diff options
| author | iximeow <me@iximeow.net> | 2021-10-26 21:37:44 -0700 |
|---|---|---|
| committer | iximeow <me@iximeow.net> | 2021-10-26 21:37:44 -0700 |
| commit | 9b3d3d5c52a619e8e090e676033ee99abe33553d (patch) | |
| tree | 109e278dd866df4f33c80dfa4cb1652e7f3aa5e7 | |
| parent | d384d7fda7589de7451e8c6a82c25e7864c5cb0f (diff) | |
add many test cases from openblas, ldnp, stnp
| -rw-r--r-- | src/armv8/a64.rs | 80 | ||||
| -rw-r--r-- | test/armv8/a64.rs | 1773 |
2 files changed, 1843 insertions, 10 deletions
diff --git a/src/armv8/a64.rs b/src/armv8/a64.rs index f3fada2..9f51a80 100644 --- a/src/armv8/a64.rs +++ b/src/armv8/a64.rs @@ -214,7 +214,7 @@ pub enum SizeCode { X, W } #[derive(Copy, Clone, Debug, PartialEq)] #[repr(u8)] -pub enum SIMDSizeCode { S, D, Q } +pub enum SIMDSizeCode { B, H, S, D, Q } #[derive(Copy, Clone, Debug, PartialEq)] #[repr(C)] @@ -914,6 +914,12 @@ impl Display for Instruction { Opcode::CRC32CX => { write!(fmt, "crc32cx")?; } + Opcode::STNP => { + write!(fmt, "stnp")?; + } + Opcode::LDNP => { + write!(fmt, "ldnp")?; + } }; if self.operands[0] != Operand::Nothing { @@ -1111,6 +1117,8 @@ pub enum Opcode { CRC32CH, CRC32CW, CRC32CX, + STNP, + LDNP, } #[derive(Copy, Clone, Debug, PartialEq)] @@ -1198,6 +1206,8 @@ impl Display for Operand { }, Operand::SIMDRegister(size, reg) => { match size { + SIMDSizeCode::B => { write!(fmt, "b{}", reg) } + SIMDSizeCode::H => { write!(fmt, "h{}", reg) } SIMDSizeCode::S => { write!(fmt, "s{}", reg) } SIMDSizeCode::D => { write!(fmt, "d{}", reg) } SIMDSizeCode::Q => { write!(fmt, "q{}", reg) } @@ -1824,7 +1834,6 @@ impl Decoder<ARMv8> for InstDecoder { ]; let compound_idx = (op << 1) | sf; - crate::armv8::a64::std::eprintln!("word: {:#08x}, compound_idx: {:x}", word, compound_idx); let (opcode, source_size, dest_size) = DATA_PROCESSING_3_SOURCE.get(compound_idx as usize) .map(std::borrow::ToOwned::to_owned) .unwrap_or(Err(DecodeError::InvalidOpcode))?; @@ -2565,16 +2574,64 @@ impl Decoder<ARMv8> for InstDecoder { 0b10000 => { // load/store no-allocate pair (offset) // V == 0 - // let opc_L = ((word >> 22) & 1) | ((word >> 29) & 0x6); - // eprintln!("C3.3.7 V==0, opc_L: {}", opc_L); - return Err(DecodeError::IncompleteDecoder); + let opc_L = ((word >> 22) & 1) | ((word >> 29) & 0x6); + const ENCODINGS: &'static [Result<(Opcode, SizeCode), DecodeError>] = &[ + Ok((Opcode::STNP, SizeCode::W)), + Ok((Opcode::LDNP, SizeCode::W)), + Err(DecodeError::InvalidOpcode), + Err(DecodeError::InvalidOpcode), + Ok((Opcode::STNP, SizeCode::X)), + Ok((Opcode::LDNP, SizeCode::X)), + Err(DecodeError::InvalidOpcode), + Err(DecodeError::InvalidOpcode), + ]; + + let (opcode, size) = ENCODINGS[opc_L as usize]?; + inst.opcode = opcode; + + let Rt = (word & 0x1f) as u16; + let Rn = ((word >> 5) & 0x1f) as u16; + let Rt2 = ((word >> 10) & 0x1f) as u16; + let imm7 = ((word >> 15) & 0x7f) as i16; + let imm7 = (imm7 << 9) >> 9; + + inst.operands = [ + Operand::Register(size, Rt), + Operand::Register(size, Rt2), + Operand::RegPreIndex(Rn, imm7), + Operand::Nothing, + ]; }, 0b10100 => { // load/store no-allocate pair (offset) // V == 1 - // let opc_L = ((word >> 22) & 1) | ((word >> 29) & 0x6); - // eprintln!("C3.3.7 V==1, opc_L: {}", opc_L); - return Err(DecodeError::IncompleteDecoder); + let opc_L = ((word >> 22) & 1) | ((word >> 29) & 0x6); + const ENCODINGS: &'static [Result<(Opcode, SIMDSizeCode), DecodeError>] = &[ + Ok((Opcode::STNP, SIMDSizeCode::S)), + Ok((Opcode::LDNP, SIMDSizeCode::S)), + Ok((Opcode::STNP, SIMDSizeCode::D)), + Ok((Opcode::LDNP, SIMDSizeCode::D)), + Ok((Opcode::STNP, SIMDSizeCode::Q)), + Ok((Opcode::LDNP, SIMDSizeCode::Q)), + Err(DecodeError::InvalidOpcode), + Err(DecodeError::InvalidOpcode), + ]; + + let (opcode, size) = ENCODINGS[opc_L as usize]?; + inst.opcode = opcode; + + let Rt = (word & 0x1f) as u16; + let Rn = ((word >> 5) & 0x1f) as u16; + let Rt2 = ((word >> 10) & 0x1f) as u16; + let imm7 = ((word >> 15) & 0x7f) as i16; + let imm7 = (imm7 << 9) >> 9; + + inst.operands = [ + Operand::SIMDRegister(size, Rt), + Operand::SIMDRegister(size, Rt2), + Operand::RegPreIndex(Rn, imm7), + Operand::Nothing, + ]; }, 0b10001 => { // load/store register pair (post-indexed) @@ -2757,7 +2814,8 @@ impl Decoder<ARMv8> for InstDecoder { 0b11001 => { /* * load/store register {unscaled immediate, immediate post-indexed, - * unprivileged, immediate pre-indexd, register offset} + * unprivileged, immediate pre-indexd, register offset, pac} + * atomic memory operations * V == 0 */ let Rt = (word & 0x1f) as u16; @@ -2951,7 +3009,8 @@ impl Decoder<ARMv8> for InstDecoder { 0b11101 => { /* * load/store register {unscaled immediate, immediate post-indexed, - * unprivileged, immediate pre-indexd, register offset} + * unprivileged, immediate pre-indexed, register offset, pac} + * also atomic memory operations * V == 1 */ let Rt = (word & 0x1f) as u16; @@ -3348,6 +3407,7 @@ impl Decoder<ARMv8> for InstDecoder { ]; } 0b01001 => { // conditional branch (imm) + // probably actually an invalid opcode? return Err(DecodeError::IncompleteDecoder); inst.opcode = Opcode::Invalid; } diff --git a/test/armv8/a64.rs b/test/armv8/a64.rs index dde9802..69039e7 100644 --- a/test/armv8/a64.rs +++ b/test/armv8/a64.rs @@ -2430,3 +2430,1776 @@ pub fn bench_60000_instrs(b: &mut Bencher) { }); } */ + +enum ErrorDesc { + IncorrectDecode([u8; 4], String, String), + FailedDecode([u8; 4], String, DecodeError), +} + +impl fmt::Display for ErrorDesc { + fn fmt(&self, f: &mut fmt::Formatter) -> fmt::Result { + match self { + ErrorDesc::IncorrectDecode(bytes, expected, actual) => { + writeln!(f, " decoded incorrect instruction from {:02x}{:02x}{:02x}{:02x}", bytes[0], bytes[1], bytes[2], bytes[3])?; + writeln!(f, " expected \"{}\"", expected)?; + writeln!(f, " got \"{}\"", actual)?; + }, + ErrorDesc::FailedDecode(bytes, expected, err) => { + writeln!(f, " decode error for {:02x}{:02x}{:02x}{:02x}", bytes[0], bytes[1], bytes[2], bytes[3])?; + writeln!(f, " expected \"{}\"", expected)?; + writeln!(f, " got \"{:?}\"", err)?; + }, + } + + Ok(()) + } +} + +fn run_tests(cases: &[([u8; 4], &'static str)]) -> Vec<ErrorDesc> { + let decoder = <ARMv8 as Arch>::Decoder::default(); + + let mut errors = Vec::new(); + + for (bytes, text) in cases { + let mut reader = yaxpeax_arch::U8Reader::new(&bytes[..]); + match decoder.decode(&mut reader) { + Ok(result) => { + if &result.to_string() != text { + errors.push(ErrorDesc::IncorrectDecode(*bytes, text.to_string(), result.to_string())); + } + }, + Err(err) => { + errors.push(ErrorDesc::FailedDecode(*bytes, text.to_string(), err)); + } + } + } + + errors +} + +#[test] +fn test_openblas_arithmetic() { + const TESTS: &[([u8; 4], &'static str)] = &[ + ([0x61, 0x06, 0x13, 0x0b], "add w1, w19, w19, lsl 1"), + ([0x05, 0x07, 0x13, 0x0b], "add w5, w24, w19, lsl 1"), + ([0x08, 0x7d, 0x48, 0x0b], "add w8, w8, w8, lsr 31"), + ([0x00, 0x04, 0x00, 0x11], "add w0, w0, 0x1"), + ([0x00, 0x04, 0x01, 0x11], "add w0, w0, 0x41"), + ([0x5b, 0xff, 0x01, 0x11], "add w27, w26, 0x7f"), + ([0x02, 0xfc, 0x3f, 0x11], "add w2, w0, 0xfff"), + ([0x00, 0x00, 0x60, 0x11], "add w0, w0, 0x800000"), + ([0x0a, 0x0c, 0x00, 0x31], "adds w10, w0, 0x3"), + ([0x82, 0x00, 0x00, 0x8b], "add x2, x4, x0"), + ([0x00, 0x35, 0x00, 0x8b], "add x0, x8, x0, lsl 13"), + ([0x00, 0x00, 0x01, 0x8b], "add x0, x0, x1"), + ([0xde, 0x03, 0x12, 0x8b], "add x30, x30, x18"), + ([0x03, 0x04, 0x12, 0x8b], "add x3, x0, x18, lsl 1"), + ([0x56, 0x06, 0x12, 0x8b], "add x22, x18, x18, lsl 1"), + ([0x12, 0x08, 0x12, 0x8b], "add x18, x0, x18, lsl 2"), + ([0x32, 0x08, 0x12, 0x8b], "add x18, x1, x18, lsl 2"), + ([0x92, 0x13, 0x12, 0x8b], "add x18, x28, x18, lsl 4"), + ([0x4f, 0x14, 0x12, 0x8b], "add x15, x2, x18, lsl 5"), + ([0x42, 0x18, 0x12, 0x8b], "add x2, x2, x18, lsl 6"), + ([0xc6, 0x1c, 0x12, 0x8b], "add x6, x6, x18, lsl 7"), + ([0x62, 0xce, 0x21, 0x8b], "add x2, x19, w1, sxtw 3"), + ([0x87, 0xd3, 0x21, 0x8b], "add x7, x28, w1, sxtw 4"), + ([0x2a, 0x41, 0x22, 0x8b], "add x10, x9, w2, uxtw 0"), + ([0x18, 0x4f, 0x22, 0x8b], "add x24, x24, w2, uxtw 3"), + ([0x82, 0x4f, 0x22, 0x8b], "add x2, x28, w2, uxtw 3"), + ([0xe2, 0x50, 0x22, 0x8b], "add x2, x7, w2, uxtw 4"), + ([0x20, 0x04, 0x84, 0x8b], "add x0, x1, x4, asr 1"), + ([0x14, 0x04, 0x94, 0x8b], "add x20, x0, x20, asr 1"), + ([0xe6, 0x62, 0x00, 0x91], "add x6, x23, 0x18"), + ([0x00, 0x80, 0x04, 0x91], "add x0, x0, 0x120"), + ([0xf8, 0x63, 0x06, 0x91], "add x24, sp, 0x198"), + ([0xb7, 0xfc, 0x3f, 0x91], "add x23, x5, 0xfff"), + ([0x5a, 0xff, 0x3f, 0x91], "add x26, x26, 0xfff"), + ([0x7b, 0xff, 0x3f, 0x91], "add x27, x27, 0xfff"), + ([0x9c, 0xff, 0x3f, 0x91], "add x28, x28, 0xfff"), + ([0xc6, 0x08, 0x40, 0x91], "add x6, x6, 0x2000"), + ([0xe8, 0x13, 0x40, 0x91], "add x8, sp, 0x4000"), + ([0xe2, 0x43, 0x40, 0x91], "add x2, sp, 0x10000"), + ([0xff, 0x0f, 0x42, 0x91], "add sp, sp, 0x83000"), + ([0xe1, 0x13, 0x42, 0x91], "add x1, sp, 0x84000"), + ([0xff, 0x13, 0x42, 0x91], "add sp, sp, 0x84000"), + ([0x63, 0x02, 0x60, 0x91], "add x3, x19, 0x800000"), + ([0x03, 0x03, 0x60, 0x91], "add x3, x24, 0x800000"), + ([0x14, 0x00, 0x13, 0xab], "adds x20, x0, x19"), + ([0x37, 0x01, 0x1c, 0xab], "adds x23, x9, x28"), + ([0x0b, 0x08, 0x00, 0x4b], "sub w11, w0, w0, lsl 2"), + ([0x00, 0x00, 0x01, 0x4b], "sub w0, w0, w1"), + ([0xfc, 0x03, 0x08, 0x4b], "neg w28, w8"), + ([0x00, 0x04, 0x08, 0x4b], "sub w0, w0, w8, lsl 1"), + ([0x81, 0x09, 0x0c, 0x4b], "sub w1, w12, w12, lsl 2"), + ([0xe0, 0x02, 0x0d, 0x4b], "sub w0, w23, w13"), + ([0xfb, 0x03, 0x15, 0x4b], "neg w27, w21"), + ([0x63, 0x04, 0x15, 0x4b], "sub w3, w3, w21, lsl 1"), + ([0xfc, 0x03, 0x1c, 0x4b], "neg w28, w28"), + ([0x00, 0x04, 0x1c, 0x4b], "sub w0, w0, w28, lsl 1"), + ([0x00, 0x7c, 0x81, 0x4b], "sub w0, w0, w1, asr 31"), + ([0x6e, 0x7c, 0x8e, 0x4b], "sub w14, w3, w14, asr 31"), + ([0x4c, 0x04, 0x00, 0x51], "sub w12, w2, 0x1"), + ([0x67, 0x0f, 0x00, 0x51], "sub w7, w27, 0x3"), + ([0x04, 0x11, 0x00, 0x51], "sub w4, w8, 0x4"), + ([0x42, 0x1c, 0x00, 0x51], "sub w2, w2, 0x7"), + ([0xc5, 0x20, 0x00, 0x51], "sub w5, w6, 0x8"), + ([0x08, 0x25, 0x00, 0x51], "sub w8, w8, 0x9"), + ([0x00, 0x2c, 0x00, 0x51], "sub w0, w0, 0xb"), + ([0x00, 0x30, 0x00, 0x51], "sub w0, w0, 0xc"), + ([0x2d, 0x30, 0x00, 0x51], "sub w13, w1, 0xc"), + ([0x21, 0x34, 0x00, 0x51], "sub w1, w1, 0xd"), + ([0xb8, 0x46, 0x00, 0x51], "sub w24, w21, 0x11"), + ([0xe0, 0x84, 0x01, 0x51], "sub w0, w7, 0x61"), + ([0x0c, 0x01, 0x00, 0x6b], "subs w12, w8, w0"), + ([0xe2, 0x03, 0x00, 0x6b], "negs w2, w0"), + ([0x21, 0x04, 0x00, 0x71], "subs w1, w1, 0x1"), + ([0x25, 0x04, 0x00, 0x71], "subs w5, w1, 0x1"), + ([0x2c, 0x04, 0x00, 0x71], "subs w12, w1, 0x1"), + ([0x3a, 0x04, 0x00, 0x71], "subs w26, w1, 0x1"), + ([0xe3, 0x13, 0x02, 0xcb], "neg x3, x2, lsl 4"), + ([0xeb, 0x0f, 0x03, 0xcb], "neg x11, x3, lsl 3"), + ([0x75, 0x08, 0x02, 0xcb], "sub x21, x3, x2, lsl 2"), + ([0xa3, 0x00, 0x03, 0xcb], "sub x3, x5, x3"), + ([0x81, 0x08, 0x03, 0xcb], "sub x1, x4, x3, lsl 2"), + ([0x70, 0x12, 0x15, 0xcb], "sub x16, x19, x21, lsl 4"), + ([0xc6, 0x40, 0x20, 0xcb], "sub x6, x6, w0, uxtw 0"), + ([0x24, 0x50, 0x20, 0xcb], "sub x4, x1, w0, uxtw 4"), + ([0x8a, 0xc1, 0x20, 0xcb], "sub x10, x12, w0, sxtw 0"), + ([0xbb, 0xcc, 0x20, 0xcb], "sub x27, x5, w0, sxtw 3"), + ([0xd6, 0x04, 0x00, 0xd1], "sub x22, x6, 0x1"), + ([0xc7, 0x40, 0x10, 0xd1], "sub x7, x6, 0x410"), + ([0x7b, 0x43, 0x10, 0xd1], "sub x27, x27, 0x410"), + ([0xff, 0xc3, 0x27, 0xd1], "sub sp, sp, 0x9f0"), + ([0xec, 0x03, 0x42, 0xd1], "sub x12, sp, 0x80000"), + ([0x97, 0x02, 0x1b, 0xeb], "subs x23, x20, x27"), + ([0x98, 0x02, 0x1b, 0xeb], "subs x24, x20, x27"), + ([0x6b, 0x05, 0x00, 0xf1], "subs x11, x11, 0x1"), + ([0x5a, 0x07, 0x00, 0xf1], "subs x26, x26, 0x1"), + ([0x08, 0x09, 0x00, 0xf1], "subs x8, x8, 0x2"), + ]; + + let errs = run_tests(TESTS); + + for err in errs.iter() { + println!