More optimizations.

Author: csyonghe

source/slang/slang-emit.cpp

@@ -220,12 +220,17 @@ struct RequiredLoweringPassSet
     bool combinedTextureSamplers;
     bool reinterpret;
     bool generics;
+    bool bindExistential;
     bool autodiff;
     bool derivativePyBindWrapper;
     bool bitcast;
     bool existentialTypeLayout;
     bool bindingQuery;
     bool meshOutput;
+    bool higherOrderFunc;
+    bool glslGlobalVar;
+    bool glslSSBO;
+    bool byteAddressBuffer;
 };
 
 void calcRequiredLoweringPassSet(RequiredLoweringPassSet& result, CodeGenContext* codeGenContext, IRInst* inst)
@@ -303,6 +308,27 @@ void calcRequiredLoweringPassSet(RequiredLoweringPassSet& result, CodeGenContext
     case kIROp_AutoPyBindCudaDecoration:
         result.derivativePyBindWrapper = true;
         break;
+    case kIROp_Param:
+        if (as<IRFuncType>(inst->getDataType()))
+            result.higherOrderFunc = true;
+        break;
+    case kIROp_GlobalInputDecoration:
+    case kIROp_GlobalOutputDecoration:
+    case kIROp_GetWorkGroupSize:
+        result.glslGlobalVar = true;
+        break;
+    case kIROp_BindExistentialSlotsDecoration:
+        result.bindExistential = true;
+        result.generics = true;
+        result.existentialTypeLayout = true;
+        break;
+    case kIROp_GLSLShaderStorageBufferType:
+        result.glslSSBO = true;
+        break;
+    case kIROp_ByteAddressBufferLoad:
+    case kIROp_ByteAddressBufferStore:
+        result.byteAddressBuffer = true;
+        break;
     }
     for (auto child : inst->getDecorationsAndChildren())
     {
@@ -348,10 +374,14 @@ Result linkAndOptimizeIR(
     // un-specialized IR.
     dumpIRIfEnabled(codeGenContext, irModule, "POST IR VALIDATION");
 
-    if(!isKhronosTarget(targetRequest))
+    // Scan the IR module and determine which lowering/legalization passes are needed.
+    RequiredLoweringPassSet requiredLoweringPassSet = {};
+    calcRequiredLoweringPassSet(requiredLoweringPassSet, codeGenContext, irModule->getModuleInst());
+
+    if(!isKhronosTarget(targetRequest) && requiredLoweringPassSet.glslSSBO)
         lowerGLSLShaderStorageBufferObjectsToStructuredBuffers(irModule, sink);
 
-    if (!targetProgram->getOptionSet().shouldPerformMinimumOptimizations())
+    if (requiredLoweringPassSet.glslGlobalVar)
         translateGLSLGlobalVar(codeGenContext, irModule);
 
     // Replace any global constants with their values.
@@ -370,7 +400,8 @@ Result linkAndOptimizeIR(
     // shader parameters for those slots, to be wired up to
     // use sites.
     //
-    bindExistentialSlots(irModule, sink);
+    if (requiredLoweringPassSet.bindExistential)
+        bindExistentialSlots(irModule, sink);
 #if 0
     dumpIRIfEnabled(codeGenContext, irModule, "EXISTENTIALS BOUND");
 #endif
@@ -450,9 +481,6 @@ Result linkAndOptimizeIR(
         break;
     }
 
-    RequiredLoweringPassSet requiredLoweringPassSet = {};
-    calcRequiredLoweringPassSet(requiredLoweringPassSet, codeGenContext, irModule->getModuleInst());
-
     if (requiredLoweringPassSet.optionalType)
         lowerOptionalType(irModule, sink);
 
@@ -540,7 +568,10 @@ Result linkAndOptimizeIR(
             return SLANG_FAIL;
         dumpIRIfEnabled(codeGenContext, irModule, "AFTER-SPECIALIZE");
 
-        applySparseConditionalConstantPropagation(irModule, codeGenContext->getSink());
+        if (changed)
+        {
+            applySparseConditionalConstantPropagation(irModule, codeGenContext->getSink());
+        }
         eliminateDeadCode(irModule, deadCodeEliminationOptions);
 
         validateIRModuleIfEnabled(codeGenContext, irModule);
@@ -564,7 +595,7 @@ Result linkAndOptimizeIR(
         // which do.
         // Specialize away these parameters
         // TODO: We should implement a proper defunctionalization pass
-        if (!targetProgram->getOptionSet().shouldPerformMinimumOptimizations())
+        if (requiredLoweringPassSet.higherOrderFunc)
             changed |= specializeHigherOrderParameters(codeGenContext, irModule);
 
         dumpIRIfEnabled(codeGenContext, irModule, "BEFORE-AUTODIFF");
@@ -673,7 +704,11 @@ Result linkAndOptimizeIR(
     // up downstream passes like type legalization, so we
     // will run a DCE pass to clean up after the specialization.
     //
-    if (!fastIRSimplificationOptions.minimalOptimization)
+    if (fastIRSimplificationOptions.minimalOptimization)
+    {
+        eliminateDeadCode(irModule, deadCodeEliminationOptions);
+    }
+    else
     {
         simplifyIR(targetProgram, irModule, defaultIRSimplificationOptions, sink);
     }
@@ -788,7 +823,9 @@ Result linkAndOptimizeIR(
     // to see if we can clean up any temporaries created by legalization.
     // (e.g., things that used to be aggregated might now be split up,
     // so that we can work with the individual fields).
-    if (!fastIRSimplificationOptions.minimalOptimization)
+    if (fastIRSimplificationOptions.minimalOptimization)
+        eliminateDeadCode(irModule, deadCodeEliminationOptions);
+    else
         simplifyIR(targetProgram, irModule, fastIRSimplificationOptions, sink);
 
