[dxv] Disallow work graphs in DXIL 1.10

docs/DXIL.rst

@@ -3418,6 +3418,7 @@ SM.RESOURCERANGEOVERLAP                               Resource ranges must not o
 SM.ROVONLYINPS                                        RasterizerOrdered objects are only allowed in 5.0+ pixel shaders.
 SM.SAMPLECOUNTONLYON2DMS                              Only Texture2DMS/2DMSArray could has sample count.
 SM.SEMANTIC                                           Semantic must be defined in target shader model
+SM.SHADERSTAGE                                        Shader stage must be supported by the target shader model
 SM.STREAMINDEXRANGE                                   Stream index (%0) must between 0 and %1.
 SM.TESSFACTORFORDOMAIN                                Required TessFactor for domain not found declared anywhere in Patch Constant data.
 SM.TESSFACTORSIZEMATCHDOMAIN                          TessFactor rows, columns (%0, %1) invalid for domain %2.  Expected %3 rows and 1 column.

docs/ReleaseNotes.md

@@ -68,7 +68,8 @@ first shipped in the 1.10.2605 preview.
 - Restricted the component types allowed in LinAlg matrices
   [#8608](https://github.com/microsoft/DirectXShaderCompiler/pull/8608).
 - Added `BFloat16` to the ComponentType enum in DxilConstants and the linalg
-  header [#8722](https://github.com/microsoft/DirectXShaderCompiler/issues/8722)
+  header [#8722](https://github.com/microsoft/DirectXShaderCompiler/issues/8722).
+- Removed work graph support from Shader Model 6.10, and DXIL 1.10 [microsoft/hlsl-specs#915](https://github.com/microsoft/hlsl-specs/issues/915).
 
 ### Version 1.9.2607
 

lib/DXIL/DxilOperations.cpp

@@ -4001,7 +4001,7 @@ void OP::GetMinShaderModelAndMask(OpCode C, bool bWithTranslation,
   if ((2147483649 <= op && op <= 2147483650)) {
     major = 6;
     minor = 10;
-    mask = SFLAG(Compute) | SFLAG(Mesh) | SFLAG(Amplification) | SFLAG(Node);
+    mask = SFLAG(Compute) | SFLAG(Mesh) | SFLAG(Amplification);
     return;
   }
   // Instructions: ClusterID=2147483651, TriangleObjectPosition=2147483655

lib/DxilValidation/DxilValidation.cpp

@@ -6663,6 +6663,12 @@ static void ValidateEntryProps(ValidationContext &ValCtx,
 
   ValidateWaveSize(ValCtx, EntryProps, F);
 
+  const ShaderModel *SM = ValCtx.DxilMod.GetShaderModel();
+  if (Props.IsNode() && SM->IsSM610Plus())
+    ValCtx.EmitFnFormatError(
+        F, ValidationRule::SmShaderStage,
+        {ShaderModel::GetKindName(ShaderType), SM->GetName()});
+
   if (ShaderType == DXIL::ShaderKind::Compute || Props.IsNode()) {
     unsigned X = Props.numThreads[0];
     unsigned Y = Props.numThreads[1];

tools/clang/test/LitDXILValidation/node-shader-sm-6-10.ll

@@ -0,0 +1,37 @@
+; REQUIRES: dxil-1-10
+; RUN: not %dxv %s 2>&1 | FileCheck %s
+
+; Node shaders are not valid in shader model 6.10, even when the entry point
+; contains no DXIL calls.
+
+; CHECK: Function: Node: error: Shader stage 'node' not valid in shader model lib_6_10.
+; CHECK: Validation failed.
+
+target datalayout = "e-m:e-p:32:32-i1:32-i8:32-i16:32-i32:32-i64:64-f16:32-f32:32-f64:64-n8:16:32:64"
+target triple = "dxil-ms-dx"
+
+define void @Node() {
+  ret void
+}
+
+!llvm.ident = !{!0}
+!dx.version = !{!1}
+!dx.valver = !{!1}
+!dx.shaderModel = !{!2}
+!dx.typeAnnotations = !{!3}
+!dx.entryPoints = !{!7, !9}
+
+!0 = !{!"custom IR"}
+!1 = !{i32 1, i32 10}
+!2 = !{!"lib", i32 6, i32 10}
+!3 = !{i32 1, void ()* @Node, !4}
+!4 = !{!5}
+!5 = !{i32 0, !6, !6}
+!6 = !{}
+!7 = !{null, !"", null, null, !8}
+!8 = !{i32 0, i64 524288}
+!9 = !{void ()* @Node, !"Node", null, null, !10}
+!10 = !{i32 8, i32 15, i32 13, i32 1, i32 15, !11, i32 16, i32 -1, i32 4, !12, i32 5, !13}
+!11 = !{!"Node", i32 0}
+!12 = !{i32 1, i32 1, i32 1}
+!13 = !{i32 0}

