[flang-commits] [flang] [llvm] [flang] Round COMPLEX(4)**INTEGER to single precision once (PR #230257)

[Chris Frisz via flang-commits]

cfrisz-nvidia wrote:

A note with respect to the Fujitsu compiler-test-suite: this PR changes two results there. Test 0668_0112 prints `(1+2.11i)**10`, and 0676_0103 checks the sum of `(1.4+2.4i)**i` for `i = 0..10` against constants with a 1-ULP tolerance for the real part and 5-ULP for the imaginary part. Both sets of reference values are less accurate than what this PR computes. For 0668_0112 at `i = 10`, the reference real part is 6 ULPs from the exact value, while the value computed by this PR is correctly rounded. For 0676_0103, the reference constants are 1.6 ULPs (real part) and 3.4 ULPs (imaginary) from the exact sum, while the result computed by this PR is within 1 ULP.

https://github.com/llvm/llvm-project/pull/230257