[DirectX] Remove the "checked" variants of dx.resource.load (#120778)

bogner · web-flow · commit 2c7c07df828e · 2025-01-07T07:18:54.000-08:00
We'd introduced separate versions of `llvm.dx.resource.load` with a
struct return to handle the CheckAccessFullyMapped case without making
the IR for the common case unnecessarily complicated. However, at this
point the common case is really `resource.getpointer`, so the ergonomics
of a simplified version of `load` don't actually gain us as much as the
cost of having multiple opcodes.

Drop the `dx.resource.loadchecked` functions and have `dx.resource.load`
consistently return `{element_type, i1}`.
diff --git a/llvm/docs/DirectX/DXILResources.rst b/llvm/docs/DirectX/DXILResources.rst
@@ -296,8 +296,8 @@ instead. That is, ``llvm.dx.resource.load.typedbuffer`` from a
 of 4 floats, and from ``Buffer<double2>`` a vector of two doubles, etc. The
 operations are then expanded out to match DXIL's format during lowering.
 
-In cases where we need ``CheckAccessFullyMapped``, we have a second intrinsic
-that returns an anonymous struct with element-0 being the contained type, and
+In order to support ``CheckAccessFullyMapped``, we need these intrinsics to
+return an anonymous struct with element-0 being the contained type, and
 element-1 being the ``i1`` result of a ``CheckAccessFullyMapped`` call. We
 don't have a separate call to ``CheckAccessFullyMapped`` at all, since that's
 the only operation that can possibly be done on this value. In practice this
@@ -317,8 +317,8 @@ HLSL source, but this actually matches DXC's behaviour in practice.
      - Description
    * - Return value
      -
-     - The contained type of the buffer
-     - The data loaded from the buffer
+     - A structure of the contained type and the check bit
+     - The data loaded from the buffer and the check bit
    * - ``%buffer``
      - 0
      - ``target(dx.TypedBuffer, ...)``
@@ -332,48 +332,22 @@ Examples:
 
 .. code-block:: llvm
 
-   %ret = call <4 x float>
+   %ret = call {<4 x float>, i1}
        @llvm.dx.resource.load.typedbuffer.v4f32.tdx.TypedBuffer_v4f32_0_0_0t(
            target("dx.TypedBuffer", <4 x float>, 0, 0, 0) %buffer, i32 %index)
-   %ret = call float
+   %ret = call {float, i1}
        @llvm.dx.resource.load.typedbuffer.f32.tdx.TypedBuffer_f32_0_0_0t(
            target("dx.TypedBuffer", float, 0, 0, 0) %buffer, i32 %index)
-   %ret = call <4 x i32>
+   %ret = call {<4 x i32>, i1}
        @llvm.dx.resource.load.typedbuffer.v4i32.tdx.TypedBuffer_v4i32_0_0_0t(
            target("dx.TypedBuffer", <4 x i32>, 0, 0, 0) %buffer, i32 %index)
-   %ret = call <4 x half>
+   %ret = call {<4 x half>, i1}
        @llvm.dx.resource.load.typedbuffer.v4f16.tdx.TypedBuffer_v4f16_0_0_0t(
            target("dx.TypedBuffer", <4 x half>, 0, 0, 0) %buffer, i32 %index)
-   %ret = call <2 x double>
+   %ret = call {<2 x double>, i1}
        @llvm.dx.resource.load.typedbuffer.v2f64.tdx.TypedBuffer_v2f64_0_0t(
            target("dx.TypedBuffer", <2 x double>, 0, 0, 0) %buffer, i32 %index)
 
