[llvm] [DXIL] Add GroupMemoryBarrierWithGroupSync intrinsic (PR #111884)

[via llvm-commits]

llvmbot wrote:


<!--LLVM PR SUMMARY COMMENT-->

@llvm/pr-subscribers-backend-directx

Author: Adam Yang (adam-yang)

<details>
<summary>Changes</summary>

partially fixes #<!-- -->70103

### Changes
- Added int_dx_group_memory_barrier_with_group_sync intrinsic in IntrinsicsDirectX.td
- Added expansion for int_dx_group_memory_barrier_with_group_sync in DXILIntrinsicExpansion.cpp`
- Added DXIL backend test case

### Related PRs
* [[clang][HLSL] Add GroupMemoryBarrierWithGroupSync intrinsic #<!-- -->111883](https://github.com/llvm/llvm-project/pull/111883)
* [[SPIRV] Add GroupMemoryBarrierWithGroupSync intrinsic #<!-- -->111888](https://github.com/llvm/llvm-project/pull/111888)

---
Full diff: https://github.com/llvm/llvm-project/pull/111884.diff


5 Files Affected:

- (modified) llvm/include/llvm/IR/IntrinsicsDirectX.td (+2) 
- (modified) llvm/lib/Target/DirectX/DXIL.td (+10) 
- (modified) llvm/lib/Target/DirectX/DXILConstants.h (+7) 
- (modified) llvm/lib/Target/DirectX/DXILOpLowering.cpp (+28) 
- (added) llvm/test/CodeGen/DirectX/group_memory_barrier_with_group_sync.ll (+8) 


``````````diff
diff --git a/llvm/include/llvm/IR/IntrinsicsDirectX.td b/llvm/include/llvm/IR/IntrinsicsDirectX.td
index f2b9e286ebb476..275a61049980e4 100644
--- a/llvm/include/llvm/IR/IntrinsicsDirectX.td
+++ b/llvm/include/llvm/IR/IntrinsicsDirectX.td
@@ -87,4 +87,6 @@ def int_dx_wave_is_first_lane : DefaultAttrsIntrinsic<[llvm_i1_ty], [], [IntrCon
 def int_dx_sign : DefaultAttrsIntrinsic<[LLVMScalarOrSameVectorWidth<0, llvm_i32_ty>], [llvm_any_ty], [IntrNoMem]>;
 def int_dx_step : DefaultAttrsIntrinsic<[LLVMMatchType<0>], [llvm_anyfloat_ty, LLVMMatchType<0>], [IntrNoMem]>;
 def int_dx_radians : DefaultAttrsIntrinsic<[llvm_anyfloat_ty], [LLVMMatchType<0>], [IntrNoMem]>;
+
+def int_dx_group_memory_barrier_with_group_sync : DefaultAttrsIntrinsic<[], [], []>;
 }
diff --git a/llvm/lib/Target/DirectX/DXIL.td b/llvm/lib/Target/DirectX/DXIL.td
index 9aa0af3e3a6b17..756621840a8453 100644
--- a/llvm/lib/Target/DirectX/DXIL.td
+++ b/llvm/lib/Target/DirectX/DXIL.td
@@ -277,6 +277,7 @@ def IsFeedback : DXILAttribute;
 def IsWave : DXILAttribute;
 def NeedsUniformInputs : DXILAttribute;
 def IsBarrier : DXILAttribute;
+def NoDuplicate : DXILAttribute;
 
 class Overloads<Version ver, list<DXILOpParamType> ols> {
   Version dxil_version = ver;
@@ -801,3 +802,12 @@ def WaveIsFirstLane :  DXILOp<110, waveIsFirstLane> {
   let stages = [Stages<DXIL1_0, [all_stages]>];
   let attributes = [Attributes<DXIL1_0, [ReadNone]>];
 }
+
+def Barrier : DXILOp<80, barrier> {
+  let Doc = "inserts a memory barrier in the shader";
+  let arguments = [Int32Ty];
+  let result = VoidTy;
+  let stages = [Stages<DXIL1_0, [compute, library]>];
+  let attributes = [Attributes<DXIL1_0, [NoDuplicate]>];
+}
+
