[clang] [llvm] [SROA] Canonicalize homogeneous structs into fixed vectors (PR #165159)

[Yaxun Liu via cfe-commits]

================
@@ -5130,6 +5130,162 @@ bool SROA::presplitLoadsAndStores(AllocaInst &AI, AllocaSlices &AS) {
   return true;
 }
 
+/// Try to canonicalize a homogeneous, tightly-packed struct to a vector type.
+///
+/// For structs where all elements have the same type and are tightly packed
+/// (no padding), we can represent them as a fixed vector which enables better
+/// optimization (e.g., vector selects instead of memcpy).
+///
+/// \param STy The struct type to try to canonicalize.
+/// \param DL The DataLayout for size/alignment queries.
+/// \returns The equivalent vector type, or nullptr if not applicable.
+static FixedVectorType *tryCanonicalizeStructToVector(StructType *STy,
+                                                      const DataLayout &DL) {
+  unsigned NumElts = STy->getNumElements();
+  if (NumElts != 2 && NumElts != 4)
+    return nullptr;
+
+  // All elements must be the same type.
+  Type *EltTy = STy->getElementType(0);
+  for (unsigned I = 1; I < NumElts; ++I)
+    if (STy->getElementType(I) != EltTy)
+      return nullptr;
+
+  // Element type must be valid for vectors.
+  if (!VectorType::isValidElementType(EltTy))
+    return nullptr;
+
+  // Only allow integer types >= 8 bits or floating point.
+  if (auto *IT = dyn_cast<IntegerType>(EltTy)) {
+    if (IT->getBitWidth() < 8)
+      return nullptr;
+  } else if (!EltTy->isFloatingPointTy()) {
+    return nullptr;
+  }
+
+  // Element size must be fixed and non-zero.
+  TypeSize EltTS = DL.getTypeAllocSize(EltTy);
+  if (!EltTS.isFixed())
+    return nullptr;
+  uint64_t EltSize = EltTS.getFixedValue();
+  if (EltSize < 1)
+    return nullptr;
+
+  const StructLayout *SL = DL.getStructLayout(STy);
+  uint64_t StructSize = SL->getSizeInBytes();
+  if (StructSize == 0)
----------------
yxsamliu wrote:

removed the separate zero-size guard and now just rely on the existing element checks together with the packed-size check below

https://github.com/llvm/llvm-project/pull/165159