gasm-sdk/disasm/naming_amd64.go

// Copyright (c) 2026 Petr Balvín <opensource@petrbalvin.org> (https://petrbalvin.org)
// SPDX-License-Identifier: BSD-3-Clause

package disasm

import "fmt"

// golang.org/x/arch decodes a handful of amd64 opcode families to no error
// and the degenerate zero instruction: no opcode, no operands, one byte.
// GoSyntax then renders the placeholder "Op(0)", decorated with whatever
// prefixes it saw, and reports a length of one.  The families are the ADCX
// and ADOX carry-propagation pair, RDSEED, RDPID, the WAITPKG trio TPAUSE,
// UMONITOR and UMWAIT, and the ENDBR pair.
//
// nameAMD64Degenerate restores the names the Go toolchain itself carries
// for these families.  The toolchain's assembler test corpus
// machine-checks the encodings (amd64enc.s's "ADCXL DX, DX // 660f38f6d2"
// and its kin), and the spellings below are the corpus's own; the parity
// fixtures pin every corpus row the table names, so a decoder bump that
// changes a length or a register reading fails there rather than silently
// moving the text.  The table is keyed by the opcode pattern: legacy
// prefixes, opcode bytes and the ModR/M shape, so it names every encoding
// of a family, not only the corpus rows.

// amd64GPRNames holds the Plan 9 spelling of the general-purpose registers
// by index.  The names are width-independent: the mnemonic's suffix carries
// the size, so RDX reads DX in both ADCXL and ADCXQ.
var amd64GPRNames = [16]string{
	"AX", "CX", "DX", "BX", "SP", "BP", "SI", "DI",
	"R8", "R9", "R10", "R11", "R12", "R13", "R14", "R15",
}

// amd64Prefixes is the legacy prefix reading of one instruction: the
// operand-size override, the rep and repne not-really-prefixes, and REX.
// n is the number of bytes the prefixes consumed.
type amd64Prefixes struct {
	osz                    bool // 66
	rep                    bool // f3
	repne                  bool // f2
	rexW, rexR, rexX, rexB bool
	n                      int
}

// scanAMD64Prefixes reads the legacy prefixes at the start of code.  It
// stops at the first byte that is not one, which is where the opcode
// begins.
func scanAMD64Prefixes(code []byte) (amd64Prefixes, bool) {
	var p amd64Prefixes
	for ; p.n < len(code); p.n++ {
		switch b := code[p.n]; {
		case b == 0x66:
			p.osz = true
		case b == 0xf3:
			p.rep = true
		case b == 0xf2:
			p.repne = true
		case b >= 0x40 && b <= 0x4f:
			p.rexW = b&0x8 != 0
			p.rexR = b&0x4 != 0
			p.rexX = b&0x2 != 0
			p.rexB = b&0x1 != 0
		default:
			return p, true
		}
	}
	return p, false // prefixes with no opcode behind them
}

// amd64RM is the ModR/M (and SIB) reading of one operand, in the shape the
// named families use: one register or one memory reference, never an
// immediate.  n counts the ModR/M byte, an SIB byte and the displacement.
type amd64RM struct {
	regForm bool // mod == 11: the operand is the r/m register
	reg     int  // reg field with REX.R applied
	rm      int  // r/m field with REX.B applied, register form
	base    int  // memory base register, -1 when absent
	index   int  // memory index register, -1 when absent
	scale   int
	disp    int32
	rip     bool // mod == 00, r/m == 101: RIP-relative
	n       int
}

// decodeAMD64RM reads the ModR/M byte at the start of code, with the SIB
// byte and displacement that mod 00 and 01 may carry behind it.  rexR,
// rexX and rexB extend the register fields to R8 through R15.
func decodeAMD64RM(code []byte, rexR, rexX, rexB bool) (amd64RM, bool) {
	if len(code) == 0 {
		return amd64RM{}, false
	}
	b := code[0]
	var rm amd64RM
	rm.n = 1
	rm.reg = int(b>>3) & 7
	if rexR {
		rm.reg += 8
	}
	mod := b >> 6
	rmb := int(b & 7)
	if mod == 3 {
		rm.regForm = true
		rm.rm = rmb
		if rexB {
			rm.rm += 8
		}
		return rm, true
	}
	rm.base, rm.index, rm.scale = -1, -1, 1
	switch rmb {
	case 4: // SIB byte follows
		if len(code) < rm.n+1 {
			return amd64RM{}, false
		}
		sib := code[rm.n]
		rm.n++
		rm.scale = 1 << (sib >> 6)
		idx := int(sib>>3) & 7
		if rexX {
			idx += 8
		}
		if idx%8 != 4 { // index 100 is the no-index encoding
			rm.index = idx
		}
		bs := int(sib & 7)
		if rexB {
			bs += 8
		}
		if mod == 0 && bs%8 == 5 {
			// base 101 with no displacement byte is disp32 alone
		} else {
			rm.base = bs
		}
	case 5:
		if mod == 0 {
			rm.rip = true
		} else {
			rm.base = rmb
			if rexB {
				rm.base += 8
			}
		}
	default:
		rm.base = rmb
		if rexB {
			rm.base += 8
		}
	}
	switch mod {
	case 1:
		if len(code) < rm.n+1 {
			return amd64RM{}, false
		}
		rm.disp = int32(int8(code[rm.n]))
		rm.n++
	case 2:
		if len(code) < rm.n+4 {
			return amd64RM{}, false
		}
		rm.disp = int32(uint32(code[rm.n]) | uint32(code[rm.n+1])<<8 |
			uint32(code[rm.n+2])<<16 | uint32(code[rm.n+3])<<24)
		rm.n += 4
	}
	return rm, true
}