-.. list-table:: ``@llvm.dx.resource.loadchecked.typedbuffer``
-   :header-rows: 1
-
-   * - Argument
-     -
-     - Type
-     - Description
-   * - Return value
-     -
-     - A structure of the contained type and the check bit
-     - The data loaded from the buffer and the check bit
-   * - ``%buffer``
-     - 0
-     - ``target(dx.TypedBuffer, ...)``
-     - The buffer to load from
-   * - ``%index``
-     - 1
-     - ``i32``
-     - Index into the buffer
-
-.. code-block:: llvm
-
-   %ret = call {<4 x float>, i1}
-       @llvm.dx.resource.loadchecked.typedbuffer.v4f32.tdx.TypedBuffer_v4f32_0_0_0t(
-           target("dx.TypedBuffer", <4 x float>, 0, 0, 0) %buffer, i32 %index)
-
 Texture and Typed Buffer Stores
 -------------------------------
 
diff --git a/llvm/include/llvm/IR/IntrinsicsDirectX.td b/llvm/include/llvm/IR/IntrinsicsDirectX.td
@@ -31,9 +31,6 @@ def int_dx_resource_getpointer
     : DefaultAttrsIntrinsic<[llvm_anyptr_ty], [llvm_any_ty, llvm_i32_ty],
                             [IntrNoMem]>;
 def int_dx_resource_load_typedbuffer
-    : DefaultAttrsIntrinsic<[llvm_any_ty], [llvm_any_ty, llvm_i32_ty],
-                            [IntrReadMem]>;
-def int_dx_resource_loadchecked_typedbuffer
     : DefaultAttrsIntrinsic<[llvm_any_ty, llvm_i1_ty],
                             [llvm_any_ty, llvm_i32_ty], [IntrReadMem]>;
 def int_dx_resource_store_typedbuffer
@@ -43,7 +40,7 @@ def int_dx_resource_store_typedbuffer
 def int_dx_resource_updatecounter
     : DefaultAttrsIntrinsic<[llvm_i32_ty], [llvm_any_ty, llvm_i8_ty],
                             [IntrInaccessibleMemOrArgMemOnly]>;
-    
+
 // Cast between target extension handle types and dxil-style opaque handles
 def int_dx_resource_casthandle : Intrinsic<[llvm_any_ty], [llvm_any_ty]>;
 
@@ -105,7 +102,7 @@ def int_dx_wave_is_first_lane : DefaultAttrsIntrinsic<[llvm_i1_ty], [], [IntrCon
 def int_dx_wave_readlane : DefaultAttrsIntrinsic<[llvm_any_ty], [LLVMMatchType<0>, llvm_i32_ty], [IntrConvergent, IntrNoMem]>;
 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_splitdouble : DefaultAttrsIntrinsic<[llvm_anyint_ty, LLVMMatchType<0>], 
+def int_dx_splitdouble : DefaultAttrsIntrinsic<[llvm_anyint_ty, LLVMMatchType<0>],
     [LLVMScalarOrSameVectorWidth<0, llvm_double_ty>], [IntrNoMem]>;
 def int_dx_radians : DefaultAttrsIntrinsic<[llvm_anyfloat_ty], [LLVMMatchType<0>], [IntrNoMem]>;
 def int_dx_discard : DefaultAttrsIntrinsic<[], [llvm_i1_ty], []>;
diff --git a/llvm/lib/Target/DirectX/DXILOpLowering.cpp b/llvm/lib/Target/DirectX/DXILOpLowering.cpp
@@ -415,8 +415,16 @@ class OpLowerer {
         }
       }
 
-      OldResult = cast<Instruction>(
-          IRB.CreateExtractValue(Op, 0, OldResult->getName()));
+      if (OldResult->use_empty()) {
+        // Only the check bit was used, so we're done here.
+        OldResult->eraseFromParent();
+        return Error::success();
+      }
+
+      assert(OldResult->hasOneUse() &&
+             isa<ExtractValueInst>(*OldResult->user_begin()) &&
+             "Expected only use to be extract of first element");
+      OldResult = cast<Instruction>(*OldResult->user_begin());
       OldTy = ST->getElementType(0);
     }
 
