AVX-512 FMA intrinsics with a non-immediate embedded-rounding operand hit Assertion failed '(op3Reg != targetReg) || (op1Reg == targetReg)' in genHWIntrinsic_R_R_R_RM under jitstress2_jitstressregs0x80 (recurrence of #128544). The FMA case in genNonTableDrivenHWIntrinsicsJumpTableFallback emitted the 213 form directly, skipping the operand swapping / target-register preferencing that genFmaIntrinsic performs. When LSRA preferenced op2/op3 onto targetReg, the leading mov targetReg, op1Reg clobbered a live source — exactly the alias the assert guards against.

Changes (src/coreclr/jit/hwintrinsiccodegenxarch.cpp)

• Make genFmaIntrinsic re-entrant from a jump-table lambda. Lifted genConsumeMultiOpOperands / genProduceReg out of genFmaIntrinsic into its caller in genAvxFamilyIntrinsic. The remaining swap+emit is deterministic on register numbers and emits a single instruction per call, so it is safe to invoke once per rounding-mode case.
• Route the FMA fallback through genFmaIntrinsic. Replaced the manual genHWIntrinsic_R_R_R_RM(ins, attr, targetReg, op1Reg, op2Reg, op3, …) in the NI_AVX512_FusedMultiply* arm of genNonTableDrivenHWIntrinsicsJumpTableFallback with genFmaIntrinsic(node, newInstOptions). The outer embedded-rounding path already brackets this fallback with consume/produce, so no extra calls are needed inside the lambda. Asserts that op1/op2/op3 are non-contained are kept as defensive checks against the embedded-rounding R-R-R invariant.

auto emitSwCase = [&](int8_t i) {
    insOpts newInstOptions = AddEmbRoundingMode(instOptions, i);
    genFmaIntrinsic(node, newInstOptions);  // was: genHWIntrinsic_R_R_R_RM(ins, attr, targetReg, op1Reg, op2Reg, op3, newInstOptions)
};

The 132/213/231-form selection in genFmaIntrinsic's no-contained branch now applies on the non-immediate-rounding path as well, eliminating the source-register clobber.