[Mlir-commits] [mlir] [mlir][arith] Add rounding mode flags to binary arithmetic operations (PR #188458)

[Renato Golin]

================
@@ -1121,9 +1126,14 @@ OpFoldResult arith::SubFOp::fold(FoldAdaptor adaptor) {
   if (matchPattern(adaptor.getRhs(), m_PosZeroFloat()))
     return getLhs();
 
+  auto rm = getRoundingmodeAttr();
   return constFoldBinaryOp<FloatAttr>(
-      adaptor.getOperands(),
-      [](const APFloat &a, const APFloat &b) { return a - b; });
+      adaptor.getOperands(), [rm](const APFloat &a, const APFloat &b) {
+        APFloat result(a);
+        result.subtract(b, rm ? convertArithRoundingModeToLLVMIR(rm.getValue())
----------------
rengolin wrote:

could we change `convertArithRoundingModeToLLVMIR` to return the default value when rm is null? Or is that bypassing too many layers of abstraction?

https://github.com/llvm/llvm-project/pull/188458