tools/clang/test/LitDXILValidation/work-graph-opcodes-sm-6-10.ll

@@ -0,0 +1,90 @@
+; REQUIRES: dxil-1-10
+; RUN: not %dxv %s 2>&1 | FileCheck %s
+
+; Work graph opcodes are not valid in shader model 6.10.
+
+; CHECK-DAG: error: Opcode AllocateNodeOutputRecords not valid in shader model lib_6_10(node).
+; CHECK-DAG: error: Opcode GetNodeRecordPtr not valid in shader model lib_6_10(node).
+; CHECK-DAG: error: Opcode IncrementOutputCount not valid in shader model lib_6_10(node).
+; CHECK-DAG: error: Opcode OutputComplete not valid in shader model lib_6_10(node).
+; CHECK-DAG: error: Opcode GetInputRecordCount not valid in shader model lib_6_10(node).
+; CHECK-DAG: error: Opcode FinishedCrossGroupSharing not valid in shader model lib_6_10(node).
+; CHECK-DAG: error: Opcode BarrierByNodeRecordHandle not valid in shader model lib_6_10(node).
+; CHECK-DAG: error: Opcode CreateNodeOutputHandle not valid in shader model lib_6_10(node).
+; CHECK-DAG: error: Opcode IndexNodeHandle not valid in shader model lib_6_10(node).
+; CHECK-DAG: error: Opcode AnnotateNodeHandle not valid in shader model lib_6_10(node).
+; CHECK-DAG: error: Opcode CreateNodeInputRecordHandle not valid in shader model lib_6_10(node).
+; CHECK-DAG: error: Opcode AnnotateNodeRecordHandle not valid in shader model lib_6_10(node).
+; CHECK-DAG: error: Opcode NodeOutputIsValid not valid in shader model lib_6_10(node).
+; CHECK-DAG: error: Opcode GetRemainingRecursionLevels not valid in shader model lib_6_10(node).
+; CHECK: Validation failed.
+
+target datalayout = "e-m:e-p:32:32-i1:32-i8:32-i16:32-i32:32-i64:64-f16:32-f32:32-f64:64-n8:16:32:64"
+target triple = "dxil-ms-dx"
+
+%dx.types.NodeHandle = type { i8* }
+%dx.types.NodeInfo = type { i32, i32 }
+%dx.types.NodeRecordHandle = type { i8* }
+%dx.types.NodeRecordInfo = type { i32, i32 }
+%struct.Record = type { i32 }
+
+define void @Node() {
+  %output = call %dx.types.NodeHandle @dx.op.createNodeOutputHandle(i32 247, i32 0)
+  %indexed = call %dx.types.NodeHandle @dx.op.indexNodeHandle(i32 248, %dx.types.NodeHandle %output, i32 0)
+  %annotatedOutput = call %dx.types.NodeHandle @dx.op.annotateNodeHandle(i32 249, %dx.types.NodeHandle %indexed, %dx.types.NodeInfo { i32 6, i32 4 })
+  %outputRecord = call %dx.types.NodeRecordHandle @dx.op.allocateNodeOutputRecords(i32 238, %dx.types.NodeHandle %annotatedOutput, i32 1, i1 true)
+  %annotatedOutputRecord = call %dx.types.NodeRecordHandle @dx.op.annotateNodeRecordHandle(i32 251, %dx.types.NodeRecordHandle %outputRecord, %dx.types.NodeRecordInfo { i32 38, i32 4 })
+  %recordPtr = call %struct.Record addrspace(6)* @dx.op.getNodeRecordPtr.struct.Record(i32 239, %dx.types.NodeRecordHandle %annotatedOutputRecord, i32 0)