("{}", err); + } + + assert!(errs.is_empty()); +} + +#[test] +fn test_openblas_bitwise() { + const TESTS: &[([u8; 4], &'static str)] = &[ + ([0x94, 0x7c, 0x01, 0x13], "asr w20, w4, 0x1"), + ([0x0a, 0x7d, 0x02, 0x13], "asr w10, w8, 0x2"), + ([0x00, 0x7c, 0x0c, 0x13], "asr w0, w0, 0xc"), + ([0x20, 0x04, 0x81, 0x13], "ror w0, w1, 0x1"), + ([0x61, 0x21, 0xc0, 0x1a], "lsl w1, w11, w0"), + ([0x00, 0x20, 0xc1, 0x1a], "lsl w0, w0, w1"), + ([0x94, 0x22, 0xc1, 0x1a], "lsl w20, w20, w1"), + ([0x63, 0x20, 0xc2, 0x1a], "lsl w3, w3, w2"), + ([0x03, 0x7c, 0x01, 0x53], "lsr w3, w0, 0x1"), + ([0x23, 0x7c, 0x02, 0x53], "lsr w3, w1, 0x2"), + ([0x00, 0x7c, 0x06, 0x53], "lsr w0, w0, 0x6"), + ([0x42, 0x7c, 0x07, 0x53], "lsr w2, w2, 0x7"), + ([0x00, 0x4c, 0x14, 0x53], "lsl w0, w0, 0xc"), + ([0xe0, 0x56, 0x16, 0x53], "lsl w0, w23, 0xa"), + ([0x20, 0x7c, 0x18, 0x53], "lsr w0, w1, 0x18"), + ([0x35, 0x6c, 0x1c, 0x53], "lsl w21, w1, 0x4"), + ([0x2c, 0x70, 0x1d, 0x53], "lsl w12, w1, 0x3"), + ([0xa8, 0x74, 0x1e, 0x53], "lsl w8, w5, 0x2"), + ([0x23, 0x7b, 0x1f, 0x53], "lsl w3, w25, 0x1"), + ([0x04, 0x7f, 0x1f, 0x53], "lsr w4, w24, 0x1f"), + ([0x05, 0xfe, 0x41, 0xd3], "lsr x5, x16, 0x1"), + ([0x22, 0x3c, 0x44, 0xd3], "ubfx x2, x1, 0x4, 0xc"), + ([0xb5, 0xfe, 0x46, 0xd3], "lsr x21, x21, 0x6"), + ([0xe7, 0xfc, 0x60, 0xd3], "lsr x7, x7, 0x20"), + ([0x21, 0xfc, 0x61, 0xd3], "lsr x1, x1, 0x21"), + ([0x00, 0xfc, 0x63, 0xd3], "lsr x0, x0, 0x23"), + ([0xa6, 0xd4, 0x76, 0xd3], "lsl x6, x5, 0xa"), + ([0x03, 0x08, 0x7a, 0xd3], "ubfiz x3, x0, 0x6, 0x3"), + ([0xa1, 0x7e, 0x79, 0x93], "sbfiz x1, x21, 0x7, 0x20"), + ([0x01, 0x7c, 0x7a, 0x93], "sbfiz x1, x0, 0x6, 0x20"), + ([0x9a, 0x7c, 0x7d, 0x93], "sbfiz x26, x4, 0x3, 0x20"), + ([0xc3, 0x0f, 0x7a, 0xd3], "ubfiz x3, x30, 0x6, 0x4"), + ([0x01, 0xe4, 0x7a, 0xd3], "lsl x1, x0, 0x6"), + ([0x65, 0xe8, 0x7b, 0xd3], "lsl x5, x3, 0x5"), + ([0x79, 0xf3, 0x7d, 0xd3], "lsl x25, x27, 0x3"), + ([0x81, 0x7d, 0x7f, 0xd3], "ubfiz x1, x12, 0x1, 0x20"), + ([0x0e, 0xf8, 0x7f, 0xd3], "lsl x14, x0, 0x1"), + ([0xe9, 0x22, 0xce, 0x9a], "lsl x9, x23, x14"), + ([0xca, 0xff, 0x7f, 0xd3], "lsr x10, x30, 0x3f"), + ([0xd3, 0x7f, 0x40, 0x93], "sxtw x19, w30"), + ([0x00, 0xfc, 0x62, 0x93], "asr x0, x0, 0x22"), + ([0x20, 0x00, 0x00, 0x0a], "and w0, w1, w0"), + ([0x28, 0x00, 0x00, 0x0a], "and w8, w1, w0"), + ([0x29, 0x00, 0x00, 0x0a], "and w9, w1, w0"), + ([0xb5, 0x02, 0x22, 0x0a], "bic w21, w21, w2"), + ([0x00, 0x2c, 0x00, 0x12], "and w0, w0, 0xfff"), + ([0x42, 0x2c, 0x00, 0x12], "and w2, w2, 0xfff"), + ([0x00, 0x00, 0x01, 0x12], "and w0, w0, 0x80000000"), + ([0x21, 0x00, 0x01, 0x12], "and w1, w1, 0x80000000"), + ([0x10, 0x00, 0x18, 0x12], "and w16, w0, 0x100"), + ([0x11, 0x00, 0x18, 0x12], "and w17, w0, 0x100"), + ([0x06, 0x79, 0x1a, 0x12], "and w6, w8, 0xffffffdf"), + ([0x08, 0x79, 0x1a, 0x12], "and w8, w8, 0xffffffdf"), + ([0x02, 0x70, 0x1d, 0x12], "and w2, w0, 0xfffffff8"), + ([0x00, 0x74, 0x1e, 0x12], "and w0, w0, 0xfffffffc"), + ([0x02, 0x74, 0x1e, 0x12], "and w2, w0, 0xfffffffc"), + ([0x00, 0x78, 0x1e, 0x12], "and w0, w0, 0xfffffffd"), + ([0x02, 0x78, 0x1e, 0x12], "and w2, w0, 0xfffffffd"), + ([0xc6, 0x78, 0x1e, 0x12], "and w6, w6, 0xfffffffd"), + ([0x10, 0x7a, 0x1e, 0x12], "and w16, w16, 0xfffffffd"), + ([0x39, 0x7b, 0x1f, 0x12], "and w25, w25, 0xfffffffe"), + ([0x63, 0x00, 0x05, 0x6a], "ands w3, w3, w5"), + ([0x00, 0x04, 0x00, 0x72], "ands w0, w0, 0x3"), + ([0xc6, 0x04, 0x00, 0x72], "ands w6, w6, 0x3"), + ([0x00, 0x78, 0x1e, 0x72], "ands w0, w0, 0xfffffffd"), + ([0x21, 0x78, 0x1e, 0x72], "ands w1, w1, 0xfffffffd"), + ([0x40, 0x00, 0x00, 0x8a], "and x0, x2, x0"), + ([0x84, 0x00, 0x1c, 0x8a], "and x4, x4, x28"), + ([0x01, 0x00, 0x40, 0x92], "and x1, x0, 0x1"), + ([0x02, 0x00, 0x40, 0x92], "and x2, x0, 0x1"), + ([0xa5, 0xf0, 0x7d, 0x92], "and x5, x5, 0xfffffffffffffff8"), + ([0x00, 0x03, 0x7e, 0x92], "and x0, x24, 0x4"), + ([0xc1, 0x06, 0x7e, 0x92], "and x1, x22, 0xc"), + ([0x42, 0xf4, 0x7e, 0x92], "and x2, x2, 0xfffffffffffffffc"), + ([0x01, 0x00, 0x7f, 0x92], "and x1, x0, 0x2"), + ([0xe0, 0x01, 0x7f, 0x92], "and x0, x15, 0x2"), + ([0x75, 0x02, 0x7f, 0x92], "and x21, x19, 0x2"), + ([0x76, 0x02, 0x7f, 0x92], "and x22, x19, 0x2"), + ([0x8e, 0x02, 0x7f, 0x92], "and x14, x20, 0x2"), + ([0x9a, 0x02, 0x7f, 0x92], "and x26, x20, 0x2"), + ([0x9b, 0x02, 0x7f, 0x92], "and x27, x20, 0x2"), + ([0xc1, 0x06, 0x7f, 0x92], "and x1, x22, 0x6"), + ([0x00, 0x07, 0x7f, 0x92], "and x0, x24, 0x6"), + ([0x01, 0x08, 0x7f, 0x92], "and x1, x0, 0xe"), + ([0xc1, 0x0a, 0x7f, 0x92], "and x1, x22, 0xe"), + ([0x6b, 0xfa, 0x7f, 0x92], "and x11, x19, 0xfffffffffffffffe"), + ([0x80, 0xfb, 0x7f, 0x92], "and x0, x28, 0xfffffffffffffffe"), + ([0x0c, 0x04, 0x40, 0xf2], "ands x12, x0, 0x3"), + ([0x88, 0x0a, 0x40, 0xf2], "ands x8, x20, 0x7"), + ([0x01, 0x10, 0x40, 0xf2], "ands x1, x0, 0x1f"), + ([0x05, 0x10, 0x40, 0xf2], "ands x5, x0, 0x1f"), + ([0x0c, 0x10, 0x40, 0xf2], "ands x12, x0, 0x1f"), + ([0x48, 0x10, 0x40, 0xf2], "ands x8, x2, 0x1f"), + ([0x05, 0x14, 0x40, 0xf2], "ands x5, x0, 0x3f"), + ([0x48, 0x18, 0x40, 0xf2], "ands x8, x2, 0x7f"), + ([0x02, 0xe8, 0x7b, 0xf2], "ands x2, x0, 0xffffffffffffffe0"), + ([0x07, 0xf4, 0x7e, 0xf2], "ands x7, x0, 0xfffffffffffffffc"), + ([0x20, 0x00, 0x00, 0x4a], "eor w0, w1, w0"), + ([0x99, 0x02, 0x04, 0x4a], "eor w25, w20, w4"), + ([0xa6, 0x00, 0x00, 0x52], "eor w6, w5, 0x1"), + ([0x48, 0x01, 0x08, 0xca], "eor x8, x10, x8"), + ([0xe7, 0xf0, 0x7d, 0xd2], "eor x7, x7, 0xfffffffffffffff8"), + ([0x81, 0x12, 0x17, 0x2a], "orr w1, w20, w23, lsl 4"), + ([0x29, 0x03, 0x1c, 0x2a], "orr w9, w25, w28"), + ([0x73, 0x7d, 0x4d, 0x2a], "orr w19, w11, w13, lsr 31"), + ([0x00, 0x7c, 0x54, 0x2a], "orr w0, w0, w20, lsr 31"), + ([0x20, 0x00, 0x00, 0x32], "orr w0, w1, 0x1"), + ([0x80, 0x02, 0x00, 0x32], "orr w0, w20, 0x1"), + ([0x12, 0x00, 0x11, 0x32], "orr w18, w0, 0x8000"), + ([0x1e, 0x00, 0x11, 0x32], "orr w30, w0, 0x8000"), + ([0x08, 0x01, 0x1c, 0x32], "orr w8, w8, 0x10"), + ([0x81, 0x02, 0x1c, 0x32], "orr w1, w20, 0x10"), + ([0x82, 0x02, 0x1c, 0x32], "orr w2, w20, 0x10"), + ([0x94, 0x02, 0x1c, 0x32], "orr w20, w20, 0x10"), + ([0x20, 0x00, 0x1e, 0x32], "orr w0, w1, 0x4"), + ([0x73, 0x02, 0x1e, 0x32], "orr w19, w19, 0x4"), + ([0x80, 0x02, 0x1e, 0x32], "orr w0, w20, 0x4"), + ([0x94, 0x02, 0x1e, 0x32], "orr w20, w20, 0x4"), + ]; + + let errs = run_tests(TESTS); + + for err in errs.iter() { + println!("{}", err); + } + + assert!(errs.is_empty()); +} + +#[test] +fn test_openblas_simd_loadstore() { + const TESTS: &[([u8; 4], &'static str)] = &[ + ([0x88, 0x28, 0x00, 0x0c], "st1 {v8.2s-v11.2s}, [x4]"), + ([0x80, 0x2c, 0x00, 0x0c], "st1 {v0.1d-v3.1d}, [x4]"), + ([0x84, 0x2c, 0x00, 0x0c], "st1 {v4.1d-v7.1d}, [x4]"), + ([0x80, 0x2d, 0x00, 0x0c], "st1 {v0.1d-v3.1d}, [x12]"), + ([0x20, 0x2e, 0x00, 0x0c], "st1 {v0.1d-v3.1d}, [x17]"), + ([0x24, 0x2e, 0x00, 0x0c], "st1 {v4.1d-v7.1d}, [x17]"), + ([0x60, 0x78, 0x00, 0x0c], "st1 {v0.2s}, [x3]"), + ([0x62, 0x78, 0x00, 0x0c], "st1 {v2.2s}, [x3]"), + ([0x88, 0x79, 0x00, 0x0c], "st1 {v8.2s}, [x12]"), + ([0xac, 0x79, 0x00, 0x0c], "st1 {v12.2s}, [x13]"), + ([0xc8, 0x79, 0x00, 0x0c], "st1 {v8.2s}, [x14]"), + ([0xa0, 0x89, 0x00, 0x0c], "st2 {v0.2s, v1.2s}, [x13]"), + ([0xa4, 0x89, 0x00, 0x0c], "st2 {v4.2s, v5.2s}, [x13]"), + ([0x88, 0xa9, 0x00, 0x0c], "st1 {v8.2s, v9.2s}, [x12]"), + ([0xa0, 0xa9, 0x00, 0x0c], "st1 {v0.2s, v1.2s}, [x13]"), + ([0xa2, 0xa9, 0x00, 0x0c], "st1 {v2.2s, v3.2s}, [x13]"), + ([0xac, 0xa9, 0x00, 0x0c], "st1 {v12.2s, v13.2s}, [x13]"), + ([0xc4, 0xa9, 0x00, 0x0c], "st1 {v4.2s, v5.2s}, [x14]"), + ([0xc6, 0xa9, 0x00, 0x0c], "st1 {v6.2s, v7.2s}, [x14]"), + ([0xc8, 0xa9, 0x00, 0x0c], "st1 {v8.2s, v9.2s}, [x14]"), + ([0x88, 0x79, 0x40, 0x0c], "ld1 {v8.2s}, [x12]"), + ([0xac, 0x79, 0x40, 0x0c], "ld1 {v12.2s}, [x13]"), + ([0x68, 0x89, 0x40, 0x0c], "ld2 {v8.2s, v9.2s}, [x11]"), + ([0x6a, 0x89, 0x40, 0x0c], "ld2 {v10.2s, v11.2s}, [x11]"), + ([0xa4, 0x89, 0x40, 0x0c], "ld2 {v4.2s, v5.2s}, [x13]"), + ([0xa2, 0xa9, 0x40, 0x0c], "ld1 {v2.2s, v3.2s}, [x13]"), + ([0xac, 0xa9, 0x40, 0x0c], "ld1 {v12.2s, v13.2s}, [x13]"), + ([0xa5, 0x79, 0x9f, 0x0c], "st1 {v5.2s}, [x13], 8"), + ([0x45, 0x79, 0xc2, 0x0c], "ld1 {v5.2s}, [x10], x2"), + ([0x20, 0x78, 0xdf, 0x0c], "ld1 {v0.2s}, [x1], 8"), + ([0xcc, 0x85, 0x00, 0x0d], "st1 {v12.d}[0], [x14]"), + ([0xa8, 0x91, 0x00, 0x0d], "st1 {v8.s}[1], [x13]"), + ([0xa0, 0x81, 0x20, 0x0d], "st2 {v0.s, v1.s}[0], [x13]"), + ([0x88, 0x81, 0x40, 0x0d], "ld1 {v8.s}[0], [x12]"), + ([0xa8, 0x91, 0x40, 0x0d], "ld1 {v8.s}[1], [x13]"), + ([0xa0, 0x85, 0x60, 0x0d], "ld2 {v0.d, v1.d}[0], [x13]"), + ([0x43, 0x81, 0x82, 0x0d], "st1 {v3.s}[0], [x10], x2"), + ([0x64, 0x90, 0x9f, 0x0d], "st1 {v4.s}[1], [x3], 0x4"), + ([0x22, 0x84, 0xc2, 0x0d], "ld1 {v2.d}[0], [x1], x2"), + ([0x61, 0x80, 0xc4, 0x0d], "ld1 {v1.s}[0], [x3], x4"), + ([0x24, 0xc9, 0xdf, 0x0d], "ld1r {v4.2s}, [x9], 0x4"), + ([0x88, 0x28, 0x00, 0x4c], "st1 {v8.4s-v11.4s}, [x4]"), + ([0x60, 0x2d, 0x00, 0x4c], "st1 {v0.2d-v3.2d}, [x11]"), + ([0x9c, 0x2e, 0x00, 0x4c], "st1 {v28.2d-v31.2d}, [x20]"), + ([0x60, 0x7c, 0x00, 0x4c], "st1 {v0.2d}, [x3]"), + ([0x62, 0x7c, 0x00, 0x4c], "st1 {v2.2d}, [x3]"), + ([0x88, 0x7d, 0x00, 0x4c], "st1 {v8.2d}, [x12]"), + ([0xac, 0x7d, 0x00, 0x4c], "st1 {v12.2d}, [x13]"), + ([0xc8, 0x7d, 0x00, 0x4c], "st1 {v8.2d}, [x14]"), + ([0xe6, 