Fixed invalid alignment of vector loads in PTX backend.

Author: m4rs-mt

Src/ILGPU/Backends/PTX/PTXBackend.cs

@@ -168,7 +168,7 @@ protected override StringBuilder CreateKernelBuilder(
                 ? PointerAlignments.Create(
                     backendContext.KernelMethod,
                     DefaultGlobalMemoryAlignment)
-                : PointerAlignments.Empty;
+                : null;
 
             data = new PTXCodeGenerator.GeneratorArgs(
                 this,

Src/ILGPU/Backends/PTX/PTXCodeGenerator.Emitter.cs

@@ -659,21 +659,27 @@ public void EmitComplexCommandWithOffsets<TEmitter>(
         /// to resolve the correct offset in bytes within a structure.
         /// </summary>
         /// <typeparam name="TEmitter">The emitter type.</typeparam>
+        /// <param name="pointerValue">The pointer to get the alignment for.</param>
+        /// <param name="safeAlignment">The safe minimum alignment in bytes.</param>
         /// <param name="command">The generic command to emit.</param>
         /// <param name="emitter">The current emitter.</param>
         /// <param name="register">The involved register.</param>
-        /// <param name="alignment">The base alignment in bytes.</param>
         [MethodImpl(MethodImplOptions.AggressiveInlining)]
         public void EmitVectorizedCommand<TEmitter>(
+            Value pointerValue,
+            int safeAlignment,
             string command,
             in TEmitter emitter,
-            Register register,
-            int alignment)
+            Register register)
             where TEmitter : IVectorizedCommandEmitter
         {
-            if (register is CompoundRegister compoundRegister)
+            if (PointerAlignments != null &&
+                register is CompoundRegister compoundRegister)
             {
                 // Check the provided alignment value to create vectorized instructions
+                int alignment = PointerAlignments.GetAlignment(
+                    pointerValue,
+                    safeAlignment);
                 var ranges = compoundRegister.Type.VectorizableFields;
                 for (int i = 0, e = ranges.Count; i < e; ++i)
                 {

Src/ILGPU/Backends/PTX/PTXCodeGenerator.Values.cs

@@ -439,15 +439,12 @@ public void GenerateCode(Load load)
             var sourceType = load.Source.Type as PointerType;
             var targetRegister = Allocate(load);
 
-            // Query alignment information to emit vectorized instructions
-            int alignment = PointerAlignments.GetAlignment(
-                load.Source,
-                sourceType.ElementType.Alignment);
             EmitVectorizedCommand(
+                load.Source,
+                sourceType.ElementType.Alignment,
                 PTXInstructions.LoadOperation,
                 new LoadEmitter(sourceType, address),
-                targetRegister,
-                alignment);
+                targetRegister);
         }
 
         /// <summary>
@@ -540,15 +537,12 @@ public void GenerateCode(Store store)
             var targetType = store.Target.Type as PointerType;
             var value = Load(store.Value);
 
-            // Query alignment information to emit vectorized instructions
-            int baseAlignment = PointerAlignments.GetAlignment(
-                store.Target,
-                targetType.ElementType.Alignment);
             EmitVectorizedCommand(
+                store.Target,
+                targetType.ElementType.Alignment,
                 PTXInstructions.StoreOperation,
                 new StoreEmitter(targetType, address),
-                value,
-                baseAlignment);
+                value);
         }
 
         /// <summary cref="IBackendCodeGenerator.GenerateCode(LoadFieldAddress)"/>

Src/ILGPU/IR/Analyses/PointerAlignments.cs

@@ -84,14 +84,15 @@ public readonly int Merge(int first, int second) =>
                 {
                     LoadFieldAddress lfa =>
                         AnalysisValue.Create(
-                            context[lfa.Source].Data +
-                            lfa.StructureType.GetOffset(lfa.FieldSpan.Access),
-                            lfa.Type),
+                            Math.Min(
+                                context[lfa.Source].Data,
+                                lfa.StructureType[lfa.FieldSpan.Access].Alignment),
+                                lfa.Type),
                     LoadElementAddress lea =>
                         AnalysisValue.Create(
-                            Math.Min(
+                            Math.Max(
                                 context[lea.Source].Data,
-                                (lea.Type as IAddressSpaceType).ElementType.Size),
+                                (lea.Type as IAddressSpaceType).ElementType.Alignment),
                             lea.Type),
                     _ => null,
                 };

