+// Test that the RWByteAddressBuffer::InterlockedCompareStoreFloatBitwise
+// member method lowers to `resource_getpointer -> cmpxchg`, for both DXIL and
+// SPIR-V targets. The float arguments become their bit patterns first, because
+// `cmpxchg` takes an integer.
+
+RWByteAddressBuffer BAB : register(u0);
+
+// CHECK-LABEL: define {{.*}}void @{{.*}}test_bab_float
+// DXCHECK:  %[[HANDLE:.*]] = load target("dx.RawBuffer", i8, 1, 0), ptr {{.*}}
+// DXCHECK:  %[[PTR:.*]] = call ptr @llvm.dx.resource.getpointer.p0.tdx.RawBuffer_i8_1_0t.i32(target("dx.RawBuffer", i8, 1, 0) %[[HANDLE]], i32 %{{.*}})
+// SPVCHECK: %[[HANDLE:.*]] = load target("spirv.VulkanBuffer", [0 x i8], 12, 1), ptr {{.*}}
+// SPVCHECK: %[[PTR:.*]] = call ptr addrspace(11) @llvm.spv.resource.getpointer.p11.tspirv.VulkanBuffer_a0i8_12_1t.i32(target("spirv.VulkanBuffer", [0 x i8], 12, 1) %[[HANDLE]], i32 %{{.*}})
+// CHECK: %[[CMP:.*]] = bitcast float %{{.*}} to i32
+// CHECK-NEXT: %[[VAL:.*]] = bitcast float %{{.*}} to i32
+// DXCHECK-NEXT:  cmpxchg ptr %[[PTR]], i32 %[[CMP]], i32 %[[VAL]] syncscope("device") monotonic monotonic
+// SPVCHECK-NEXT: cmpxchg ptr addrspace(11) %[[PTR]], i32 %[[CMP]], i32 %[[VAL]] syncscope("device") monotonic monotonic
+export void test_bab_float(uint off, float cmp, float v) {
+  BAB.InterlockedCompareStoreFloatBitwise(off, cmp, v);
+}
diff --git a/clang/test/CodeGenHLSL/builtins/RasterizerOrderedByteAddressBuffer-InterlockedCompareStoreFloatBitwise.hlsl b/clang/test/CodeGenHLSL/builtins/RasterizerOrderedByteAddressBuffer-InterlockedCompareStoreFloatBitwise.hlsl
new file mode 100644
index 0000000000000..5ddc6a4956f15
--- /dev/null
+++ b/clang/test/CodeGenHLSL/builtins/RasterizerOrderedByteAddressBuffer-InterlockedCompareStoreFloatBitwise.hlsl
@@ -0,0 +1,21 @@
+// RUN: %clang_cc1 -std=hlsl202x -finclude-default-header -triple \
+// RUN:   dxil-pc-shadermodel6.6-library %s -emit-llvm -disable-llvm-passes -o - | \
+// RUN:   FileCheck %s --check-prefixes=CHECK,DXCHECK
+
+// SPIR-V codegen for RasterizerOrderedByteAddressBuffer is not implemented
+// yet (asserts in clang/lib/CodeGen/Targets/SPIR.cpp on
+// `!ResAttrs.IsROV && "Rasterizer order views not implemented for SPIR-V yet"`).
+// Add a `spirv-pc-vulkan1.3-library` RUN line here when SPIR-V ROV support
+// lands.