@@ -723,9 +731,6 @@ class OpLowerer {
         HasErrors |= lowerGetPointer(F);
         break;
       case Intrinsic::dx_resource_load_typedbuffer:
-        HasErrors |= lowerTypedBufferLoad(F, /*HasCheckBit=*/false);
-        break;
-      case Intrinsic::dx_resource_loadchecked_typedbuffer:
         HasErrors |= lowerTypedBufferLoad(F, /*HasCheckBit=*/true);
         break;
       case Intrinsic::dx_resource_store_typedbuffer:
diff --git a/llvm/lib/Target/DirectX/DXILResourceAccess.cpp b/llvm/lib/Target/DirectX/DXILResourceAccess.cpp
@@ -30,6 +30,9 @@ static void replaceTypedBufferAccess(IntrinsicInst *II,
          "Unexpected typed buffer type");
   Type *ContainedType = HandleType->getTypeParameter(0);
 
+  Type *LoadType =
+      StructType::get(ContainedType, Type::getInt1Ty(II->getContext()));
+
   // We need the size of an element in bytes so that we can calculate the offset
   // in elements given a total offset in bytes later.
   Type *ScalarType = ContainedType->getScalarType();
@@ -81,13 +84,15 @@ static void replaceTypedBufferAccess(IntrinsicInst *II,
         // We're storing a scalar, so we need to load the current value and only
         // replace the relevant part.
         auto *Load = Builder.CreateIntrinsic(
-            ContainedType, Intrinsic::dx_resource_load_typedbuffer,
+            LoadType, Intrinsic::dx_resource_load_typedbuffer,
             {II->getOperand(0), II->getOperand(1)});
+        auto *Struct = Builder.CreateExtractValue(Load, {0});
+
         // If we have an offset from seeing a GEP earlier, use it.
         Value *IndexOp = Current.Index
                              ? Current.Index
                              : ConstantInt::get(Builder.getInt32Ty(), 0);
-        V = Builder.CreateInsertElement(Load, V, IndexOp);
+        V = Builder.CreateInsertElement(Struct, V, IndexOp);
       } else {
         llvm_unreachable("Store to typed resource has invalid type");
       }
@@ -101,8 +106,10 @@ static void replaceTypedBufferAccess(IntrinsicInst *II,
     } else if (auto *LI = dyn_cast<LoadInst>(Current.Access)) {
       IRBuilder<> Builder(LI);
       Value *V = Builder.CreateIntrinsic(
-          ContainedType, Intrinsic::dx_resource_load_typedbuffer,
+          LoadType, Intrinsic::dx_resource_load_typedbuffer,
           {II->getOperand(0), II->getOperand(1)});
+      V = Builder.CreateExtractValue(V, {0});
+
       if (Current.Index)
         V = Builder.CreateExtractElement(V, Current.Index);
 
diff --git a/llvm/test/CodeGen/DirectX/BufferLoad.ll b/llvm/test/CodeGen/DirectX/BufferLoad.ll
@@ -17,8 +17,9 @@ define void @loadv4f32() {
   ; CHECK-NOT: %dx.resource.casthandle
 
   ; CHECK: [[DATA0:%.*]] = call %dx.types.ResRet.f32 @dx.op.bufferLoad.f32(i32 68, %dx.types.Handle [[HANDLE]], i32 0, i32 undef)
-  %data0 = call <4 x float> @llvm.dx.resource.load.typedbuffer(
+  %load0 = call {<4 x float>, i1} @llvm.dx.resource.load.typedbuffer(
       target("dx.TypedBuffer", <4 x float>, 0, 0, 0) %buffer, i32 0)
+  %data0 = extractvalue {<4 x float>, i1} %load0, 0
 
   ; The extract order depends on the users, so don't enforce that here.
   ; CHECK-DAG: [[VAL0_0:%.*]] = extractvalue %dx.types.ResRet.f32 [[DATA0]], 0
@@ -34,8 +35,9 @@ define void @loadv4f32() {
   call void @scalar_user(float %data0_2)
 
