[flang-commits] [flang] [flang][runtime] Address PRODUCT numeric discrepancy, folding vs runtime (PR #90125)

[Peter Klausler via flang-commits]

================
@@ -188,7 +188,8 @@ void RTDEF(CppSumComplex16)(CppTypeFor<TypeCategory::Complex, 16> &result,
 void RTDEF(SumDim)(Descriptor &result, const Descriptor &x, int dim,
     const char *source, int line, const Descriptor *mask) {
   TypedPartialNumericReduction<IntegerSumAccumulator, RealSumAccumulator,
-      ComplexSumAccumulator>(result, x, dim, source, line, mask, "SUM");
+      ComplexSumAccumulator, /*MIN_REAL_KIND=*/4>(
----------------
klausler wrote:

They should both use 32-bit floats with Kahan summation.  I'll double-check them and confirm.

https://github.com/llvm/llvm-project/pull/90125