0x8d, 0x00, 0x4c], "st2 {v6.2d, v7.2d}, [x15]"), + ([0x00, 0xa3, 0x00, 0x4c], "st1 {v0.16b, v1.16b}, [x24]"), + ([0x60, 0xa3, 0x00, 0x4c], "st1 {v0.16b, v1.16b}, [x27]"), + ([0x00, 0xa9, 0x00, 0x4c], "st1 {v0.4s, v1.4s}, [x8]"), + ([0x44, 0xad, 0x00, 0x4c], "st1 {v4.2d, v5.2d}, [x10]"), + ([0x66, 0xad, 0x00, 0x4c], "st1 {v6.2d, v7.2d}, [x11]"), + ([0xc8, 0xad, 0x00, 0x4c], "st1 {v8.2d, v9.2d}, [x14]"), + ([0xec, 0xad, 0x00, 0x4c], "st1 {v12.2d, v13.2d}, [x15]"), + ([0x21, 0x28, 0x40, 0x4c], "ld1 {v1.4s-v4.4s}, [x1]"), + ([0x22, 0x2c, 0x40, 0x4c], "ld1 {v2.2d-v5.2d}, [x1]"), + ([0x61, 0x2c, 0x40, 0x4c], "ld1 {v1.2d-v4.2d}, [x3]"), + ([0xb0, 0x2c, 0x40, 0x4c], "ld1 {v16.2d-v19.2d}, [x5]"), + ([0x80, 0x2d, 0x40, 0x4c], "ld1 {v0.2d-v3.2d}, [x12]"), + ([0xa4, 0x2d, 0x40, 0x4c], "ld1 {v4.2d-v7.2d}, [x13]"), + ([0x68, 0x79, 0x40, 0x4c], "ld1 {v8.4s}, [x11]"), + ([0x6c, 0x79, 0x40, 0x4c], "ld1 {v12.4s}, [x11]"), + ([0x00, 0x7e, 0x40, 0x4c], "ld1 {v0.2d}, [x16]"), + ([0x30, 0x88, 0x40, 0x4c], "ld2 {v16.4s, v17.4s}, [x1]"), + ([0x32, 0x88, 0x40, 0x4c], "ld2 {v18.4s, v19.4s}, [x1]"), + ([0x7e, 0x8c, 0x40, 0x4c], "ld2 {v30.2d, v31.2d}, [x3]"), + ([0x68, 0x8d, 0x40, 0x4c], "ld2 {v8.2d, v9.2d}, [x11]"), + ([0x6a, 0x8d, 0x40, 0x4c], "ld2 {v10.2d, v11.2d}, [x11]"), + ([0x6c, 0x8d, 0x40, 0x4c], "ld2 {v12.2d, v13.2d}, [x11]"), + ([0x6e, 0x8d, 0x40, 0x4c], "ld2 {v14.2d, v15.2d}, [x11]"), + ([0x80, 0x8d, 0x40, 0x4c], "ld2 {v0.2d, v1.2d}, [x12]"), + ([0x82, 0x8d, 0x40, 0x4c], "ld2 {v2.2d, v3.2d}, [x12]"), + ([0xa0, 0x8d, 0x40, 0x4c], "ld2 {v0.2d, v1.2d}, [x13]"), + ([0xa4, 0x8d, 0x40, 0x4c], "ld2 {v4.2d, v5.2d}, [x13]"), + ([0xa6, 0x8d, 0x40, 0x4c], "ld2 {v6.2d, v7.2d}, [x13]"), + ([0xa3, 0x7c, 0x86, 0x4c], "st1 {v3.2d}, [x5], x6"), + ([0x61, 0x2c, 0x9f, 0x4c], "st1 {v1.2d-v4.2d}, [x3], 64"), + ([0xb0, 0x2c, 0x9f, 0x4c], "st1 {v16.2d-v19.2d}, [x5], 64"), + ([0x24, 0x78, 0x9f, 0x4c], "st1 {v4.4s}, [x1], 0x10"), + ([0xa5, 0x7d, 0x9f, 0x4c], "st1 {v5.2d}, [x13], 0x10"), + ([0xa4, 0x88, 0x9f, 0x4c], "st2 {v4.4s, v5.4s}, [x5], 32"), + ([0xc4, 0x88, 0x9f, 0x4c], "st2 {v4.4s, v5.4s}, [x6], 32"), + ([0xb0, 0xad, 0x9f, 0x4c], "st1 {v16.2d, v17.2d}, [x13], 32"), + ([0x20, 0x7c, 0xc2, 0x4c], "ld1 {v0.2d}, [x1], x2"), + ([0x46, 0x7d, 0xc6, 0x4c], "ld1 {v6.2d}, [x10], x6"), + ([0x20, 0x0c, 0xdf, 0x4c], "ld4 {v0.2d-v3.2d}, [x1], 64"), + ([0x51, 0x2d, 0xdf, 0x4c], "ld1 {v17.2d-v20.2d}, [x10], 64"), + ([0x20, 0x78, 0xdf, 0x4c], "ld1 {v0.4s}, [x1], 0x10"), + ([0x21, 0x78, 0xdf, 0x4c], "ld1 {v1.4s}, [x1], 0x10"), + ([0x46, 0x7d, 0xdf, 0x4c], "ld1 {v6.2d}, [x10], 0x10"), + ([0x20, 0x88, 0xdf, 0x4c], "ld2 {v0.4s, v1.4s}, [x1], 32"), + ([0x50, 0xad, 0xdf, 0x4c], "ld1 {v16.2d, v17.2d}, [x10], 32"), + ([0xa8, 0x85, 0x00, 0x4d], "st1 {v8.d}[1], [x13]"), + ([0xac, 0x85, 0x00, 0x4d], "st1 {v12.d}[1], [x13]"), + ([0xec, 0x85, 0x00, 0x4d], "st1 {v12.d}[1], [x15]"), + ([0x62, 0x84, 0x40, 0x4d], "ld1 {v2.d}[1], [x3]"), + ([0xa8, 0x85, 0x40, 0x4d], "ld1 {v8.d}[1], [x13]"), + ([0xec, 0x85, 0x40, 0x4d], "ld1 {v12.d}[1], [x15]"), + ([0x64, 0x84, 0x84, 0x4d], "st1 {v4.d}[1], [x3], x4"), + ([0x64, 0x84, 0x9f, 0x4d], "st1 {v4.d}[1], [x3], 8"), + ([0x24, 0xcd, 0xdf, 0x4d], "ld1r {v4.2d}, [x9], 8"), + ([0x60, 0x04, 0x81, 0x3c], "str q0, [x3], 0x10"), + ([0x61, 0x00, 0x9f, 0x3c], "stur q1, [x3, -0x10]"), + ([0xa0, 0x00, 0x9f, 0x3c], "stur q0, [x5, -0x10]"), + ([0x30, 0x00, 0xc0, 0x3c], "ldur q16, [x1]"), + ([0x68, 0x01, 0xc0, 0x3c], "ldur q8, [x11]"), + ([0x6c, 0x01, 0xc0, 0x3c], "ldur q12, [x11]"), + ([0x04, 0x02, 0xc0, 0x3c], "ldur q4, [x16]"), + ([0x32, 0x00, 0xc1, 0x3c], "ldur q18, [x1, 0x10]"), + ([0x6c, 0x01, 0xc1, 0x3c], "ldur q12, [x11, 0x10]"), + ([0x05, 0x02, 0xc1, 0x3c], "ldur q5, [x16, 0x10]"), + ([0x20, 0x04, 0xc1, 0x3c], "ldr q0, [x1], 0x10"), + ([0x04, 0x06, 0xc1, 0x3c], "ldr q4, [x16], 0x10"), + ([0x05, 0x06, 0xc1, 0x3c], "ldr q5, [x16], 0x10"), + ([0x34, 0x00, 0xc2, 0x3c], "ldur q20, [x1, 32]"), + ([0x3c, 0x00, 0xc6, 0x3c], "ldur q28, [x1, 96]"), + ([0x06, 0x02, 0xc6, 0x3c], "ldur q6, [x16, 96]"), + ([0x3e, 0x00, 0xc7, 0x3c], "ldur q30, [x1, 112]"), + ([0x07, 0x02, 0xc7, 0x3c], "ldur q7, [x16, 112]"), + ([0x44, 0x00, 0xdf, 0x3c], "ldur q4, [x2, -0x10]"), + ([0x60, 0x00, 0x80, 0x3d], "str q0, [x3]"), + ([0x61, 0x00, 0x80, 0x3d], "str q1, [x3]"), + ([0xa0, 0x00, 0x80, 0x3d], "str q0, [x5]"), + ([0x60, 0x01, 0x80, 0x3d], "str q0, [x11]"), + ([0x80, 0x01, 0x80, 0x3d], "str q0, [x12]"), + ([0xa1, 0x01, 0x80, 0x3d], "str q1, [x13]"), + ([0xc2, 0x01, 0x80, 0x3d], "str q2, [x14]"), + ([0xe0, 0x01, 0x80, 0x3d], "str q0, [x15]"), + ([0xe3, 0x01, 0x80, 0x3d], "str q3, [x15]"), + ([0x00, 0x02, 0x80, 0x3d], "str q0, [x16]"), + ([0xe0, 0x27, 0x80, 0x3d], "str q0, [sp, 144]"), + ([0xe1, 0x2b, 0x80, 0x3d], "str q1, [sp, 160]"), + ([0xe2, 0x2f, 0x80, 0x3d], "str q2, [sp, 176]"), + ([0xe3, 0x33, 0x80, 0x3d], "str q3, [sp, 192]"), + ([0xe4, 0x37, 0x80, 0x3d], "str q4, [sp, 208]"), + ([0xe5, 0x3b, 0x80, 0x3d], "str q5, [sp, 224]"), + ([0xe6, 0x3f, 0x80, 0x3d], "str q6, [sp, 240]"), + ([0xe7, 0x43, 0x80, 0x3d], "str q7, [sp, 256]"), + ([0x20, 0x00, 0xc0, 0x3d], "ldr q0, [x1]"), + ([0x21, 0x00, 0xc0, 0x3d], "ldr q1, [x1]"), + ([0x24, 0x00, 0xc0, 0x3d], "ldr q4, [x1]"), + ([0x30, 0x00, 0xc0, 0x3d], "ldr q16, [x1]"), + ([0x60, 0x00, 0xc0, 0x3d], "ldr q0, [x3]"), + ([0x78, 0x00, 0xc0, 0x3d], "ldr q24, [x3]"), + ([0xa1, 0x00, 0xc0, 0x3d], "ldr q1, [x5]"), + ([0xc0, 0x00, 0xc0, 0x3d], "ldr q0, [x6]"), + ([0xe1, 0x00, 0xc0, 0x3d], "ldr q1, [x7]"), + ([0xe4, 0x00, 0xc0, 0x3d], "ldr q4, [x7]"), + ([0x02, 0x01, 0xc0, 0x3d], "ldr q2, [x8]"), + ([0x23, 0x01, 0xc0, 0x3d], "ldr q3, [x9]"), + ([0x44, 0x01, 0xc0, 0x3d], "ldr q4, [x10]"), + ([0xc1, 0x05, 0xc0, 0x3d], "ldr q1, [x14, 0x10]"), + ([0xe5, 0x05, 0xc0, 0x3d], "ldr q5, [x15, 0x10]"), + ([0xc2, 0x09, 0xc0, 0x3d], "ldr q2, [x14, 32]"), + ([0xe6, 0x09, 0xc0, 0x3d], "ldr q6, [x15, 32]"), + ([0x33, 0x0c, 0xc0, 0x3d], "ldr q19, [x1, 0x30]"), + ([0x85, 0x0c, 0xc0, 0x3d], "ldr q5, [x4, 0x30]"), + ([0x83, 0x0d, 0xc0, 0x3d], "ldr q3, [x12, 0x30]"), + ([0xa7, 0x0d, 0xc0, 0x3d], "ldr q7, [x13, 0x30]"), + ([0xc3, 0x0d, 0xc0, 0x3d], "ldr q3, [x14, 0x30]"), + ([0xe7, 0x0d, 0xc0, 0x3d], "ldr q7, [x15, 0x30]"), + ([0x00, 0x06, 0xc1, 0xac], "ldp q0, q1, [x16], 32"), + ([0xbe, 0x7d, 0xc1, 0xac], "ldp q30, q31, [x13], 32"), + ([0xb0, 0x44, 0x00, 0xad], "stp q16, q17, [x5]"), + ([0x82, 0x0d, 0x01, 0xad], "stp q2, q3, [x12, 32]"), + ([0xc2, 0x0d, 0x01, 0xad], "stp q2, q3, [x14, 32]"), + ([0xb8, 0x64, 0x02, 0xad], "stp q24, q25, [x5, 64]"), + ([0xba, 0x6c, 0x03, 0xad], "stp q26, q27, [x5, 96]"), + ([0xc0, 0x04, 0x40, 0xad], "ldp q0, q1, [x6]"), + ([0xb0, 0x44, 0x40, 0xad], "ldp q16, q17, [x5]"), + ([0x02, 0x0e, 0x41, 0xad], "ldp q2, q3, [x16, 32]"), + ([0x68, 0x25, 0x41, 0xad], "ldp q8, q9, [x11, 32]"), + ([0x6c, 0x35, 0x41, 0xad], "ldp q12, q13, [x11, 32]"), + ([0x7c, 0x74, 0x42, 0xad], "ldp q28, q29, [x3, 64]"), + ([0x02, 0x0e, 0x43, 0xad], "ldp q2, q3, [x16, 96]"), + ([0x06, 0x1e, 0x43, 0xad], "ldp q6, q7, [x16, 96]"), + ([0x30, 0x44, 0x43, 0xad], "ldp q16, q17, [x1, 96]"), + ([0x3e, 0x7c, 0x47, 0xad], "ldp q30, q31, [x1, 224]"), + ([0x6e, 0x3d, 0xc1, 0x6c], "ldp d14, d15, [x11], 0x10"), + ([0xe8, 0x27, 0xc4, 0x6c], "ldp d8, d9, [sp], 64"), + ([0x81, 0x01, 0x00, 0x6d], "stp d1, d0, [x12]"), + ([0x39, 0x60, 0x00, 0x6d], "stp d25, d24, [x1]"), + ([0xdd, 0x71, 0x01, 0x6d], "stp d29, d28, [x14, 0x10]"), + ([0xfd, 0x70, 0x31, 0x6d], "stp d29, d28, [x7, -0xf0]"), + ([0x86, 0xd8, 0x3f, 0x6d], "stp d6, d22, [x4, -0x8]"), + ([0x24, 0x00, 0x40, 0x6d], "ldp d4, d0, [x1]"), + ([0xe1, 0x02, 0x40, 0x6d], "ldp d1, d0, [x23]"), + ([0xb2, 0x45, 0x7f, 0x6d], "ldp d18, d17, [x13, -0x10]"), + ([0x32, 0xc4, 0x7f, 0x6d], "ldp d18, d17, [x1, -0x8]"), + ([0x33, 0xd0, 0x7f, 0x6d], "ldp d19, d20, [x1, -0x8]"), + ([0x00, 0x84, 0x00, 0xfc], "str d0, [x0], 0x8"), + ([0xc2, 0x87, 0x00, 0xfc], "str d2, [x30], 0x8"), + ([0x40, 0x8f, 0x00, 0xfc], "str d0, [x26, 0x8]!"), + ([0x15, 0x04, 0x01, 0xfc], "str d21, [x0], 0x10"), + ([0xc2, 0x07, 0x01, 0xfc], "str d2, [x30], 0x10"), + ([0x20, 0x84, 0x02, 0xfc], "str d0, [x1], 0x28"), + ([0x66, 0x00, 0x15, 0xfc], "stur d6, [x3, -0xb0]"), + ([0x29, 0x03, 0x1f, 0xfc], "stur d9, [x25, -0x10]"), + ([0xe8, 0x0f, 0x1f, 0xfc], "str d8, [sp, -0x10]!"), + ([0x6d, 0x83, 0x1f, 0xfc], "stur d13, [x27, -0x8]"), + ([0xc3, 0x83, 0x1f, 0xfc], "stur d3, [x30, -0x8]"), + ([0x0d, 0x84, 0x1f, 0xfc], "str d13, [x0], -0x8"), + ([0x61, 0x8e, 0x1f, 0xfc], "str d1, [x19, -0x8]!"), + ([0xa3, 0x68, 0x20, 0xfc], "str d3, [x5, x0]"), + ([0xd2, 0x68, 0x20, 0xfc], "str d18, [x6, x0]"), + ([0x94, 0x6b, 0x20, 0xfc], "str d20, [x28, x0]"), + ([0x48, 0x78, 0x20, 0xfc], "str d8, [x2, x0, lsl 3]"), + ([0x83, 0x7b, 0x77, 0xfc], "ldr d3, [x28, x23, lsl 3]"), + ([0x80, 0xda, 0x77, 0xfc], "ldr d0, [x20, w23, sxtw 3]"), + ([0x22, 0x6b, 0x78, 0xfc], "ldr d2, [x25, x24]"), + ([0x08, 0x78, 0x78, 0xfc], "ldr d8, [x0, x24, lsl 3]"), + ([0xf3, 0x9b, 0x41, 0xfd], "ldr d19, [sp, 0x330]"), + ([0x4a, 0x44, 0x47, 0xfd], "ldr d10, [x2, 3720]"), + ([0x01, 0xd0, 0x47, 0xfd], "ldr d1, [x0, 4000]"), + ([0xe8, 0x27, 0xbc, 0x6d], "stp d8, d9, [sp, -0x40]!"), + ([0xe8, 0x27, 0xbf, 0x6d], "stp d8, d9, [sp, -0x10]!"), + ([0xa2, 0x0f, 0x32, 0x3d], "str b2, [x29, 3203]"), + ([0x97, 0xff, 0xff, 0x1c], "ldr s23, 8b95f8 <zscal_k_TSV110@@Base+0x10>"), + ([0x6e, 0x3d, 0xc1, 0x2c], "ldp s14, s15, [x11], 8"), + ([0xef, 0x7e, 0x40, 0x2d], "ldp s15, s31, [x23]"), + ([0x41, 0x80, 0x40, 0x2d], "ldp s1, s0, [x2, 0x4]"), + ([0x33, 0xd0, 0x7f, 0x2d], "ldp s19, s20, [x1, -0x4]"), + ([0x60, 0x47, 0x00, 0xbc], "str s0, [x27], 0x4"), + ([0x81, 0x47, 0x00, 0xbc], "str s1, [x28], 0x4"), + ([0xc2, 0x47, 0x00, 0xbc], "str s2, [x30], 0x4"), + ([0x00, 0x4d, 0x00, 0xbc], "str s0, [x8, 0x4]!"), + ([0x40, 0x4f, 0x00, 0xbc], "str s0, [x26, 0x4]!"), + ([0x15, 0x84, 0x00, 0xbc], "str s21, [x0], 0x8"), + ([0x80, 0x87, 0x00, 0xbc], "str s0, [x28], 0x8"), + ([0xf1, 0x00, 0x1c, 0xbc], "stur s17, [x7, -0x40]"), + ([0x01, 0x04, 0x1f, 0xbc], "str s1, [x0], -0x10"), + ([0x8d, 0xc3, 0x1f, 0xbc], "stur s13, [x28, -0x4]"), + ([0x0d, 0xc4, 0x1f, 0xbc], "str s13, [x0], -0x4"), + ([0x61, 0xce, 0x1f, 0xbc], "str s1, [x19, -0x4]!"), + ([0x82, 0x6b, 0x20, 0xbc], "str s2, [x28, x0]"), + ([0xad, 0x7a, 0x20, 0xbc], "str s13, [x21, x0, lsl 2]"), + ([0x88, 0x7b, 0x21, 0xbc], "str s8, [x28, x1, lsl 2]"), + ([0x60, 0xda, 0x21, 0xbc], "str s0, [x19, w1, sxtw 2]"), + ([0xc0, 0x5a, 0x32, 0xbc], "str s0, [x22, w18, uxtw 2]"), + ([0xb3, 0x85, 0x40, 0xbc], "ldr s19, [x13], 8"), + ([0x02, 0x8f, 0x40, 0xbc], "ldr s2, [x24, 