[llvm] [DXIL] Add GroupMemoryBarrierWithGroupSync intrinsic (PR #112641)

Adam Yang via llvm-commits llvm-commits at lists.llvm.org
Wed Oct 16 17:59:21 PDT 2024


https://github.com/adam-yang created https://github.com/llvm/llvm-project/pull/112641

Example implementation of lowering GroupMemoryBarrierWithGroupSync via new tablegen based method.

>From 6fb6bd74848180a5cf007723580f7302b07699a6 Mon Sep 17 00:00:00 2001
From: Adam Yang <31109344+adam-yang at users.noreply.github.com>
Date: Tue, 8 Oct 2024 00:53:11 -0700
Subject: [PATCH 1/9] Added GroupMemoryBarrierWithGroupSync intrinsic for DXIL

---
 llvm/include/llvm/IR/IntrinsicsDirectX.td     |  2 ++
 llvm/lib/Target/DirectX/DXIL.td               |  9 ++++++
 llvm/lib/Target/DirectX/DXILConstants.h       |  7 +++++
 llvm/lib/Target/DirectX/DXILOpLowering.cpp    | 31 +++++++++++++++++++
 .../GroupMemoryBarrierWithGroupSync.ll        |  8 +++++
 5 files changed, 57 insertions(+)
 create mode 100644 llvm/test/CodeGen/DirectX/GroupMemoryBarrierWithGroupSync.ll

diff --git a/llvm/include/llvm/IR/IntrinsicsDirectX.td b/llvm/include/llvm/IR/IntrinsicsDirectX.td
index 1cf6acbf126475..32ff2c0de04662 100644
--- a/llvm/include/llvm/IR/IntrinsicsDirectX.td
+++ b/llvm/include/llvm/IR/IntrinsicsDirectX.td
@@ -88,4 +88,6 @@ 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_groupMemoryBarrierWithGroupSync : DefaultAttrsIntrinsic<[], [], []>;
 }
diff --git a/llvm/lib/Target/DirectX/DXIL.td b/llvm/lib/Target/DirectX/DXIL.td
index e8f56b18730d71..a4ffc38d2af297 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;
@@ -810,3 +811,11 @@ def WaveGetLaneIndex : DXILOp<111, waveGetLaneIndex> {
   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..93b9aa49379f3e 100644
--- a/llvm/lib/Target/DirectX/DXILOpLowering.cpp
+++ b/llvm/lib/Target/DirectX/DXILOpLowering.cpp
@@ -210,6 +210,34 @@ class OpLowerer {
     });
   }
 
+  [[nodiscard]] bool lowerBarrier(Function& F, Intrinsic::ID IntrId) {
+    IRBuilder<> &IRB = OpBuilder.getIRB();
+    return replaceFunction(F, [&](CallInst *CI) -> Error {
+      unsigned BarrierMode = 0;
+      switch (IntrId) {
+      default:
+        report_fatal_error("Unhandled barrier operation type.");
+        break;
+      case Intrinsic::dx_groupMemoryBarrierWithGroupSync:
+        BarrierMode = (unsigned)dxil::BarrierMode::TGSMFence | (unsigned)dxil::BarrierMode::SyncThreadGroup;
+        break;
+      }
+
+      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 +505,9 @@ class OpLowerer {
     HasErrors |= replaceFunctionWithOp(F, OpCode);                             \
     break;
 #include "DXILOperation.inc"
+      case Intrinsic::dx_groupMemoryBarrierWithGroupSync:
+        HasErrors |= lowerBarrier(F, ID);
+        break;
       case Intrinsic::dx_handle_fromBinding:
         HasErrors |= lowerHandleFromBinding(F);
         break;
diff --git a/llvm/test/CodeGen/DirectX/GroupMemoryBarrierWithGroupSync.ll b/llvm/test/CodeGen/DirectX/GroupMemoryBarrierWithGroupSync.ll
new file mode 100644
index 00000000000000..a99c6757814f3b
--- /dev/null
+++ b/llvm/test/CodeGen/DirectX/GroupMemoryBarrierWithGroupSync.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.groupMemoryBarrierWithGroupSync()
+  ret void
+}
\ No newline at end of file

>From ea6e1b3f6d0b92f4b324cd8a535eb0012221c496 Mon Sep 17 00:00:00 2001
From: Adam Yang <hanbyang at microsoft.com>
Date: Thu, 10 Oct 2024 12:25:26 -0700
Subject: [PATCH 2/9] Changed naming convention and fixed formatting

