Fix existential specialization of mutable buffer loads. (#1671)

* Fix existential specialization of mutable buffer loads.

* fix

Co-authored-by: Yong He <yhe@nvidia.com>

4 files changed, 72 insertions, 16 deletions

diff --git a/source/core/slang-dictionary.h b/source/core/slang-dictionary.h
index df5ee520d..4c352d55b 100644
--- a/source/core/slang-dictionary.h
+++ b/source/core/slang-dictionary.h
@@ -1003,7 +1003,17 @@ namespace Slang
     };
 
     template <typename T> class OrderedHashSet : public HashSetBase<T, OrderedDictionary<T, _DummyClass>>
-    {};
+    {
+    public:
+        T& getLast()
+        {
+            return this->dict.Last().Key;
+        }
+        void removeLast()
+        {
+            this->Remove(getLast());
+        }
+    };
 }
 
 #endif
diff --git a/source/slang/slang-ir-specialize.cpp b/source/slang/slang-ir-specialize.cpp
index 91852ff88..7fb6a06bd 100644
--- a/source/slang/slang-ir-specialize.cpp
+++ b/source/slang/slang-ir-specialize.cpp
@@ -107,9 +107,7 @@ struct SpecializationContext
     // to be considered for specialization or simplification,
     // whether generic, existential, etc.
     //
-    List<IRInst*> workList;
-    HashSet<IRInst*> workListSet;
-
+    OrderedHashSet<IRInst*> workList;
     HashSet<IRInst*> cleanInsts;
 
     void addToWorkList(
@@ -125,14 +123,12 @@ struct SpecializationContext
                 return;
         }
 
-        if(workListSet.Contains(inst))
-            return;
-
-        workList.add(inst);
-        workListSet.Add(inst);
-        cleanInsts.Remove(inst);
+        if (workList.Add(inst))
+        {
+            cleanInsts.Remove(inst);
 
-        addUsersToWorkList(inst);
+            addUsersToWorkList(inst);
+        }
     }
 
     // When a transformation makes a change to an instruction,
@@ -682,17 +678,17 @@ struct SpecializationContext
         //
         addToWorkList(module->getModuleInst());
 
-        while(workList.getCount() != 0)
+        while(workList.Count() != 0)
         {
 
         // We will then iterate until our work list goes dry.
         //
-        while(workList.getCount() != 0)
+        while(workList.Count() != 0)
         {
             IRInst* inst = workList.getLast();
 
             workList.removeLast();
-            workListSet.Remove(inst);
+
             cleanInsts.Add(inst);
 
             // For each instruction we process, we want to perform
@@ -838,15 +834,29 @@ struct SpecializationContext
                     // we should update it to be `specialize(.operator[], elementType)`, so the return type
                     // of the load call is `elementType`.
                     auto oldCallee = inst->getCallee();
-                    auto newCallee = getNewSpecializedBufferLoadCallee(inst->getCallee(), sbType, elementType);
-                    auto newCall = builder.emitCallInst(elementType, newCallee, args);
+
+                    // A subscript operation on mutable buffers returns a ptr type instead of a value type.
+                    // We need to make sure the pointer-ness is preserved correctly.
+                    auto innerResultType = elementType;
+                    if (auto ptrResultType = as<IRPtrType>(inst->getDataType()))
+                    {
+                        innerResultType = builder.getPtrType(elementType);
+                    }
+                    auto newCallee = getNewSpecializedBufferLoadCallee(inst->getCallee(), sbType, innerResultType);
+                    auto newCall = builder.emitCallInst(innerResultType, newCallee, args);
                     auto newWrapExistential = builder.emitWrapExistential(
                         resultType, newCall, slotOperandCount, slotOperands.getBuffer());
                     inst->replaceUsesWith(newWrapExistential);
+                    workList.Remove(inst);
                     inst->removeAndDeallocate();
                     SLANG_ASSERT(!oldCallee->hasUses());
+                    workList.Remove(oldCallee);
                     oldCallee->removeAndDeallocate();
                     addUsersToWorkList(newWrapExistential);
+
+                    workList.Remove(wrapExistential);
+                    SLANG_ASSERT(!wrapExistential->hasUses());
+                    wrapExistential->removeAndDeallocate();
                     return true;
                 }
             }
diff --git a/tests/bugs/rwstructuredbuffer-existential-in-struct.slang b/tests/bugs/rwstructuredbuffer-existential-in-struct.slang
new file mode 100644
index 000000000..c763dd3c1
--- /dev/null
+++ b/tests/bugs/rwstructuredbuffer-existential-in-struct.slang
@@ -0,0 +1,32 @@
+//TEST(compute):COMPARE_COMPUTE:-dx11 -shaderobj
+
+[anyValueSize(8)]
+interface IMaterial
+{
+    float eval();
+}
+
+struct MaterialImpl : IMaterial
+{
+    float eval() { return 0.0f; }
+};
+
+struct Parameters
+{
+    uint ordinary;
+
+//TEST_INPUT: entryPointExistentialType MaterialImpl
+//TEST_INPUT:ubuffer(data=[0 0 0 0 0 0], stride=4):name=params.materials
+    RWStructuredBuffer<IMaterial> materials;
+
+//TEST_INPUT:ubuffer(data=[0 0 0 0], stride=4):out,name=params.result
+    RWStructuredBuffer<float> result;
+};
+
+[numthreads(4, 1, 1)]
+void computeMain(uint3 dispatchThreadID : SV_DispatchThreadID, uniform Parameters params, uniform int id)
+{
+    uint tid = dispatchThreadID.x;
+    float rs = params.materials[0].eval();
+    params.result[tid] = rs;
+}
diff --git a/tests/bugs/rwstructuredbuffer-existential-in-struct.slang.expected.txt b/tests/bugs/rwstructuredbuffer-existential-in-struct.slang.expected.txt
new file mode 100644
index 000000000..44e0be8e3
--- /dev/null
+++ b/tests/bugs/rwstructuredbuffer-existential-in-struct.slang.expected.txt
@@ -0,0 +1,4 @@
+0
+0
+0
+0