8]!"), + ([0x71, 0x05, 0x41, 0xbc], "ldr s17, [x11], 0x10"), + ([0xe0, 0x42, 0x5d, 0xbc], "ldur s0, [x23, -0x2c]"), + ([0x27, 0x80, 0x5d, 0xbc], "ldur s7, [x1, -0x28]"), + ([0x21, 0x84, 0x5f, 0xbc], "ldr s1, [x1], -0x8"), + ([0x22, 0x87, 0x5f, 0xbc], "ldr s2, [x25], -0x8"), + ([0xe0, 0x8e, 0x5f, 0xbc], "ldr s0, [x23, -0x8]!"), + ([0x79, 0x6a, 0x60, 0xbc], "ldr s25, [x19, x0]"), + ([0x6a, 0x78, 0x60, 0xbc], "ldr s10, [x3, x0, lsl 2]"), + ([0xeb, 0x7a, 0x60, 0xbc], "ldr s11, [x23, x0, lsl 2]"), + ([0x96, 0x7b, 0x61, 0xbc], "ldr s22, [x28, x1, lsl 2]"), + ([0x68, 0xdb, 0x61, 0xbc], "ldr s8, [x27, w1, sxtw 2]"), + ([0x49, 0x03, 0x00, 0xbd], "str s9, [x26]"), + ([0x62, 0x78, 0x00, 0xbd], "str s2, [x3, 0x78]"), + ([0x6c, 0x78, 0x00, 0xbd], "str s12, [x3, 0x78]"), + ([0xea, 0x7f, 0x00, 0xbd], "str s10, [sp, 124]"), + ([0xe9, 0x87, 0x04, 0xbd], "str s9, [sp, 1156]"), + ([0xc0, 0x8a, 0x0a, 0xbd], "str s0, [x22, 2696]"), + ([0x41, 0x58, 0x4f, 0xbd], "ldr s1, [x2, 3928]"), + ([0x88, 0x7c, 0x00, 0x2d], "stp s8, s31, [x4]"), + ([0x03, 0x84, 0x00, 0x2d], "stp s3, s1, [x0, 0x4]"), + ([0x13, 0xdc, 0x3f, 0x2d], "stp s19, s23, [x0, -0x4]"), + ([0x03, 0xe4, 0x00, 0x2f], "movi d3, 0x0"), + ([0x00, 0xd8, 0x21, 0x5e], "scvtf s0, s0"), + ([0x82, 0xd8, 0x61, 0x5e], "scvtf d2, d4"), + ([0x01, 0x00, 0x62, 0x9e], "scvtf d1, x0"), + ([0x03, 0x00, 0x62, 0x9e], "scvtf d3, x0"), + ([0x69, 0x03, 0x62, 0x9e], "scvtf d9, x27"), + ([0x88, 0x03, 0x62, 0x9e], "scvtf d8, x28"), + ([0x22, 0x00, 0x22, 0x1e], "scvtf s2, w1"), + ([0xac, 0x02, 0x22, 0x1e], "scvtf s12, w21"), + ([0x00, 0x00, 0x62, 0x1e], "scvtf d0, w0"), + ([0x8a, 0x03, 0x62, 0x1e], "scvtf d10, w28"), + ]; + + let errs = run_tests(TESTS); + + for err in errs.iter() { + println!("{}", err); + } + + assert!(errs.is_empty()); +} + +#[test] +fn test_openblas_loadstore() { + const TESTS: &[([u8; 4], &'static str)] = &[ + ([0x83, 0x68, 0x60, 0x38], "ldrb w3, [x4, x0]"), + ([0x63, 0x03, 0x40, 0x29], "ldp w3, w0, [x27]"), + ([0x49, 0x00, 0x40, 0x39], "ldrb w9, [x2]"), + ([0x4a, 0x00, 0x40, 0x39], "ldrb w10, [x2]"), + ([0x82, 0x08, 0x40, 0x39], "ldrb w2, [x4, 0x2]"), + ([0x82, 0x0a, 0x40, 0x39], "ldrb w2, [x20, 0x2]"), + ([0x41, 0x0b, 0x40, 0x39], "ldrb w1, [x26, 0x2]"), + ([0x88, 0x10, 0x40, 0x39], "ldrb w8, [x4, 0x4]"), + ([0xc1, 0x16, 0x40, 0x39], "ldrb w1, [x22, 0x5]"), + ([0xc2, 0x16, 0x40, 0x39], "ldrb w2, [x22, 0x5]"), + ([0x00, 0x40, 0x40, 0x39], "ldrb w0, [x0, 0x10]"), + ([0xe1, 0x43, 0x41, 0x39], "ldrb w1, [sp, 0x50]"), + ([0xe0, 0x83, 0x41, 0x39], "ldrb w0, [sp, 0x60]"), + ([0xe1, 0x8b, 0x41, 0x39], "ldrb w1, [sp, 0x62]"), + ([0xe2, 0x8b, 0x41, 0x39], "ldrb w2, [sp, 0x62]"), + ([0xe2, 0xa3, 0x41, 0x39], "ldrb w2, [sp, 0x68]"), + ([0xe1, 0xb7, 0x41, 0x39], "ldrb w1, [sp, 0x6d]"), + ([0xe2, 0xb7, 0x41, 0x39], "ldrb w2, [sp, 0x6d]"), + ([0xf3, 0xcb, 0x41, 0x39], "ldrb w19, [sp, 0x72]"), + ([0xe2, 0xeb, 0x41, 0x39], "ldrb w2, [sp, 0x7a]"), + ([0xe1, 0x13, 0x42, 0x39], "ldrb w1, [sp, 0x84]"), + ([0xe3, 0x13, 0x42, 0x39], "ldrb w3, [sp, 0x84]"), + ([0xe7, 0x23, 0x42, 0x39], "ldrb w7, [sp, 0x88]"), + ([0xe3, 0x37, 0x42, 0x39], "ldrb w3, [sp, 0x8d]"), + ([0xe1, 0x4f, 0x42, 0x39], "ldrb w1, [sp, 0x93]"), + ([0x60, 0x02, 0x72, 0x39], "ldrb w0, [x19, 0xc80]"), + ([0x22, 0x04, 0x40, 0x69], "ldpsw x2, x1, [x1]"), + ([0x38, 0x15, 0x40, 0x69], "ldpsw x24, x5, [x9]"), + ([0x38, 0x1d, 0x40, 0x69], "ldpsw x24, x7, [x9]"), + ([0x43, 0x80, 0x40, 0x69], "ldpsw x3, x0, [x2, 0x4]"), + ([0x05, 0x90, 0x40, 0x69], "ldpsw x5, x4, [x0, 0x4]"), + ([0x39, 0x90, 0x40, 0x69], "ldpsw x25, x4, [x1, 0x4]"), + ([0x05, 0x98, 0x40, 0x69], "ldpsw x5, x6, [x0, 0x4]"), + ([0xe1, 0x13, 0x5f, 0x69], "ldpsw x1, x4, [sp, 0xf8]"), + ([0xe3, 0x33, 0x45, 0x78], "ldurh w3, [sp, 0x53]"), + ([0xe1, 0x13, 0x46, 0x78], "ldurh w1, [sp, 0x61]"), + ([0xe2, 0x93, 0x46, 0x78], "ldurh w2, [sp, 0x69]"), + ([0xe2, 0xb3, 0x46, 0x78], "ldurh w2, [sp, 0x6b]"), + ([0xe5, 0xb3, 0x48, 0x78], "ldurh w5, [sp, 0x8b]"), + ([0xe5, 0x8b, 0x00, 0x79], "strh w5, [sp, 0x44]"), + ([0xe2, 0xa3, 0x00, 0x79], "strh w2, [sp, 0x50]"), + ([0xe0, 0xab, 0x00, 0x79], "strh w0, [sp, 0x54]"), + ([0xe1, 0xb3, 0x00, 0x79], "strh w1, [sp, 0x58]"), + ([0xe2, 0xc3, 0x00, 0x79], "strh w2, [sp, 0x60]"), + ([0xe5, 0xf3, 0x00, 0x79], "strh w5, [sp, 0x78]"), + ([0x01, 0x00, 0x40, 0x79], "ldrh w1, [x0]"), + ([0x02, 0x00, 0x40, 0x79], "ldrh w2, [x0]"), + ([0x22, 0x00, 0x40, 0x79], "ldrh w2, [x1]"), + ([0x83, 0x02, 0x40, 0x79], "ldrh w3, [x20]"), + ([0x42, 0x03, 0x40, 0x79], "ldrh w2, [x26]"), + ([0x20, 0x08, 0x40, 0x79], "ldrh w0, [x1, 0x4]"), + ([0xe0, 0xa3, 0x40, 0x79], "ldrh w0, [sp, 0x50]"), + ([0xe1, 0xa3, 0x40, 0x79], "ldrh w1, [sp, 0x50]"), + ([0xe0, 0xb3, 0x40, 0x79], "ldrh w0, [sp, 0x58]"), + ([0xe0, 0xc3, 0x40, 0x79], "ldrh w0, [sp, 0x60]"), + ([0xe2, 0xc3, 0x40, 0x79], "ldrh w2, [sp, 0x60]"), + ([0xe3, 0xf3, 0x40, 0x79], "ldrh w3, [sp, 0x78]"), + ([0xfd, 0x7b, 0xc1, 0xa8], "ldp x29, x30, [sp], 0x10"), + ([0xfd, 0x7b, 0xd9, 0xa8], "ldp x29, x30, [sp], 0x190"), + ([0xfd, 0x7b, 0xda, 0xa8], "ldp x29, x30, [sp], 0x1a0"), + ([0x3c, 0x61, 0x00, 0xa9], "stp x28, x24, [x9]"), + ([0xba, 0x61, 0x00, 0xa9], "stp x26, x24, [x13]"), + ([0xed, 0x6f, 0x00, 0xa9], "stp x13, x27, [sp]"), + ([0xea, 0x9f, 0x1c, 0xa9], "stp x10, x7, [sp, 0x1c8]"), + ([0xff, 0x7c, 0x3e, 0xa9], "stp xzr, xzr, [x7, -0x20]"), + ([0x03, 0x00, 0x40, 0xa9], "ldp x3, x0, [x0]"), + ([0x2a, 0x00, 0x40, 0xa9], "ldp x10, x0, [x1]"), + ([0x5a, 0x28, 0x40, 0xa9], "ldp x26, x10, [x2]"), + ([0x16, 0x87, 0x7f, 0xa9], "ldp x22, x1, [x24, -0x8]"), + ([0xc1, 0x46, 0x00, 0xb8], "str w1, [x22], 0x4"), + ([0x3f, 0x47, 0x00, 0xb8], "str wzr, [x25], 0x4"), + ([0x7f, 0x4c, 0x00, 0xb8], "str wzr, [x3, 0x4]!"), + ([0x1f, 0x4d, 0x00, 0xb8], "str wzr, [x8, 0x4]!"), + ([0x3f, 0x84, 0x00, 0xb8], "str wzr, [x1], 0x8"), + ([0x7f, 0x0e, 0x04, 0xb8], "str wzr, [x19, 0x40]!"), + ([0x9f, 0x0e, 0x04, 0xb8], "str wzr, [x20, 0x40]!"), + ([0xe9, 0x93, 0x08, 0xb8], "stur w9, [sp, 0x89]"), + ([0x1f, 0x40, 0x10, 0xb8], "stur wzr, [x0, -0xfc]"), + ([0x1f, 0xc0, 0x11, 0xb8], "stur wzr, [x0, -0xe4]"), + ([0xff, 0x82, 0x14, 0xb8], "stur wzr, [x23, -0xb8]"), + ([0x0b, 0x83, 0x14, 0xb8], "stur w11, [x24, -0xb8]"), + ([0xff, 0x69, 0x20, 0xb8], "str wzr, [x15, x0]"), + ([0x57, 0x78, 0x20, 0xb8], "str w23, [x2, x0, lsl 2]"), + ([0x7f, 0x68, 0x21, 0xb8], "str wzr, [x3, x1]"), + ([0xa8, 0x68, 0x21, 0xb8], "str w8, [x5, x1]"), + ([0xe0, 0xda, 0x34, 0xb8], "str w0, [x23, w20, sxtw 2]"), + ([0x48, 0x7b, 0x35, 0xb8], "str w8, [x26, x21, lsl 2]"), + ([0x23, 0x44, 0x40, 0xb8], "ldr w3, [x1], 0x4"), + ([0xe1, 0x13, 0x45, 0xb8], "ldur w1, [sp, 0x51]"), + ([0xe2, 0xf3, 0x48, 0xb8], "ldur w2, [sp, 0x8f]"), + ([0x27, 0x80, 0x5f, 0xb8], "ldur w7, [x1, -0x8]"), + ([0xe3, 0x68, 0x61, 0xb8], "ldr w3, [x7, x1]"), + ([0x0e, 0x79, 0x61, 0xb8], "ldr w14, [x8, x1, lsl 2]"), + ([0xf2, 0x6a, 0x62, 0xb8], "ldr w18, [x23, x2]"), + ([0x63, 0x7b, 0x7c, 0xb8], "ldr w3, [x27, x28, lsl 2]"), + ([0xe5, 0x44, 0x80, 0xb8], "ldrsw x5, [x7], 0x4"), + ([0xa3, 0xc0, 0x9f, 0xb8], "ldursw x3, [x5, -0x4]"), + ([0x04, 0x78, 0xa2, 0xb8], "ldrsw x4, [x0, x2, lsl 2]"), + ([0xe4, 0x68, 0xa3, 0xb8], "ldrsw x4, [x7, x3]"), + ([0xe5, 0x6a, 0xb3, 0xb8], "ldrsw x5, [x23, x19]"), + ([0x65, 0x02, 0x00, 0xb9], "str w5, [x19]"), + ([0xff, 0x03, 0x00, 0xb9], "str wzr, [sp]"), + ([0xe9, 0x23, 0x00, 0xb9], "str w9, [sp, 0x20]"), + ([0x01, 0x3c, 0x0c, 0xb9], "str w1, [x0, 0xc3c]"), + ([0x60, 0x4a, 0x0c, 0xb9], "str w0, [x19, 0xc48]"), + ([0x7f, 0x4a, 0x0c, 0xb9], "str wzr, [x19, 0xc48]"), + ([0x40, 0xf0, 0x0d, 0xb9], "str w0, [x2, 0xdf0]"), + ([0x89, 0x00, 0x40, 0xb9], "ldr w9, [x4]"), + ([0xa0, 0x6f, 0x40, 0xb9], "ldr w0, [x29, 0x6c]"), + ([0xe4, 0x6f, 0x40, 0xb9], "ldr w4, [sp, 0x6c]"), + ([0xea, 0xc3, 0x6e, 0xb9], "ldr w10, [sp, 0x2ec0]"), + ([0xeb, 0xc3, 0x6e, 0xb9], "ldr w11, [sp, 0x2ec0]"), + ([0xea, 0xcb, 0x6e, 0xb9], "ldr w10, [sp, 0x2ec8]"), + ([0xf9, 0x03, 0x6f, 0xb9], "ldr w25, [sp, 0x2f00]"), + ([0xf9, 0x0b, 0x6f, 0xb9], "ldr w25, [sp, 0x2f08]"), + ([0xf7, 0x33, 0x6f, 0xb9], "ldr w23, [sp, 0x2f30]"), + ([0xf7, 0x3b, 0x6f, 0xb9], "ldr w23, [sp, 0x2f38]"), + ([0xec, 0x1b, 0x71, 0xb9], "ldr w12, [sp, 0x3118]"), + ([0xeb, 0x23, 0x71, 0xb9], "ldr w11, [sp, 0x3120]"), + ([0xe0, 0x73, 0x71, 0xb9], "ldr w0, [sp, 0x3170]"), + ([0xe0, 0x7b, 0x71, 0xb9], "ldr w0, [sp, 0x3178]"), + ([0xe0, 0x83, 0x71, 0xb9], "ldr w0, [sp, 0x3180]"), + ([0xe0, 0x8b, 0x71, 0xb9], "ldr w0, [sp, 0x3188]"), + ([0xe9, 0x1b, 0x72, 0xb9], "ldr w9, [sp, 0x3218]"), + ([0xeb, 0x23, 0x72, 0xb9], "ldr w11, [sp, 0x3220]"), + ([0xea, 0x2b, 0x72, 0xb9], "ldr w10, [sp, 0x3228]"), + ([0xeb, 0x33, 0x72, 0xb9], "ldr w11, [sp, 0x3230]"), + ([0x24, 0x03, 0x80, 0xb9], "ldrsw x4, [x25]"), + ([0x34, 0x10, 0x80, 0xb9], "ldrsw x20, [x1, 0x10]"), + ([0xc8, 0x7c, 0x89, 0xb9], "ldrsw x8, [x6, 0x97c]"), + ([0x7f, 0xfe, 0x01, 0xc8], "stlxr w1, xzr, [x19]"), + ([0xbf, 0xfe, 0x01, 0xc8], "stlxr w1, xzr, [x21]"), + ([0x64, 0x7e, 0x03, 0xc8], "stxr w3, x4, [x19]"), + ([0x60, 0xfe, 0x5f, 0xc8], "ldaxr x0, [x19]"), + ([0x62, 0xfe, 0x5f, 0xc8], "ldaxr x2, [x19]"), + ([0x3f, 0xfc, 0x9f, 0xc8], "stlr xzr, [x1]"), + ([0x81, 0xfd, 0x9f, 0xc8], "stlr x1, [x12]"), + ([0x77, 0xfe, 0xdf, 0xc8], "ldar x23, [x19]"), + ([0xff, 0x42, 0x00, 0xf8], "stur xzr, [x23, 0x4]"), + ([0x3f, 0x84, 0x00, 0xf8], "str xzr, [x1], 0x8"), + ([0x3f, 0x87, 0x00, 0xf8], "str xzr, [x25], 0x8"), + ([0x1f, 0x8c, 0x00, 0xf8], "str xzr, [x0, 0x8]!"), + ([0x02, 0x8f, 0x00, 0xf8], "str x2, [x24, 0x8]!"), + ([0x20, 0xc0, 0x00, 0xf8], "stur x0, [x1, 0xc]"), + ([0x3f, 0x04, 0x01, 0xf8], "str xzr, [x1], 0x10"), + ([0xe1, 0xc3, 0x07, 0xf8], "stur x1, [sp, 0x7c]"), + ([0x7f, 0x80, 0x10, 0xf8], "stur xzr, [x3, -0xf8]"), + ([0xc3, 0x00, 0x1c, 0xf8], "stur x3, [x6, -0x40]"), + ([0x02, 0x87, 0x1f, 0xf8], "str x2, [x24], -0x8"), + ([0x1f, 0xc3, 0x1f, 0xf8], "stur xzr, [x24, -0x4]"), + ([0xbf, 0x68, 0x20, 0xf8], "str xzr, [x5, x0]"), + ([0xff, 0x6a, 0x38, 0xf8], "str xzr, [x23, x24]"), + ([0x7f, 0x7b, 0x3c, 0xf8], "str xzr, [x27, x28, lsl 3]"), + ([0xc7, 0x40, 0x40, 0xf8], "ldur x7, [x6, 0x4]"), + ([0x01, 0x87, 0x40, 0xf8], "ldr x1, [x24], 0x8"), + ([0x98, 0x83, 0x5f, 0xf8], "ldur x24, [x28, -0x8]"), + ([0x80, 0x86, 0x5f, 0xf8], "ldr x0, [x20], -0x8"), + ([0xe3, 0x6a, 0x60, 0xf8], "ldr x3, [x23, x0]"), + ([0xc5, 0x6a, 0x7b, 0xf8], "ldr x5, [x22, x27]"), + ([0xa0, 0x7a, 0x7c, 0xf8], "ldr x0, [x21, x28, lsl 3]"), + ([0x08, 0xdb, 0x7c, 0xf8], "ldr x8, [x24, w28, sxtw 3]"), + ([0x5f, 0x00, 0x00, 0xf9], "str xzr, [x2]"), + ([0x2a, 0x6d, 0x04, 0xf9], "str x10, [x9, 0x8d8]"), + ([0xeb, 0x1f, 