Files
Odin/core/rexcode/isa/arm64/decoder.odin
Brendan Punsky d40d9e687a rexcode/arm64: model SME's ZA tiles and tile slices
A ZA tile printed as the bare number it is encoded as -- `addha #0,
p0/m, p0/m, z0.s` -- and a tile slice printed as one too, where the
syntax is `za0h.b[w12, 0]`: a tile, taken along its rows or columns,
addressed by one of W12..W15 plus an offset. Neither was anything an
assembler would take.

Tiles are a register class now (ZA0..ZA15, viewed at an element size),
so they print through the same path as every other register. A slice is
its own operand kind holding the four things it is made of, rather than
one packed immediate that only the encoder understood.

Getting that right needed the field layout, and the layout is not what
the encoding table implied: the tile number and the offset share the low
nibble, and how it splits follows the element size -- a byte tile has no
tile bits at all and four of offset, while a quadword tile is all tile
and none. Reading a fixed four bits as the tile made every byte slice
come back as tile 4.

LD1Q/ST1Q scale their index by 16, which the mnemonic's last letter
does not spell the way B/H/W/D do; they were left unscaled when the
other 38 forms were fixed. That also wanted a .q element shape, which
nothing had needed before.

SVE/SME2 decode entries against llvm-mc: 559 byte-exact and 0
mismatched, from 211 and 78 at the start of the session.

What is left is mostly one structural limit: SME's outer products take
five operands (`umopa za0.s, p0/m, p1/m, z0.b, z1.b`) and Instruction
holds four. Widening it would fit -- five Operands is 55 bytes of the 64
-- but it reaches through Encoding, the table, both codecs and the
builders, so it wants its own change.

Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_018UmHLRF11EoWwNWCJ7JGaA
2026-08-28 16:58:33 -04:00

853 lines
32 KiB
Odin

// rexcode · Brendan Punsky (dotbmp@github), original author
package rexcode_arm64
import "core:rexcode/isa"
// =============================================================================
// AArch64 DECODER
// =============================================================================
//
// Two passes, mirroring riscv/decoder.odin. Specifics:
//
// * Single-level dispatch by op0 (bits[28:25], 4 bits = 16 slots);
// linear scan within each bucket. Entries are sorted by mask-
// popcount descending so the most-specific encoding form wins.
//
// * SP-vs-ZR reconstruction is contextual: the decoder reads hw 0-31
// and emits an X / W register; if the form expects WSP_REG/XSP_REG
// it emits a REG_WSP/REG_XSP at hw 31 instead of ZR.
//
// * .RM extraction is form-dependent: SHIFTED_REG and EXTENDED_REG
// operand types pull both the register hw and the shift/extend bits.
Instruction_Info :: struct {
offset: u32,
decode_entry: u16,
_: u16,
}
#assert(size_of(Instruction_Info) == 8)
decode :: proc(
data: []u8,
relocs: []Relocation,
instructions: ^[dynamic]Instruction,
inst_info: ^[dynamic]Instruction_Info,
label_defs: ^[dynamic]Label_Definition,
errors: ^[dynamic]Error,
endianness: Endianness = .LITTLE,
) -> (byte_count: u32, ok: bool) {
n_bytes := u32(len(data)) & ~u32(3)
errors_start := u32(len(errors))
pending_branches: [dynamic]isa.Branch_Target
defer delete(pending_branches)
for byte_count < n_bytes {
word := read_u32(data, byte_count, endianness)
inst: Instruction
info: Instruction_Info
entry_idx := decode_one_inline(word, byte_count, &inst, &info)
if entry_idx < 0 {
append(errors, Error{inst_idx = byte_count, code = .INVALID_OPCODE})
inst = Instruction{mnemonic = .INVALID, length = 4}
info = Instruction_Info{offset = byte_count}
} else {
inst_idx_for_branches := u32(len(instructions))
for slot in 0..<inst.operand_count {
op := &inst.ops[slot]
if op.kind == .RELATIVE && op.relative >= 0 {
append(&pending_branches, isa.Branch_Target{
inst_idx = inst_idx_for_branches,
op_idx = slot,
target = u32(op.relative),
})
}
}
}
append(instructions, inst)
append(inst_info, info)
byte_count += 4
}
isa.infer_labels_from_branches(pending_branches[:], byte_count, label_defs, relocs)
ok = u32(len(errors)) == errors_start
return
}
// =============================================================================
// Internal
// =============================================================================
@(private="file")
decode_one_inline :: #force_inline proc "contextless" (
word: u32, pc: u32, inst: ^Instruction, info: ^Instruction_Info,
) -> int {
op0 := (word >> 25) & 0xF
range := DECODE_INDEX_OP0[op0]
if range.count == 0 { return -1 }
base := int(range.start)
cnt := int(range.count)
matched_idx := -1
for i in 0..<cnt {
e := &DECODE_ENTRIES[base + i]
if (word & e.mask) == e.bits {
// RN_RM is one operand filling both source slots, which is what
// makes cinc/cinv/cneg an alias at all -- the encoding is only
// theirs when the two register fields actually agree. No mask can
// say that, so it is checked here; the branch costs nothing,
// since it is only reached on a match.
