[Mlir-commits] [mlir] [mlir][tosa] Avoid introducing int <-> float casts (PR #195882)

[Ian Tayler Lessa]

================
@@ -941,6 +941,13 @@ struct NonNarrowingCastsOptimization : public OpRewritePattern<tosa::CastOp> {
                   "legal in TOSA");
     }
 
+    if (innerInputType.getElementType().isInteger() !=
+        outerOutputType.getElementType().isInteger()) {
+      return rewriter.notifyMatchFailure(
+          castOp, "integer to float and float to integer casts are not "
+                  "supported to avoid introducing illegal type combinations");
+    }
----------------
IanTaylerLessa-arm wrote:

There's also `i64` which doesn't support floating point casts at all (directly).

I'll rewrite it so that we allow the legal cases, though. It might look at bit convoluted, though.

I also need to reject optimization of i8 -> i32 -> i64 (long-standing bug) and f16 -> f32 -> bf16 (new-ish bug)

https://github.com/llvm/llvm-project/pull/195882