Address feedback for predicated load/store

Address feedback from the previous patch for predicated load/store.
Also disable promotion to predicated load/store, when
StatelessToStatefull optimization is enabled.

Author: YuriPlyakhin

IGC/Compiler/CISACodeGen/EmitVISAPass.cpp

@@ -19576,6 +19576,7 @@ void EmitPass::emitLSCVectorLoad_subDW(LSC_CACHE_OPTS CacheOpts, bool UseA32,
             finalPredicate = IGC_IS_FLAG_ENABLED(UseVMaskPredicateForLoads) ? GetCombinedVMaskPred(flag) : flag;
 
         if (inputPredicate && finalPredicate) {
+            m_pCtx->EmitError("Predicated load is not expected.", nullptr);
             IGC_ASSERT_MESSAGE(false, "Not expected scenario. Predicated load should not be used.");
             m_encoder->And(finalPredicate, finalPredicate, inputPredicate);
             m_encoder->Push();
@@ -19763,7 +19764,7 @@ void EmitPass::emitLSCVectorLoad(Instruction* inst,
         eOffset = TruncatePointer(eOffset);
     }
 
-    // if the merge value is dead after predicated load and is not immidiate, and
+    // if the merge value is dead after predicated load and is not immediate, and
     // has other properties same as destination -> use it as destination
     bool isDestReplacedWithMerge = false;
     CVariable* mergeVar = nullptr;
@@ -19912,6 +19913,7 @@ void EmitPass::emitLSCVectorLoad(Instruction* inst,
             if(predicatedLoad) {
                 CVariable* cond = GetSymbol(inst->getOperand(2));
                 if (pred) {
+                    m_pCtx->EmitError("Predicated load is not expected.", inst);
                     IGC_ASSERT_MESSAGE(false, "Not expected scenario. Predicated load should not be used.");
                     m_encoder->And(pred, pred, cond);
                     m_encoder->Push();
@@ -20018,7 +20020,10 @@ void EmitPass::emitLSCVectorStore_subDW(LSC_CACHE_OPTS CacheOpts, bool UseA32,
             // setPredicateForDiscard used earlier is used only when compiler needs to create the
             // resource loop, which is never expected for simple LLVM store or PredicatedStore.
             // Hence, these predicates should never be used together.
-            IGC_ASSERT_MESSAGE(doUniformStore || flag == nullptr, "Unexpected scenario: resource loop with predicated store!");
+            if(!doUniformStore && flag != nullptr) {
+                m_pCtx->EmitError("Predicated store is not expected.", nullptr);
+                IGC_ASSERT_MESSAGE(false, "Unexpected scenario: resource loop with predicated store!");
+            }
             m_encoder->SetPredicate(GetSymbol(predicate));
         }
 
@@ -20286,7 +20291,10 @@ void EmitPass::emitLSCVectorStore(Value *Ptr,
                 // setPredicateForDiscard used earlier is used only when compiler needs to create the
                 // resource loop, which is never expected for simple LLVM store or PredicatedStore.
                 // Hence, these predicates should never be used together.
-                IGC_ASSERT_MESSAGE(flag == nullptr, "Unexpected scenario: resource loop with predicated store!");
+                if (flag != nullptr) {
+                    m_pCtx->EmitError("Predicated store is not expected.", nullptr);
+                    IGC_ASSERT_MESSAGE(false, "Unexpected scenario: resource loop with predicated store!");
+                }
                 m_encoder->SetPredicate(GetSymbol(predicate));
             }
 

IGC/Compiler/CISACodeGen/ShaderCodeGen.cpp

@@ -411,15 +411,6 @@ static void UpdateInstTypeHint(CodeGenContext& ctx)
 // forward declaration
 llvm::ModulePass* createPruneUnusedArgumentsPass();
 
-static bool useStatelessToStateful(CodeGenContext& ctx)
-{
-    return (ctx.m_instrTypes.hasLoadStore &&
-        ctx.m_DriverInfo.SupportsStatelessToStatefulBufferTransformation() &&
-        !ctx.getModuleMetaData()->compOpt.GreaterThan4GBBufferRequired &&
-        IGC_IS_FLAG_ENABLED(EnableStatelessToStateful) &&
-        !ctx.m_instrTypes.hasInlineAsmPointerAccess);
-}
-
 void AddLegalizationPasses(CodeGenContext& ctx, IGCPassManager& mpm, PSSignature* pSignature)
 {
     COMPILER_TIME_START(&ctx, TIME_CG_Add_Legalization_Passes);
@@ -802,7 +793,7 @@ void AddLegalizationPasses(CodeGenContext& ctx, IGCPassManager& mpm, PSSignature
         }
     }
 
-    if (!isOptDisabled && useStatelessToStateful(ctx))
+    if (!isOptDisabled && ctx.useStatelessToStateful())
     {
         mpm.add(new StatelessToStateful(TargetAddressing::BINDFUL));
     }
@@ -1510,7 +1501,7 @@ void OptimizeIR(CodeGenContext* const pContext)
                 bool runGEPLSR = IGC_IS_FLAG_ENABLED(EnableGEPLSR) &&
                     pContext->type == ShaderType::OPENCL_SHADER &&
                     pContext->platform.getPlatformInfo().eProductFamily == IGFX_PVC &&
-                    !useStatelessToStateful(*pContext) &&
+                    !pContext->useStatelessToStateful() &&
                     pContext->m_retryManager.IsFirstTry();
 
                 if (runGEPLSR && IGC_IS_FLAG_DISABLED(RunGEPLSRAfterLICM))

IGC/Compiler/CodeGenPublic.h

@@ -1297,6 +1297,15 @@ namespace IGC
         {
             return (mscratchSpaceUsedBySpills > m_spillAllowed);
         }
+
+        bool useStatelessToStateful()
+        {
+            return (m_instrTypes.hasLoadStore &&
+                    m_DriverInfo.SupportsStatelessToStatefulBufferTransformation() &&
+                    !getModuleMetaData()->compOpt.GreaterThan4GBBufferRequired &&
+                    IGC_IS_FLAG_ENABLED(EnableStatelessToStateful) &&
+                    !m_instrTypes.hasInlineAsmPointerAccess);
+        }
     };
 
     struct SComputeShaderSecondCompileInput

IGC/Compiler/Optimizer/PromoteToPredicatedMemoryAccess.cpp

@@ -47,8 +47,11 @@ namespace IGC {
 
 bool PromoteToPredicatedMemoryAccess::runOnFunction(Function &F) {
   CodeGenContext* pCtx = getAnalysis<CodeGenContextWrapper>().getCodeGenContext();
-  if (!pCtx->platform.hasLSC() || !pCtx->platform.LSCEnabled() ||
-      (pCtx->type == ShaderType::OPENCL_SHADER && static_cast<OpenCLProgramContext*>(pCtx)->m_InternalOptions.PromoteStatelessToBindless))
+  if (!pCtx->platform.hasLSC() ||
+      !pCtx->platform.LSCEnabled() ||
+      (pCtx->type == ShaderType::OPENCL_SHADER &&
+               static_cast<OpenCLProgramContext*>(pCtx)->m_InternalOptions.PromoteStatelessToBindless) ||
+      pCtx->useStatelessToStateful())
     return false;
 
   SmallVector<std::pair<BranchInst *, bool>, 8> WorkList;