[Mlir-commits] [mlir] [MLIR] Preserve operand segment sizes when removing dead inputs (PR #223154)

[Ivan Butygin]

================
@@ -244,6 +244,37 @@ void RewriterBase::eraseBlock(Block *block) {
   block->erase();
 }
 
+void RewriterBase::eraseOperands(Operation *op, const BitVector &eraseIndices) {
+  assert(eraseIndices.size() == op->getNumOperands() &&
+         "expected one bit per operand");
+  if (eraseIndices.none())
+    return;
+
+  modifyOpInPlace(op, [&]() {
+    op->eraseOperands(eraseIndices);
+    if (!op->hasTrait<OpTrait::AttrSizedOperandSegments>())
+      return;
+
+    // TODO: Add an interface to update the operand segment-size property
+    // directly, without converting it to and from an attribute.
+    auto attrName = StringAttr::get(
----------------
Hardcode84 wrote:

`StringRef` will be converted into `StringAttr` internally on each call, and we will end up with 2 `StringAttr` constructions instead of 1. As a side comment, this is a common antipattern and I'd rather move into nuking `StringRef` overloads entirely.

https://github.com/llvm/llvm-project/pull/223154