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

[Luke Hutton]

================
@@ -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");
+    }
----------------
lhutton1 wrote:

This seems to block optimization of legitimate cases such as:
```
i8 -> cast -> i16 -> cast -> fp16
```
Looking at the spec, integer inputs have full coverage for casts to floating point types. It's support for bool  that was missing in 1.0. I wonder if we can relax this check and apply the above logic only when the input type is bool?

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