   ; CHECK: [[DATA4:%.*]] = call %dx.types.ResRet.f32 @dx.op.bufferLoad.f32(i32 68, %dx.types.Handle [[HANDLE]], i32 4, i32 undef)
-  %data4 = call <4 x float> @llvm.dx.resource.load.typedbuffer(
+  %load4 = call {<4 x float>, i1} @llvm.dx.resource.load.typedbuffer(
       target("dx.TypedBuffer", <4 x float>, 0, 0, 0) %buffer, i32 4)
+  %data4 = extractvalue {<4 x float>, i1} %load4, 0
 
   ; CHECK: extractvalue %dx.types.ResRet.f32 [[DATA4]], 0
   ; CHECK: extractvalue %dx.types.ResRet.f32 [[DATA4]], 1
@@ -48,8 +50,9 @@ define void @loadv4f32() {
   call void @vector_user(<4 x float> %data4)
 
   ; CHECK: [[DATA12:%.*]] = call %dx.types.ResRet.f32 @dx.op.bufferLoad.f32(i32 68, %dx.types.Handle [[HANDLE]], i32 12, i32 undef)
-  %data12 = call <4 x float> @llvm.dx.resource.load.typedbuffer(
+  %load12 = call {<4 x float>, i1} @llvm.dx.resource.load.typedbuffer(
       target("dx.TypedBuffer", <4 x float>, 0, 0, 0) %buffer, i32 12)
+  %data12 = extractvalue {<4 x float>, i1} %load12, 0
 
   ; CHECK: [[DATA12_3:%.*]] = extractvalue %dx.types.ResRet.f32 [[DATA12]], 3
   %data12_3 = extractelement <4 x float> %data12, i32 3
@@ -70,8 +73,9 @@ define void @index_dynamic(i32 %bufindex, i32 %elemindex) {
           i32 0, i32 0, i32 1, i32 0, i1 false)
 
   ; CHECK: [[LOAD:%.*]] = call %dx.types.ResRet.f32 @dx.op.bufferLoad.f32(i32 68, %dx.types.Handle [[HANDLE]], i32 %bufindex, i32 undef)
-  %load = call <4 x float> @llvm.dx.resource.load.typedbuffer(
+  %load = call {<4 x float>, i1} @llvm.dx.resource.load.typedbuffer(
       target("dx.TypedBuffer", <4 x float>, 0, 0, 0) %buffer, i32 %bufindex)
+  %data = extractvalue {<4 x float>, i1} %load, 0
 
   ; CHECK: [[ALLOCA:%.*]] = alloca [4 x float]
   ; CHECK: [[V0:%.*]] = extractvalue %dx.types.ResRet.f32 [[LOAD]], 0
@@ -89,10 +93,10 @@ define void @index_dynamic(i32 %bufindex, i32 %elemindex) {
   ;
   ; CHECK: [[PTR:%.*]] = getelementptr inbounds [4 x float], ptr [[ALLOCA]], i32 0, i32 %elemindex
   ; CHECK: [[X:%.*]] = load float, ptr [[PTR]]
-  %data = extractelement <4 x float> %load, i32 %elemindex
+  %x = extractelement <4 x float> %data, i32 %elemindex
 
   ; CHECK: call void @scalar_user(float [[X]])
-  call void @scalar_user(float %data)
+  call void @scalar_user(float %x)
 
   ret void
 }
@@ -105,8 +109,9 @@ define void @loadf32() {
           i32 0, i32 0, i32 1, i32 0, i1 false)
 
   ; CHECK: [[DATA0:%.*]] = call %dx.types.ResRet.f32 @dx.op.bufferLoad.f32(i32 68, %dx.types.Handle [[HANDLE]], i32 0, i32 undef)
-  %data0 = call float @llvm.dx.resource.load.typedbuffer(
+  %load0 = call {float, i1} @llvm.dx.resource.load.typedbuffer(
       target("dx.TypedBuffer", float, 0, 0, 0) %buffer, i32 0)
+  %data0 = extractvalue {float, i1} %load0, 0
 