+
+RasterizerOrderedByteAddressBuffer ROVB : register(u1);
+
+// CHECK-LABEL: define void @{{.*}}test_rovb_float
+// DXCHECK: %[[HANDLE:.*]] = load target("dx.RawBuffer", i8, 1, 1), ptr {{.*}}
+// DXCHECK: %[[PTR:.*]] = call ptr @llvm.dx.resource.getpointer.p0.tdx.RawBuffer_i8_1_1t.i32(target("dx.RawBuffer", i8, 1, 1) %[[HANDLE]], i32 %{{.*}})
+// DXCHECK: %[[CMP:.*]] = bitcast float %{{.*}} to i32
+// DXCHECK-NEXT: %[[VAL:.*]] = bitcast float %{{.*}} to i32
+// DXCHECK-NEXT: cmpxchg ptr %[[PTR]], i32 %[[CMP]], i32 %[[VAL]] syncscope("device") monotonic monotonic
+export void test_rovb_float(uint off, float cmp, float v) {
+  ROVB.InterlockedCompareStoreFloatBitwise(off, cmp, v);
+}
diff --git a/clang/test/SemaHLSL/BuiltIns/ByteAddressBuffer-InterlockedCompareStoreFloatBitwise-sm60.hlsl b/clang/test/SemaHLSL/BuiltIns/ByteAddressBuffer-InterlockedCompareStoreFloatBitwise-sm60.hlsl
new file mode 100644
index 0000000000000..5954884163670
--- /dev/null
+++ b/clang/test/SemaHLSL/BuiltIns/ByteAddressBuffer-InterlockedCompareStoreFloatBitwise-sm60.hlsl
@@ -0,0 +1,39 @@
+// RUN: %clang_cc1 -std=hlsl202x -finclude-default-header \
+// RUN:   -triple dxil-pc-shadermodel6.0-library %s -fsyntax-only -verify \
+// RUN:   -verify-ignore-unexpected=warning
+
+// The float-bitwise compare reuses the 32-bit integer DXIL operation, so it
+// needs no capability bits and works from SM 6.0. The 64-bit compare-store
+// needs SM 6.6. This file checks both halves, so it proves the two are gated
+// differently.
+
+RWByteAddressBuffer BAB : register(u0);
+RasterizerOrderedByteAddressBuffer ROVB : register(u1);
+groupshared float gs_f32;
+groupshared uint64_t gs_u64;
+
+void sm60_bab_float_bitwise_ok(uint off, float cmp, float v) {
+  BAB.InterlockedCompareStoreFloatBitwise(off, cmp, v);
+}
+
+void sm60_rovb_float_bitwise_ok(uint off, float cmp, float v) {
+  ROVB.InterlockedCompareStoreFloatBitwise(off, cmp, v);
+}
+
+void sm60_free_function_ok(float cmp, float v) {
+  InterlockedCompareStoreFloatBitwise(gs_f32, cmp, v);
+}
+
+void sm60_direct_builtin_ok(float cmp, float v) {
+  __builtin_hlsl_interlocked_compare_store_float_bitwise(gs_f32, cmp, v);
+}
+
+void sm60_no_bab_compare_store64(uint off, uint64_t cmp, uint64_t v) {
+  BAB.InterlockedCompareStore64(off, cmp, v);
+  // expected-error at -1 {{no member named 'InterlockedCompareStore64' in 'hlsl::RWByteAddressBuffer'}}
+}
+
+void sm60_no_direct_builtin_u64(uint64_t cmp, uint64_t v) {
+  __builtin_hlsl_interlocked_compare_store(gs_u64, cmp, v);
+  // expected-error at -1 {{'__builtin_hlsl_interlocked_compare_store' requires shader model 6.6 or newer}}
+}
diff --git a/clang/test/SemaHLSL/BuiltIns/InterlockedCompareStoreFloatBitwise-errors.hlsl b/clang/test/SemaHLSL/BuiltIns/InterlockedCompareStoreFloatBitwise-errors.hlsl
new file mode 100644
index 0000000000000..f291c6264f099
--- /dev/null
+++ b/clang/test/SemaHLSL/BuiltIns/InterlockedCompareStoreFloatBitwise-errors.hlsl
@@ -0,0 +1,100 @@
+// RUN: %clang_cc1 -std=hlsl202x -finclude-default-header \
+// RUN:   -triple dxil-pc-shadermodel6.6-library %s -emit-llvm-only \
+// RUN:   -disable-llvm-passes -verify
+
+// InterlockedCompareStoreFloatBitwise compares the bit pattern of a 32-bit
+// float, so it is provided as a float-only overload set (groupshared/device).