0x18, 0xf9], "str x11, [sp, 0x3038]"), + ([0xff, 0x7f, 0x18, 0xf9], "str xzr, [sp, 0x30f8]"), + ([0xe9, 0x2f, 0x25, 0xf9], "str x9, [sp, 0x4a58]"), + ([0xe1, 0xd7, 0x26, 0xf9], "str x1, [sp, 0x4da8]"), + ([0x42, 0x01, 0x40, 0xf9], "ldr x2, [x10]"), + ([0xf7, 0xef, 0x66, 0xf9], "ldr x23, [sp, 0x4dd8]"), + ([0x01, 0x00, 0x00, 0xb0], "adrp x1, $+0x1000"), + ([0xea, 0x6e, 0x00, 0xb0], "adrp x10, $+0xddd000"), + ([0x13, 0x70, 0x00, 0xb0], "adrp x19, $+0xe01000"), + ([0x20, 0x00, 0x75, 0x10], "adr x0, $+0xea004"), + ([0x01, 0x00, 0x00, 0x90], "adrp x1, $+0x0"), + ([0xc7, 0x6e, 0x00, 0x90], "adrp x7, $+0xdd8000"), + ([0x01, 0x00, 0x00, 0xd0], "adrp x1, $+0x2000"), + ([0x13, 0x70, 0x00, 0xd0], "adrp x19, $+0xe02000"), + ([0xfc, 0xff, 0xff, 0xf0], "adrp x28, $-0x1000"), + ([0xfd, 0x7b, 0xa1, 0xa9], "stp x29, x30, [sp, -0x1f0]!"), + ([0xfd, 0x7b, 0xa2, 0xa9], "stp x29, x30, [sp, -0x1e0]!"), + ([0xfd, 0x7b, 0xbf, 0xa9], "stp x29, x30, [sp, -0x10]!"), + ([0x23, 0xf0, 0x1f, 0x38], "sturb w3, [x1, -0x1]"), + ([0x01, 0x00, 0x00, 0x39], "strb w1, [x0]"), + ([0x20, 0x00, 0x00, 0x39], "strb w0, [x1]"), + ([0x80, 0x00, 0x00, 0x39], "strb w0, [x4]"), + ([0x60, 0x02, 0x00, 0x39], "strb w0, [x19]"), + ([0x80, 0x02, 0x00, 0x39], "strb w0, [x20]"), + ([0xa0, 0x02, 0x00, 0x39], "strb w0, [x21]"), + ([0xe0, 0x02, 0x00, 0x39], "strb w0, [x23]"), + ([0x20, 0x03, 0x00, 0x39], "strb w0, [x25]"), + ([0x40, 0x03, 0x00, 0x39], "strb w0, [x26]"), + ([0x68, 0x13, 0x00, 0x39], "strb w8, [x27, 0x4]"), + ([0xe0, 0x03, 0x01, 0x39], "strb w0, [sp, 0x40]"), + ([0xe1, 0x43, 0x01, 0x39], "strb w1, [sp, 0x50]"), + ([0xe1, 0x8b, 0x01, 0x39], "strb w1, [sp, 0x62]"), + ([0xe2, 0xa3, 0x01, 0x39], "strb w2, [sp, 0x68]"), + ([0xe5, 0x03, 0x0c, 0x39], "strb w5, [sp, 0x300]"), + ([0xe0, 0xe3, 0x0d, 0x39], "strb w0, [sp, 0x378]"), + ([0xe1, 0xe3, 0x0d, 0x39], "strb w1, [sp, 0x378]"), + ([0xe2, 0xe3, 0x0d, 0x39], "strb w2, [sp, 0x378]"), + ([0xe0, 0x03, 0x0e, 0x39], "strb w0, [sp, 0x380]"), + ([0xe0, 0x23, 0x0e, 0x39], "strb w0, [sp, 0x388]"), + ([0xe1, 0x23, 0x0e, 0x39], "strb w1, [sp, 0x388]"), + ([0xe2, 0x23, 0x0e, 0x39], "strb w2, [sp, 0x388]"), + ([0xe0, 0x43, 0x0e, 0x39], "strb w0, [sp, 0x390]"), + ([0xe0, 0x23, 0x0f, 0x39], "strb w0, [sp, 0x3c8]"), + ([0xe1, 0x23, 0x0f, 0x39], "strb w1, [sp, 0x3c8]"), + ([0xe2, 0x23, 0x0f, 0x39], "strb w2, [sp, 0x3c8]"), + ([0xe0, 0x43, 0x0f, 0x39], "strb w0, [sp, 0x3d0]"), + ([0xe0, 0xa3, 0x0f, 0x39], "strb w0, [sp, 0x3e8]"), + ([0xe1, 0xa3, 0x0f, 0x39], "strb w1, [sp, 0x3e8]"), + ([0xe2, 0xa3, 0x0f, 0x39], "strb w2, [sp, 0x3e8]"), + ([0xe0, 0xc3, 0x0f, 0x39], "strb w0, [sp, 0x3f0]"), + ([0x60, 0x02, 0x32, 0x39], "strb w0, [x19, 0xc80]"), + ([0x13, 0xb8, 0x00, 0x29], "stp w19, w14, [x0, 0x4]"), + ([0xeb, 0x9f, 0x1f, 0x29], "stp w11, w7, [sp, 0xfc]"), + ([0x7f, 0x7c, 0x3f, 0x29], "stp wzr, wzr, [x3, -0x8]"), + ]; + + let errs = run_tests(TESTS); + + for err in errs.iter() { + println!("{}", err); + } + + assert!(errs.is_empty()); +} + +#[test] +fn test_openblas_data_ops() { + const TESTS: &[([u8; 4], &'static str)] = &[ + ([0xe5, 0x03, 0x1e, 0x2a], "mov w5, w30"), + ([0xe7, 0x03, 0x1e, 0x2a], "mov w7, w30"), + ([0xe0, 0x03, 0x1f, 0x2a], "mov w0, wzr"), + ([0x20, 0x0f, 0x80, 0x52], "mov w0, 0x79"), + ([0x41, 0x0f, 0x80, 0x52], "mov w1, 0x7a"), + ([0x01, 0x10, 0x80, 0x52], "mov w1, 0x80"), + ([0x02, 0x10, 0x80, 0x52], "mov w2, 0x80"), + ([0x08, 0x10, 0x80, 0x52], "mov w8, 0x80"), + ([0x0a, 0x10, 0x80, 0x52], "mov w10, 0x80"), + ([0x13, 0x10, 0x80, 0x52], "mov w19, 0x80"), + ([0x15, 0x10, 0x80, 0x52], "mov w21, 0x80"), + ([0x60, 0x10, 0x80, 0x52], "mov w0, 0x83"), + ([0x81, 0x10, 0x80, 0x52], "mov w1, 0x84"), + ([0x05, 0x16, 0x80, 0x52], "mov w5, 0xb0"), + ([0x06, 0x16, 0x80, 0x52], "mov w6, 0xb0"), + ([0x08, 0x16, 0x80, 0x52], "mov w8, 0xb0"), + ([0x00, 0x18, 0x80, 0x52], "mov w0, 0xc0"), + ([0x82, 0x18, 0x80, 0x52], "mov w2, 0xc4"), + ([0x60, 0x1a, 0x80, 0x52], "mov w0, 0xd3"), + ([0x80, 0x1a, 0x80, 0x52], "mov w0, 0xd4"), + ([0xa0, 0x1a, 0x80, 0x52], "mov w0, 0xd5"), + ([0xc0, 0x1a, 0x80, 0x52], "mov w0, 0xd6"), + ([0x22, 0x1e, 0x80, 0x52], "mov w2, 0xf1"), + ([0x00, 0x20, 0x80, 0x52], "mov w0, 0x100"), + ([0x02, 0x20, 0x80, 0x52], "mov w2, 0x100"), + ([0x20, 0x20, 0x80, 0x52], "mov w0, 0x101"), + ([0x22, 0x20, 0x80, 0x52], "mov w2, 0x101"), + ([0x80, 0x20, 0x80, 0x52], "mov w0, 0x104"), + ([0xa0, 0x20, 0x80, 0x52], "mov w0, 0x105"), + ([0xc2, 0x21, 0x80, 0x52], "mov w2, 0x10e"), + ([0x42, 0x22, 0x80, 0x52], "mov w2, 0x112"), + ([0xc6, 0x31, 0x80, 0x52], "mov w6, 0x18e"), + ([0x25, 0x60, 0x80, 0x52], "mov w5, 0x301"), + ([0x00, 0x82, 0x80, 0x52], "mov w0, 0x410"), + ([0x20, 0x82, 0x80, 0x52], "mov w0, 0x411"), + ([0x80, 0x82, 0x80, 0x52], "mov w0, 0x414"), + ([0xa0, 0x82, 0x80, 0x52], "mov w0, 0x415"), + ([0x60, 0x9e, 0x80, 0x52], "mov w0, 0x4f3"), + ([0x02, 0x00, 0x81, 0x52], "mov w2, 0x800"), + ([0x00, 0x02, 0x81, 0x52], "mov w0, 0x810"), + ([0x20, 0x02, 0x81, 0x52], "mov w0, 0x811"), + ([0x80, 0x02, 0x81, 0x52], "mov w0, 0x814"), + ([0xa0, 0x02, 0x81, 0x52], "mov w0, 0x815"), + ([0x00, 0x82, 0x81, 0x52], "mov w0, 0xc10"), + ([0x20, 0x82, 0x81, 0x52], "mov w0, 0xc11"), + ([0x80, 0x82, 0x81, 0x52], "mov w0, 0xc14"), + ([0xa0, 0x82, 0x81, 0x52], "mov w0, 0xc15"), + ([0x60, 0xfb, 0x81, 0x52], "mov w0, 0xfdb"), + ([0x64, 0xfb, 0x81, 0x52], "mov w4, 0xfdb"), + ([0x01, 0x00, 0x82, 0x52], "mov w1, 0x1000"), + ([0xe2, 0x07, 0x82, 0x52], "mov w2, 0x103f"), + ([0xe3, 0x07, 0x82, 0x52], "mov w3, 0x103f"), + ([0xe4, 0x07, 0x82, 0x52], "mov w4, 0x103f"), + ([0xea, 0x07, 0x82, 0x52], "mov w10, 0x103f"), + ([0xeb, 0x07, 0x82, 0x52], "mov w11, 0x103f"), + ([0x01, 0x08, 0x82, 0x52], "mov w1, 0x1040"), + ([0x02, 0x08, 0x82, 0x52], "mov w2, 0x1040"), + ([0x03, 0x08, 0x82, 0x52], "mov w3, 0x1040"), + ([0x80, 0x46, 0x82, 0x52], "mov w0, 0x1234"), + ([0xe0, 0x4d, 0x82, 0x52], "mov w0, 0x126f"), + ([0xe2, 0xed, 0x82, 0x52], "mov w2, 0x176f"), + ([0x04, 0x00, 0x84, 0x52], "mov w4, 0x2000"), + ([0x05, 0x00, 0x84, 0x52], "mov w5, 0x2000"), + ([0x25, 0x00, 0x84, 0x52], "mov w5, 0x2001"), + ([0x85, 0x00, 0x84, 0x52], "mov w5, 0x2004"), + ([0xa5, 0x00, 0x84, 0x52], "mov w5, 0x2005"), + ([0xa0, 0x18, 0x84, 0x52], "mov w0, 0x20c5"), + ([0xa1, 0x18, 0x84, 0x52], "mov w1, 0x20c5"), + ([0xa2, 0x18, 0x84, 0x52], "mov w2, 0x20c5"), + ([0x03, 0xe2, 0x84, 0x52], "mov w3, 0x2710"), + ([0x05, 0xe2, 0x84, 0x52], "mov w5, 0x2710"), + ([0x06, 0x00, 0x86, 0x52], "mov w6, 0x3000"), + ([0x61, 0x66, 0x86, 0x52], "mov w1, 0x3333"), + ([0x06, 0x00, 0x88, 0x52], "mov w6, 0x4000"), + ([0x80, 0x29, 0x88, 0x52], "mov w0, 0x414c"), + ([0x81, 0x29, 0x88, 0x52], "mov w1, 0x414c"), + ([0xe0, 0x48, 0x88, 0x52], "mov w0, 0x4247"), + ([0x00, 0x4a, 0x88, 0x52], "mov w0, 0x4250"), + ([0xe1, 0xa8, 0x88, 0x52], "mov w1, 0x4547"), + ([0xe2, 0xa8, 0x88, 0x52], "mov w2, 0x4547"), + ([0x00, 0xa9, 0x88, 0x52], "mov w0, 0x4548"), + ([0xe0, 0xe8, 0x88, 0x52], "mov w0, 0x4747"), + ([0xc0, 0xf5, 0x88, 0x52], "mov w0, 0x47ae"), + ([0x21, 0x8a, 0x89, 0x52], "mov w1, 0x4c51"), + ([0xa0, 0xca, 0x89, 0x52], "mov w0, 0x4e55"), + ([0x02, 0xea, 0x89, 0x52], "mov w2, 0x4f50"), + ([0x81, 0x29, 0x8a, 0x52], "mov w1, 0x514c"), + ([0x41, 0x2a, 0x8a, 0x52], "mov w1, 0x5152"), + ([0x41, 0x48, 0x8a, 0x52], "mov w1, 0x5242"), + ([0x01, 0x49, 0x8a, 0x52], "mov w1, 0x5248"), + ([0xe0, 0x49, 0x8a, 0x52], "mov w0, 0x524f"), + ([0x21, 0x4a, 0x8a, 0x52], "mov w1, 0x5251"), + ([0x80, 0x4a, 0x8a, 0x52], "mov w0, 0x5254"), + ([0x81, 0x4a, 0x8a, 0x52], "mov w1, 0x5254"), + ([0x60, 0x8a, 0x8a, 0x52], "mov w0, 0x5453"), + ([0xc0, 0xaa, 0x8a, 0x52], "mov w0, 0x5556"), + ([0xc1, 0xaa, 0x8a, 0x52], "mov w1, 0x5556"), + ([0xc4, 0xaa, 0x8a, 0x52], "mov w4, 0x5556"), + ([0xc6, 0xaa, 0x8a, 0x52], "mov w6, 0x5556"), + ([0xc8, 0xaa, 0x8a, 0x52], "mov w8, 0x5556"), + ([0xca, 0xaa, 0x8a, 0x52], "mov w10, 0x5556"), + ([0xd7, 0xaa, 0x8a, 0x52], "mov w23, 0x5556"), + ([0xd8, 0xaa, 0x8a, 0x52], "mov w24, 0x5556"), + ([0xd9, 0xaa, 0x8a, 0x52], "mov w25, 0x5556"), + ([0x61, 0x2a, 0x8b, 0x52], "mov w1, 0x5953"), + ([0x62, 0x2a, 0x8b, 0x52], "mov w2, 0x5953"), + ([0xc0, 0xcc, 0x8c, 0x52], "mov w0, 0x6666"), + ([0xc1, 0xcc, 0x8c, 0x52], "mov w1, 0x6666"), + ([0xe0, 0xcc, 0x8c, 0x52], "mov w0, 0x6667"), + ([0x00, 0x43, 0x8e, 0x52], "mov w0, 0x7218"), + ([0x02, 0x43, 0x8e, 0x52], "mov w2, 0x7218"), + ([0x40, 0xdf, 0x8f, 0x52], "mov w0, 0x7efa"), + ([0x13, 0x00, 0x90, 0x52], "mov w19, 0x8000"), + ([0x20, 0xc7, 0x91, 0x52], "mov w0, 0x8e39"), + ([0x23, 0xc7, 0x91, 0x52], "mov w3, 0x8e39"), + ([0x3b, 0xc7, 0x91, 0x52], "mov w27, 0x8e39"), + ([0x42, 0x55, 0x95, 0x52], "mov w2, 0xaaaa"), + ([0x64, 0x55, 0x95, 0x52], "mov w4, 0xaaab"), + ([0xe3, 0xce, 0x97, 0x52], "mov w3, 0xbe77"), + ([0xa2, 0x99, 0x99, 0x52], "mov w2, 0xcccd"), + ([0x42, 0xe1, 0x9a, 0x52], "mov w2, 0xd70a"), + ([0x62, 0x0f, 0x9e, 0x52], "mov w2, 0xf07b"), + ([0x05, 0x70, 0xa6, 0x52], "mov w5, 0x33800000"), + ([0x04, 0x30, 0xa7, 0x52], "mov w4, 0x39800000"), + ([0x00, 0x50, 0xa8, 0x52], "mov w0, 0x42800000"), + ([0x02, 0x59, 0xa8, 0x52], "mov w2, 0x42c80000"), + ([0x40, 0x9f, 0xa8, 0x52], "mov w0, 0x44fa0000"), + ([0x04, 0xb0, 0xa8, 0x52], "mov w4, 0x45800000"), + ([0x05, 0x70, 0xa9, 0x52], "mov w5, 0x4b800000"), + ([0x00, 0xf8, 0xaf, 0x52], "mov w0, 0x7fc00000"), + ([0x41, 0x5f, 0xb8, 0x52], "mov w1, 0xc2fa0000"), + ([0x06, 0x00, 0x80, 0x92], "mov x6, 0xffffffffffffffff"), + ([0xfb, 0x01, 0x80, 0x92], "mov x27, 0xfffffffffffffff0"), + ([0xe3, 0x02, 0x80, 0x92], "mov x3, 0xffffffffffffffe8"), + ([0xf8, 0x02, 0x80, 0x92], "mov x24, 0xffffffffffffffe8"), + ([0xfa, 0x03, 0x80, 0x92], "mov x26, 0xffffffffffffffe0"), + ([0x6b, 0x21, 0x88, 0x92], "mov x11, 0xffffffffffffbef4"), + ([0xe7, 0x41, 0x90, 0x92], "mov x7, 0xffffffffffff7df0"), + ([0xe8, 0x41, 0x90, 0x92], "mov x8, 0xffffffffffff7df0"), + ([0x00, 0x02, 0xe0, 0x92], "mov x0, 0xffefffffffffffff"), + ([0x00, 0x02, 0xf0, 0x92], "mov x0, 0x7fefffffffffffff"), + ([0xe0, 0x03, 0x00, 0xb2], "mov x0, 0x100000001"), + ([0xe1, 0x03, 0x00, 0xb2], "mov x1, 0x100000001"), + ([0xe2, 0x03, 0x00, 0xb2], "mov x2, 0x100000001"), + ([0xe4, 0x03, 0x00, 0xb2], "mov x4, 0x100000001"), + ([0xe1, 0xe7, 0x03, 0xb2], "mov x1, 0x6666666666666666"), + ([0xe0, 0x1b, 0x09, 0xb2], "mov x0, 0x3f8000003f800000"), + ([0xe0, 0x7f, 0x60, 0xb2], "mov x0, 0xffffffff00000000"), + ([0x1b, 0x00, 0x80, 0xd2], "mov x27, 0x0"), + ([0x8e, 0x01, 0x80, 0xd2], "mov x14, 0xc"), + ([0x12, 0x10, 0x80, 0xd2], "mov x18, 0x80"), + ([0x05, 0x14, 0x80, 0xd2], "mov x5, 0xa0"), + ([0x60, 0x1d, 0x80, 0xd2], "mov x0, 0xeb"), + ([0x02, 0x20, 0x80, 