---
 llvm/include/llvm/IR/IntrinsicsDirectX.td     |  2 +-
 llvm/lib/Target/DirectX/DXILOpLowering.cpp    | 23 ++++++++-----------
 ...> group_memory_barrier_with_group_sync.ll} |  2 +-
 3 files changed, 12 insertions(+), 15 deletions(-)
 rename llvm/test/CodeGen/DirectX/{GroupMemoryBarrierWithGroupSync.ll => group_memory_barrier_with_group_sync.ll} (80%)

diff --git a/llvm/include/llvm/IR/IntrinsicsDirectX.td b/llvm/include/llvm/IR/IntrinsicsDirectX.td
index 32ff2c0de04662..d891809da01bcb 100644
--- a/llvm/include/llvm/IR/IntrinsicsDirectX.td
+++ b/llvm/include/llvm/IR/IntrinsicsDirectX.td
@@ -89,5 +89,5 @@ def int_dx_sign : DefaultAttrsIntrinsic<[LLVMScalarOrSameVectorWidth<0, llvm_i32
 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_groupMemoryBarrierWithGroupSync : DefaultAttrsIntrinsic<[], [], []>;
+def int_dx_group_memory_barrier_with_group_sync : DefaultAttrsIntrinsic<[], [], []>;
 }
diff --git a/llvm/lib/Target/DirectX/DXILOpLowering.cpp b/llvm/lib/Target/DirectX/DXILOpLowering.cpp
index 93b9aa49379f3e..f3ff372af8d201 100644
--- a/llvm/lib/Target/DirectX/DXILOpLowering.cpp
+++ b/llvm/lib/Target/DirectX/DXILOpLowering.cpp
@@ -210,19 +210,14 @@ class OpLowerer {
     });
   }
 
-  [[nodiscard]] bool lowerBarrier(Function& F, Intrinsic::ID IntrId) {
+  [[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 {
-      unsigned BarrierMode = 0;
-      switch (IntrId) {
-      default:
-        report_fatal_error("Unhandled barrier operation type.");
-        break;
-      case Intrinsic::dx_groupMemoryBarrierWithGroupSync:
-        BarrierMode = (unsigned)dxil::BarrierMode::TGSMFence | (unsigned)dxil::BarrierMode::SyncThreadGroup;
-        break;
-      }
-
       std::array<Value *, 1> Args{IRB.getInt32(BarrierMode)};
 
       IRB.SetInsertPoint(CI);
@@ -505,8 +500,10 @@ class OpLowerer {
     HasErrors |= replaceFunctionWithOp(F, OpCode);                             \
     break;
 #include "DXILOperation.inc"
-      case Intrinsic::dx_groupMemoryBarrierWithGroupSync:
-        HasErrors |= lowerBarrier(F, ID);
+      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);
diff --git a/llvm/test/CodeGen/DirectX/GroupMemoryBarrierWithGroupSync.ll b/llvm/test/CodeGen/DirectX/group_memory_barrier_with_group_sync.ll
similarity index 80%
rename from llvm/test/CodeGen/DirectX/GroupMemoryBarrierWithGroupSync.ll
rename to llvm/test/CodeGen/DirectX/group_memory_barrier_with_group_sync.ll
index a99c6757814f3b..48907647c660f8 100644
--- a/llvm/test/CodeGen/DirectX/GroupMemoryBarrierWithGroupSync.ll
+++ b/llvm/test/CodeGen/DirectX/group_memory_barrier_with_group_sync.ll
@@ -3,6 +3,6 @@
 define void @test_group_memory_barrier_with_group_sync() {
 entry:
   ; CHECK: call void @dx.op.barrier(i32 80, i32 9)
-  call void @llvm.dx.groupMemoryBarrierWithGroupSync()
+  call void @llvm.dx.group.memory.barrier.with.group.sync()
   ret void
 }
\ No newline at end of file

>From 7ee8d540c359757eab545e04d8091e52f6b932dc Mon Sep 17 00:00:00 2001
From: Adam Yang <hanbyang at microsoft.com>
Date: Thu, 10 Oct 2024 12:59:52 -0700
Subject: [PATCH 3/9] Got rid of the noduplicate attr

