[Mlir-commits] [mlir] Fix TOSA FP16->INT16 CAST lowering (PR #79299)

[Jakub Kuderski]

================
@@ -480,23 +480,53 @@ createLinalgBodyCalculationForElementwiseOp(Operation *op, ValueRange args,
     }
 
     if (arith::FPToSIOp::areCastCompatible(srcTy, dstTy)) {
-      auto intMin = rewriter.create<arith::ConstantOp>(
+      auto intMinFP = rewriter.create<arith::ConstantOp>(
           loc, rewriter.getFloatAttr(
                    getElementTypeOrSelf(srcTy),
                    APInt::getSignedMinValue(dstTy.getIntOrFloatBitWidth())
                        .getSExtValue()));
 
-      auto intMax = rewriter.create<arith::ConstantOp>(
+      auto rounded = rewriter.create<math::RoundEvenOp>(loc, args[0]);
+
+      // The input floating-point type has enough mantissa bits to represent
+      // the max int value so just clamp the input in the floating-point
+      // domain and convert to int. Note: the min value can be represented
+      // because it consists of a mantissa with only the lsb set.
+      if (cast<FloatType>(srcTy).getFPMantissaWidth() >=
+          dstTy.getIntOrFloatBitWidth() - 1) {
----------------
kuhar wrote:

Can you expand on the INT_MIN case? I'd think that the minimum 16-bit int would be -2^15 and if the FP format has a dedicated sign bit (and in turn +0 and -0), the mantissa needs to have 16 bits as well to represent 2^15 (i.e., sign bit set, msb mantissa bit set).

I haven't thought about this too much so maybe I'm missing something, but shouldn't the condition be?
```
if (MantissaWidth() >= IntBitWidth())
```

https://github.com/llvm/llvm-project/pull/79299