if e.enc[1] == .RN_RM && ((word >> 16) & 0x1F) != ((word >> 5) & 0x1F) {
continue
}
matched_idx = base + i
break
}
}
if matched_idx < 0 { return -1 }
entry := &DECODE_ENTRIES[matched_idx]
inst.mnemonic = entry.mnemonic
inst.length = 4
inst.flags = {}
cnt_used: u8 = 0
if entry.ops[0] != .NONE {
inst.ops[0] = extract_operand_inline(word, pc, entry.ops[0], entry.enc[0])
cnt_used = 1
if entry.ops[1] != .NONE {
inst.ops[1] = extract_operand_inline(word, pc, entry.ops[1], entry.enc[1])
cnt_used = 2
if entry.ops[2] != .NONE {
inst.ops[2] = extract_operand_inline(word, pc, entry.ops[2], entry.enc[2])
cnt_used = 3
if entry.ops[3] != .NONE {
inst.ops[3] = extract_operand_inline(word, pc, entry.ops[3], entry.enc[3])
cnt_used = 4
}
}
}
}
inst.operand_count = cnt_used
info.offset = pc
info.decode_entry = u16(matched_idx)
return matched_idx
}
@(private="file")
extract_operand_inline :: #force_inline proc "contextless" (
word: u32, pc: u32, ot: Operand_Type, en: Operand_Encoding,
) -> Operand {
#partial switch en {
case .NONE, .IMPL:
// For IMPL on .COND_HI/etc. cases the operand stays NONE.
return {}
// ---- Register slots ----------------------------------------------------
case .RD, .RT:
return reg_from_field(word, 0, ot)
case .RN, .RN_RM:
return reg_from_field(word, 5, ot)
case .RT2, .RA:
return reg_from_field(word, 10, ot)
case .RM:
// Three flavours per operand type: plain / shifted / extended.
#partial switch ot {
case .W_SHIFTED, .X_SHIFTED:
hw := u8((word >> 16) & 0x1F)
return Operand{
shifted = Shifted_Reg{
reg = ot == .X_SHIFTED ? Register(REG_X | u16(hw)) : Register(REG_W | u16(hw)),
type = Shift_Type((word >> 22) & 0x3),
amount = u8((word >> 10) & 0x3F),
},
kind = .SHIFTED_REG, size = 4,
}
case .W_EXTENDED, .X_EXTENDED:
hw := u8((word >> 16) & 0x1F)
return Operand{
extended = Extended_Reg{
reg = ot == .X_EXTENDED ? Register(REG_X | u16(hw)) : Register(REG_W | u16(hw)),
extend = Extend((word >> 13) & 0x7),
amount = u8((word >> 10) & 0x7),
},
kind = .EXTENDED_REG, size = 4,
}
case:
return reg_from_field(word, 16, ot)
}
// ---- Immediates --------------------------------------------------------
case .IMM12: return Operand{immediate = i64((word >> 10) & 0xFFF), kind = .IMMEDIATE, size = 2}
case .IMM16: return Operand{immediate = i64((word >> 5) & 0xFFFF), kind = .IMMEDIATE, size = 2}
case .IMM6: return Operand{immediate = i64((word >> 10) & 0x3F), kind = .IMMEDIATE, size = 1}
case .IMM9:
v := i32((word >> 12) & 0x1FF)
if v & (1 << 8) != 0 { v |= ~i32(0x1FF) } // sign-extend from bit 8
return Operand{immediate = i64(v), kind = .IMMEDIATE, size = 1}
case .IMM_HW: return Operand{immediate = i64((word >> 21) & 0x3), kind = .IMMEDIATE, size = 1}
case .IMM_SH12: return Operand{immediate = i64((word >> 22) & 0x1), kind = .IMMEDIATE, size = 1}
case .SHIFT_TYPE: return Operand{immediate = i64((word >> 22) & 0x3), kind = .IMMEDIATE, size = 1}
case .EXT_OPT: return Operand{immediate = i64((word >> 13) & 0x7), kind = .IMMEDIATE, size = 1}
case .EXT_IMM3: return Operand{immediate = i64((word >> 10) & 0x7), kind = .IMMEDIATE, size = 1}
case .COND_HI:
return Operand{cond = u8((word >> 12) & 0xF), kind = .COND, size = 1}
case .COND_HI_INV:
return Operand{cond = u8(((word >> 12) & 0xF) ~ 1), kind = .COND, size = 1}
case .COND_LO:
return Operand{cond = u8(word & 0xF), kind = .COND, size = 1}
case .NZCV_FIELD:
return Operand{immediate = i64(word & 0xF), kind = .IMMEDIATE, size = 1}
case .SYS_FIELD:
return Operand{sysreg = System_Register((word >> 5) & 0x7FFF), kind = .SYSTEM_REGISTER, size = 2}
case .HINT_FIELD:
return Operand{immediate = i64((word >> 5) & 0x7F), kind = .IMMEDIATE, size = 1}
case .BARRIER_FIELD:
return Operand{immediate = i64((word >> 8) & 0xF), kind = .IMMEDIATE, size = 1}
// ---- NEON shift-by-immediate: recover the amount from immh:immb ---------
case .NEON_SHL_IMM, .NEON_SHR_IMM:
immh := (word >> 19) & 0xF
esize: i64 = 8
if immh >= 8 { esize = 64 }
else if immh >= 4 { esize = 32 }
else if immh >= 2 { esize = 16 }
val := i64((word >> 16) & 0x7F)
amt := val - esize
if en == .NEON_SHR_IMM { amt = 2 * esize - val }
return Operand{immediate = amt, kind = .IMMEDIATE, size = 1}
// ---- NEON copy/permute index fields ------------------------------------
case .VN_VM_DUP:
return Operand{reg = Register(REG_V | u16((word >> 5) & 0x1F)), kind = .REGISTER, size = reg_size_for_type(ot)}
case .NEON_IDX5:
// imm5 = index << (markerbit+1) | (1 << markerbit); marker = lowest set bit.