---
 llvm/lib/Target/DirectX/DXIL.td | 3 +--
 1 file changed, 1 insertion(+), 2 deletions(-)

diff --git a/llvm/lib/Target/DirectX/DXIL.td b/llvm/lib/Target/DirectX/DXIL.td
index a4ffc38d2af297..bff9fa3ca3b68e 100644
--- a/llvm/lib/Target/DirectX/DXIL.td
+++ b/llvm/lib/Target/DirectX/DXIL.td
@@ -277,7 +277,6 @@ 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;
@@ -817,5 +816,5 @@ def Barrier : DXILOp<80, barrier> {
   let arguments = [Int32Ty];
   let result = VoidTy;
   let stages = [Stages<DXIL1_0, [compute, library]>];
-  let attributes = [Attributes<DXIL1_0, [NoDuplicate]>];
+  let attributes = [Attributes<DXIL1_0, []>];
 }

>From c7d83cf8ae32d98b0677ff1a88f74fe4827dd61f Mon Sep 17 00:00:00 2001
From: Adam Yang <hanbyang at microsoft.com>
Date: Thu, 10 Oct 2024 13:03:52 -0700
Subject: [PATCH 4/9] Shader model changed to 6.0

---
 .../CodeGen/DirectX/group_memory_barrier_with_group_sync.ll     | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

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
index 48907647c660f8..7cacbe778ac952 100644
--- a/llvm/test/CodeGen/DirectX/group_memory_barrier_with_group_sync.ll
+++ b/llvm/test/CodeGen/DirectX/group_memory_barrier_with_group_sync.ll
@@ -1,4 +1,4 @@
-; RUN: opt -S  -dxil-op-lower -mtriple=dxil-pc-shadermodel6.3-library < %s | FileCheck %s --check-prefix=CHECK
+; RUN: opt -S  -dxil-op-lower -mtriple=dxil-pc-shadermodel6.0-library < %s | FileCheck %s --check-prefix=CHECK
 
 define void @test_group_memory_barrier_with_group_sync() {
 entry:

>From 279fe4550883f8941b113641653971222e16dec6 Mon Sep 17 00:00:00 2001
From: Adam Yang <hanbyang at microsoft.com>
Date: Thu, 10 Oct 2024 13:07:03 -0700
Subject: [PATCH 5/9] Fixed the incorrect arguments list

---
 llvm/lib/Target/DirectX/DXIL.td | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

diff --git a/llvm/lib/Target/DirectX/DXIL.td b/llvm/lib/Target/DirectX/DXIL.td
index bff9fa3ca3b68e..ae55eb9f02cb55 100644
--- a/llvm/lib/Target/DirectX/DXIL.td
+++ b/llvm/lib/Target/DirectX/DXIL.td
@@ -813,7 +813,7 @@ def WaveGetLaneIndex : DXILOp<111, waveGetLaneIndex> {
 
 def Barrier : DXILOp<80, barrier> {
   let Doc = "inserts a memory barrier in the shader";
-  let arguments = [Int32Ty];
+  let arguments = [];
   let result = VoidTy;
   let stages = [Stages<DXIL1_0, [compute, library]>];
   let attributes = [Attributes<DXIL1_0, []>];

>From a4258a8280822af40490f4f4e260fa8ff1c38f0d Mon Sep 17 00:00:00 2001
From: Adam Yang <hanbyang at microsoft.com>
Date: Thu, 10 Oct 2024 13:13:21 -0700
Subject: [PATCH 6/9] Revert "Shader model changed to 6.0"

This reverts commit c7d83cf8ae32d98b0677ff1a88f74fe4827dd61f.
---
 .../CodeGen/DirectX/group_memory_barrier_with_group_sync.ll     | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

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
index 7cacbe778ac952..48907647c660f8 100644
--- a/llvm/test/CodeGen/DirectX/group_memory_barrier_with_group_sync.ll
+++ b/llvm/test/CodeGen/DirectX/group_memory_barrier_with_group_sync.ll
@@ -1,4 +1,4 @@
-; RUN: opt -S  -dxil-op-lower -mtriple=dxil-pc-shadermodel6.0-library < %s | FileCheck %s --check-prefix=CHECK
+; 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:

