Fix SROA incorrect vector element index for padded min precision types

SROA's getNaturalGEPRecursively used getTypeSizeInBits (primitive size)
for vector element offsets, but GEP offset calculation uses
getTypeAllocSize (padded size). With DXC's data layout (i16:32, f16:32),
this mismatch caused byte offset 4 (element 1) to map to vector index 2
instead of index 1 (4/2=2 vs 4/4=1). This led SROA to misplace or
eliminate stores to vector elements 1 and 2, producing incorrect code.

Fix: Use getTypeAllocSizeInBits consistently for vector element sizes in
getNaturalGEPRecursively, isVectorPromotionViable, and the
AllocaSliceRewriter constructor.

Co-authored-by: Copilot <[email protected]>

Author: alsepkow

lib/Transforms/Scalar/SROA.cpp

@@ -1671,7 +1671,11 @@ static Value *getNaturalGEPRecursively(IRBuilderTy &IRB, const DataLayout &DL,
   // extremely poorly defined currently. The long-term goal is to remove GEPing
   // over a vector from the IR completely.
   if (VectorType *VecTy = dyn_cast<VectorType>(Ty)) {
-    unsigned ElementSizeInBits = DL.getTypeSizeInBits(VecTy->getScalarType());
+    // HLSL Change: Use alloc size instead of primitive type size for vector
+    // elements. DXC's data layout pads min precision types (i16:32, f16:32),
+    // so getTypeAllocSize matches the GEP offset stride while
+    // getTypeSizeInBits returns the unpadded primitive width.
+    unsigned ElementSizeInBits = DL.getTypeAllocSizeInBits(VecTy->getScalarType());
     if (ElementSizeInBits % 8 != 0) {
       // GEPs over non-multiple of 8 size vector elements are invalid.
       return nullptr;
@@ -2134,7 +2138,8 @@ static VectorType *isVectorPromotionViable(AllocaSlices::Partition &P,
 
   // Try each vector type, and return the one which works.
   auto CheckVectorTypeForPromotion = [&](VectorType *VTy) {
-    uint64_t ElementSize = DL.getTypeSizeInBits(VTy->getElementType());
+    // HLSL Change: Use alloc size to match GEP offset stride for padded types.
+    uint64_t ElementSize = DL.getTypeAllocSizeInBits(VTy->getElementType());
 
     // While the definition of LLVM vectors is bitpacked, we don't support sizes
     // that aren't byte sized.
@@ -2492,12 +2497,13 @@ class AllocaSliceRewriter : public InstVisitor<AllocaSliceRewriter, bool> {
                   : nullptr),
         VecTy(PromotableVecTy),
         ElementTy(VecTy ? VecTy->getElementType() : nullptr),
-        ElementSize(VecTy ? DL.getTypeSizeInBits(ElementTy) / 8 : 0),
+        // HLSL Change: Use alloc size to match GEP offset stride for padded types.
+        ElementSize(VecTy ? DL.getTypeAllocSizeInBits(ElementTy) / 8 : 0),
         BeginOffset(), EndOffset(), IsSplittable(), IsSplit(), OldUse(),
         OldPtr(), PHIUsers(PHIUsers), SelectUsers(SelectUsers),
         IRB(NewAI.getContext(), ConstantFolder()) {
     if (VecTy) {
-      assert((DL.getTypeSizeInBits(ElementTy) % 8) == 0 &&
+      assert((DL.getTypeAllocSizeInBits(ElementTy) % 8) == 0 &&
              "Only multiple-of-8 sized vector elements are viable");
       ++NumVectorized;
     }