[Mlir-commits] [mlir] [MLIR][LLVM][Mem2Reg] Extends support for partial stores (PR #89740)

[Tobias Gysi]

================
@@ -224,20 +245,67 @@ static Value createConversionSequence(RewriterBase &rewriter, Location loc,
   }
 
   // Now cast the integer to the actual target type if required.
-  if (isa<LLVM::LLVMPointerType>(targetType))
-    replacement =
-        rewriter.createOrFold<LLVM::IntToPtrOp>(loc, targetType, replacement);
-  else if (replacement.getType() != targetType)
-    replacement =
-        rewriter.createOrFold<LLVM::BitcastOp>(loc, targetType, replacement);
-
-  return replacement;
+  return convertIntValueToType(rewriter, loc, replacement, targetType);
 }
 
 Value LLVM::StoreOp::getStored(const MemorySlot &slot, RewriterBase &rewriter,
+                               Value reachingDef,
                                const DataLayout &dataLayout) {
-  return createConversionSequence(rewriter, getLoc(), getValue(), slot.elemType,
-                                  dataLayout);
+  uint64_t valueTypeSize = dataLayout.getTypeSizeInBits(getValue().getType());
+  uint64_t slotTypeSize = dataLayout.getTypeSizeInBits(slot.elemType);
+  if (slotTypeSize <= valueTypeSize)
+    return createConversionSequence(rewriter, getLoc(), getValue(),
----------------
gysit wrote:

Would it make sense to call this function `createExtractAndCast` while the code below could be in a `createInsertAndCast` or similar.

Does the < case here make sense. If the slot is smaller than the stored value, it sounds like we drop some information?

https://github.com/llvm/llvm-project/pull/89740