>From 70688f663e91c5cf3a8d57aa1e636bf34dbac1cc Mon Sep 17 00:00:00 2001
From: Adam Yang <31109344+adam-yang at users.noreply.github.com>
Date: Tue, 15 Oct 2024 15:56:50 -0700
Subject: [PATCH 7/9] Added another intermediate memory barrier intrinsic that
 maps directly to the barrier dxil op

---
 llvm/include/llvm/IR/IntrinsicsDirectX.td     |  1 +
 llvm/lib/Target/DirectX/DXIL.td               |  3 +-
 .../Target/DirectX/DXILIntrinsicExpansion.cpp | 26 +++++++++++++++++
 llvm/lib/Target/DirectX/DXILOpLowering.cpp    | 28 -------------------
 .../group_memory_barrier_with_group_sync.ll   |  2 +-
 5 files changed, 30 insertions(+), 30 deletions(-)

diff --git a/llvm/include/llvm/IR/IntrinsicsDirectX.td b/llvm/include/llvm/IR/IntrinsicsDirectX.td
index d891809da01bcb..af60f5e2136f99 100644
--- a/llvm/include/llvm/IR/IntrinsicsDirectX.td
+++ b/llvm/include/llvm/IR/IntrinsicsDirectX.td
@@ -89,5 +89,6 @@ def int_dx_sign : DefaultAttrsIntrinsic<[LLVMScalarOrSameVectorWidth<0, llvm_i32
 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_memory_barrier : DefaultAttrsIntrinsic<[], [llvm_i32_ty], []>;
 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 ae55eb9f02cb55..b596aa8239a0e2 100644
--- a/llvm/lib/Target/DirectX/DXIL.td
+++ b/llvm/lib/Target/DirectX/DXIL.td
@@ -813,7 +813,8 @@ def WaveGetLaneIndex : DXILOp<111, waveGetLaneIndex> {
 
 def Barrier : DXILOp<80, barrier> {
   let Doc = "inserts a memory barrier in the shader";
-  let arguments = [];
+  let LLVMIntrinsic = int_dx_memory_barrier;
+  let arguments = [Int32Ty];
   let result = VoidTy;
   let stages = [Stages<DXIL1_0, [compute, library]>];
   let attributes = [Attributes<DXIL1_0, []>];
diff --git a/llvm/lib/Target/DirectX/DXILIntrinsicExpansion.cpp b/llvm/lib/Target/DirectX/DXILIntrinsicExpansion.cpp
index 1e84a7216013da..2bbd5c8e3d7cf0 100644
--- a/llvm/lib/Target/DirectX/DXILIntrinsicExpansion.cpp
+++ b/llvm/lib/Target/DirectX/DXILIntrinsicExpansion.cpp
@@ -10,6 +10,7 @@
 //  opcodes in DirectX Intermediate Language (DXIL).
 //===----------------------------------------------------------------------===//
 
+#include "DXILConstants.h"
 #include "DXILIntrinsicExpansion.h"
 #include "DirectX.h"
 #include "llvm/ADT/STLExtras.h"
@@ -65,6 +66,7 @@ static bool isIntrinsicExpansion(Function &F) {
   case Intrinsic::dx_sign:
   case Intrinsic::dx_step:
   case Intrinsic::dx_radians:
+  case Intrinsic::dx_group_memory_barrier_with_group_sync:
     return true;
   }
   return false;
@@ -451,6 +453,27 @@ static Value *expandRadiansIntrinsic(CallInst *Orig) {
   return Builder.CreateFMul(X, PiOver180);
 }
 
