We've had a couple issues with math/big assembly code routines (e.g., #31084, #42838), so it's probably worth revisiting this issue and seeing if we can get some basic fuzzing tests implemented. There's a rather minimal set of low-level arithmetic routines to worry about (https://github.com/golang/go/blob/master/src/math/big/arith_decl.go), and they all have known-good reference implementations that can be easily compared against.
I think the tricky part here is getting full coverage of the assembly routines. Can the fuzzer fuzz assembly? Getting full coverage on the reference Go implementation is a good start, but might miss some paths through the assembly.
I was thinking as a start just randomly/exhaustively generating inputs (e.g., random word values, but probably exhaustive shift values from 0 through 63), feeding them through both the reference and assembly implementations, and making sure they produce identical outputs.
Edit: I'm suggesting we just write some manual differential fuzzers as regular Go tests. Using go-fuzz or something would be nice, but I don't think it's necessary.