imm5 := (word >> 16) & 0x1F
mb: u32 = 0
if imm5 & 0x1 != 0 { mb = 0 }
else if imm5 & 0x2 != 0 { mb = 1 }
else if imm5 & 0x4 != 0 { mb = 2 }
else { mb = 3 }
return Operand{immediate = i64(imm5 >> (mb + 1)), kind = .IMMEDIATE, size = LANE_INDEX}
case .NEON_IDX4:
// imm4 = index << markerbit; recover markerbit from imm5 in the word.
imm5 := (word >> 16) & 0x1F
mb: u32 = 0
if imm5 & 0x1 != 0 { mb = 0 }
else if imm5 & 0x2 != 0 { mb = 1 }
else if imm5 & 0x4 != 0 { mb = 2 }
else { mb = 3 }
return Operand{immediate = i64(((word >> 11) & 0xF) >> mb), kind = .IMMEDIATE, size = LANE_INDEX}
case .NEON_EXT_IDX:
return Operand{immediate = i64((word >> 11) & 0xF), kind = .IMMEDIATE, size = 1}
case .IMM5_HI:
return Operand{immediate = i64((word >> 16) & 0x1F), kind = .IMMEDIATE, size = 1}
case .MSR_PSTATE:
v := ((word >> 16) & 0x7) << 3 | ((word >> 5) & 0x7)
return Operand{immediate = i64(v), kind = .IMMEDIATE, size = 1}
case .FMOV_SCALAR_IMM:
return Operand{immediate = i64((word >> 13) & 0xFF), kind = .IMMEDIATE, size = 1}
case .PG4_PM_DUP:
return Operand{reg = Register(REG_P | u16((word >> 10) & 0xF)), kind = .REGISTER, size = 4}
case .PN_PM_DUP, .PN_PG_PM_DUP:
return Operand{reg = Register(REG_P | u16((word >> 5) & 0xF)), kind = .REGISTER, size = 4}
case .ZD_ZM_DUP:
return Operand{reg = Register(REG_Z | u16(word & 0x1F)), kind = .REGISTER, size = reg_size_for_type(ot)}
case .SVE_EXT_IMM:
v := ((word >> 16) & 0x1F) << 3 | ((word >> 10) & 0x7)
return Operand{immediate = i64(v), kind = .IMMEDIATE, size = 1}
case .ZA_TILE_LOW:
return Operand{reg = Register(REG_ZA | u16(word & 0x7)), kind = .REGISTER,
size = za_elem_for_type(ot)}
case .NEON_LANE_B:
i := ((word >> 30) & 0x1) << 3 | ((word >> 12) & 0x1) << 2 | ((word >> 10) & 0x3)
return Operand{immediate = i64(i), kind = .IMMEDIATE, size = LANE_INDEX}
case .NEON_LANE_H:
i := ((word >> 30) & 0x1) << 2 | ((word >> 12) & 0x1) << 1 | ((word >> 11) & 0x1)
return Operand{immediate = i64(i), kind = .IMMEDIATE, size = LANE_INDEX}
case .NEON_LANE_S:
i := ((word >> 30) & 0x1) << 1 | ((word >> 12) & 0x1)
return Operand{immediate = i64(i), kind = .IMMEDIATE, size = LANE_INDEX}
case .NEON_LANE_D:
return Operand{immediate = i64((word >> 30) & 0x1), kind = .IMMEDIATE, size = LANE_INDEX}
case .SVE_XAR_SHIFT:
return Operand{immediate = i64(sve_tsz_shift(sve_tsz_field(word))), kind = .IMMEDIATE, size = 1}
// ---- Memory operand variants ------------------------------------------
case .OFFSET_BASE_U12:
size := u32(1) << ((word >> 30) & 0x3)
base_hw := u8((word >> 5) & 0x1F)
imm12 := u32((word >> 10) & 0xFFF)
return Operand{
mem = Memory{
base = Register(REG_X | u16(base_hw)),
index = NONE,
disp = i32(imm12 * size),
mode = .OFFSET,
},
kind = .MEMORY, size = 4,
}
case .OFFSET_BASE_S9:
base_hw := u8((word >> 5) & 0x1F)
imm9 := i32((word >> 12) & 0x1FF)
if imm9 & (1 << 8) != 0 { imm9 |= ~i32(0x1FF) }
return Operand{
mem = Memory{
base = Register(REG_X | u16(base_hw)),
index = NONE,
disp = imm9,
mode = .OFFSET,
},
kind = .MEMORY, size = 4,
}
case .OFFSET_BASE_PRE:
base_hw := u8((word >> 5) & 0x1F)
imm9 := i32((word >> 12) & 0x1FF)
if imm9 & (1 << 8) != 0 { imm9 |= ~i32(0x1FF) }
return Operand{
mem = Memory{
base = Register(REG_X | u16(base_hw)),
index = NONE,
disp = imm9,
mode = .PRE_INDEXED,
},
kind = .MEMORY, size = 4,
}
case .OFFSET_BASE_POST:
base_hw := u8((word >> 5) & 0x1F)
imm9 := i32((word >> 12) & 0x1FF)
if imm9 & (1 << 8) != 0 { imm9 |= ~i32(0x1FF) }
return Operand{
mem = Memory{
base = Register(REG_X | u16(base_hw)),
index = NONE,
disp = imm9,
mode = .POST_INDEXED,
},
kind = .MEMORY, size = 4,
}
case .OFFSET_BASE_A:
// [Xn] only: no displacement, no index.
