The GNUSyntax uses . to denote the address of the next instruction, but that doesn't match GNU as' definition. This makes it impossible to round trip the disassembly (GNU as will not be able to assemble je .-0x80 into two-byte short jump.)

package main
import "golang.org/x/arch/x86/x86asm"
func main() {
	text := []byte{0x74, 0x80, 0xeb, 0x80, 0x0f, 0x87, 0x80, 0xff, 0xff, 0xff, 0xe9, 0x80, 0xff, 0xff, 0xff}
	for len(text) > 0 {
		inst, err := x86asm.Decode(text[:], 64)
		if err != nil {
			panic(err)
		}
		println("GNU:", x86asm.GNUSyntax(inst))
		println("Go :", x86asm.GoSyntax(inst, 0, nil))
		text = text[inst.Len:]
	}
}

Output:

GNU: je .-0x80
Go : JE .-128
GNU: jmp .-0x80
Go : JMP .-128
GNU: ja .-0x80
Go : JA .-128
GNU: jmpq .-0x80
Go : JMP .-128

However, in AT&T syntax, . means the address of the current instruction, whereas the encoding actually means (conditional) jumps to the address of next instruction - 0x80.

I think to better test the disassembler, we need to do round-trip test (disassemble -> use GNU assembler to assemble the instruction -> disassemble again and verify the two Insts are deep equal.)

I want to fix this issue, but the current behavior is locked down the the current testdata/decode.txt.
Should I change the data?

/cc @rsc