   ; CHECK: [[VAL0:%.*]] = extractvalue %dx.types.ResRet.f32 [[DATA0]], 0
   ; CHECK: call void @scalar_user(float [[VAL0]])
@@ -123,7 +128,7 @@ define void @loadv2f32() {
           i32 0, i32 0, i32 1, i32 0, i1 false)
 
   ; CHECK: [[DATA0:%.*]] = call %dx.types.ResRet.f32 @dx.op.bufferLoad.f32(i32 68, %dx.types.Handle [[HANDLE]], i32 0, i32 undef)
-  %data0 = call <2 x float> @llvm.dx.resource.load.typedbuffer(
+  %data0 = call {<2 x float>, i1} @llvm.dx.resource.load.typedbuffer(
       target("dx.TypedBuffer", <2 x float>, 0, 0, 0) %buffer, i32 0)
 
   ret void
@@ -137,7 +142,7 @@ define void @loadv4f32_checkbit() {
           i32 0, i32 0, i32 1, i32 0, i1 false)
 
   ; CHECK: [[DATA0:%.*]] = call %dx.types.ResRet.f32 @dx.op.bufferLoad.f32(i32 68, %dx.types.Handle [[HANDLE]], i32 0, i32 undef)
-  %data0 = call {<4 x float>, i1} @llvm.dx.resource.loadchecked.typedbuffer.f32(
+  %data0 = call {<4 x float>, i1} @llvm.dx.resource.load.typedbuffer.f32(
       target("dx.TypedBuffer", <4 x float>, 0, 0, 0) %buffer, i32 0)
 
   ; CHECK: [[STATUS:%.*]] = extractvalue %dx.types.ResRet.f32 [[DATA0]], 4
@@ -158,7 +163,7 @@ define void @loadv4i32() {
           i32 0, i32 0, i32 1, i32 0, i1 false)
 
   ; CHECK: [[DATA0:%.*]] = call %dx.types.ResRet.i32 @dx.op.bufferLoad.i32(i32 68, %dx.types.Handle [[HANDLE]], i32 0, i32 undef)
-  %data0 = call <4 x i32> @llvm.dx.resource.load.typedbuffer(
+  %data0 = call {<4 x i32>, i1} @llvm.dx.resource.load.typedbuffer(
       target("dx.TypedBuffer", <4 x i32>, 0, 0, 0) %buffer, i32 0)
 
   ret void
@@ -172,7 +177,7 @@ define void @loadv4f16() {
           i32 0, i32 0, i32 1, i32 0, i1 false)
 
   ; CHECK: [[DATA0:%.*]] = call %dx.types.ResRet.f16 @dx.op.bufferLoad.f16(i32 68, %dx.types.Handle [[HANDLE]], i32 0, i32 undef)
-  %data0 = call <4 x half> @llvm.dx.resource.load.typedbuffer(
+  %data0 = call {<4 x half>, i1} @llvm.dx.resource.load.typedbuffer(
       target("dx.TypedBuffer", <4 x half>, 0, 0, 0) %buffer, i32 0)
 
   ret void
@@ -186,7 +191,7 @@ define void @loadv4i16() {
           i32 0, i32 0, i32 1, i32 0, i1 false)
 