base_hw := u8((word >> 5) & 0x1F)
return Operand{
mem = Memory{
base = Register(REG_X | u16(base_hw)),
index = NONE,
mode = .OFFSET,
},
kind = .MEMORY, size = 4,
}
case .OFFSET_PAIR_4, .OFFSET_PAIR_8, .OFFSET_PAIR_16:
// LDP/STP: signed imm7 at 21:15, scaled by the transfer size. The
// addressing mode is bits[24:23] of the word (01 post, 11 pre,
// 10 signed offset / 00 no-allocate), not part of the encoding.
base_hw := u8((word >> 5) & 0x1F)
imm7 := i32((word >> 15) & 0x7F)
if imm7 & (1 << 6) != 0 {
imm7 |= ~i32(0x7F)
}
scale := i32(4)
if en == .OFFSET_PAIR_8 {
scale = 8
} else if en == .OFFSET_PAIR_16 {
scale = 16
}
mode := Address_Mode.OFFSET
switch (word >> 23) & 0x3 {
case 0b01: mode = .POST_INDEXED
case 0b11: mode = .PRE_INDEXED
}
return Operand{
mem = Memory{
base = Register(REG_X | u16(base_hw)),
index = NONE,
disp = imm7 * scale,
mode = mode,
},
kind = .MEMORY, size = 4,
}
case .OFFSET_REG, .OFFSET_EXT:
base_hw := u8((word >> 5) & 0x1F)
idx_hw := u8((word >> 16) & 0x1F)
option := Extend((word >> 13) & 0x7)
s := u8((word >> 12) & 0x1)
idx_cls := u16(REG_X)
if option == .UXTW || option == .SXTW { idx_cls = REG_W }
return Operand{
mem = Memory{
base = Register(REG_X | u16(base_hw)),
index = Register(idx_cls | u16(idx_hw)),
extend = option,
shift = s,
mode = en == .OFFSET_EXT ? .EXT_REG_OFFSET : .REG_OFFSET,
},
kind = .MEMORY, size = 4,
}
// ---- PC-relative branches ---------------------------------------------
case .BRANCH_26:
v := i32(word & 0x03FFFFFF)
if v & (1 << 25) != 0 { v |= ~i32(0x03FFFFFF) }
target := u32(i32(pc) + (v << 2))
return Operand{relative = i64(target), kind = .RELATIVE, size = 4}
case .BRANCH_19:
v := i32((word >> 5) & 0x7FFFF)
if v & (1 << 18) != 0 { v |= ~i32(0x7FFFF) }
target := u32(i32(pc) + (v << 2))
return Operand{relative = i64(target), kind = .RELATIVE, size = 4}
case .BRANCH_14:
v := i32((word >> 5) & 0x3FFF)
if v & (1 << 13) != 0 { v |= ~i32(0x3FFF) }
target := u32(i32(pc) + (v << 2))
return Operand{relative = i64(target), kind = .RELATIVE, size = 4}
case .BRANCH_PG21:
// Sign-extended 21-bit value reassembled from immlo/immhi.
lo := (word >> 29) & 0x3
hi := (word >> 5) & 0x7FFFF
v := i32((hi << 2) | lo)
if v & (1 << 20) != 0 { v |= ~i32(0x1FFFFF) }
// For ADR (op=0 bit 31) target = PC + imm21.
// For ADRP (op=1) target = (PC & ~0xFFF) + (imm21 << 12).
if (word >> 31) & 1 != 0 {
// ADRP
target := (i64(pc) & ~i64(0xFFF)) + (i64(v) << 12)
return Operand{relative = target, kind = .RELATIVE, size = 4}
} else {
target := u32(i32(pc) + v)
return Operand{relative = i64(target), kind = .RELATIVE, size = 4}
}
case .TBZ_BIT:
// Reassemble bit position: b5 at bit 31, b40 at bits 23-19.
b5 := (word >> 31) & 0x1
b40 := (word >> 19) & 0x1F
return Operand{immediate = i64((b5 << 5) | b40), kind = .IMMEDIATE, size = 1}
// ---- Bitmask logical immediate (round-trip back to the raw mask) ----
case .BITMASK_FIELD:
is_64 := (word >> 31) & 1 != 0
n_bit := u8((word >> 22) & 1)
immr := u8((word >> 16) & 0x3F)
imms := u8((word >> 10) & 0x3F)
value, ok := decode_bitmask_imm(n_bit, immr, imms, is_64)
if !ok { return {} }
return Operand{immediate = i64(value), kind = .IMMEDIATE, size = is_64 ? 8 : 4}
// ---- NEON / SIMD register slots ----
// The class comes from the operand TYPE, not from the encoding: SVE forms
// use these same Vd/Vn/Vm slots with Z_REG_* operands, so hardcoding
// REG_V here decoded `add z0.d, z0.d, z0.d` as a V register.