+  call void @dx.op.incrementOutputCount(i32 240, %dx.types.NodeHandle %annotatedOutput, i32 1, i1 true)
+  call void @dx.op.outputComplete(i32 241, %dx.types.NodeRecordHandle %annotatedOutputRecord)
+  %inputRecord = call %dx.types.NodeRecordHandle @dx.op.createNodeInputRecordHandle(i32 250, i32 0)
+  %inputCount = call i32 @dx.op.getInputRecordCount(i32 242, %dx.types.NodeRecordHandle %inputRecord)
+  %finished = call i1 @dx.op.finishedCrossGroupSharing(i32 243, %dx.types.NodeRecordHandle %inputRecord)
+  call void @dx.op.barrierByNodeRecordHandle(i32 246, %dx.types.NodeRecordHandle %inputRecord, i32 0)
+  %valid = call i1 @dx.op.nodeOutputIsValid(i32 252, %dx.types.NodeHandle %annotatedOutput)
+  %levels = call i32 @dx.op.getRemainingRecursionLevels(i32 253)
+  ret void
+}
+
+declare %dx.types.NodeRecordHandle @dx.op.allocateNodeOutputRecords(i32, %dx.types.NodeHandle, i32, i1)
+declare %struct.Record addrspace(6)* @dx.op.getNodeRecordPtr.struct.Record(i32, %dx.types.NodeRecordHandle, i32)
+declare void @dx.op.incrementOutputCount(i32, %dx.types.NodeHandle, i32, i1)
+declare void @dx.op.outputComplete(i32, %dx.types.NodeRecordHandle)
+declare i32 @dx.op.getInputRecordCount(i32, %dx.types.NodeRecordHandle)
+declare i1 @dx.op.finishedCrossGroupSharing(i32, %dx.types.NodeRecordHandle)
+declare void @dx.op.barrierByNodeRecordHandle(i32, %dx.types.NodeRecordHandle, i32)
+declare %dx.types.NodeHandle @dx.op.createNodeOutputHandle(i32, i32)
+declare %dx.types.NodeHandle @dx.op.indexNodeHandle(i32, %dx.types.NodeHandle, i32)
+declare %dx.types.NodeHandle @dx.op.annotateNodeHandle(i32, %dx.types.NodeHandle, %dx.types.NodeInfo)
+declare %dx.types.NodeRecordHandle @dx.op.createNodeInputRecordHandle(i32, i32)
+declare %dx.types.NodeRecordHandle @dx.op.annotateNodeRecordHandle(i32, %dx.types.NodeRecordHandle, %dx.types.NodeRecordInfo)
+declare i1 @dx.op.nodeOutputIsValid(i32, %dx.types.NodeHandle)
+declare i32 @dx.op.getRemainingRecursionLevels(i32)
+
+!llvm.ident = !{!0}
+!dx.version = !{!1}
+!dx.valver = !{!1}
+!dx.shaderModel = !{!2}
+!dx.typeAnnotations = !{!3}
+!dx.entryPoints = !{!7, !9}
+
+!0 = !{!"custom IR"}
+!1 = !{i32 1, i32 10}
+!2 = !{!"lib", i32 6, i32 10}
+!3 = !{i32 1, void ()* @Node, !4}
+!4 = !{!5}
+!5 = !{i32 0, !6, !6}
+!6 = !{}
+!7 = !{null, !"", null, null, !8}
+!8 = !{i32 0, i64 524288}
+!9 = !{void ()* @Node, !"Node", null, null, !10}
+!10 = !{i32 8, i32 15, i32 13, i32 3, i32 15, !11, i32 16, i32 -1, i32 20, !12, i32 21, !15, i32 4, !18, i32 5, !19}
+!11 = !{!"Node", i32 0}
+!12 = !{!13}
+!13 = !{i32 1, i32 37, i32 2, !14}
+!14 = !{i32 0, i32 4, i32 2, i32 4}
+!15 = !{!16}
+!16 = !{i32 1, i32 6, i32 2, !14, i32 3, i32 1, i32 0, !17}
+!17 = !{!"output", i32 0}
+!18 = !{i32 1, i32 1, i32 1}
+!19 = !{i32 0}