+static Value *expandMemoryBarrier(CallInst *Orig, Intrinsic::ID IntrinsicId) {
+  assert(IntrinsicId == Intrinsic::dx_group_memory_barrier_with_group_sync);
+  unsigned BarrierMode = 0;
+  switch (IntrinsicId) {
+  case Intrinsic::dx_group_memory_barrier_with_group_sync:
+    BarrierMode = (unsigned)dxil::BarrierMode::TGSMFence |
+                  (unsigned)dxil::BarrierMode::SyncThreadGroup;
+    break;
+  default:
+    report_fatal_error(Twine("Unexpected memory barrier intrinsic."),
+                       /* gen_crash_diag=*/false);
+    break;
+  }
+
+  IRBuilder<> Builder(Orig);
+  return Builder.CreateIntrinsic(
+      Builder.getVoidTy(), Intrinsic::dx_memory_barrier,
+      ArrayRef<Value *>{Builder.getInt32(BarrierMode)}, nullptr,
+      Orig->getName());
+}
+
 static Intrinsic::ID getMaxForClamp(Type *ElemTy,
                                     Intrinsic::ID ClampIntrinsic) {
   if (ClampIntrinsic == Intrinsic::dx_uclamp)
@@ -574,6 +597,9 @@ static bool expandIntrinsic(Function &F, CallInst *Orig) {
   case Intrinsic::dx_radians:
     Result = expandRadiansIntrinsic(Orig);
     break;
+  case Intrinsic::dx_group_memory_barrier_with_group_sync:
+    Result = expandMemoryBarrier(Orig, IntrinsicId);
+    break;
   }
   if (Result) {
     Orig->replaceAllUsesWith(Result);
diff --git a/llvm/lib/Target/DirectX/DXILOpLowering.cpp b/llvm/lib/Target/DirectX/DXILOpLowering.cpp
index f3ff372af8d201..c0f8d433833ee7 100644
--- a/llvm/lib/Target/DirectX/DXILOpLowering.cpp
+++ b/llvm/lib/Target/DirectX/DXILOpLowering.cpp
@@ -210,29 +210,6 @@ 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();
 
@@ -500,11 +477,6 @@ 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
index 48907647c660f8..c43625755d6efc 100644
--- a/llvm/test/CodeGen/DirectX/group_memory_barrier_with_group_sync.ll
+++ b/llvm/test/CodeGen/DirectX/group_memory_barrier_with_group_sync.ll
@@ -1,4 +1,4 @@
-; RUN: opt -S  -dxil-op-lower -mtriple=dxil-pc-shadermodel6.3-library < %s | FileCheck %s --check-prefix=CHECK
+; RUN: opt -S -dxil-intrinsic-expansion -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:

>From e41187a1a32115b313ccd1972e895835fc06ac92 Mon Sep 17 00:00:00 2001
From: Adam Yang <31109344+adam-yang at users.noreply.github.com>
Date: Tue, 15 Oct 2024 16:27:13 -0700
Subject: [PATCH 8/9] Format

---
 llvm/lib/Target/DirectX/DXILIntrinsicExpansion.cpp | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

diff --git a/llvm/lib/Target/DirectX/DXILIntrinsicExpansion.cpp b/llvm/lib/Target/DirectX/DXILIntrinsicExpansion.cpp
index 2bbd5c8e3d7cf0..e24b04f913be52 100644
--- a/llvm/lib/Target/DirectX/DXILIntrinsicExpansion.cpp
+++ b/llvm/lib/Target/DirectX/DXILIntrinsicExpansion.cpp
@@ -10,8 +10,8 @@
 //  opcodes in DirectX Intermediate Language (DXIL).
 //===----------------------------------------------------------------------===//
 
-#include "DXILConstants.h"
 #include "DXILIntrinsicExpansion.h"
+#include "DXILConstants.h"
 #include "DirectX.h"
 #include "llvm/ADT/STLExtras.h"
 #include "llvm/ADT/SmallVector.h"

>From 36dbb21745a6e1c2581ef35e6af3428d9253de25 Mon Sep 17 00:00:00 2001
From: Adam Yang <hanbyang at microsoft.com>
Date: Wed, 16 Oct 2024 17:48:25 -0700
Subject: [PATCH 9/9] Tablegen simple

---
 llvm/include/llvm/IR/IntrinsicsDirectX.td     |  1 -
 llvm/lib/Target/DirectX/DXIL.td               | 14 +++-
 .../Target/DirectX/DXILIntrinsicExpansion.cpp | 26 --------
 llvm/lib/Target/DirectX/DXILOpLowering.cpp    | 25 +++----
 .../group_memory_barrier_with_group_sync.ll   |  2 +-
 llvm/utils/TableGen/DXILEmitter.cpp           | 66 +++++++++++++++----
 6 files changed, 80 insertions(+), 54 deletions(-)

diff --git a/llvm/include/llvm/IR/IntrinsicsDirectX.td b/llvm/include/llvm/IR/IntrinsicsDirectX.td
index af60f5e2136f99..d891809da01bcb 100644
--- a/llvm/include/llvm/IR/IntrinsicsDirectX.td
+++ b/llvm/include/llvm/IR/IntrinsicsDirectX.td
@@ -89,6 +89,5 @@ def int_dx_sign : DefaultAttrsIntrinsic<[LLVMScalarOrSameVectorWidth<0, llvm_i32
 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_memory_barrier : DefaultAttrsIntrinsic<[], [llvm_i32_ty], []>;
 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 b596aa8239a0e2..bf5b9ba5006ce8 100644
--- a/llvm/lib/Target/DirectX/DXIL.td
+++ b/llvm/lib/Target/DirectX/DXIL.td
@@ -293,6 +293,11 @@ class Attributes<Version ver = DXIL1_0, list<DXILAttribute> attrs> {
   list<DXILAttribute> op_attrs = attrs;
 }
 
