Add new GenISA_AsyncStackPtrPlaceHolder intrinsic

Add new GenISA_AsyncStackPtrPlaceHolder intrinsic

Author: JaroszPiotr

IGC/AdaptorCommon/RayTracing/MemRegionAnalysis.cpp

@@ -52,6 +52,7 @@ static std::optional<RTMemRegion> getIntrinsicRegion(
     case GenISAIntrinsic::GenISA_LocalBufferPointer:
         return RTMemRegion::LocalArgs;
     case GenISAIntrinsic::GenISA_AsyncStackPtr:
+    case GenISAIntrinsic::GenISA_AsyncStackPtrPlaceHolder:
         return RTMemRegion::RTAsyncStack;
     case GenISAIntrinsic::GenISA_SyncStackPtr:
         return RTMemRegion::RTSyncStack;

IGC/Compiler/CISACodeGen/LSCCacheOptimizationPass.cpp

@@ -456,7 +456,9 @@ bool LSCCacheOptimizationPass::create_48_wide_store(Function& function)
             // which would look something like this:
             // %perLaneAsyncStackPointer24 = call noalias align 128 dereferenceable(256) %"struct.RTStackFormat::RTStack" addrspace(1)* @"llvm.genx.GenISA.AsyncStackPtr.p1struct.RTStackFormat::RTStack.i64"(i64 %19)
             if (auto* intrinsic_call = dyn_cast<GenIntrinsicInst>(i)) {
-                if (intrinsic_call->getIntrinsicID() == llvm::GenISAIntrinsic::GenISA_AsyncStackPtr) {
+                if ((intrinsic_call->getIntrinsicID() == llvm::GenISAIntrinsic::GenISA_AsyncStackPtr) ||
+                    (intrinsic_call->getIntrinsicID() == llvm::GenISAIntrinsic::GenISA_AsyncStackPtrPlaceHolder))
+                {
                     // Create an IRBuilder with an insertion point set to the given intrinsic_call instruction.
                     // IRBuilder automatically inserts instructions when it creates them,
                     // and the inserted instructions (dynamically allocated) are deleted when the function is destroyed.

IGC/GenISAIntrinsics/generator/input/Intrinsic_definitions.yml

@@ -10745,6 +10745,25 @@ intrinsics:
         memory_effects:
             - !<MemoryRestriction>
                 memory_access: !MemoryAccessType NoModRef
+    - !<IntrinsicDefinition>
+        name: "GenISA_AsyncStackPtrPlaceHolder"
+        comment: "Raytracing: Intrinsic used as marker in later passes to find per\
+            \ lane async stack base. It is just a cast and will be lowered prior to\
+            \ codegen"
+        return_definition: !<ReturnDefinition>
+            type_definition: *p_any_
+            comment: ""
+        arguments:
+            - !<ArgumentDefinition>
+                name: Arg0
+                type_definition: *any_int
+                comment: ""
+        attributes:
+            - !AttributeID "NoDuplicate"
+            - !AttributeID "NoUnwind"
+        memory_effects:
+            - !<MemoryRestriction>
+                memory_access: !MemoryAccessType NoModRef
     - !<IntrinsicDefinition>
         name: "GenISA_SyncStackPtr"
         comment: "Raytracing: Intrinsic used as marker in later passes to find per\