case .VD:
return reg_from_field(word, 0, ot)
// The element size is not in the static pattern -- it shares the tsz field
// with the shift -- so it has to be read out of the word.
case .VD_TSZ:
return Operand{reg = Register(REG_Z | u16(word & 0x1F)), kind = .REGISTER,
size = sve_esize_code(sve_tsz_esize(sve_tsz_field(word)))}
case .VN_TSZ:
return Operand{reg = Register(REG_Z | u16((word >> 5) & 0x1F)), kind = .REGISTER,
size = sve_esize_code(sve_tsz_esize(sve_tsz_field(word)))}
case .VN:
return reg_from_field(word, 5, ot)
case .VD_LIST1, .VD_LIST2, .VD_LIST3, .VD_LIST4:
op := reg_from_field(word, 0, ot)
op.list_count = u8(int(en) - int(Operand_Encoding.VD_LIST1)) + 1
return op
case .VN_LIST1, .VN_LIST2, .VN_LIST3, .VN_LIST4:
op := reg_from_field(word, 5, ot)
op.list_count = u8(int(en) - int(Operand_Encoding.VN_LIST1)) + 1
return op
case .VM:
return reg_from_field(word, 16, ot)
case .VA:
return reg_from_field(word, 10, ot)
// ---- NEON / SVE indexed/immediate fields ----
case .NEON_IMM8_FMOV:
v := ((word >> 16) & 0x7) << 5 | ((word >> 5) & 0x1F)
return Operand{immediate = i64(v), kind = .IMMEDIATE, size = 1}
case .NEON_INDEX_H:
return Operand{immediate = i64((word >> 19) & 0x3), kind = .IMMEDIATE, size = 1}
case .NEON_INDEX_S:
v := ((word >> 21) & 0x1) | ((word >> 11) & 0x1) << 1
return Operand{immediate = i64(v), kind = .IMMEDIATE, size = 1}
case .NEON_INDEX_D:
return Operand{immediate = i64((word >> 11) & 0x1), kind = .IMMEDIATE, size = 1}
// ---- LSE atomic register slots ----
case .ATOMIC_RS:
return reg_from_field(word, 16, ot)
case .ATOMIC_RT:
return reg_from_field(word, 0, ot)
case .ATOMIC_RN:
// Memory operand: only the base register is encoded in the word,
// displacement is always zero (atomic addressing).
base_hw := u8((word >> 5) & 0x1F)
return Operand{
mem = Memory{
base = Register(REG_X | u16(base_hw)),
index = NONE,
mode = .OFFSET,
},
kind = .MEMORY, size = 4,
}
// ---- SVE predicate slots ----
case .PD:
return Operand{reg = Register(REG_P | u16(word & 0xF)), kind = .REGISTER, size = pqual_for_type(ot)}
case .PN:
return Operand{reg = Register(REG_P | u16((word >> 5) & 0xF)), kind = .REGISTER, size = pqual_for_type(ot)}
case .PM:
return Operand{reg = Register(REG_P | u16((word >> 16) & 0xF)), kind = .REGISTER, size = pqual_for_type(ot)}
case .ENC_ZT0:
return Operand{reg = ZT0, kind = .REGISTER, size = 0}
case .LUTI_IDX:
return Operand{immediate = i64((word >> 15) & 0x3), kind = .IMMEDIATE, size = LANE_INDEX}
case .PNG:
return Operand{reg = Register(REG_PN | u16(8 + ((word >> 10) & 0x7))), kind = .REGISTER, size = pqual_for_type(ot)}
case .PG:
return Operand{reg = Register(REG_P | u16((word >> 10) & 0x7)), kind = .REGISTER, size = pqual_for_type(ot)}
case .PG4:
return Operand{reg = Register(REG_P | u16((word >> 10) & 0xF)), kind = .REGISTER, size = pqual_for_type(ot)}
case .PM3:
return Operand{reg = Register(REG_P | u16((word >> 13) & 0x7)), kind = .REGISTER, size = pqual_for_type(ot)}
// ---- SVE immediates ----
case .SVE_IMM8:
v := i32((word >> 5) & 0xFF)
if v & 0x80 != 0 { v |= ~i32(0xFF) }
return Operand{immediate = i64(v), kind = .IMMEDIATE, size = 1}
case .SVE_IMM5A:
v := i64((word >> 5) & 0x1F)
if v & 0x10 != 0 { v |= ~i64(0x1F) }
return Operand{immediate = v, kind = .IMMEDIATE, size = 1}
case .SVE_IMM5:
return Operand{immediate = i64((word >> 16) & 0x1F), kind = .IMMEDIATE, size = 1}
case .SVE_SHIFT_TSZ_IMM:
return Operand{immediate = i64((word >> 16) & 0x7F), kind = .IMMEDIATE, size = 1}
case .SVE_PATTERN:
return Operand{immediate = i64((word >> 5) & 0x1F), kind = .IMMEDIATE, size = 1}
// ---- SVE memory operands ----
case .SVE_OFFSET_BASE_SS, .SVE_OFFSET_BASE_SS1, .SVE_OFFSET_BASE_SS2, .SVE_OFFSET_BASE_SS3,
.SVE_OFFSET_BASE_SS4:
base_hw := u8((word >> 5) & 0x1F)
idx_hw := u8((word >> 16) & 0x1F)