 #if 0
@@ -849,6 +886,7 @@ Result linkAndOptimizeIR(
     // of aggregate types from/to byte-address buffers into
     // stores of individual scalar or vector values.
     //
+    if (requiredLoweringPassSet.byteAddressBuffer)
     {
         ByteAddressBufferLegalizationOptions byteAddressBufferOptions;
 

source/slang/slang-ir-insts.h

@@ -3681,6 +3681,13 @@ struct IRBuilder
     {
         return emitCallInst(type, func, args.getCount(), args.getBuffer());
     }
+    IRCall* emitCallInst(
+        IRType* type,
+        IRInst* func,
+        ArrayView<IRInst*> args)
+    {
+        return emitCallInst(type, func, args.getCount(), args.getBuffer());
+    }
 
     IRInst* emitTryCallInst(
         IRType* type,

source/slang/slang-ir-legalize-types.cpp

@@ -212,7 +212,7 @@ struct LegalCallBuilder
     IRCall* m_call = nullptr;
 
         /// The legalized arguments for the call
-    List<IRInst*> m_args;
+    ShortList<IRInst*> m_args;
 
         /// Add a logical argument to the call (which may map to zero or mmore actual arguments)
     void addArg(
@@ -463,7 +463,7 @@ struct LegalCallBuilder
             resultType,
             m_call->getCallee(),
             m_args.getCount(),
-            m_args.getBuffer());
+            m_args.getArrayView().getBuffer());
     }
 };
 
@@ -706,7 +706,7 @@ static LegalVal legalizeRetVal(
     return LegalVal();
 }
 
-static void _addVal(List<IRInst*>& rs, const LegalVal& legalVal)
+static void _addVal(ShortList<IRInst*>& rs, const LegalVal& legalVal)
 {
     switch (legalVal.flavor)
     {
@@ -733,7 +733,7 @@ static LegalVal legalizeUnconditionalBranch(
     ArrayView<LegalVal>        args,
     IRUnconditionalBranch*     branchInst)
 {
-    List<IRInst*> newArgs;
+    ShortList<IRInst*> newArgs;
     for (auto arg : args)
     {
         switch (arg.flavor)
@@ -757,7 +757,7 @@ static LegalVal legalizeUnconditionalBranch(
                 SLANG_UNIMPLEMENTED_X("Unknown legalized val flavor.");
         }
     }
-    context->builder->emitIntrinsicInst(nullptr, branchInst->getOp(), newArgs.getCount(), newArgs.getBuffer());
+    context->builder->emitIntrinsicInst(nullptr, branchInst->getOp(), newArgs.getCount(), newArgs.getArrayView().getBuffer());
     return LegalVal();
 }
 
@@ -861,7 +861,7 @@ static LegalVal legalizeDebugVar(IRTypeLegalizationContext* context, LegalType t
 static LegalVal legalizeDebugValue(IRTypeLegalizationContext* context, LegalVal debugVar, LegalVal debugValue, IRDebugValue* originalInst)
 {
     // For now we just discard any special part and keep the ordinary part.
-    List<IRInst*> accessChain;
+    ShortList<IRInst*> accessChain;
     for (UInt i = 0; i < originalInst->getAccessChainCount(); i++)
     {
         accessChain.add(originalInst->getAccessChain(i));
@@ -873,7 +873,7 @@ static LegalVal legalizeDebugValue(IRTypeLegalizationContext* context, LegalVal
             context->builder->emitDebugValue(
                 debugVar.getSimple(),
                 debugValue.getSimple(),
-                accessChain.getArrayView()));
+                accessChain.getArrayView().arrayView));
     case LegalType::Flavor::none:
         return LegalVal();
     case LegalType::Flavor::pair:
@@ -2205,7 +2205,7 @@ static LegalVal legalizeInst(
     // value of each, and collect them in an array for subsequent use.
     //
     auto argCount = inst->getOperandCount();
-    List<LegalVal> legalArgs;
+    ShortList<LegalVal> legalArgs;
     //
     // Along the way we will also note whether there were any operands
     // with non-simple legalized values.
@@ -2277,7 +2277,7 @@ static LegalVal legalizeInst(
         context,
         inst,
         legalType,
-        legalArgs.getArrayView());
+        legalArgs.getArrayView().arrayView);
 
     if (legalVal.flavor == LegalVal::Flavor::simple)
     {

source/slang/slang-ir-specialize-resources.cpp

@@ -1189,8 +1189,11 @@ bool specializeResourceUsage(
             // and turned into SSA temporaries. Such optimization may enable
             // the following passes to "see" and specialize more cases.
             //
-            simplifyIR(codeGenContext->getTargetProgram(), irModule,
-                IRSimplificationOptions::getFast(codeGenContext->getTargetProgram()));
+            if (changed)
+            {
+                simplifyIR(codeGenContext->getTargetProgram(), irModule,
+                    IRSimplificationOptions::getFast(codeGenContext->getTargetProgram()));
+            }
             result |= changed;
         }
         if (unspecializableFuncs.getCount() == 0)