Fix regressions.

Author: csyonghe

source/slang/hlsl.meta.slang

@@ -150,7 +150,7 @@ struct ByteAddressBuffer
         {
         case hlsl: __intrinsic_asm ".Load2";
         default:
-            return __byteAddressBufferLoad<uint2>(this, location, __alignOf_intrinsic<uint2>());
+            return __byteAddressBufferLoad<uint2>(this, location, __naturalStrideOf<uint2>());
         }
     }
 
@@ -192,7 +192,7 @@ struct ByteAddressBuffer
         {
         case hlsl: __intrinsic_asm ".Load3";
         default:
-            return __byteAddressBufferLoad<uint3>(this, location, __alignOf_intrinsic<uint3>());
+            return __byteAddressBufferLoad<uint3>(this, location, __naturalStrideOf<uint3>());
         }
     }
 
@@ -234,7 +234,7 @@ struct ByteAddressBuffer
         {
         case hlsl: __intrinsic_asm ".Load4";
         default:
-            return __byteAddressBufferLoad<uint4>(this, location, __alignOf_intrinsic<uint4>());
+            return __byteAddressBufferLoad<uint4>(this, location, __naturalStrideOf<uint4>());
         }
     }
 
@@ -259,7 +259,7 @@ struct ByteAddressBuffer
     [ForceInline]
     T LoadAligned<T>(int location)
     {
-        return __byteAddressBufferLoad<T>(this, location, __alignOf_intrinsic<T>());
+        return __byteAddressBufferLoad<T>(this, location, __naturalStrideOf<T>());
     }
 };
 
@@ -3758,7 +3758,7 @@ struct $(item.name)
         {
         case hlsl: __intrinsic_asm ".Load2";
         default:
-            return __byteAddressBufferLoad<uint2>(this, location, __alignOf_intrinsic<uint2>());
+            return __byteAddressBufferLoad<uint2>(this, location, __naturalStrideOf<uint2>());
         }
     }
 
@@ -3800,7 +3800,7 @@ struct $(item.name)
         {
         case hlsl: __intrinsic_asm ".Load3";
         default:
-            return __byteAddressBufferLoad<uint3>(this, location, __alignOf_intrinsic<uint3>());
+            return __byteAddressBufferLoad<uint3>(this, location, __naturalStrideOf<uint3>());
         }
     }
 
@@ -3842,7 +3842,7 @@ struct $(item.name)
         {
         case hlsl: __intrinsic_asm ".Load4";
         default:
-            return __byteAddressBufferLoad<uint4>(this, location, __alignOf_intrinsic<uint4>());
+            return __byteAddressBufferLoad<uint4>(this, location, __naturalStrideOf<uint4>());
         }
     }
 
@@ -3870,7 +3870,7 @@ struct $(item.name)
     [require(cpp_cuda_glsl_hlsl_spirv, byteaddressbuffer_rw)]
     T LoadAligned<T>(int location)
     {
-        return __byteAddressBufferLoad<T>(this, location, __alignOf_intrinsic<T>());
+        return __byteAddressBufferLoad<T>(this, location, __naturalStrideOf<T>());
     }
 
 ${{{{
@@ -4763,7 +4763,7 @@ ${{{{
         {
         case hlsl: __intrinsic_asm ".Store2";
         default:
-            __byteAddressBufferStore(this, address, __alignOf_intrinsic<uint2>(), value);
+            __byteAddressBufferStore(this, address, __naturalStrideOf<uint2>(), value);
         }
     }
 
@@ -4800,7 +4800,7 @@ ${{{{
         {
         case hlsl: __intrinsic_asm ".Store3";
         default:
-            __byteAddressBufferStore(this, address, __alignOf_intrinsic<uint3>(), value);
+            __byteAddressBufferStore(this, address, __naturalStrideOf<uint3>(), value);
         }
     }
 
@@ -4837,7 +4837,7 @@ ${{{{
         {
         case hlsl: __intrinsic_asm ".Store4";
         default:
-            __byteAddressBufferStore(this, address, __alignOf_intrinsic<uint4>(), value);
+            __byteAddressBufferStore(this, address, __naturalStrideOf<uint4>(), value);
         }
     }
 