shift: u8 = 0
#partial switch en {
case .SVE_OFFSET_BASE_SS1: shift = 1
case .SVE_OFFSET_BASE_SS2: shift = 2
case .SVE_OFFSET_BASE_SS3: shift = 3
case .SVE_OFFSET_BASE_SS4: shift = 4
}
return Operand{
mem = Memory{
base = Register(REG_X | u16(base_hw)),
index = Register(REG_X | u16(idx_hw)),
shift = shift,
mode = .REG_OFFSET,
},
kind = .MEMORY, size = 4,
}
case .SVE_OFFSET_BASE_SI:
base_hw := u8((word >> 5) & 0x1F)
imm := i32((word >> 16) & 0xF)
if imm & 0x8 != 0 { imm |= ~i32(0xF) }
return Operand{
mem = Memory{
base = Register(REG_X | u16(base_hw)),
index = NONE,
disp = imm,
mode = .OFFSET,
},
kind = .MEMORY, size = 4,
}
// ---- SME ZA tile fields ----
case .ZA_TILE_NUM_B:
return Operand{reg = Register(REG_ZA | 0), kind = .REGISTER,
size = za_elem_for_type(ot)}
case .ZA_TILE_NUM_H:
return Operand{reg = Register(REG_ZA | u16((word >> 22) & 0x1)), kind = .REGISTER,
size = za_elem_for_type(ot)}
case .ZA_TILE_NUM_S:
return Operand{reg = Register(REG_ZA | u16((word >> 22) & 0x3)), kind = .REGISTER,
size = za_elem_for_type(ot)}
case .ZA_TILE_NUM_D:
return Operand{reg = Register(REG_ZA | u16((word >> 21) & 0x7)), kind = .REGISTER,
size = za_elem_for_type(ot)}
case .SME_PATTERN_FIELD:
return Operand{immediate = i64((word >> 5) & 0xF), kind = .IMMEDIATE, size = 1}
// ---- SVE gather/scatter + vector-base memory ----
case .SVE_OFFSET_BASE_VEC, .SVE_OFFSET_BASE_VEC_S, .SVE_OFFSET_BASE_VEC_D,
.SVE_OFFSET_BASE_VECST_S, .SVE_OFFSET_BASE_VECST_D:
base_hw := u8((word >> 5) & 0x1F)
idx_hw := u8((word >> 16) & 0x1F)
m := Memory{
base = Register(REG_X | u16(base_hw)),
index = Register(REG_Z | u16(idx_hw)),
mode = .REG_OFFSET,
}
// These forms all extend a 32-bit-wide index, and bit 22 says which way.
// The index's own element size is not derivable from the word, so it
// travels in the operand's size -- Memory has no bits left.
shape := u8(4)
#partial switch en {
case .SVE_OFFSET_BASE_VEC_S, .SVE_OFFSET_BASE_VEC_D:
m.mode = .EXT_REG_OFFSET
m.extend = (word >> 22) & 1 != 0 ? .SXTW : .UXTW
shape = en == .SVE_OFFSET_BASE_VEC_D ? ZSHAPE_D : ZSHAPE_S
case .SVE_OFFSET_BASE_VECST_S, .SVE_OFFSET_BASE_VECST_D:
m.mode = .EXT_REG_OFFSET
m.extend = (word >> 14) & 1 != 0 ? .SXTW : .UXTW
shape = en == .SVE_OFFSET_BASE_VECST_D ? ZSHAPE_D : ZSHAPE_S
}
return Operand{mem = m, kind = .MEMORY, size = shape}
case .SVE_OFFSET_VEC_BASE:
base_hw := u8((word >> 5) & 0x1F)
imm := i32((word >> 16) & 0x1F)
return Operand{
mem = Memory{
base = Register(REG_Z | u16(base_hw)),
disp = imm,
mode = .OFFSET,
},
kind = .MEMORY, size = 4,
}
// ---- SVE indexed lane field ----
case .SVE_FMLA_IDX_H:
v := ((word >> 22) & 0x1) << 2 | ((word >> 19) & 0x3)
return Operand{immediate = i64(v), kind = .IMMEDIATE, size = LANE_INDEX}
case .SVE_FMLA_IDX_S:
v := (word >> 19) & 0x3
return Operand{immediate = i64(v), kind = .IMMEDIATE, size = LANE_INDEX}
case .SVE_FMLA_IDX_D:
v := (word >> 20) & 0x1
return Operand{immediate = i64(v), kind = .IMMEDIATE, size = LANE_INDEX}
// ---- SME tile slice descriptor (round-trip back to the packed form) ----
//
// Decode is the inverse of the packer: tile_num and imm bits live in
// instruction bits 3:0 (packed per element size), Ws at bits 14:13,
// V flag at bit 15.
case .SME_SLICE_B:
// The tile number and the offset share the low nibble; how it
// splits follows the element size.
return op_za_slice(u8((word >> 4) & 0x0), u8((word >> 13) & 0x3),
u8(word & 0xF), ZSHAPE_B, (word >> 15) & 0x1 != 0)
case .SME_SLICE_H:
// The tile number and the offset share the low nibble; how it
// splits follows the element size.
return op_za_slice(u8((word >> 3) & 0x1), u8((word >> 13) & 0x3),
u8(word & 0x7), ZSHAPE_H, (word >> 15) & 0x1 != 0)
case .SME_SLICE_W:
// The tile number and the offset share the low nibble; how it
// splits follows the element size.