+// It reports nothing, so it has a single 3-argument form and no out parameter.
+
+groupshared float gs_f32;
+groupshared int gs_i32;
+groupshared double gs_f64;
+groupshared half gs_f16;
+struct S { float x; };
+groupshared S gs_s;
+
+void too_few(float cmp) {
+  InterlockedCompareStoreFloatBitwise(gs_f32, cmp); // expected-error{{no matching function for call to 'InterlockedCompareStoreFloatBitwise'}}
+  // expected-note@*:* 2 {{candidate function}}
+}
+
+void too_many(float cmp, float v, float extra) {
+  InterlockedCompareStoreFloatBitwise(gs_f32, cmp, v, extra); // expected-error{{no matching function for call to 'InterlockedCompareStoreFloatBitwise'}}
+  // expected-note@*:* 2 {{candidate function}}
+}
+
+void local_dest(float cmp, float v) {
+  float dest;
+  InterlockedCompareStoreFloatBitwise(dest, cmp, v); // expected-error{{no matching function for call to 'InterlockedCompareStoreFloatBitwise'}}
+  // expected-note@*:* 2 {{candidate function}}
+}
+
+// The bitwise compare is defined for 32-bit float alone, so there is no
+// integer, half or double overload.
+void int_dest(int cmp, int v) {
+  InterlockedCompareStoreFloatBitwise(gs_i32, cmp, v); // expected-error{{no matching function for call to 'InterlockedCompareStoreFloatBitwise'}}
+  // expected-note@*:* 2 {{candidate function}}
+}
+
+void double_dest(double cmp, double v) {
+  InterlockedCompareStoreFloatBitwise(gs_f64, cmp, v); // expected-error{{no matching function for call to 'InterlockedCompareStoreFloatBitwise'}}
+  // expected-note@*:* 2 {{candidate function}}
+}
+
+void half_dest(half cmp, half v) {
+  InterlockedCompareStoreFloatBitwise(gs_f16, cmp, v); // expected-error{{no matching function for call to 'InterlockedCompareStoreFloatBitwise'}}
+  // expected-note@*:* 2 {{candidate function}}
+}
+
+void struct_dest(float cmp, float v) {
+  InterlockedCompareStoreFloatBitwise(gs_s, cmp, v); // expected-error{{no matching function for call to 'InterlockedCompareStoreFloatBitwise'}}
+  // expected-note@*:* 2 {{candidate function}}
+}
+
+void direct_too_few(float cmp) {
+  __builtin_hlsl_interlocked_compare_store_float_bitwise(gs_f32, cmp);
+  // expected-error at -1 {{too few arguments to function call, expected 3, have 2}}
+}
+
+void direct_too_many(float cmp, float v, float extra) {
+  __builtin_hlsl_interlocked_compare_store_float_bitwise(gs_f32, cmp, v, extra);
+  // expected-error at -1 {{too many arguments to function call, expected 3, have 4}}
+}
+
+void direct_integer_dest(int cmp, int v) {
+  __builtin_hlsl_interlocked_compare_store_float_bitwise(gs_i32, cmp, v);
+  // expected-error at -1 {{1st argument must be a scalar 32 bit floating-point type (was 'int')}}
+}
+
+void direct_double_dest(double cmp, double v) {
+  __builtin_hlsl_interlocked_compare_store_float_bitwise(gs_f64, cmp, v);
+  // expected-error at -1 {{1st argument must be a scalar 32 bit floating-point type (was 'double')}}
+}
+
+void direct_half_dest(half cmp, half v) {
+  __builtin_hlsl_interlocked_compare_store_float_bitwise(gs_f16, cmp, v);
+  // expected-error at -1 {{1st argument must be a scalar 32 bit floating-point type (was 'half')}}
+}
+
+void direct_non_scalar_dest() {
+  S local_s;
+  __builtin_hlsl_interlocked_compare_store_float_bitwise(local_s, 1.0f, 2.0f);
+  // expected-error at -1 {{1st argument must be a scalar 32 bit floating-point type (was 'S')}}
+}
+
+void direct_nonlvalue_dest(float cmp, float v) {
+  __builtin_hlsl_interlocked_compare_store_float_bitwise(1.0f, cmp, v);
+  // expected-error at -1 {{cannot bind non-lvalue argument '1.F' to out parameter}}
+}
+
+void direct_default_as_dest(float cmp, float v) {
+  float local;
+  __builtin_hlsl_interlocked_compare_store_float_bitwise(local, cmp, v);
+  // expected-error at -1 {{1st argument to atomic builtin must reference groupshared or device memory (was 'float')}}
+}
+
+// The last argument is the new value rather than an out parameter, so an
+// rvalue is accepted here.
