[LoadStoreOpToLLVM] Reuse 2D block IO load lowering for both regular pointer and block pointer to clean up duplicate code.

[chengjunlu]

Use the common code to lower the load to 2D block IO with linear layout utils.

1. It supports more layouts other than DPAS for block pointer.
2. It supports load transposed matrix for block pointer for general case.

[Copilot]

Pull Request Overview

This PR refactors the 2D block I/O load lowering code to consolidate duplicate logic between regular pointer and block pointer handling. The refactoring creates shared utilities for unpacking block pointer structures and computing memory access parameters, enabling both pointer types to use the same code path.

Key Changes:

• Introduces helper functions (unpackLLBlockPointer, getBases, getPitch, getBaseOffsets) to extract and process pointer metadata uniformly
• Removes the specialized rewriteTensorPointerLoad function (~800 lines) in favor of the unified lowering path
• Extends support to more layout types and transposed matrix loads for block pointers

---

💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.

[Copilot]

Duplicate assignment to baseWidth where the first assignment on line 1444 is immediately overwritten by line 1445-1446. The first assignment should be removed as it has no effect.

Suggested change

baseWidth = b.i32_val(std::max<unsigned>(64u, vBlocks * tileWidth));

[Copilot]

[nitpick] The multi-line comment explaining the offsetX calculation should be placed above the argument rather than inline to improve code readability and follow common commenting practices.

Suggested change

	/*x*/ b.udiv(offsetX, b.i32_val(numPackedVals)), // The offsetX is the
	// number of original
	// elements. The 2d
	// block io requires
	// the offsetX is the
	// number of packed
	// elements.
	// The offsetX is the number of original elements. The 2d
	// block io requires the offsetX is the number of packed
	// elements.
	/*x*/ b.udiv(offsetX, b.i32_val(numPackedVals)),

[Copilot]

[nitpick] Inconsistent formatting: this multi-line function call should align with the assignment pattern used elsewhere in the file (e.g., line 1605-1606) for consistency.

Suggested change

	newVal = targetInfo.shuffleIdx(
	rewriter, loc, oldVal,
	b.urem(threadId, b.i32_val(tileWidth)));
	Value shuffleIdx = b.urem(threadId, b.i32_val(tileWidth));
	newVal = targetInfo.shuffleIdx(
	rewriter, loc, oldVal, shuffleIdx);