return op_za_slice(u8((word >> 2) & 0x3), u8((word >> 13) & 0x3),
u8(word & 0x3), ZSHAPE_S, (word >> 15) & 0x1 != 0)
case .SME_SLICE_D:
// The tile number and the offset share the low nibble; how it
// splits follows the element size.
return op_za_slice(u8((word >> 1) & 0x7), u8((word >> 13) & 0x3),
u8(word & 0x1), ZSHAPE_D, (word >> 15) & 0x1 != 0)
case .SME_SLICE_Q:
// The tile number and the offset share the low nibble; how it
// splits follows the element size.
return op_za_slice(u8((word >> 0) & 0xF), u8((word >> 13) & 0x3),
u8(word & 0x0), ZSHAPE_Q, (word >> 15) & 0x1 != 0)
case .NEON_IDX2:
return Operand{immediate = i64((word >> 12) & 0x3), kind = .IMMEDIATE, size = LANE_INDEX}
case .ENC_FCMLA_ROT:
return Operand{immediate = i64((word >> 12) & 0x3), kind = .IMMEDIATE, size = 1}
case .ENC_FCADD_ROT:
return Operand{immediate = i64((word >> 12) & 0x1), kind = .IMMEDIATE, size = 1}
case .ENC_SVE_PRFOP:
return Operand{immediate = i64(word & 0xF), kind = .IMMEDIATE, size = 1}
case .ENC_LDRAA_IMM10:
v := i32((word >> 12) & 0x3FF)
if v & 0x200 != 0 { v |= ~i32(0x3FF) }
return Operand{immediate = i64(v << 3), kind = .IMMEDIATE, size = 2}
// ---- Batch 5 ----
case .ENC_LSL_IMM_W:
// Recover shift from imms: imms = 31 - imm.
imms := (word >> 10) & 0x1F
return Operand{immediate = i64((31 - imms) & 0x1F), kind = .IMMEDIATE, size = 1}
case .ENC_LSL_IMM_X:
imms := (word >> 10) & 0x3F
return Operand{immediate = i64((63 - imms) & 0x3F), kind = .IMMEDIATE, size = 1}
case .ENC_IMM6_LO:
v := i32((word >> 5) & 0x3F)
if v & (1 << 5) != 0 { v |= ~i32(0x3F) } // sign-extend from bit 5
return Operand{immediate = i64(v), kind = .IMMEDIATE, size = 1}
case .ENC_SHIFT_IMMR:
// LSR/ASR immediate: the shift is immr verbatim (bits 21:16).
return Operand{immediate = i64((word >> 16) & 0x3F), kind = .IMMEDIATE, size = 1}
case .ENC_DUAL_RN_RM:
// Take the Rn slot (9:5) as the source register.
return Operand{reg = Register(REG_X | u16((word >> 5) & 0x1F)), kind = .REGISTER, size = 4}
case .ENC_ROR_SHIFT:
return Operand{immediate = i64((word >> 10) & 0x3F), kind = .IMMEDIATE, size = 1}
case .ENC_Z_PAIR_VD, .ENC_Z_QUAD_VD:
return Operand{reg = Register(REG_Z | u16(word & 0x1F)), kind = .REGISTER,
size = reg_size_for_type(ot), list_count = en == .ENC_Z_PAIR_VD ? 2 : 4}
case .ENC_Z_PAIR_VN, .ENC_Z_QUAD_VN:
return Operand{reg = Register(REG_Z | u16((word >> 5) & 0x1F)), kind = .REGISTER,
size = reg_size_for_type(ot), list_count = en == .ENC_Z_PAIR_VN ? 2 : 4}
case .ENC_Z_PAIR_VM, .ENC_Z_QUAD_VM:
return Operand{reg = Register(REG_Z | u16((word >> 16) & 0x1F)), kind = .REGISTER, size = reg_size_for_type(ot)}
}
return {}
}
// reg_from_field reconstructs a Register from a 5-bit hw field at `shift`,
// choosing the right class per the form's Operand_Type. SP/WSP variants
// use the REG_XSP/REG_WSP class at hw=31; everything else uses REG_X/REG_W.
@(private="file")
reg_from_field :: #force_inline proc "contextless" (
word: u32, shift: u8, ot: Operand_Type,
) -> Operand {
hw := u16((word >> shift) & 0x1F)
cls: u16 = REG_X
#partial switch ot {
case .W_REG: cls = REG_W
case .X_REG: cls = REG_X
case .WSP_REG: cls = hw == 31 ? REG_WSP : REG_W
case .XSP_REG: cls = hw == 31 ? REG_XSP : REG_X
case .B_REG: cls = REG_B
case .H_REG: cls = REG_H
case .S_REG: cls = REG_S
case .D_REG: cls = REG_D
case .Q_REG: cls = REG_Q
case .V_REG,
.V_8B, .V_16B, .V_4H, .V_8H, .V_2S, .V_4S, .V_1D, .V_2D,
.V_4H_FP16, .V_8H_FP16, .V_1Q,
.V_ELEM_B, .V_ELEM_H, .V_ELEM_S, .V_ELEM_D:
cls = REG_V
case .Z_REG_B, .Z_REG_H, .Z_REG_S, .Z_REG_D, .Z_REG_ANY:
cls = REG_Z
case .P_REG, .P_REG_MERGE, .P_REG_ZERO, .P_REG_GOV,
.P_REG_B, .P_REG_H, .P_REG_S, .P_REG_D:
cls = REG_P
case .PN_REG, .PN_REG_ZERO:
cls = REG_PN
case .ZT_REG:
cls = REG_ZT
case .Z_PAIR_B, .Z_PAIR_H, .Z_PAIR_S, .Z_PAIR_D,
.Z_QUAD_B, .Z_QUAD_H, .Z_QUAD_S, .Z_QUAD_D:
cls = REG_Z
}
// SP class needs the special hw=31 marker; everything else uses the
// raw hw with the chosen class.