utils/hct/hctdb.py

@@ -123,6 +123,7 @@ def __init__(self, name, **kwargs):
         self.is_barrier = False  # whether this is a barrier operation
         self.shader_stages = ()  # shader stages to which this applies, empty for all.
         self.shader_model = 6, 0  # minimum shader model required
+        self.shader_model_max = ()  # maximum shader model allowed, empty for no maximum
         self.inst_helper_prefix = None
         self.fully_qualified_name_prefix = "hlsl::OP::OpCode"
         self.shader_model_translated = ()  # minimum shader model required with translation by linker
@@ -1037,6 +1038,10 @@ def populate_categories_and_models(self):
             self.name_idx[i].category = "Work Graph intrinsics"
             self.name_idx[i].shader_model = 6, 8
             self.name_idx[i].shader_stages = ("node",)
+        for i in (
+            "AllocateNodeOutputRecords,GetNodeRecordPtr,IncrementOutputCount,OutputComplete,GetInputRecordCount,FinishedCrossGroupSharing,BarrierByNodeRecordHandle,CreateNodeOutputHandle,IndexNodeHandle,AnnotateNodeHandle,CreateNodeInputRecordHandle,AnnotateNodeRecordHandle,NodeOutputIsValid,GetRemainingRecursionLevels"
+        ).split(","):
+            self.name_idx[i].shader_model_max = 6, 9
         # All barrier ops:
         for i in "Barrier".split(","):
             self.name_idx[i].category = "Synchronization"
@@ -1110,7 +1115,7 @@ def populate_categories_and_models_ExperimentalOps(self):
         for i in insts("GetGroupWaveIndex,GetGroupWaveCount"):
             i.category = "Group Wave Ops"
             i.shader_model = experimental_sm
-            i.shader_stages = ("compute", "mesh", "amplification", "node")
+            i.shader_stages = ("compute", "mesh", "amplification")
             i.is_wave = True
 
         # Clustered Geometry
@@ -8790,6 +8795,11 @@ def build_valrules(self):
             "Opcode must be defined in target shader model",
             "Opcode %0 not valid in shader model %1.",
         )
+        self.add_valrule_msg(
+            "Sm.ShaderStage",
+            "Shader stage must be supported by the target shader model",
+            "Shader stage '%0' not valid in shader model %1.",
+        )
         self.add_valrule(
             "Sm.Operand", "Operand must be defined in target shader model."
         )

utils/hct/hctdb_instrhelp.py

@@ -1609,14 +1609,21 @@ def flush_instrs(grouped_instrs, last_model, last_model_translated, last_stage):
 def get_valopcode_sm_text():
     db = get_db_dxil()
     instrs = sorted(
-        db.get_dxil_ops(), key=lambda v: (v.shader_model, v.shader_stages, v.dxil_opid)
+        db.get_dxil_ops(),
+        key=lambda v: (
+            v.shader_model,
+            v.shader_model_max,
+            v.shader_stages,
+            v.dxil_opid,
+        ),
     )
     last_model = None
+    last_model_max = None
     last_stage = None
     grouped_instrs = []
     code = ""
 
-    def flush_instrs(grouped_instrs, last_model, last_stage):
+    def flush_instrs(grouped_instrs, last_model, last_model_max, last_stage):
         if len(grouped_instrs) == 0:
             return ""
         result = format_comment(
@@ -1631,34 +1638,51 @@ def flush_instrs(grouped_instrs, last_model, last_stage):
         )
         result += "  return "
 
-        model_cond = stage_cond = None
+        model_cond = model_max_cond = stage_cond = None
         if last_model != (6, 0):
             model_cond = "major > %d || (major == %d && minor >= %d)" % (
                 last_model[0],
                 last_model[0],
                 last_model[1],
             )
+        if last_model_max:
+            model_max_cond = "major < %d || (major == %d && minor <= %d)" % (
+                last_model_max[0],
+                last_model_max[0],
+                last_model_max[1],
+            )
         if last_stage:
             stage_cond = " || ".join(
                 [check_pSM_for_shader_stage[c] for c in last_stage]
             )
-        if model_cond or stage_cond:
+        if model_cond or model_max_cond or stage_cond:
             result += "\n      && ".join(
-                ["(%s)" % expr for expr in (model_cond, stage_cond) if expr]
+                [
+                    "(%s)" % expr
+                    for expr in (model_cond, model_max_cond, stage_cond)
+                    if expr
+                ]
             )
             return result + ";\n"
         else:
             # don't write these out, instead fall through
             return ""
 
     for i in instrs:
-        if (i.shader_model, i.shader_stages) != (last_model, last_stage):
-            code += flush_instrs(grouped_instrs, last_model, last_stage)
+        if (i.shader_model, i.shader_model_max, i.shader_stages) != (
+            last_model,
+            last_model_max,
+            last_stage,
+        ):
+            code += flush_instrs(
+                grouped_instrs, last_model, last_model_max, last_stage
+            )
             grouped_instrs = []
             last_model = i.shader_model
+            last_model_max = i.shader_model_max
             last_stage = i.shader_stages
         grouped_instrs.append(i)
-    code += flush_instrs(grouped_instrs, last_model, last_stage)
+    code += flush_instrs(grouped_instrs, last_model, last_model_max, last_stage)
     code += "return true;\n"
     return code