   ; CHECK: [[DATA0:%.*]] = call %dx.types.ResRet.i16 @dx.op.bufferLoad.i16(i32 68, %dx.types.Handle [[HANDLE]], i32 0, i32 undef)
-  %data0 = call <4 x i16> @llvm.dx.resource.load.typedbuffer(
+  %data0 = call {<4 x i16>, i1} @llvm.dx.resource.load.typedbuffer(
       target("dx.TypedBuffer", <4 x i16>, 0, 0, 0) %buffer, i32 0)
 
   ret void
diff --git a/llvm/test/CodeGen/DirectX/ResourceAccess/load_typedbuffer.ll b/llvm/test/CodeGen/DirectX/ResourceAccess/load_typedbuffer.ll
@@ -15,17 +15,19 @@ define void @load_float4(i32 %index, i32 %elemindex) {
   %ptr = call ptr @llvm.dx.resource.getpointer(
       target("dx.TypedBuffer", <4 x float>, 1, 0, 0) %buffer, i32 %index)
 
-  ; CHECK: %[[VALUE:.*]] = call <4 x float> @llvm.dx.resource.load.typedbuffer.v4f32.tdx.TypedBuffer_v4f32_1_0_0t(target("dx.TypedBuffer", <4 x float>, 1, 0, 0) %buffer, i32 %index)
+  ; CHECK: %[[VALUE:.*]] = call { <4 x float>, i1 } @llvm.dx.resource.load.typedbuffer.v4f32.tdx.TypedBuffer_v4f32_1_0_0t(target("dx.TypedBuffer", <4 x float>, 1, 0, 0) %buffer, i32 %index)
   %vec_data = load <4 x float>, ptr %ptr
   call void @use_float4(<4 x float> %vec_data)
 
-  ; CHECK: %[[VALUE:.*]] = call <4 x float> @llvm.dx.resource.load.typedbuffer.v4f32.tdx.TypedBuffer_v4f32_1_0_0t(target("dx.TypedBuffer", <4 x float>, 1, 0, 0) %buffer, i32 %index)
+  ; CHECK: %[[LOAD:.*]] = call { <4 x float>, i1 } @llvm.dx.resource.load.typedbuffer.v4f32.tdx.TypedBuffer_v4f32_1_0_0t(target("dx.TypedBuffer", <4 x float>, 1, 0, 0) %buffer, i32 %index)
+  ; CHECK: %[[VALUE:.*]] = extractvalue { <4 x float>, i1 } %[[LOAD]], 0
   ; CHECK: extractelement <4 x float> %[[VALUE]], i32 1
   %y_ptr = getelementptr inbounds <4 x float>, ptr %ptr, i32 0, i32 1
   %y_data = load float, ptr %y_ptr
   call void @use_float(float %y_data)
 
-  ; CHECK: %[[VALUE:.*]] = call <4 x float> @llvm.dx.resource.load.typedbuffer.v4f32.tdx.TypedBuffer_v4f32_1_0_0t(target("dx.TypedBuffer", <4 x float>, 1, 0, 0) %buffer, i32 %index)
+  ; CHECK: %[[LOAD:.*]] = call { <4 x float>, i1 } @llvm.dx.resource.load.typedbuffer.v4f32.tdx.TypedBuffer_v4f32_1_0_0t(target("dx.TypedBuffer", <4 x float>, 1, 0, 0) %buffer, i32 %index)
+  ; CHECK: %[[VALUE:.*]] = extractvalue { <4 x float>, i1 } %[[LOAD]], 0
   ; CHECK: extractelement <4 x float> %[[VALUE]], i32 %elemindex
   %dynamic = getelementptr inbounds <4 x float>, ptr %ptr, i32 0, i32 %elemindex
   %dyndata = load float, ptr %dynamic
diff --git a/llvm/test/CodeGen/DirectX/ResourceAccess/store_typedbuffer.ll b/llvm/test/CodeGen/DirectX/ResourceAccess/store_typedbuffer.ll
diff --git a/llvm/test/CodeGen/DirectX/ResourceGlobalElimination.ll b/llvm/test/CodeGen/DirectX/ResourceGlobalElimination.ll
diff --git a/llvm/test/CodeGen/DirectX/ShaderFlags/typed-uav-load-additional-formats.ll b/llvm/test/CodeGen/DirectX/ShaderFlags/typed-uav-load-additional-formats.ll