if (ot == .WSP_REG && hw == 31) || (ot == .XSP_REG && hw == 31) {
return Operand{reg = Register(cls | 31), kind = .REGISTER, size = 4}
}
// Vector operands carry their arrangement / element view in `size`, using
// the same codes op_v_*/op_z_* produce (see operands.odin). Without this a
// decoded V register would come back as a bare `v0` with no `.4s`, so a
// disassembly could not be fed back to an assembler.
return Operand{reg = Register(cls | hw), kind = .REGISTER, size = reg_size_for_type(ot)}
}
// The `size` marker an operand of this type carries: the NEON arrangement
// (multiples of 8), an element view (odd), or an SVE element width. 4 is the
// neutral "no vector shape" value used by every scalar class.
@(private="file", require_results)
reg_size_for_type :: #force_inline proc "contextless" (ot: Operand_Type) -> u8 {
#partial switch ot {
case .V_8B: return 8
case .V_16B: return 16
case .V_4H, .V_4H_FP16: return 24
case .V_8H, .V_8H_FP16: return 32
case .V_2S: return 40
case .V_4S: return 48
case .V_1D: return 56
case .V_2D: return 64
case .V_1Q: return 72
case .V_ELEM_B: return 1
case .V_ELEM_H: return 3
case .V_ELEM_S: return 5
case .V_ELEM_D: return 7
case .Z_REG_B: return 1
case .Z_REG_H: return 2
case .Z_REG_S: return 4
case .Z_REG_D: return 8
case .Z_REG_ANY: return 0 // no element size in the syntax
case .P_REG_ZERO, .PN_REG_ZERO: return PQUAL_ZERO
case .P_REG_MERGE: return PQUAL_MERGE
case .P_REG_B: return PSHAPE_B
case .P_REG_H: return PSHAPE_H
case .P_REG_S: return PSHAPE_S
case .P_REG_D: return PSHAPE_D
case .Z_PAIR_B, .Z_QUAD_B: return 1
case .Z_PAIR_H, .Z_QUAD_H: return 2
case .Z_PAIR_S, .Z_QUAD_S: return 4
case .Z_PAIR_D, .Z_QUAD_D: return 8
}
return 4
}
// A predicate operand's governing qualifier, which the form -- not the caller
// -- decides: an SVE load zeroes, a predicated add merges.
@(private="file", require_results)
pqual_for_type :: #force_inline proc "contextless" (ot: Operand_Type) -> u8 {
#partial switch ot {
case .P_REG_ZERO, .PN_REG_ZERO: return PQUAL_ZERO
case .P_REG_MERGE: return PQUAL_MERGE
case .P_REG_B: return PSHAPE_B
case .P_REG_H: return PSHAPE_H
case .P_REG_S: return PSHAPE_S
case .P_REG_D: return PSHAPE_D
}
return PQUAL_NONE
}
// -----------------------------------------------------------------------------
// Buffer-Sizing Helpers (let callers pre-size so the decode hot path never
// reallocates; allocates no new buffers -- only the caller's arrays grow).
// -----------------------------------------------------------------------------
// Exact instruction-count ceiling for `data` (AArch64 instructions are 4 bytes).
@(require_results)
decode_max_instruction_count :: #force_inline proc "contextless" (data: []u8) -> int {
return len(data) / 4
}
// Typical-case estimate (AArch64 is fixed 4 bytes/instruction, so this is exact).
@(require_results)
decode_estimate_instruction_count :: #force_inline proc "contextless" (data: []u8) -> int {
return len(data) / 4 + 8
}
// Pre-size the caller's decode output arrays for `data` (reserves on top of any
// existing elements; nil to skip; exact=true for the ceiling, else the estimate).
decode_reserve :: proc(instructions: ^[dynamic]Instruction, inst_info: ^[dynamic]Instruction_Info, label_defs: ^[dynamic]Label_Definition, data: []u8, exact: bool = false) {
n := exact ? decode_max_instruction_count(data) : decode_estimate_instruction_count(data)
if instructions != nil { reserve(instructions, len(instructions) + n) }
if inst_info != nil { reserve(inst_info, len(inst_info) + n) }
if label_defs != nil { reserve(label_defs, len(label_defs) + n) }
}
// The packed tszh:tszl:imm3 value an SVE shift-by-immediate carries.
@(private="file", require_results)
sve_tsz_field :: #force_inline proc "contextless" (word: u32) -> u32 {
return ((word >> 22) & 0x3) << 5 | ((word >> 19) & 0x3) << 3 | ((word >> 16) & 0x7)
}
// The element width a ZA tile operand is viewed at.
@(private="file", require_results)
za_elem_for_type :: #force_inline proc "contextless" (ot: Operand_Type) -> u8 {
#partial switch ot {
case .ZA_TILE_H: return ZSHAPE_H
case .ZA_TILE_S: return ZSHAPE_S
case .ZA_TILE_D: return ZSHAPE_D
}
return ZSHAPE_B
}