// text renders the operand the way the GoSyntax renderer prints a memory
// reference: the displacement in hex (a zero displacement printed as 0),
// then the base, then the scaled index.
func (rm amd64RM) text() string {
	if rm.regForm {
		return amd64GPRNames[rm.rm]
	}
	s := "0"
	if rm.disp != 0 {
		s = fmt.Sprintf("%#x", rm.disp)
	}
	if rm.rip {
		// The renderer names the instruction pointer IP in a memory
		// reference.
		return s + "(IP)"
	}
	if rm.base >= 0 {
		s += "(" + amd64GPRNames[rm.base] + ")"
	}
	if rm.index >= 0 {
		s += fmt.Sprintf("(%s*%d)", amd64GPRNames[rm.index], rm.scale)
	}
	return s
}

// nameAMD64Degenerate names an amd64 encoding the decoder returned as the
// degenerate zero instruction.  It reports the rendered text, the
// instruction's length in bytes and whether the bytes matched a family.
// Unmatched bytes keep the renderer's own placeholder output.
func nameAMD64Degenerate(code []byte) (string, int, bool) {
	p, ok := scanAMD64Prefixes(code)
	if !ok || len(code) < p.n+3 || code[p.n] != 0x0f {
		return "", 0, false
	}
	tail := code[p.n+1:]
	switch tail[0] {
	case 0x38:
		return nameAMD64Carry(p, tail[1:])
	case 0xc7:
		return nameAMD64RNG(p, tail[1:])
	case 0xae:
		return nameAMD64Wait(p, tail[1:])
	case 0x1e:
		return nameAMD64Endbr(p, tail[1:])
	}
	return "", 0, false
}

// nameAMD64Carry names the ADCX and ADOX pair, 0F 38 F6 /r.  The operand
// size override carries ADCX and rep carries ADOX; the length suffix
// follows REX.W.  The destination is the reg field and the source the
// r/m operand, printed source first.
func nameAMD64Carry(p amd64Prefixes, tail []byte) (string, int, bool) {
	var name string
	switch {
	case p.osz && !p.rep && !p.repne:
		name = "ADCX"
	case p.rep && !p.osz && !p.repne:
		name = "ADOX"
	default:
		return "", 0, false
	}
	if len(tail) < 2 || tail[0] != 0xf6 {
		return "", 0, false
	}
	rm, ok := decodeAMD64RM(tail[1:], p.rexR, p.rexX, p.rexB)
	if !ok {
		return "", 0, false
	}
	suffix := "L"
	if p.rexW {
		suffix = "Q"
	}
	text := name + suffix + " " + rm.text() + ", " + amd64GPRNames[rm.reg]
	return text, p.n + 3 + rm.n, true
}

// nameAMD64RNG names RDSEED, 0F C7 /7, and its rep-prefixed sibling
// RDPID.  Both take the destination register in the r/m field and exist
// only in the register form; the memory forms of the same opcode are
// CLFLUSH and its descendants, which the decoder names itself.  The size
// suffix follows REX.W and the operand-size override.
func nameAMD64RNG(p amd64Prefixes, tail []byte) (string, int, bool) {
	var name string
	switch {
	case p.rep && !p.osz && !p.repne:
		name = "RDPID"
	case !p.rep && !p.repne:
		name = "RDSEED"
	default:
		return "", 0, false
	}
	if len(tail) < 1 {
		return "", 0, false
	}
	rm, ok := decodeAMD64RM(tail, p.rexR, p.rexX, p.rexB)
	if !ok || !rm.regForm || rm.reg != 7 {
		return "", 0, false
	}
	if name == "RDPID" {
		return name + " " + amd64GPRNames[rm.rm], p.n + 2 + rm.n, true
	}
	suffix := "L"
	switch {
	case p.rexW:
		suffix = "Q"
	case p.osz:
		suffix = "W"
	}
	return name + suffix + " " + amd64GPRNames[rm.rm], p.n + 2 + rm.n, true
}

// nameAMD64Wait names the WAITPKG and monitor trio on 0F AE /6: TPAUSE
// under the operand-size override, UMONITOR under rep and UMWAIT under
// repne.  Each takes one 32-bit register in the r/m field.
func nameAMD64Wait(p amd64Prefixes, tail []byte) (string, int, bool) {
	var name string
	switch {
	case p.osz && !p.rep && !p.repne:
		name = "TPAUSE"
	case p.rep && !p.osz && !p.repne:
		name = "UMONITOR"
	case p.repne && !p.osz && !p.rep:
		name = "UMWAIT"
	default:
		return "", 0, false
	}
	if len(tail) < 1 {
		return "", 0, false
	}
	rm, ok := decodeAMD64RM(tail, p.rexR, p.rexX, p.rexB)
	if !ok || !rm.regForm || rm.reg != 6 {
		return "", 0, false
	}
	return name + " " + amd64GPRNames[rm.rm], p.n + 2 + rm.n, true
}

// nameAMD64Endbr names the ENDBR pair, F3 0F 1E with the fixed ModR/M
// bytes FA for the 64-bit variant and FB for the 32-bit one.  The
// instructions take no operands.
func nameAMD64Endbr(p amd64Prefixes, tail []byte) (string, int, bool) {
	if !p.rep || p.osz || p.repne {
		return "", 0, false
	}
	if len(tail) < 1 {
		return "", 0, false
	}
	switch tail[0] {
	case 0xfa:
		return "ENDBR64", p.n + 3, true
	case 0xfb:
		return "ENDBR32", p.n + 3, true
	}
	return "", 0, false
}