In scalar_4x64_impl.h, function secp256k1_scalar_reduce_512. The first asm block is missing early clobber for m0~m2.

The input %rsi is used after writing to m0, m1 and m2. Any of those write can thus re-use %rsi (which will most likely result in segfault).

Also, field_5x52_asm_impl.h is missing similar early clobbers for tmp1~tmp3