0xd2], "mov x2, 0x100"), + ([0x15, 0x40, 0x80, 0xd2], "mov x21, 0x200"), + ([0x14, 0xc0, 0x81, 0xd2], "mov x20, 0xe00"), + ([0x00, 0x00, 0x82, 0xd2], "mov x0, 0x1000"), + ([0x0c, 0x3a, 0x82, 0xd2], "mov x12, 0x11d0"), + ([0x00, 0x4a, 0x82, 0xd2], "mov x0, 0x1250"), + ([0x05, 0x4b, 0x82, 0xd2], "mov x5, 0x1258"), + ([0x03, 0x4d, 0x82, 0xd2], "mov x3, 0x1268"), + ([0x00, 0x4f, 0x82, 0xd2], "mov x0, 0x1278"), + ([0x02, 0x4f, 0x82, 0xd2], "mov x2, 0x1278"), + ([0x02, 0x51, 0x82, 0xd2], "mov x2, 0x1288"), + ([0xe1, 0x7f, 0x82, 0xd2], "mov x1, 0x13ff"), + ([0x07, 0x90, 0x82, 0xd2], "mov x7, 0x1480"), + ([0x08, 0xdb, 0x82, 0xd2], "mov x8, 0x16d8"), + ([0xe1, 0xff, 0x82, 0xd2], "mov x1, 0x17ff"), + ([0x09, 0xcf, 0x83, 0xd2], "mov x9, 0x1e78"), + ([0xe5, 0xff, 0x83, 0xd2], "mov x5, 0x1fff"), + ([0x03, 0x00, 0x84, 0xd2], "mov x3, 0x2000"), + ([0x0c, 0x4a, 0x84, 0xd2], "mov x12, 0x2250"), + ([0xe2, 0x7f, 0x84, 0xd2], "mov x2, 0x23ff"), + ([0x01, 0x80, 0x84, 0xd2], "mov x1, 0x2400"), + ([0x05, 0xe2, 0x84, 0xd2], "mov x5, 0x2710"), + ([0x00, 0xec, 0x84, 0xd2], "mov x0, 0x2760"), + ([0x0c, 0xee, 0x84, 0xd2], "mov x12, 0x2770"), + ([0x0c, 0x8a, 0x85, 0xd2], "mov x12, 0x2c50"), + ([0x0c, 0xc6, 0x85, 0xd2], "mov x12, 0x2e30"), + ([0x0c, 0xd6, 0x85, 0xd2], "mov x12, 0x2eb0"), + ([0x0c, 0xd8, 0x85, 0xd2], "mov x12, 0x2ec0"), + ([0x0c, 0xdc, 0x85, 0xd2], "mov x12, 0x2ee0"), + ([0x0c, 0xde, 0x85, 0xd2], "mov x12, 0x2ef0"), + ([0x0c, 0xe0, 0x85, 0xd2], "mov x12, 0x2f00"), + ([0x0c, 0xe4, 0x85, 0xd2], "mov x12, 0x2f20"), + ([0x0c, 0xe6, 0x85, 0xd2], "mov x12, 0x2f30"), + ([0x0c, 0x08, 0x86, 0xd2], "mov x12, 0x3040"), + ([0x0c, 0x20, 0x86, 0xd2], "mov x12, 0x3100"), + ([0x02, 0x21, 0x86, 0xd2], "mov x2, 0x3108"), + ([0x0c, 0x22, 0x86, 0xd2], "mov x12, 0x3110"), + ([0x01, 0x29, 0x86, 0xd2], "mov x1, 0x3148"), + ([0x0c, 0x2a, 0x86, 0xd2], "mov x12, 0x3150"), + ([0x00, 0x2d, 0x86, 0xd2], "mov x0, 0x3168"), + ([0x0c, 0x2e, 0x86, 0xd2], "mov x12, 0x3170"), + ([0x0c, 0x30, 0x86, 0xd2], "mov x12, 0x3180"), + ([0x0c, 0x42, 0x86, 0xd2], "mov x12, 0x3210"), + ([0x0c, 0x44, 0x86, 0xd2], "mov x12, 0x3220"), + ([0x06, 0x4d, 0x86, 0xd2], "mov x6, 0x3268"), + ([0x0c, 0x4e, 0x86, 0xd2], "mov x12, 0x3270"), + ([0xe1, 0xff, 0x87, 0xd2], "mov x1, 0x3fff"), + ([0xe2, 0xff, 0x87, 0xd2], "mov x2, 0x3fff"), + ([0x08, 0x02, 0x88, 0xd2], "mov x8, 0x4010"), + ([0x04, 0x51, 0x88, 0xd2], "mov x4, 0x4288"), + ([0x0c, 0x52, 0x88, 0xd2], "mov x12, 0x4290"), + ([0x04, 0x53, 0x88, 0xd2], "mov x4, 0x4298"), + ([0x0c, 0x54, 0x88, 0xd2], "mov x12, 0x42a0"), + ([0x8c, 0x58, 0x88, 0xd2], "mov x12, 0x42c4"), + ([0x02, 0x59, 0x88, 0xd2], "mov x2, 0x42c8"), + ([0x06, 0x59, 0x88, 0xd2], "mov x6, 0x42c8"), + ([0x08, 0x59, 0x88, 0xd2], "mov x8, 0x42c8"), + ([0x09, 0x59, 0x88, 0xd2], "mov x9, 0x42c8"), + ([0x0c, 0x6a, 0x88, 0xd2], "mov x12, 0x4350"), + ([0x0c, 0x4a, 0x89, 0xd2], "mov x12, 0x4a50"), + ([0x0c, 0x4c, 0x89, 0xd2], "mov x12, 0x4a60"), + ([0x00, 0xb4, 0x89, 0xd2], "mov x0, 0x4da0"), + ([0x0c, 0xb6, 0x89, 0xd2], "mov x12, 0x4db0"), + ([0x00, 0x0a, 0x90, 0xd2], "mov x0, 0x8050"), + ([0x01, 0x20, 0x90, 0xd2], "mov x1, 0x8100"), + ([0x0d, 0x20, 0x90, 0xd2], "mov x13, 0x8100"), + ([0x0a, 0x7a, 0x90, 0xd2], "mov x10, 0x83d0"), + ([0x0c, 0x7a, 0x90, 0xd2], "mov x12, 0x83d0"), + ([0x08, 0x7c, 0x90, 0xd2], "mov x8, 0x83e0"), + ([0x07, 0x7d, 0x90, 0xd2], "mov x7, 0x83e8"), + ([0x01, 0x46, 0x93, 0xd2], "mov x1, 0x9a30"), + ([0x0c, 0x48, 0x93, 0xd2], "mov x12, 0x9a40"), + ([0x01, 0x20, 0xa0, 0xd2], "mov x1, 0x1000000"), + ([0x00, 0xf0, 0xa7, 0xd2], "mov x0, 0x3f800000"), + ([0x1c, 0xf0, 0xa7, 0xd2], "mov x28, 0x3f800000"), + ([0x00, 0xf0, 0xb7, 0xd2], "mov x0, 0xbf800000"), + ([0x05, 0xf0, 0xb7, 0xd2], "mov x5, 0xbf800000"), + ([0x00, 0x00, 0xc8, 0xd2], "mov x0, 0x400000000000"), + ([0x00, 0x00, 0xd0, 0xd2], "mov x0, 0x800000000000"), + ([0x00, 0x00, 0xdd, 0xd2], "mov x0, 0xe80000000000"), + ([0x00, 0x02, 0xe0, 0xd2], "mov x0, 0x10000000000000"), + ([0x01, 0x94, 0xe7, 0xd2], "mov x1, 0x3ca0000000000000"), + ([0x01, 0x96, 0xe7, 0xd2], "mov x1, 0x3cb0000000000000"), + ([0x02, 0xce, 0xe7, 0xd2], "mov x2, 0x3e70000000000000"), + ([0x00, 0xe6, 0xe7, 0xd2], "mov x0, 0x3f30000000000000"), + ([0x13, 0xf0, 0xe7, 0xd2], "mov x19, 0x3f80000000000000"), + ([0x00, 0x0a, 0xe8, 0xd2], "mov x0, 0x4050000000000000"), + ([0x20, 0x0b, 0xe8, 0xd2], "mov x0, 0x4059000000000000"), + ([0x00, 0x12, 0xe8, 0xd2], "mov x0, 0x4090000000000000"), + ([0x04, 0x16, 0xe8, 0xd2], "mov x4, 0x40b0000000000000"), + ([0x40, 0x18, 0xe8, 0xd2], "mov x0, 0x40c2000000000000"), + ([0x01, 0x1c, 0xe8, 0xd2], "mov x1, 0x40e0000000000000"), + ([0x01, 0x22, 0xe8, 0xd2], "mov x1, 0x4110000000000000"), + ([0x00, 0x2e, 0xe8, 0xd2], "mov x0, 0x4170000000000000"), + ([0x00, 0xff, 0xef, 0xd2], "mov x0, 0x7ff8000000000000"), + ([0x00, 0x00, 0xf0, 0xd2], "mov x0, 0x8000000000000000"), + ([0x02, 0xfe, 0xff, 0xd2], "mov x2, 0xfff0000000000000"), + ([0x1b, 0x00, 0x80, 0x12], "mov w27, 0xffffffff"), + ([0x84, 0x00, 0x80, 0x12], "mov w4, 0xfffffffb"), + ([0x00, 0x19, 0x80, 0x12], "mov w0, 0xffffff37"), + ([0xe0, 0x07, 0x82, 0x12], "mov w0, 0xffffefc0"), + ([0xe1, 0x07, 0x82, 0x12], "mov w1, 0xffffefc0"), + ([0x00, 0x10, 0xb0, 0x12], "mov w0, 0x7f7fffff"), + ([0xf9, 0x03, 0x0a, 0xaa], "mov x25, x10"), + ([0x01, 0x20, 0xa0, 0xf2], "movk x1, 0x100, lsl 16"), + ([0x02, 0x0e, 0xc0, 0xf2], "movk x2, 0x70, lsl 32"), + ([0x03, 0x0e, 0xc0, 0xf2], "movk x3, 0x70, lsl 32"), + ([0x02, 0x0f, 0xc0, 0xf2], "movk x2, 0x78, lsl 32"), + ([0x04, 0x10, 0xc0, 0xf2], "movk x4, 0x80, lsl 32"), + ([0x05, 0x10, 0xc0, 0xf2], "movk x5, 0x80, lsl 32"), + ([0x04, 0x1c, 0xc0, 0xf2], "movk x4, 0xe0, lsl 32"), + ([0x07, 0x1e, 0xc0, 0xf2], "movk x7, 0xf0, lsl 32"), + ([0x06, 0x2c, 0xc0, 0xf2], "movk x6, 0x160, lsl 32"), + ([0xc1, 0xfd, 0xe7, 0xf2], "movk x1, 0x3fee, lsl 48"), + ([0x40, 0x09, 0xe8, 0xf2], "movk x0, 0x404a, lsl 48"), + ([0xe0, 0x13, 0xe8, 0xf2], "movk x0, 0x409f, lsl 48"), + ([0xe0, 0x11, 0xf8, 0xf2], "movk x0, 0xc08f, lsl 48"), + ([0x42, 0x55, 0xa5, 0x72], "movk w2, 0x2aaa, lsl 16"), + ([0x60, 0x1c, 0xa7, 0x72], "movk w0, 0x38e3, lsl 16"), + ([0x63, 0x1c, 0xa7, 0x72], "movk w3, 0x38e3, lsl 16"), + ([0x80, 0x99, 0xb9, 0x72], "movk w0, 0xcccc, lsl 16"), + ([0xfa, 0x03, 0x3c, 0xaa], "mvn x26, x28"), + ([0xfb, 0x03, 0x3c, 0xaa], "mvn x27, x28"), + ([0xfe, 0x03, 0x3e, 0xaa], "mvn x30, x30"), + ([0xe3, 0x03, 0x3b, 0x2a], "mvn w3, w27"), + ]; + + let errs = run_tests(TESTS); + + for err in errs.iter() { + println!("{}", err); + } + + assert!(errs.is_empty()); +} + +#[test] +fn test_openblas_cmp_ops() { + const TESTS: &[([u8; 4], &'static str)] = &[ + ([0xff, 0x01, 0x01, 0xeb], "cmp x15, x1"), + ([0x1f, 0x02, 0x01, 0xeb], "cmp x16, x1"), + ([0x9f, 0x02, 0x1b, 0xeb], "cmp x20, x27"), + ([0xbf, 0x06, 0x1b, 0xeb], "cmp x21, x27, lsl 1"), + ([0x7f, 0xc3, 0x20, 0xeb], "cmp x27, w0, sxtw 0"), + ([0xff, 0x63, 0x21, 0xeb], "cmp sp, x1"), + ([0x1f, 0xc0, 0x21, 0xeb], "cmp x0, w1, sxtw 0"), + ([0x5f, 0xc0, 0x21, 0xeb], "cmp x2, w1, sxtw 0"), + ([0x5f, 0x0b, 0x00, 0xf1], "cmp x26, 0x2"), + ([0x7f, 0x0d, 0x00, 0xf1], "cmp x11, 0x3"), + ([0xdf, 0x0f, 0x00, 0xf1], "cmp x30, 0x3"), + ([0x7f, 0x13, 0x00, 0xf1], "cmp x27, 0x4"), + ([0xdf, 0x16, 0x00, 0xf1], "cmp x22, 0x5"), + ([0x3f, 0x1b, 0x00, 0xf1], "cmp x25, 0x6"), + ([0x7f, 0x1c, 0x00, 0xf1], "cmp x3, 0x7"), + ([0x1f, 0x20, 0x00, 0xf1], "cmp x0, 0x8"), + ([0x9f, 0x23, 0x00, 0xf1], "cmp x28, 0x8"), + ([0x9f, 0x27, 0x00, 0xf1], "cmp x28, 0x9"), + ([0x9f, 0x2b, 0x00, 0xf1], "cmp x28, 0xa"), + ([0x1f, 0x2c, 0x00, 0xf1], "cmp x0, 0xb"), + ([0x9f, 0x2f, 0x00, 0xf1], "cmp x28, 0xb"), + ([0x1f, 0x30, 0x00, 0xf1], "cmp x0, 0xc"), + ([0x9f, 0x33, 0x00, 0xf1], "cmp x28, 0xc"), + ([0x1f, 0x34, 0x00, 0xf1], "cmp x0, 0xd"), + ([0xbf, 0x36, 0x00, 0xf1], "cmp x21, 0xd"), + ([0xdf, 0x36, 0x00, 0xf1], "cmp x22, 0xd"), + ([0x1f, 0x3d, 0x00, 0xf1], "cmp x8, 0xf"), + ([0xbf, 0x3e, 0x00, 0xf1], "cmp x21, 0xf"), + ([0xdf, 0x3e, 0x00, 0xf1], "cmp x22, 0xf"), + ([0x1f, 0x41, 0x00, 0xf1], "cmp x8, 0x10"), + ([0x5f, 0x41, 0x00, 0xf1], "cmp x10, 0x10"), + ([0xff, 0x80, 0x00, 0xf1], "cmp x7, 0x20"), + ([0x3f, 0xfc, 0x00, 0xf1], "cmp x1, 0x3f"), + ([0x1f, 0x01, 0x01, 0xf1], "cmp x8, 0x40"), + ([0x3f, 0xfc, 0x07, 0xf1], "cmp x1, 0x1ff"), + ([0x3f, 0xfc, 0x0f, 0xf1], "cmp x1, 0x3ff"), + ([0x7f, 0xfc, 0x0f, 0xf1], "cmp x3, 0x3ff"), + ([0x1f, 0x00, 0x10, 0xf1], "cmp x0, 0x400"), + ([0x3f, 0x00, 0x10, 0xf1], "cmp x1, 0x400"), + ([0x3f, 0x00, 0x24, 0xf1], "cmp x1, 0x900"), + ([0x1f, 0xfc, 0x3f, 0xf1], "cmp x0, 0xfff"), + ([0x3f, 0xfc, 0x3f, 0xf1], "cmp x1, 0xfff"), + ([0x1f, 0x08, 0x40, 0xf1], "cmp x0, 0x2000"), + ([0x3f, 0x08, 0x40, 0xf1], "cmp x1, 0x2000"), + ([0x7f, 0x01, 0x00, 0xf1], "cmp x11, 0x0"), + ([0x3f, 0x07, 0x00, 0xf1], "cmp x25, 0x1"), + ([0x1f, 0x00, 0x00, 0x71], "cmp w0, 0x0"), + ([0x3f, 0x04, 0x00, 0x6b], "cmp w1, w0, lsl 1"), + ([0x5f, 0x05, 0x80, 0x6b], "cmp w10, w0, asr 1"), + ([0x3f, 0x04, 0x00, 0x71], "cmp w1, 0x1"), + ([0x5f, 0x24, 0x34, 0x71], "cmp w2, 0xd09"), + ([0x7f, 0x0a, 0x40, 0x71], "cmp w19, 0x2000"), + ([0x1f, 0x80, 0x40, 0x71], "cmp w0, 0x20000"), + ([0x1f, 0x00, 0x44, 0x71], "cmp w0, 0x100000"), + ([0x9f, 0x02, 0x1a, 0x2b], "cmn w20, w26"), + ([0x1f, 0x58, 0x00, 0x31], "cmn w0, 0x16"), + ([0x1f, 0x04, 0x00, 0xb1], "cmn x0, 0x1"), + ([0xdf, 0x38, 0x00, 0xb1], "cmn x6, 0xe"), + ]; + + let errs = run_tests(TESTS); + + for err in errs.iter() { + println!("{}", err); + } + + assert!(errs.is_empty()); +} + +#[test] +fn test_openblas_tst_ops() { + const TESTS: &[([u8; 4], &'static str)] = &[ + ([0x9f, 0x02, 0x05, 0x6a], "tst w20, w5"), + ([0x1f, 0x78, 0x1e, 0x72], "tst w0, 0xfffffffd"), + ([0x3f, 0x78, 0x1e, 0x72], "tst w1, 0xfffffffd"), + ([0x5f, 0x03, 0x13, 0xea], "tst x26, x19"), + ([0x3f, 0x01, 0x40, 0xf2], "tst x9, 0x1"), + ([0x5f, 0x01, 0x40, 0xf2], "tst x10, 0x1"), + ([0x1f, 0x04, 0x40, 0xf2], "tst x0, 0x3"), + ([0x3f, 0x09, 0x40, 0xf2], "tst x9, 0x7"), + ([0x3f, 0x01, 0x7e, 0xf2], "tst x9, 0x4"), + ([0x5f, 0x04, 0x7f, 0xf2], "tst x2, 0x6"), + ]; + + let errs = run_tests(TESTS); + + for err in errs.iter() { + println!("{}", err); + } + + assert!