+class IntrinsicSelect<Intrinsic intr, list<string> extra_args> {
+  Intrinsic Intr = intr;
+  list<string> ExtraArgs = extra_args;
+}
+
 // Abstraction DXIL Operation
 class DXILOp<int opcode, DXILOpClass opclass> {
   // A short description of the operation
@@ -307,6 +312,8 @@ class DXILOp<int opcode, DXILOpClass opclass> {
   // LLVM Intrinsic DXIL Operation maps to
   Intrinsic LLVMIntrinsic = ?;
 
+  list<IntrinsicSelect> intrinsic_selects = [];
+
   // Result type of the op
   DXILOpParamType result;
 
@@ -813,7 +820,12 @@ def WaveGetLaneIndex : DXILOp<111, waveGetLaneIndex> {
 
 def Barrier : DXILOp<80, barrier> {
   let Doc = "inserts a memory barrier in the shader";
-  let LLVMIntrinsic = int_dx_memory_barrier;
+  let intrinsic_selects = [
+    IntrinsicSelect<
+        int_dx_group_memory_barrier_with_group_sync,
+        [ "OpBuilder.getIRB().getInt32((unsigned)BarrierMode::SyncThreadGroup | (unsigned)BarrierMode::TGSMFence)" ]>,
+  ];
+
   let arguments = [Int32Ty];
   let result = VoidTy;
   let stages = [Stages<DXIL1_0, [compute, library]>];
diff --git a/llvm/lib/Target/DirectX/DXILIntrinsicExpansion.cpp b/llvm/lib/Target/DirectX/DXILIntrinsicExpansion.cpp
index e24b04f913be52..1e84a7216013da 100644
--- a/llvm/lib/Target/DirectX/DXILIntrinsicExpansion.cpp
+++ b/llvm/lib/Target/DirectX/DXILIntrinsicExpansion.cpp
@@ -11,7 +11,6 @@
 //===----------------------------------------------------------------------===//
 
 #include "DXILIntrinsicExpansion.h"
-#include "DXILConstants.h"
 #include "DirectX.h"
 #include "llvm/ADT/STLExtras.h"
 #include "llvm/ADT/SmallVector.h"
@@ -66,7 +65,6 @@ static bool isIntrinsicExpansion(Function &F) {
   case Intrinsic::dx_sign:
   case Intrinsic::dx_step:
   case Intrinsic::dx_radians:
-  case Intrinsic::dx_group_memory_barrier_with_group_sync:
     return true;
   }
   return false;
@@ -453,27 +451,6 @@ static Value *expandRadiansIntrinsic(CallInst *Orig) {
   return Builder.CreateFMul(X, PiOver180);
 }
 
