[Bug] TopMoveReluAheadConcatPattern loses `relu_limit` attribute when creating new ReluOp

[henaixin]

Describe the bug

Describe the bug

In tpu-mlir/lib/Dialect/Top/Canonicalize/Relu.cpp, the conversion pattern TopMoveReluAheadConcatPattern moves the Relu operation ahead of Concat.

However, during this optimization pass, when a new ReluOp is created inside the loop, the original relu_limit attribute is lost. This causes unexpected behavior or bugs when the model relies on a specific relu_limit.

Source Code Location

• File: tpu-mlir/lib/Dialect/Top/Canonicalize/Relu.cpp
• Pattern: TopMoveReluAheadConcatPattern::matchAndRewriteImpl

At line 72, the relu_limit is fetched:

auto relu_limit = op.getReluLimit();

But at line 90, the new ReluOp is created without passing this attribute:

auto newOp = rewriter.create<ReluOp>(
    NameLoc::get(nameAttr), formerOp->getOperand(i).getType(),
    ArrayRef<Value>{formerOp->getOperand(i)});

Suggested Fix

The relu_limit should be forwarded when creating the new ReluOp. It should be updated to something like:

auto newOp = rewriter.create<ReluOp>(
    NameLoc::get(nameAttr), formerOp->getOperand(i).getType(),
    ArrayRef<Value>{formerOp->getOperand(i)},
    relu_limit); // Pass the original relu_limit here

Steps to reproduce

1. Run model conversion pipeline with a model containing a Relu layer that has a specific `relu_limit` value.
2. The `TopMoveReluAheadConcatPattern` canonicalization pass is triggered.
3. The newly created `ReluOp` loses its original `relu_limit` attribute.

Expected behavior

The TopMoveReluAheadConcatPattern optimization pass should preserve and forward the original relu_limit attribute when recreating the ReluOp ahead of Concat.

Error logs / stack trace

No specific crash log, but a logic bug in `tpu-mlir/lib/Dialect/Top/Canonicalize/Relu.cpp`:

At line 72, the `relu_limit` is fetched:
auto relu_limit = op.getReluLimit();

But at line 90, the new `ReluOp` is created without passing this attribute:
auto newOp = rewriter.create<ReluOp>(
    NameLoc::get(nameAttr), formerOp->getOperand(i).getType(),
    ArrayRef<Value>{formerOp->getOperand(i)});

TPU-MLIR version

commit 18c49f9

Target chip

bm1684x

Source framework / model

No response

OS / Docker image

No response