[Mlir-commits] [mlir] [mlir][Interfaces] Track and infer no-overflow flags in integer ranges (PR #191777)

[Mehdi Amini]

================
@@ -65,8 +176,11 @@ void arith::ConstantOp::inferResultRanges(ArrayRef<ConstantIntRanges> argRanges,
 
 void arith::AddIOp::inferResultRanges(ArrayRef<ConstantIntRanges> argRanges,
                                       SetIntRangeFn setResultRange) {
-  setResultRange(getResult(), inferAdd(argRanges, convertArithOverflowFlags(
-                                                      getOverflowFlags())));
+  OverflowFlags declaredFlags = convertArithOverflowFlags(getOverflowFlags());
+  ConstantIntRanges range = inferAdd(argRanges, declaredFlags);
+  OverflowFlags overflowFlags =
+      proveNoOverflowForAdd(argRanges) | declaredFlags;
----------------
joker-eph wrote:

> would it be possible that declared flags differ from the inferred one?

Of course, but why does it matter when you do an or? 

Basically if the bit is 1 already, in the declared one, why would you infer it? Whether it is 1 or 0 the result is still 1.

https://github.com/llvm/llvm-project/pull/191777