-static Value *expandMemoryBarrier(CallInst *Orig, Intrinsic::ID IntrinsicId) {
-  assert(IntrinsicId == Intrinsic::dx_group_memory_barrier_with_group_sync);
-  unsigned BarrierMode = 0;
-  switch (IntrinsicId) {
-  case Intrinsic::dx_group_memory_barrier_with_group_sync:
-    BarrierMode = (unsigned)dxil::BarrierMode::TGSMFence |
-                  (unsigned)dxil::BarrierMode::SyncThreadGroup;
-    break;
-  default:
-    report_fatal_error(Twine("Unexpected memory barrier intrinsic."),
-                       /* gen_crash_diag=*/false);
-    break;
-  }
-
-  IRBuilder<> Builder(Orig);
-  return Builder.CreateIntrinsic(
-      Builder.getVoidTy(), Intrinsic::dx_memory_barrier,
-      ArrayRef<Value *>{Builder.getInt32(BarrierMode)}, nullptr,
-      Orig->getName());
-}
-
 static Intrinsic::ID getMaxForClamp(Type *ElemTy,
                                     Intrinsic::ID ClampIntrinsic) {
   if (ClampIntrinsic == Intrinsic::dx_uclamp)
@@ -597,9 +574,6 @@ static bool expandIntrinsic(Function &F, CallInst *Orig) {
   case Intrinsic::dx_radians:
     Result = expandRadiansIntrinsic(Orig);
     break;
-  case Intrinsic::dx_group_memory_barrier_with_group_sync:
-    Result = expandMemoryBarrier(Orig, IntrinsicId);
-    break;
   }
   if (Result) {
     Orig->replaceAllUsesWith(Result);
diff --git a/llvm/lib/Target/DirectX/DXILOpLowering.cpp b/llvm/lib/Target/DirectX/DXILOpLowering.cpp
index c0f8d433833ee7..1be20cf112bb6f 100644
--- a/llvm/lib/Target/DirectX/DXILOpLowering.cpp
+++ b/llvm/lib/Target/DirectX/DXILOpLowering.cpp
@@ -105,20 +105,21 @@ class OpLowerer {
     return false;
   }
 
-  [[nodiscard]]
-  bool replaceFunctionWithOp(Function &F, dxil::OpCode DXILOp) {
+  [[nodiscard]] bool replaceFunctionWithOp(Function &F, dxil::OpCode DXILOp,
+                                           ArrayRef<Value *> ExtraArgs) {
     bool IsVectorArgExpansion = isVectorArgExpansion(F);
     return replaceFunction(F, [&](CallInst *CI) -> Error {
-      SmallVector<Value *> Args;
-      OpBuilder.getIRB().SetInsertPoint(CI);
+      SmallVector<Value *> NewArgs;
       if (IsVectorArgExpansion) {
-        SmallVector<Value *> NewArgs = argVectorFlatten(CI, OpBuilder.getIRB());
-        Args.append(NewArgs.begin(), NewArgs.end());
+        NewArgs = argVectorFlatten(CI, OpBuilder.getIRB());
       } else
-        Args.append(CI->arg_begin(), CI->arg_end());
+        NewArgs.append(CI->arg_begin(), CI->arg_end());
 
-      Expected<CallInst *> OpCall =
-          OpBuilder.tryCreateOp(DXILOp, Args, CI->getName(), F.getReturnType());
+      NewArgs.append(ExtraArgs.begin(), ExtraArgs.end());
+
+      OpBuilder.getIRB().SetInsertPoint(CI);
+      Expected<CallInst *> OpCall = OpBuilder.tryCreateOp(
+          DXILOp, NewArgs, CI->getName(), F.getReturnType());
       if (Error E = OpCall.takeError())
         return E;
 
@@ -472,9 +473,9 @@ class OpLowerer {
       switch (ID) {
       default:
         continue;
-#define DXIL_OP_INTRINSIC(OpCode, Intrin)                                      \
-  case Intrin:                                                                 \
-    HasErrors |= replaceFunctionWithOp(F, OpCode);                             \
+#define DXIL_OP_INTRINSIC(OpCode, Intrin, ExtraArgs)                         \
+  case Intrin:                                                               \
+    HasErrors |= replaceFunctionWithOp(F, OpCode, ExtraArgs);                \
     break;
 #include "DXILOperation.inc"
       case Intrinsic::dx_handle_fromBinding:
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
index c43625755d6efc..cd588d464e302b 100644
--- a/llvm/test/CodeGen/DirectX/group_memory_barrier_with_group_sync.ll
+++ b/llvm/test/CodeGen/DirectX/group_memory_barrier_with_group_sync.ll
@@ -1,4 +1,4 @@
-; RUN: opt -S -dxil-intrinsic-expansion -dxil-op-lower -mtriple=dxil-pc-shadermodel6.3-library < %s | FileCheck %s --check-prefix=CHECK
+; 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:
diff --git a/llvm/utils/TableGen/DXILEmitter.cpp b/llvm/utils/TableGen/DXILEmitter.cpp
index 06bf7a0c0a8372..ffb5892975a7fc 100644
--- a/llvm/utils/TableGen/DXILEmitter.cpp
+++ b/llvm/utils/TableGen/DXILEmitter.cpp
@@ -33,6 +33,11 @@ using namespace llvm::dxil;
 