@@ -4856,7 +4856,7 @@ ${{{{
     [ForceInline]
     void StoreAligned<T>(int offset, T value)
     {
-        __byteAddressBufferStore(this, offset, __alignOf_intrinsic<T>(), value);
+        __byteAddressBufferStore(this, offset, __naturalStrideOf<T>(), value);
     }
 };
 

source/slang/slang-emit.cpp

@@ -259,6 +259,9 @@ void calcRequiredLoweringPassSet(RequiredLoweringPassSet& result, CodeGenContext
         }
         break;
     case kIROp_PseudoPtrType:
+    case kIROp_BoundInterfaceType:
+    case kIROp_BindExistentialsType:
+        result.generics = true;
         result.existentialTypeLayout = true;
         break;
     case kIROp_GetRegisterIndex:
@@ -267,6 +270,7 @@ void calcRequiredLoweringPassSet(RequiredLoweringPassSet& result, CodeGenContext
         break;
     case kIROp_BackwardDifferentiate:
     case kIROp_ForwardDifferentiate:
+    case kIROp_MakeDifferentialPairUserCode:
         result.autodiff = true;
         break;
     case kIROp_VerticesType:
@@ -581,6 +585,9 @@ Result linkAndOptimizeIR(
 
     finalizeSpecialization(irModule);
 
+    requiredLoweringPassSet = {};
+    calcRequiredLoweringPassSet(requiredLoweringPassSet, codeGenContext, irModule->getModuleInst());
+
     switch (target)
     {
     case CodeGenTarget::PyTorchCppBinding:
@@ -642,9 +649,6 @@ Result linkAndOptimizeIR(
         SLANG_RETURN_ON_FAIL(checkGetStringHashInsts(irModule, sink));
     }
 
-    requiredLoweringPassSet = {};
-    calcRequiredLoweringPassSet(requiredLoweringPassSet, codeGenContext, irModule->getModuleInst());
-
     // For targets that supports dynamic dispatch, we need to lower the
     // generics / interface types to ordinary functions and types using
     // function pointers.

source/slang/slang-ir-inline.cpp

@@ -698,12 +698,12 @@ struct MandatoryEarlyInliningPass : InliningPassBase
 };
 
 
-void performMandatoryEarlyInlining(IRModule* module)
+bool performMandatoryEarlyInlining(IRModule* module)
 {
     SLANG_PROFILE;
 
     MandatoryEarlyInliningPass pass(module);
-    pass.considerAllCallSites();
+    return pass.considerAllCallSites();
 }
 
 namespace { // anonymous

source/slang/slang-ir-inline.h

@@ -15,7 +15,7 @@ namespace Slang
     Result performTypeInlining(IRModule* module, DiagnosticSink* sink);
 
         /// Inline any call sites to functions marked `[unsafeForceInlineEarly]`
-    void performMandatoryEarlyInlining(IRModule* module);
+    bool performMandatoryEarlyInlining(IRModule* module);
 
         /// Inline any call sites to functions marked `[ForceInline]`
     void performForceInlining(IRModule* module);

source/slang/slang-ir-lower-generics.cpp

@@ -277,8 +277,7 @@ namespace Slang
         sharedContext.targetProgram = program;
         sharedContext.sink = sink;
 
-        cleanUpRTTIHandleTypes(&sharedContext);
-        cleanUpInterfaceTypes(&sharedContext);
+        specializeRTTIObjects(&sharedContext, sink);
 
         lowerTuples(module, sink);
         if (sink->getErrorCount() != 0)

source/slang/slang-ir-peephole.cpp

@@ -261,8 +261,12 @@ struct PeepholeContext : InstPassBase
 
                 // Save the alignment information and exit early if it is invalid
                 IRSizeAndAlignment sizeAlignment;
-                auto alignOfInst = as<IRAlignOf>(inst);
-                auto baseType = alignOfInst->getBaseOp()->getDataType();
+                IRType* baseType = nullptr;
+                if (auto t = as<IRType>(inst->getOperand(0)))
+                    baseType = t;
+                else
+                    baseType = inst->getOperand(0)->getDataType();
+
                 if (SLANG_FAILED(getNaturalSizeAndAlignment(targetProgram->getOptionSet(), baseType, &sizeAlignment)))
                     break;
                 if (sizeAlignment.size == 0)

source/slang/slang-lower-to-ir.cpp

@@ -10883,11 +10883,12 @@ RefPtr<IRModule> generateIRForTranslationUnit(
 
     // Optionally, run another optimization pass to clean up things.
     //
-    if (!sharedContextStorage.m_linkage->m_optionSet.getBoolOption(CompilerOptionName::MinimumSlangOptimization))
+    for (;;)
     {
-        for (;;)
+        bool changed = false;
+        changed |= performMandatoryEarlyInlining(module);
+        if (!sharedContextStorage.m_linkage->m_optionSet.getBoolOption(CompilerOptionName::MinimumSlangOptimization))
         {
-            bool changed = false;
             changed |= constructSSA(module);
             simplifyCFG(module, CFGSimplificationOptions::getDefault());
             changed |= applySparseConditionalConstantPropagation(module, compileRequest->getSink());
@@ -10897,9 +10898,9 @@ RefPtr<IRModule> generateIRForTranslationUnit(
                 if (auto func = as<IRGlobalValueWithCode>(inst))
                     eliminateDeadCode(func);
             }
-            if (!changed)
-                break;
         }
+        if (!changed)
+            break;
     }