[Mlir-commits] [mlir] Rewrites for I2 to I8 signed and unsigned extension (PR #121298)

[Andrzej Warzyński]

================
@@ -1233,6 +1233,117 @@ static Value rewriteI4ToI8UnsignedExt(PatternRewriter &rewriter, Location loc,
   return rewriter.create<vector::InterleaveOp>(loc, low, high);
 }
 
+/// Rewrite the i2 -> i8 signed extension into a sequence of shuffles and
+/// bitwise ops that take advantage of high-level information to avoid leaving
+/// LLVM to scramble with peephole optimizations.
----------------
banach-space wrote:

> take advantage of high-level information

There's not much "high-level" info in arith.extsi ;-)

>  avoid leaving  LLVM to scramble with peephole optimizations

IMO, this is a perfect opportunity for the LLVM backend to shine. Once LLVM's optimizer can handle this, other frontends (i.e., not just MLIR + Vector) will benefit as well.

Just to clarify, I'm not against having these patterns here - quite the opposite! I'm very thankful that you're adding them. To me, this is more about working around LLVM's current limitations and filling the gap while LLVM catches up.

https://github.com/llvm/llvm-project/pull/121298