diff --git a/llvm/lib/Target/DirectX/DXILConstants.h b/llvm/lib/Target/DirectX/DXILConstants.h
index 022cd57795a063..38984727761bb3 100644
--- a/llvm/lib/Target/DirectX/DXILConstants.h
+++ b/llvm/lib/Target/DirectX/DXILConstants.h
@@ -30,6 +30,13 @@ enum class OpParamType : unsigned {
 #include "DXILOperation.inc"
 };
 
+enum class BarrierMode : unsigned {
+  SyncThreadGroup = 0x00000001,
+  UAVFenceGlobal = 0x00000002,
+  UAVFenceThreadGroup = 0x00000004,
+  TGSMFence = 0x00000008,
+};
+
 } // namespace dxil
 } // namespace llvm
 
diff --git a/llvm/lib/Target/DirectX/DXILOpLowering.cpp b/llvm/lib/Target/DirectX/DXILOpLowering.cpp
index c0f8d433833ee7..f3ff372af8d201 100644
--- a/llvm/lib/Target/DirectX/DXILOpLowering.cpp
+++ b/llvm/lib/Target/DirectX/DXILOpLowering.cpp
@@ -210,6 +210,29 @@ class OpLowerer {
     });
   }
 
+  [[nodiscard]] bool lowerBarrier(Function &F, Intrinsic::ID IntrId,
+                                  ArrayRef<dxil::BarrierMode> BarrierModes) {
+    unsigned BarrierMode = 0;
+    for (const dxil::BarrierMode B : BarrierModes) {
+      BarrierMode |= (unsigned)B;
+    }
+    IRBuilder<> &IRB = OpBuilder.getIRB();
+    return replaceFunction(F, [&](CallInst *CI) -> Error {
+      std::array<Value *, 1> Args{IRB.getInt32(BarrierMode)};
+
+      IRB.SetInsertPoint(CI);
+      Expected<CallInst *> OpCall =
+          OpBuilder.tryCreateOp(OpCode::Barrier, Args, CI->getName());
+      if (Error E = OpCall.takeError())
+        return E;
+
+      CI->replaceAllUsesWith(OpCall.get());
+      CI->eraseFromParent();
+
+      return Error::success();
+    });
+  }
+
   [[nodiscard]] bool lowerToBindAndAnnotateHandle(Function &F) {
     IRBuilder<> &IRB = OpBuilder.getIRB();
 
@@ -477,6 +500,11 @@ class OpLowerer {
     HasErrors |= replaceFunctionWithOp(F, OpCode);                             \
     break;
 #include "DXILOperation.inc"
+      case Intrinsic::dx_group_memory_barrier_with_group_sync:
+        HasErrors |= lowerBarrier(
+            F, ID,
+            {dxil::BarrierMode::TGSMFence, dxil::BarrierMode::SyncThreadGroup});
+        break;
       case Intrinsic::dx_handle_fromBinding:
         HasErrors |= lowerHandleFromBinding(F);
         break;
diff --git a/llvm/test/CodeGen/DirectX/group_memory_barrier_with_group_sync.ll b/llvm/test/CodeGen/DirectX/group_memory_barrier_with_group_sync.ll
new file mode 100644
index 00000000000000..48907647c660f8
--- /dev/null
+++ b/llvm/test/CodeGen/DirectX/group_memory_barrier_with_group_sync.ll
@@ -0,0 +1,8 @@
+; RUN: opt -S  -dxil-op-lower -mtriple=dxil-pc-shadermodel6.3-library < %s | FileCheck %s --check-prefix=CHECK
+
+define void @test_group_memory_barrier_with_group_sync() {
+entry:
+  ; CHECK: call void @dx.op.barrier(i32 80, i32 9)
+  call void @llvm.dx.group.memory.barrier.with.group.sync()
+  ret void
+}
\ No newline at end of file

``````````

</details>


https://github.com/llvm/llvm-project/pull/111884