(errs.is_empty()); +} + +#[test] +fn test_openblas_conditional_ops() { + const TESTS: &[([u8; 4], &'static str)] = &[ + ([0x82, 0xd2, 0x80, 0x1a], "csel w2, w20, w0, le"), + ([0x94, 0xd2, 0x80, 0x1a], "csel w20, w20, w0, le"), + ([0xe2, 0x13, 0x81, 0x1a], "csel w2, wzr, w1, ne"), + ([0x73, 0xa2, 0x81, 0x1a], "csel w19, w19, w1, ge"), + ([0xe1, 0xa3, 0x81, 0x1a], "csel w1, wzr, w1, ge"), + ([0xe1, 0xa7, 0x81, 0x1a], "csinc w1, wzr, w1, ge"), + ([0x21, 0xb3, 0x81, 0x1a], "csel w1, w25, w1, lt"), + ([0x9c, 0xd3, 0x81, 0x1a], "csel w28, w28, w1, le"), + ([0x00, 0x00, 0x82, 0x1a], "csel w0, w0, w2, eq"), + ([0x00, 0x30, 0x82, 0x1a], "csel w0, w0, w2, lo"), + ([0x42, 0x54, 0x82, 0x1a], "cinc w2, w2, mi"), + ([0x62, 0x80, 0x82, 0x1a], "csel w2, w3, w2, hi"), + ([0x02, 0xa0, 0x82, 0x1a], "csel w2, w0, w2, ge"), + ([0x44, 0xa4, 0x84, 0x1a], "csinc w4, w2, w4, ge"), + ([0x81, 0xc7, 0x9f, 0x1a], "csinc w1, w28, wzr, gt"), + ([0x9c, 0xc7, 0x9f, 0x1a], "csinc w28, w28, wzr, gt"), + ([0xf4, 0xc7, 0x9f, 0x1a], "cset w20, le"), + ([0x94, 0xd2, 0x9f, 0x1a], "csel w20, w20, wzr, le"), + ([0x00, 0xd4, 0x9f, 0x1a], "csinc w0, w0, wzr, le"), + ([0x21, 0xd4, 0x9f, 0x1a], "csinc w1, w1, wzr, le"), + ([0xfa, 0xd7, 0x9f, 0x1a], "cset w26, gt"), + ([0x16, 0x04, 0x80, 0x5a], "cneg w22, w0, ne"), + ([0x00, 0x44, 0x82, 0x5a], "csneg w0, w0, w2, mi"), + ([0x9c, 0x17, 0x9c, 0x5a], "cneg w28, w28, eq"), + ([0x9c, 0x03, 0x9f, 0x5a], "csinv w28, w28, wzr, eq"), + ([0xfc, 0x03, 0x9f, 0x5a], "csetm w28, ne"), + ([0x7b, 0x13, 0x9f, 0x5a], "csinv w27, w27, wzr, ne"), + ([0xc6, 0xb0, 0x9f, 0x5a], "csinv w6, w6, wzr, lt"), + ([0x63, 0xc0, 0x9f, 0x5a], "csinv w3, w3, wzr, gt"), + ([0x84, 0x1b, 0x40, 0xfa], "ccmp x28, 0x0, 0x4, ne"), + ([0x84, 0xca, 0x40, 0xfa], "ccmp x20, 0x0, 0x4, gt"), + ([0x80, 0x08, 0x41, 0xfa], "ccmp x4, 0x1, 0x0, eq"), + ([0x64, 0x12, 0x43, 0xfa], "ccmp x19, x3, 0x4, ne"), + ([0x04, 0x10, 0x45, 0xfa], "ccmp x0, x5, 0x4, ne"), + ([0x24, 0xd3, 0x5b, 0xfa], "ccmp x25, x27, 0x4, le"), + ([0x00, 0x08, 0x40, 0x7a], "ccmp w0, 0x0, 0x0, eq"), + ([0x01, 0xc0, 0x5c, 0x7a], "ccmp w0, w28, 0x1, gt"), + ([0x04, 0x08, 0x41, 0x3a], "ccmn w0, 0x1, 0x4, eq"), + ([0x84, 0xbb, 0x41, 0x3a], "ccmn w28, 0x1, 0x4, lt"), + ([0x04, 0xc8, 0x41, 0x3a], "ccmn w0, 0x1, 0x4, gt"), + ([0xa4, 0xc8, 0x41, 0x3a], "ccmn w5, 0x1, 0x4, gt"), + ([0x84, 0xdb, 0x41, 0x3a], "ccmn w28, 0x1, 0x4, le"), + ([0x84, 0x1b, 0x4a, 0x3a], "ccmn w28, 0xa, 0x4, ne"), + ([0x08, 0x41, 0x80, 0x9a], "csel x8, x8, x0, mi"), + ([0x8c, 0xb1, 0x8f, 0x9a], "csel x12, x12, x15, lt"), + ([0xc6, 0x50, 0x97, 0x9a], "csel x6, x6, x23, pl"), + ([0x08, 0x51, 0x97, 0x9a], "csel x8, x8, x23, pl"), + ([0x77, 0xd0, 0x97, 0x9a], "csel x23, x3, x23, le"), + ([0x9c, 0xa3, 0x9f, 0x9a], "csel x28, x28, xzr, ge"), + ([0x00, 0xc0, 0x9f, 0x9a], "csel x0, x0, xzr, gt"), + ([0x21, 0xc0, 0x9f, 0x9a], "csel x1, x1, xzr, gt"), + ([0x33, 0xc0, 0x9f, 0x9a], "csel x19, x1, xzr, gt"), + ([0x63, 0xc0, 0x9f, 0x9a], "csel x3, x3, xzr, gt"), + ([0x39, 0xc3, 0x9f, 0x9a], "csel x25, x25, xzr, gt"), + ([0x05, 0xc4, 0x9f, 0x9a], "csinc x5, x0, xzr, gt"), + ([0x18, 0xc7, 0x9f, 0x9a], "csinc x24, x24, xzr, gt"), + ([0x39, 0xc7, 0x9f, 0x9a], "csinc x25, x25, xzr, gt"), + ([0xf7, 0xc7, 0x9f, 0x9a], "cset x23, le"), + ([0x1b, 0xd0, 0x9f, 0x9a], "csel x27, x0, xzr, le"), + ([0x00, 0xd3, 0x9f, 0x9a], "csel x0, x24, xzr, le"), + ([0xe0, 0xd7, 0x9f, 0x9a], "cset x0, gt"), + ]; + + let errs = run_tests(TESTS); + + for err in errs.iter() { + println!("{}", err); + } + + assert!(errs.is_empty()); +} + +#[test] +fn test_openblas_muldiv_ops() { + const TESTS: &[([u8; 4], &'static str)] = &[ + ([0x39, 0x08, 0xc0, 0x1a], "udiv w25, w1, w0"), + ([0x3a, 0x09, 0xc1, 0x1a], "udiv w26, w9, w1"), + ([0xa0, 0x0e, 0xc1, 0x1a], "sdiv w0, w21, w1"), + ([0x00, 0x08, 0xc2, 0x1a], "udiv w0, w0, w2"), + ([0x61, 0x0b, 0xc2, 0x1a], "udiv w1, w27, w2"), + ([0x00, 0x0c, 0xc2, 0x1a], "sdiv w0, w0, w2"), + ([0x01, 0x0d, 0xc2, 0x1a], "sdiv w1, w8, w2"), + ([0x00, 0x08, 0xc3, 0x1a], "udiv w0, w0, w3"), + ([0x6b, 0x0d, 0xc4, 0x1a], "sdiv w11, w11, w4"), + ([0x73, 0x0e, 0xdc, 0x1a], "sdiv w19, w19, w28"), + ([0xa1, 0x72, 0x00, 0x1b], "madd w1, w21, w0, w28"), + ([0x3c, 0xa8, 0x00, 0x1b], "msub w28, w1, w0, w10"), + ([0x80, 0x7f, 0x01, 0x1b], "mul w0, w28, w1"), + ([0x00, 0x0c, 0xd3, 0x9a], "sdiv x0, x0, x19"), + ([0x80, 0x0c, 0xda, 0x9a], "sdiv x0, x4, x26"), + ([0x00, 0x0c, 0xdb, 0x9a], "sdiv x0, x0, x27"), + ([0x00, 0x00, 0x00, 0x9b], "madd x0, x0, x0, x0"), + ([0x9c, 0x7f, 0x1b, 0x9b], "mul x28, x28, x27"), + ([0x00, 0xd4, 0x1b, 0x9b], "msub x0, x0, x27, x21"), + ([0x20, 0x00, 0x1c, 0x9b], "madd x0, x1, x28, x0"), + ([0x81, 0x7b, 0x1e, 0x9b], "madd x1, x28, x30, x30"), + ([0xad, 0x7d, 0x1e, 0x9b], "mul x13, x13, x30"), + ([0x02, 0x7f, 0x3a, 0x9b], "smull x2, w24, w26"), + ([0x42, 0x64, 0x3c, 0x9b], "smaddl x2, w2, w28, x25"), + ([0x29, 0x7c, 0x3c, 0x9b], "smull x9, w1, w28"), + ([0x20, 0x7c, 0xa0, 0x9b], "umull x0, w1, w0"), + ([0x42, 0x0c, 0xa4, 0x9b], "umaddl x2, w2, w4, x3"), + ([0x21, 0x7c, 0xa4, 0x9b], "umull x1, w1, w4"), + ([0x63, 0x98, 0xa4, 0x9b], "umsubl x3, w3, w4, x6"), + ([0x85, 0x00, 0xa5, 0x9b], "umaddl x5, w4, w5, x0"), + ([0x10, 0xa8, 0xb0, 0x9b], "umsubl x16, w0, w16, x10"), + ]; + + let errs = run_tests(TESTS); + + for err in errs.iter() { + println!("{}", err); + } + + assert!(errs.is_empty()); +} + +#[test] +fn test_openblas_branch_ops() { + const TESTS: &[([u8; 4], &'static str)] = &[ + ([0x40, 0x00, 0x00, 0x34], "cbz w0, $+0x8"), + ([0xcd, 0xff, 0xff, 0x34], "cbz w13, $-0x8"), + ([0x40, 0x00, 0x00, 0x35], "cbnz w0, $+0x8"), + ([0x19, 0xfd, 0xff, 0x35], "cbnz w25, $-0x60"), + ([0xc1, 0xff, 0xff, 0x35], "cbnz w1, $-0x8"), + ([0x61, 0x00, 0x00, 0x36], "tbz w1, 0x0, $+0xc"), + ([0x20, 0x01, 0x00, 0x36], "tbz w0, 0x0, $+0x24"), + ([0x73, 0x00, 0xf8, 0x36], "tbz w19, 0x1f, $+0xc"), + ([0x27, 0xff, 0xff, 0x36], "tbz w7, 0x1f, $-0x1c"), + ([0xf3, 0x05, 0x00, 0x37], "tbnz w19, 0x0, $+0xbc"), + ([0xfb, 0xff, 0xff, 0x17], "b $-0x14"), + ([0xfd, 0xff, 0xff, 0x17], "b $-0xc"), + ([0x6d, 0x02, 0x00, 0x54], "b.le $+0x4c"), + ([0x6d, 0xff, 0xff, 0x54], "b.le $-0x14"), + ([0x97, 0x1e, 0x00, 0x94], "bl $+0x7a5c"), + ([0xd2, 0xff, 0xff, 0x97], "bl $-0xb8"), + ([0x53, 0x00, 0x00, 0xb4], "cbz x19, $+0x8"), + ([0x7c, 0x00, 0x00, 0xb4], "cbz x28, $+0xc"), + ([0x96, 0x00, 0x00, 0xb5], "cbnz x22, $+0x10"), + ([0x98, 0x00, 0x00, 0xb5], "cbnz x24, $+0x10"), + ([0xa6, 0x00, 0xf8, 0xb6], "tbz x6, 0x3f, $+0x14"), + ([0x73, 0x03, 0xf8, 0xb7], "tbnz x19, 0x3f, $+0x6c"), + ([0x23, 0xf8, 0xff, 0xb7], "tbnz x3, 0x3f, $-0xfc"), + ([0x40, 0x00, 0x1f, 0xd6], "br x2"), + ([0x00, 0x02, 0x1f, 0xd6], "br x16"), + ([0x20, 0x02, 0x1f, 0xd6], "br x17"), + ([0x00, 0x00, 0x3f, 0xd6], "blr x0"), + ([0x20, 0x00, 0x3f, 0xd6], "blr x1"), + ([0x80, 0x03, 0x3f, 0xd6], "blr x28"), + ([0xc0, 0x03, 0x5f, 0xd6], "ret"), + ]; + + let errs = run_tests(TESTS); + + for err in errs.iter() { + println!("{}", err); + } + + assert!(errs.is_empty()); +} + +#[test] +fn test_openblas_simd_ops() { + const TESTS: &[([u8; 4], &'static str)] = &[ + ([0x64, 0xcc, 0x21, 0x0e], "fmla v4.2s, v3.2s, v1.2s"), + ([0x00, 0xd4, 0x21, 0x0e], "fadd v0.2s, v0.2s, v1.2s"), + ([0x03, 0xcc, 0x22, 0x0e], "fmla v3.2s, v0.2s, v2.2s"), + ([0x21, 0xd4, 0x23, 0x0e], "fadd v1.2s, v1.2s, v3.2s"), + ([0x29, 0xd5, 0x35, 0x0e], "fadd v9.2s, v9.2s, v21.2s"), + ([0x4a, 0xd5, 0x35, 0x0e], "fadd v10.2s, v10.2s, v21.2s"), + ([0x42, 0x78, 0x61, 0x0e], "fcvtl v2.2d, v2.2s"), + ([0x63, 0x78, 0x61, 0x0e], "fcvtl v3.2d, v3.2s"), + ([0x00, 0xf8, 0xa0, 0x0e], "fabs v0.2s, v0.2s"), + ([0x25, 0xcc, 0xa2, 0x0e], "fmls v5.2s, v1.2s, v2.2s"), + ([0x10, 0x11, 0x80, 0x0f], "fmla v16.2s, v8.2s, v0.s[0]"), + ([0x39, 0x58, 0x88, 0x0f], "fmls v25.2s, v1.2s, v8.s[2]"), + ([0x08, 0x92, 0x8a, 0x0f], "fmul v8.2s, v16.2s, v10.s[0]"), + ([0x01, 0x12, 0x8b, 0x0f], "fmla v1.2s, v16.2s, v11.s[0]"), + ([0x8c, 0x12, 0x8e, 0x0f], "fmla v12.2s, v20.2s, v14.s[0]"), + ([0xc6, 0x12, 0x8e, 0x0f], "fmla v6.2s, v22.2s, v14.s[0]"), + ([0x3c, 0x58, 0xa9, 0x0f], "fmls v28.2s, v1.2s, v9.s[3]"), + ([0x00, 0x06, 0x20, 0x1e], "fccmp s16, s0, 0x0, eq"), + ([0x54, 0xc2, 0x20, 0x1e], "fabs s20, s18"), + ([0x23, 0x40, 0x21, 0x1e], "fneg s3, s1"), + ([0x05, 0x08, 0x22, 0x1e], "fmul s5, s0, s2"), + ([0x35, 0x88, 0x24, 0x1e], "fnmul s21, s1, s4"), + ([0x42, 0x88, 0x24, 0x1e], "fnmul s2, s2, s4"), + ([0x08, 0x90, 0x26, 0x1e], "fmov s8, 2.000000000000000000e+01"), + ([0x00, 0xdc, 0x26, 0x1e], "fcsel s0, s0, s6, le"), + ([0x4e, 0x0c, 0x2e, 0x1e], "fcsel s14, s2, s14, eq"), + ([0x00, 0x10, 0x2e, 0x1e], "fmov s0, 1.000000000000000000e+00"), + ([0x41, 0x18, 0x2e, 0x1e], "fdiv s1, s2, s14"), + ([0xd3, 0x5e, 0x33, 0x1e], "fcsel s19, s22, s19, pl"), + ([0xe4, 0x3b, 0x3b, 0x1e], "fsub s4, s31, s27"), + ([0x08, 0x90, 0x3b, 0x1e], "fmov s8, -4.375000000000000000e-01"), + ([0xa1, 0x08, 0x3c, 0x1e], "fmul s1, s5, s28"), + ([0x01, 0x10, 0x3c, 0x1e], "fmov s1, -5.000000000000000000e-01"), + ([0xb8, 0x1a, 0x3c, 0x1e], "fdiv s24, s21, s28"), + ([0x5c, 0x39, 0x3c, 0x1e], "fsub s28, s10, s28"), + ([0x76, 0x08, 0x3d, 0x1e], "fmul s22, s3, s29"), + ([0x59, 0x18, 0x3d, 0x1e], "fdiv s25, s2, s29"), + ([0x26, 0x08, 0x3e, 0x1e], "fmul s6, s1, s30"), + ([0x00, 0x10, 0x3e, 0x1e], "fmov s0, -1.000000000000000000e+00"), + ([0x13, 0x10, 0x3e, 0x1e], "fmov s19, -1.000000000000000000e+00"), + ([0x39, 0x3b, 0x3f, 0x1e], "fsub s25, s25, s31"), + ([0x20, 0x04, 0x60, 0x1e], "fccmp d1, d0, 0x0, eq"), + ([0x20, 0x06, 0x60, 0x1e], "fccmp d17, d0, 0x0, eq"), + ([0x00, 0x08, 0x60, 0x1e], "fmul d0, d0, d0"), + ([0xe8, 0x0b, 0x60, 0x1e], "fmul d8, d31, d0"), + ([0xf3, 0x0b, 0x60, 0x1e], "fmul d19, d31, d0"), + ([0x40, 0x0c, 0x60, 0x1e], "fcsel d0, d2, d0, eq"), + ([0x08, 0x0d, 0x60, 0x1e], "fcsel d8, d8, d0, eq"), + ([0x14, 0x10, 0x60, 0x1e], "fmov d20, 2.000000000000000000e+00"), + ([0xc4, 0x14, 0x60, 0x1e], "fccmp d6, d0, 0x4, ne"), + ([0x72, 0x15, 0x60, 0x1e], "fccmpe d11, d0, 0x2, ne"), + ([0xa4, 0x15, 0x60, 0x1e], "fccmp d13, d0, 0x4, ne"), + ([0xe4, 0x15, 0x60, 0x1e], "fccmp d15, d0, 0x4, ne"), + ([0x20, 0x18, 0x60, 0x1e], "fdiv d0, d1, d0"), + ([0xc6, 0x1a, 0x60, 0x1e], "fdiv d6, d22, d0"), + ([0x20, 0x1c, 0x60, 0x1e], "fcsel d0, d1, d0, ne"), + ([0xef, 0x1d, 0x60, 0x1e], "fcsel d15, d15, d0, ne"), + ([0x70, 0x22, 0x60, 0x1e], "fcmpe d19, d0"), + ([0xd8, 0x22, 0x60, 0x1e], "fcmpe d22, 0.0"), + ([0x08, 0x23, 0x60, 0x1e], "fcmp d24, 0.0"), + ([0x80, 0x28, 0x60, 0x1e], "fadd d0, d4, d0"), + ([0xfe, 0x43, 0x60, 0x1e], "fmov d30, d31"), + ([0x22, 0x4c, 0x60, 0x1e], "fcsel d2, d1, d0, mi"), + ([0x21, 0x5c, 0x60, 0x1e], "fcsel d1, d1, d0, pl"), + ([0x20, 0x68, 0x60, 0x1e], "fmaxnm d0, d1, d0"), + ([0xab, 0x6a, 0x60, 0x1e], "fmaxnm d11, d21, d0"), + ([0xaf, 0x79, 0x60, 0x1e], "fminnm d15, d13, d0"), + ([0x82, 0x8b, 0x60, 0x1e], "fnmul d2, d28, d0"), + ([0xc0, 0x8b, 0x60, 0x1e], "fnmul d0, d30, d0"), + ([0x22, 0x9c, 0x60, 0x1e], "fcsel d2, d1, d0, ls"), + ([0x08, 0xad, 0x60, 0x1e], "fcsel d8, d8, d0, ge"), + ([0xf7, 0xc3, 0x60, 0x1e], "fabs