[Bug tree-optimization/124802] New: [15/16 Regression] Wrong code at -O2 with AVX2 complex double fmaddsub pattern

[Sarvesh.Chandra at amd dot com via Gcc-bugs]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=124802

            Bug ID: 124802
           Summary: [15/16 Regression] Wrong code at -O2 with AVX2 complex
                    double fmaddsub pattern
           Product: gcc
           Version: 15.2.0
            Status: UNCONFIRMED
          Keywords: wrong-code
          Severity: normal
          Priority: P3
         Component: tree-optimization
          Assignee: unassigned at gcc dot gnu.org
          Reporter: Sarvesh.Chandra at amd dot com
                CC: vekumar at gcc dot gnu.org
  Target Milestone: ---
              Host: x86_64-linux-gnu
            Target: x86_64-linux-gnu

GCC 15/16 generates wrong code at -O2/-O3 for a complex double
multiply-accumulate kernel using _mm256_fmaddsub_pd.  The attached
testcase aborts at runtime.

Compiles and runs correctly with GCC 14.2.0.
Adding -fno-tree-slp-vectorize works around the bug.

Bisected to r15-2816-gab18785840d "Rearrange SLP nodes with
duplicate statements [PR98138]".  The parent r15-2815-g71aebb36174
produces correct results.

$ g++ -O2 -mavx2 -mfma pr98138-wrongcode.c -o test && ./test
Aborted

Tested on x86_64-linux-gnu (AMD Zen 4).