Src/ILGPU/IR/Types/StructureType.cs

@@ -18,7 +18,6 @@
 using System.Diagnostics.CodeAnalysis;
 using System.Reflection;
 using System.Runtime.CompilerServices;
-using System.Security.Permissions;
 using System.Text;
 
 namespace ILGPU.IR.Types
@@ -410,19 +409,23 @@ public readonly struct VectorizableFieldCollection
             /// </summary>
             public struct Entry
             {
+                #region Instance
+
                 /// <summary>
                 /// Constructs a new entry.
                 /// </summary>
-                internal Entry(TypeNode type, int index, int offset, int count = 1)
+                internal Entry(
+                    TypeNode type,
+                    int index,
+                    int offset,
+                    int count = 1)
                 {
                     Type = type;
                     Index = index;
                     Count = count;
                     Offset = offset;
                 }
 
-                #region Instance
-
                 #endregion
 
                 #region Properties
@@ -440,17 +443,39 @@ internal Entry(TypeNode type, int index, int offset, int count = 1)
                 /// <summary>
                 /// Returns the number of fields.
                 /// </summary>
-                public int Count { get; private set; }
+                public int Count { readonly get; private set; }
 
                 /// <summary>
                 /// Returns the base offset in bytes from the beginning of the field.
                 /// </summary>
                 public int Offset { get; }
 
+                /// <summary>
+                /// Returns the required alignment in bytes.
+                /// </summary>
+                public readonly int RequiredAlignment => Count * Type.Size;
+
                 #endregion
 
                 #region Methods
 
+                /// <summary>
+                /// Splits the current entry into two parts.
+                /// </summary>
+                /// <param name="first">The first part.</param>
+                /// <param name="second">The second part.</param>
+                internal void Split(out Entry first, out Entry second)
+                {
+                    Type.Assert(Count > 1);
+                    int firstCount = Count >> 1 + Count % 2;
+                    first = new Entry(Type, Index, Offset, firstCount);
+                    second = new Entry(
+                        Type,
+                        Index + firstCount,
+                        Offset + Type.Size * firstCount,
+                        Count - firstCount);
+                }
+
                 /// <summary>
                 /// Adds a field to this entry.
                 /// </summary>
@@ -463,7 +488,22 @@ internal Entry(TypeNode type, int index, int offset, int count = 1)
                 /// <param name="alignment">The underlying alignment in bytes.</param>
                 /// <returns>True, if the range is properly aligned.</returns>
                 public readonly bool IsAligned(int alignment) =>
-                    (alignment + Offset) % (Type.Size * Count) == 0;
+                    // Check for a proper alignment of the base address
+                    alignment % RequiredAlignment == 0;
+
+                /// <summary>
+                /// Returns true if this entry can be properly aligned.
+                /// </summary>
+                /// <param name="parentType">The parent structure type.</param>
+                internal readonly bool CanBeAligned(StructureType parentType)
+                {
+                    int requiredAlignment = RequiredAlignment;
+                    return 
+                        // Check for a relative alignment inside the structure
+                        Offset % requiredAlignment == 0 &&
+                        // Check for a relative alignment of odd structure accesses
+                        (Offset + parentType.Size) % requiredAlignment == 0;
+                }
 
                 #endregion
             }
@@ -499,7 +539,7 @@ internal VectorizableFieldCollection(StructureType structureType)
                         currentOffset + nextType.Size != nextOffset)
                     {
                         // Register the current vectorizable entry
-                        RegisterRange(current);
+                        RegisterRange(structureType, current);
                         current = new Entry(
                             nextType,
                             i,
@@ -514,14 +554,17 @@ internal VectorizableFieldCollection(StructureType structureType)
                 }
 
                 // Add the last entry
-                RegisterRange(current);
+                RegisterRange(structureType, current);
             }
 
             /// <summary>
             /// Registers the given range entry.
             /// </summary>
+            /// <param name="structureType">The parent structure type.</param>
             /// <param name="entry">The entry to register.</param>
-            private void RegisterRange(in Entry entry)
+            private void RegisterRange(
+                StructureType structureType,
+                in Entry entry)
             {
                 int offset = entry.Offset;
                 for (
@@ -531,12 +574,22 @@ private void RegisterRange(in Entry entry)
                 {
                     for (; index + stepSize <= entry.Count; index += stepSize)
                     {
-                        ranges.Add(
-                            new Entry(
-                                entry.Type,
-                                index + entry.Index,
-                                offset,
-                                stepSize));
+                        var newEntry = new Entry(
+                            entry.Type,
+                            index + entry.Index,
+                            offset,
+                            stepSize);
+                        if (newEntry.Count > 1 && !newEntry.CanBeAligned(structureType))
+                        {
+                            newEntry.Split(out var first, out var second);
+                            RegisterRange(structureType, first);
+                            RegisterRange(structureType, second);
+                        }
+                        else
+                        {
+                            // The entry is properly aligned
+                            ranges.Add(newEntry);
+                        }
                         offset += entry.Type.Size * stepSize;
                     }
                 }