+void direct_rvalue_value_ok() {
+  __builtin_hlsl_interlocked_compare_store_float_bitwise(gs_f32, 1.0f, 2.0f);
+}
diff --git a/llvm/test/CodeGen/DirectX/ResourceAtomicCompareStoreFloat.ll b/llvm/test/CodeGen/DirectX/ResourceAtomicCompareStoreFloat.ll
new file mode 100644
index 0000000000000..d544976f4e129
--- /dev/null
+++ b/llvm/test/CodeGen/DirectX/ResourceAtomicCompareStoreFloat.ll
@@ -0,0 +1,30 @@
+; RUN: opt -S -dxil-resource-access -dxil-op-lower -mtriple=dxil-pc-shadermodel6.6-compute %s | FileCheck %s
+
+; InterlockedCompareStoreFloatBitwise compares the bit pattern of a float, so
+; clang emits an i32 `cmpxchg` on a float resource. Lowering must key the DXIL
+; op off the operand type rather than the resource element type.
+
+target triple = "dxil-pc-shadermodel6.6-compute"
+
+; CHECK-LABEL: define void @typed_buffer_float_compare_store
+define void @typed_buffer_float_compare_store(i32 %index, i32 %cmp, i32 %val) {
+  %buffer = call target("dx.TypedBuffer", float, 1, 0, 0)
+      @llvm.dx.resource.handlefrombinding(i32 0, i32 0, i32 1, i32 0, ptr null)
+  %ptr = call ptr @llvm.dx.resource.getpointer(
+      target("dx.TypedBuffer", float, 1, 0, 0) %buffer, i32 %index)
+  ; CHECK: call i32 @dx.op.atomicCompareExchange.i32(i32 79, %dx.types.Handle %{{.*}}, i32 %index, i32 poison, i32 0, i32 %cmp, i32 %val)
+  %old = cmpxchg ptr %ptr, i32 %cmp, i32 %val monotonic monotonic
+  ret void
+}
+
+; The raw buffer path carries no element type, so the same i32 op applies.
+; CHECK-LABEL: define void @raw_buffer_float_compare_store
+define void @raw_buffer_float_compare_store(i32 %offset, i32 %cmp, i32 %val) {
+  %buffer = call target("dx.RawBuffer", i8, 1, 0, 0)
+      @llvm.dx.resource.handlefrombinding(i32 0, i32 0, i32 1, i32 0, ptr null)
+  %ptr = call ptr @llvm.dx.resource.getpointer(
+      target("dx.RawBuffer", i8, 1, 0, 0) %buffer, i32 %offset)
+  ; CHECK: call i32 @dx.op.atomicCompareExchange.i32(i32 79, %dx.types.Handle %{{.*}}, i32 %offset, i32 poison, i32 0, i32 %cmp, i32 %val)
+  %old = cmpxchg ptr %ptr, i32 %cmp, i32 %val monotonic monotonic
+  ret void
+}



More information about the llvm-branch-commits mailing list