Thank you for the review. I verified the behaviour on the current GCC master for a pre-r6 MSA target. -ffp-contract=off produces separate fmul and fadd instructions, while -ffp-contract=fast allows fmadd, as expected. The test was therefore based on an invalid expectation, and the proposed restriction of the standard FMA pattern is not correct. We will drop this patch from the v4 series. The v4 series is available here:
https://sourceware.org/pipermail/gcc-patches/2026-July/724124.html Thanks again for the review, Eldar.