 namespace {
 
+struct DXILIntrinsicSelect {
+  StringRef Intrinsic;
+  SmallVector<StringRef, 4> ExtraArgs;
+};
+
 struct DXILOperationDesc {
   std::string OpName; // name of DXIL operation
   int OpCode;         // ID of DXIL operation
@@ -43,8 +48,7 @@ struct DXILOperationDesc {
   SmallVector<const Record *> OverloadRecs;
   SmallVector<const Record *> StageRecs;
   SmallVector<const Record *> AttrRecs;
-  StringRef Intrinsic; // The llvm intrinsic map to OpName. Default is "" which
-                       // means no map exists
+  SmallVector<DXILIntrinsicSelect> IntrinsicSelects;
   SmallVector<StringRef, 4>
       ShaderStages; // shader stages to which this applies, empty for all.
   int OverloadParamIndex;             // Index of parameter with overload type.
@@ -158,14 +162,43 @@ DXILOperationDesc::DXILOperationDesc(const Record *R) {
                            OpName);
   }
 
-  const RecordVal *RV = R->getValue("LLVMIntrinsic");
-  if (RV && RV->getValue()) {
-    if (DefInit *DI = dyn_cast<DefInit>(RV->getValue())) {
-      auto *IntrinsicDef = DI->getDef();
-      auto DefName = IntrinsicDef->getName();
-      assert(DefName.starts_with("int_") && "invalid intrinsic name");
-      // Remove the int_ from intrinsic name.
-      Intrinsic = DefName.substr(4);
+  auto GetIntrinsicName = [](const RecordVal *RV) -> StringRef {
+    if (RV && RV->getValue()) {
+      if (DefInit *DI = dyn_cast<DefInit>(RV->getValue())) {
+        auto *IntrinsicDef = DI->getDef();
+        auto DefName = IntrinsicDef->getName();
+        assert(DefName.starts_with("int_") && "invalid intrinsic name");
+        // Remove the int_ from intrinsic name.
+        return DefName.substr(4);
+      }
+    }
+    return "";
+  };
+
+  {
+    DXILIntrinsicSelect IntrSelect;
+    IntrSelect.Intrinsic = GetIntrinsicName(R->getValue("LLVMIntrinsic"));
+    if (IntrSelect.Intrinsic.size())
+      IntrinsicSelects.emplace_back(std::move(IntrSelect));
+  }
+
+  Recs = R->getValueAsListOfDefs("intrinsic_selects");
+  if (Recs.size()) {
+    if (IntrinsicSelects.size()) {
+      PrintFatalError(R,
+                      Twine("LLVMIntrinsic and intrinsic_match cannot be both "
+                            "defined for DXIL operation - ") +
+                          OpName);
+    } else {
+      for (const Record *R : Recs) {
+        DXILIntrinsicSelect IntrSelect;
+        IntrSelect.Intrinsic = GetIntrinsicName(R->getValue("Intr"));
+        auto ExtraArgs = R->getValueAsListOfStrings("ExtraArgs");
+        for (StringRef Arg : ExtraArgs) {
+          IntrSelect.ExtraArgs.push_back(Arg);
+        }
+        IntrinsicSelects.emplace_back(std::move(IntrSelect));
+      }
     }
   }
 }
@@ -378,10 +411,17 @@ static void emitDXILIntrinsicMap(ArrayRef<DXILOperationDesc> Ops,
   OS << "#ifdef DXIL_OP_INTRINSIC\n";
   OS << "\n";
   for (const auto &Op : Ops) {
-    if (Op.Intrinsic.empty())
+    if (Op.IntrinsicSelects.empty()) {
       continue;
-    OS << "DXIL_OP_INTRINSIC(dxil::OpCode::" << Op.OpName
-       << ", Intrinsic::" << Op.Intrinsic << ")\n";
+    }
+    for (const DXILIntrinsicSelect &MappedIntr : Op.IntrinsicSelects) {
+      OS << "DXIL_OP_INTRINSIC(dxil::OpCode::" << Op.OpName
+         << ", Intrinsic::" << MappedIntr.Intrinsic << ", (ArrayRef<Value *> {";
+      for (const StringRef &Arg : MappedIntr.ExtraArgs) {
+        OS << Arg << ", ";
+      }
+      OS << "}))\n";
+    }
   }
   OS << "\n";
   OS << "#undef DXIL_OP_INTRINSIC\n";



More information about the llvm-commits mailing list