[Mlir-commits] [mlir] [MLIR][ARITH] Adds missing foldings for truncf (PR #128096)

[Jakub Kuderski]

================
@@ -1517,6 +1517,28 @@ LogicalResult arith::TruncIOp::verify() {
 /// Perform safe const propagation for truncf, i.e., only propagate if FP value
 /// can be represented without precision loss.
 OpFoldResult arith::TruncFOp::fold(FoldAdaptor adaptor) {
+  if (auto extOp = getOperand().getDefiningOp<arith::ExtFOp>()) {
+    Value src = extOp.getIn();
+    Type srcType = getElementTypeOrSelf(src.getType());
+    Type dstType = getElementTypeOrSelf(getType());
+    // truncf(extf(a)) -> truncf(a)
+    if (llvm::cast<FloatType>(srcType).getWidth() >
+        llvm::cast<FloatType>(dstType).getWidth()) {
+      setOperand(src);
+      return getResult();
+    }
+
+    // truncf(extf(a)) -> a
+    if (srcType == dstType)
+      return src;
+  }
+
+  // truncf(truncf(a)) -> truncf(a)
----------------
kuhar wrote:

I meant all three fold, not just the last one. I'd think this is tricky to get right in the presence of weird small fp types, and why I suggested looking at `APFloatBase::isRepresentableBy` and types without NaNs/Signed zero/Inf.

https://github.com/llvm/llvm-project/pull/128096