d23, d31"), + ([0x20, 0xcc, 0x60, 0x1e], "fcsel d0, d1, d0, gt"), + ([0x6b, 0xdd, 0x60, 0x1e], "fcsel d11, d11, d0, le"), + ([0xec, 0x0b, 0x61, 0x1e], "fmul d12, d31, d1"), + ([0x41, 0x0c, 0x61, 0x1e], "fcsel d1, d2, d1, eq"), + ([0x18, 0x10, 0x61, 0x1e], "fmov d24, 3.000000000000000000e+00"), + ([0x04, 0x14, 0x61, 0x1e], "fccmp d0, d1, 0x4, ne"), + ([0xe1, 0x1b, 0x61, 0x1e], "fdiv d1, d31, d1"), + ([0x00, 0x1c, 0x61, 0x1e], "fcsel d0, d0, d1, ne"), + ([0xd0, 0x22, 0x61, 0x1e], "fcmpe d22, d1"), + ([0x41, 0x2b, 0x61, 0x1e], "fadd d1, d26, d1"), + ([0x00, 0x38, 0x61, 0x1e], "fsub d0, d0, d1"), + ([0x81, 0x3b, 0x61, 0x1e], "fsub d1, d28, d1"), + ([0x06, 0x40, 0x61, 0x1e], "fneg d6, d0"), + ([0x01, 0x4c, 0x61, 0x1e], "fcsel d1, d0, d1, mi"), + ([0xef, 0x5d, 0x61, 0x1e], "fcsel d15, d15, d1, pl"), + ([0x6c, 0x68, 0x61, 0x1e], "fmaxnm d12, d3, d1"), + ([0x61, 0x78, 0x61, 0x1e], "fminnm d1, d3, d1"), + ([0x06, 0x88, 0x61, 0x1e], "fnmul d6, d0, d1"), + ([0x01, 0x9c, 0x61, 0x1e], "fcsel d1, d0, d1, ls"), + ([0x02, 0xac, 0x61, 0x1e], "fcsel d2, d0, d1, ge"), + ([0x84, 0xbc, 0x61, 0x1e], "fcsel d4, d4, d1, lt"), + ([0x00, 0xc0, 0x61, 0x1e], "fsqrt d0, d0"), + ([0x02, 0xc2, 0x61, 0x1e], "fsqrt d2, d16"), + ([0x04, 0xc4, 0x61, 0x1e], "fccmp d0, d1, 0x4, gt"), + ([0x11, 0xc4, 0x61, 0x1e], "fccmpe d0, d1, 0x1, gt"), + ([0x50, 0xcc, 0x61, 0x1e], "fcsel d16, d2, d1, gt"), + ([0x92, 0xcc, 0x61, 0x1e], "fcsel d18, d4, d1, gt"), + ([0x00, 0xdc, 0x61, 0x1e], "fcsel d0, d0, d1, le"), + ([0x20, 0x04, 0x62, 0x1e], "fccmp d1, d2, 0x0, eq"), + ([0x00, 0x08, 0x62, 0x1e], "fmul d0, d0, d2"), + ([0x8b, 0x09, 0x62, 0x1e], "fmul d11, d12, d2"), + ([0x04, 0x14, 0x6a, 0x1e], "fccmp d0, d10, 0x4, ne"), + ([0x72, 0x15, 0x6a, 0x1e], "fccmpe d11, d10, 0x2, ne"), + ([0x00, 0x18, 0x6a, 0x1e], "fdiv d0, d0, d10"), + ([0xf5, 0x5e, 0x75, 0x1e], "fcsel d21, d23, d21, pl"), + ([0xa5, 0x68, 0x75, 0x1e], "fmaxnm d5, d5, d21"), + ([0x42, 0x88, 0x75, 0x1e], "fnmul d2, d2, d21"), + ([0x20, 0x06, 0x76, 0x1e], "fccmp d17, d22, 0x0, eq"), + ([0x41, 0x18, 0x76, 0x1e], "fdiv d1, d2, d22"), + ([0xa5, 0x18, 0x76, 0x1e], "fdiv d5, d5, d22"), + ([0xf0, 0x22, 0x76, 0x1e], "fcmpe d23, d22"), + ([0x00, 0x28, 0x76, 0x1e], "fadd d0, d0, d22"), + ([0xf7, 0x39, 0x77, 0x1e], "fsub d23, d15, d23"), + ([0x10, 0x3a, 0x77, 0x1e], "fsub d16, d16, d23"), + ([0x01, 0x68, 0x77, 0x1e], "fmaxnm d1, d0, d23"), + ([0x42, 0x88, 0x77, 0x1e], "fnmul d2, d2, d23"), + ([0x00, 0x00, 0x78, 0x1e], "fcvtzs w0, d0"), + ([0x85, 0x01, 0x78, 0x1e], "fcvtzs w5, d12"), + ([0x13, 0x10, 0x7e, 0x1e], "fmov d19, -1.000000000000000000e+00"), + ([0x00, 0x18, 0x7e, 0x1e], "fdiv d0, d0, d30"), + ([0x42, 0x18, 0x7e, 0x1e], "fdiv d2, d2, d30"), + ([0x84, 0x18, 0x7e, 0x1e], "fdiv d4, d4, d30"), + ([0x52, 0x1a, 0x7e, 0x1e], "fdiv d18, d18, d30"), + ([0xbd, 0x1b, 0x7e, 0x1e], "fdiv d29, d29, d30"), + ([0x21, 0x28, 0x7e, 0x1e], "fadd d1, d1, d30"), + ([0x63, 0x38, 0x7e, 0x1e], "fsub d3, d3, d30"), + ([0x7c, 0x38, 0x7e, 0x1e], "fsub d28, d3, d30"), + ([0xe7, 0x38, 0x7e, 0x1e], "fsub d7, d7, d30"), + ([0x73, 0x3a, 0x7e, 0x1e], "fsub d19, d19, d30"), + ([0x03, 0x08, 0x7f, 0x1e], "fmul d3, d0, d31"), + ([0x23, 0x08, 0x7f, 0x1e], "fmul d3, d1, d31"), + ([0x24, 0x08, 0x7f, 0x1e], "fmul d4, d1, d31"), + ([0x29, 0x08, 0x7f, 0x1e], "fmul d9, d1, d31"), + ([0x0c, 0xb2, 0x01, 0x1f], "fmsub s12, s16, s1, s12"), + ([0x39, 0xe6, 0x74, 0x1f], "fnmsub d25, d17, d20, d25"), + ([0x82, 0x0a, 0x75, 0x1f], "fnmadd d2, d20, d21, d2"), + ([0xc6, 0xf4, 0x30, 0x4e], "fmax v6.4s, v6.4s, v16.4s"), + ([0x64, 0xcc, 0x31, 0x4e], "fmla v4.4s, v3.4s, v17.4s"), + ([0xc6, 0xd4, 0x3c, 0x4e], "fadd v6.4s, v6.4s, v28.4s"), + ([0x82, 0xce, 0x3d, 0x4e], "fmla v2.4s, v20.4s, v29.4s"), + ([0xdf, 0x7b, 0x61, 0x4e], "fcvtl2 v31.2d, v30.4s"), + ([0x18, 0xf7, 0x79, 0x4e], "fmax v24.2d, v24.2d, v25.2d"), + ([0xde, 0xd7, 0x7f, 0x4e], "fadd v30.2d, v30.2d, v31.2d"), + ([0x21, 0xf8, 0xa0, 0x4e], "fabs v1.4s, v1.4s"), + ([0x00, 0xd4, 0xa1, 0x4e], "fsub v0.4s, v0.4s, v1.4s"), + ([0x65, 0xcc, 0xb0, 0x4e], "fmls v5.4s, v3.4s, v16.4s"), + ([0x64, 0xcc, 0xb1, 0x4e], "fmls v4.4s, v3.4s, v17.4s"), + ([0xff, 0xfb, 0xe0, 0x4e], "fabs v31.2d, v31.2d"), + ([0x94, 0xd6, 0xe6, 0x4e], "fsub v20.2d, v20.2d, v6.2d"), + ([0x30, 0xce, 0xf4, 0x4e], "fmls v16.2d, v17.2d, v20.2d"), + ([0x4f, 0xce, 0xf4, 0x4e], "fmls v15.2d, v18.2d, v20.2d"), + ([0x5a, 0x18, 0x88, 0x4f], "fmla v26.4s, v2.4s, v8.s[2]"), + ([0x1d, 0x98, 0xaa, 0x4f], "fmul v29.4s, v0.4s, v10.s[3]"), + ([0xfe, 0x18, 0xae, 0x4f], "fmla v30.4s, v7.4s, v14.s[3]"), + ([0xfe, 0x50, 0xaf, 0x4f], "fmls v30.4s, v7.4s, v15.s[1]"), + ([0x31, 0x50, 0xc8, 0x4f], "fmls v17.2d, v1.2d, v8.d[0]"), + ([0xc7, 0x93, 0xcb, 0x4f], "fmul v7.2d, v30.2d, v11.d[0]"), + ([0xb0, 0x18, 0xcc, 0x4f], "fmla v16.2d, v5.2d, v12.d[1]"), + ([0xf7, 0x58, 0xcc, 0x4f], "fmls v23.2d, v7.2d, v12.d[1]"), + ([0xa0, 0xb9, 0xa1, 0x5e], "fcvtzs s0, s13"), + ([0x40, 0x10, 0x82, 0x5f], "fmla s0, s2, v2.s[0]"), + ([0x84, 0x12, 0x8a, 0x5f], "fmla s4, s20, v10.s[0]"), + ([0xa5, 0x12, 0x8a, 0x5f], "fmla s5, s21, v10.s[0]"), + ([0x00, 0x13, 0x8a, 0x5f], "fmla s0, s24, v10.s[0]"), + ([0x21, 0x13, 0x8a, 0x5f], "fmla s1, s25, v10.s[0]"), + ([0x84, 0x13, 0x8a, 0x5f], "fmla s4, s28, v10.s[0]"), + ([0xa5, 0x13, 0x8a, 0x5f], "fmla s5, s29, v10.s[0]"), + ([0x00, 0x92, 0x8a, 0x5f], "fmul s0, s16, v10.s[0]"), + ([0x84, 0x92, 0x8a, 0x5f], "fmul s4, s20, v10.s[0]"), + ([0x00, 0x93, 0x8a, 0x5f], "fmul s0, s24, v10.s[0]"), + ([0x84, 0x93, 0x8a, 0x5f], "fmul s4, s28, v10.s[0]"), + ([0x01, 0x12, 0x8b, 0x5f], "fmla s1, s16, v11.s[0]"), + ([0x85, 0x12, 0x8b, 0x5f], "fmla s5, s20, v11.s[0]"), + ([0x01, 0x13, 0x8b, 0x5f], "fmla s1, s24, v11.s[0]"), + ([0x85, 0x13, 0x8b, 0x5f], "fmla s5, s28, v11.s[0]"), + ([0x20, 0x52, 0x8b, 0x5f], "fmls s0, s17, v11.s[0]"), + ([0x20, 0x52, 0xb8, 0x5f], "fmls s0, s17, v24.s[1]"), + ([0x40, 0x10, 0xc2, 0x5f], "fmla d0, d2, v2.d[0]"), + ([0x03, 0x90, 0xc2, 0x5f], "fmul d3, d0, v2.d[0]"), + ([0x20, 0x5a, 0xd8, 0x5f], "fmls d0, d17, v24.d[1]"), + ([0x21, 0xdc, 0x60, 0x6e], "fmul v1.2d, v1.2d, v0.2d"), + ([0x42, 0xdc, 0x60, 0x6e], "fmul v2.2d, v2.2d, v0.2d"), + ([0x44, 0xdc, 0x60, 0x6e], "fmul v4.2d, v2.2d, v0.2d"), + ([0x63, 0xdc, 0x60, 0x6e], "fmul v3.2d, v3.2d, v0.2d"), + ([0x84, 0xdc, 0x60, 0x6e], "fmul v4.2d, v4.2d, v0.2d"), + ([0x42, 0xdc, 0x61, 0x6e], "fmul v2.2d, v2.2d, v1.2d"), + ([0x02, 0xdc, 0x62, 0x6e], "fmul v2.2d, v0.2d, v2.2d"), + ([0x04, 0xdc, 0x62, 0x6e], "fmul v4.2d, v0.2d, v2.2d"), + ([0x06, 0xdc, 0x62, 0x6e], "fmul v6.2d, v0.2d, v2.2d"), + ([0x42, 0xd4, 0x63, 0x6e], "faddp v2.2d, v2.2d, v3.2d"), + ([0x05, 0xdc, 0x63, 0x6e], "fmul v5.2d, v0.2d, v3.2d"), + ([0x07, 0xdc, 0x63, 0x6e], "fmul v7.2d, v0.2d, v3.2d"), + ([0x42, 0xdc, 0x63, 0x6e], "fmul v2.2d, v2.2d, v3.2d"), + ([0x81, 0xd4, 0x64, 0x6e], "faddp v1.2d, v4.2d, v4.2d"), + ([0x10, 0xdc, 0x64, 0x6e], "fmul v16.2d, v0.2d, v4.2d"), + ([0x42, 0xdc, 0x64, 0x6e], "fmul v2.2d, v2.2d, v4.2d"), + ([0x83, 0xd4, 0x65, 0x6e], "faddp v3.2d, v4.2d, v5.2d"), + ([0x84, 0xd4, 0x65, 0x6e], "faddp v4.2d, v4.2d, v5.2d"), + ([0x11, 0xdc, 0x65, 0x6e], "fmul v17.2d, v0.2d, v5.2d"), + ([0x63, 0xdc, 0x65, 0x6e], "fmul v3.2d, v3.2d, v5.2d"), + ([0x84, 0xdc, 0x65, 0x6e], "fmul v4.2d, v4.2d, v5.2d"), + ([0xc6, 0xd4, 0x67, 0x6e], "faddp v6.2d, v6.2d, v7.2d"), + ([0x10, 0xde, 0x68, 0x6e], "fmul v16.2d, v16.2d, v8.2d"), + ([0x31, 0xde, 0x68, 0x6e], "fmul v17.2d, v17.2d, v8.2d"), + ([0x52, 0xde, 0x68, 0x6e], "fmul v18.2d, v18.2d, v8.2d"), + ([0x73, 0xde, 0x68, 0x6e], "fmul v19.2d, v19.2d, v8.2d"), + ([0x94, 0xde, 0x68, 0x6e], "fmul v20.2d, v20.2d, v8.2d"), + ([0xb5, 0xde, 0x68, 0x6e], "fmul v21.2d, v21.2d, v8.2d"), + ([0xd6, 0xde, 0x68, 0x6e], "fmul v22.2d, v22.2d, v8.2d"), + ([0xf7, 0xde, 0x68, 0x6e], "fmul v23.2d, v23.2d, v8.2d"), + ([0x44, 0xdc, 0x70, 0x6e], "fmul v4.2d, v2.2d, v16.2d"), + ([0x66, 0xdc, 0x70, 0x6e], "fmul v6.2d, v3.2d, v16.2d"), + ([0x18, 0xde, 0x70, 0x6e], "fmul v24.2d, v16.2d, v16.2d"), + ([0x54, 0xde, 0x70, 0x6e], "fmul v20.2d, v18.2d, v16.2d"), + ([0x66, 0xde, 0x70, 0x6e], "fmul v6.2d, v19.2d, v16.2d"), + ([0x04, 0xd6, 0x71, 0x6e], "faddp v4.2d, v16.2d, v17.2d"), + ([0x10, 0xd6, 0x71, 0x6e], "faddp v16.2d, v16.2d, v17.2d"), + ([0x45, 0xdc, 0x71, 0x6e], "fmul v5.2d, v2.2d, v17.2d"), + ([0x66, 0xdc, 0x71, 0x6e], "fmul v6.2d, v3.2d, v17.2d"), + ([0x39, 0xde, 0x71, 0x6e], "fmul v25.2d, v17.2d, v17.2d"), + ([0x55, 0xde, 0x71, 0x6e], "fmul v21.2d, v18.2d, v17.2d"), + ([0x66, 0xde, 0x71, 0x6e], "fmul v6.2d, v19.2d, v17.2d"), + ([0x18, 0xf6, 0x71, 0x6e], "fmaxp v24.2d, v16.2d, v17.2d"), + ([0x5a, 0xde, 0x72, 0x6e], "fmul v26.2d, v18.2d, v18.2d"), + ([0x45, 0xd6, 0x73, 0x6e], "faddp v5.2d, v18.2d, v19.2d"), + ([0x52, 0xd6, 0x73, 0x6e], "faddp v18.2d, v18.2d, v19.2d"), + ([0x7b, 0xde, 0x73, 0x6e], "fmul v27.2d, v19.2d, v19.2d"), + ([0x59, 0xf6, 0x73, 0x6e], "fmaxp v25.2d, v18.2d, v19.2d"), + ([0x94, 0xd6, 0x75, 0x6e], "faddp v20.2d, v20.2d, v21.2d"), + ([0x9a, 0xf6, 0x75, 0x6e], "fmaxp v26.2d, v20.2d, v21.2d"), + ([0xd6, 0xd6, 0x77, 0x6e], "faddp v22.2d, v22.2d, v23.2d"), + ([0xdb, 0xf6, 0x77, 0x6e], "fmaxp v27.2d, v22.2d, v23.2d"), + ([0x18, 0xf7, 0x78, 0x6e], "fmaxp v24.2d, v24.2d, v24.2d"), + ([0x18, 0xd7, 0x79, 0x6e], "faddp v24.2d, v24.2d, v25.2d"), + ([0x18, 0xf7, 0x79, 0x6e], "fmaxp v24.2d, v24.2d, v25.2d"), + ([0x18, 0xf7, 0x7a, 0x6e], "fmaxp v24.2d, v24.2d, v26.2d"), + ([0x5a, 0xf7, 0x7b, 0x6e], "fmaxp v26.2d, v26.2d, v27.2d"), + ([0x00, 0xd8, 0x30, 0x7e], "faddp s0, v0.2s"), + ([0x21, 0xd8, 0x30, 0x7e], "faddp s1, v1.2s"), + ([0x29, 0xd9, 0x30, 0x7e], "faddp s9, v9.2s"), + ([0x4a, 0xd9, 0x30, 0x7e], "faddp s10, v10.2s"), + ([0x00, 0xd8, 0x70, 0x7e], "faddp d0, v0.2d"), + ([0x21, 0xd8, 0x70, 0x7e], "faddp d1, v1.2d"), + ([0x29, 0xd9, 0x70, 0x7e], "faddp d9, v9.2d"), + ([0x4a, 0xd9, 0x70, 0x7e], "faddp d10, v10.2d"), + ([0x18, 0xdb, 0x70, 0x7e], "faddp d24, v24.2d"), + ([0x00, 0xf8, 0x70, 0x7e], "fmaxp d0, v0.2d"), + ([0x41, 0xf8, 0x70, 0x7e], "fmaxp d1, v2.2d"), + ([0x20, 0xd4, 0xa0, 0x7e], "fabd s0, s1, s0"), + ([0x21, 0xd4, 0xa0, 0x7e], "fabd s1, s1, s0"), + ([0xd6, 0xd6, 0xf9, 0x7e], "fabd d22, d22, d25"), + ([0x21, 0xd4, 0xfa, 0x7e], "fabd d1, d1, d26"), + ([0x06, 0x00, 0x66, 0x9e], "fmov x6, d0"), + ([0x11, 0x00, 0x66, 0x9e], "fmov x17, d0"), + ([0x4b, 0x02, 0x67, 0x9e], "fmov d11, x18"), + ([0xfc, 0x03, 0x67, 0x9e], "fmov d28, xzr"), + ([0x06, 0x01, 0x78, 0x9e], "fcvtzs x6, d8"), + ([0x42, 0xdc, 0x20, 0x2e], "fmul v2.2s, v2.2s, v0.2s"), + ([0x44, 0xdc, 0x20, 0x2e], "fmul v4.2s, v2.2s, v0.2s"), + ([0x42, 0xdc, 0x21, 0x2e], "fmul v2.2s, v2.2s, v1.2s"), + ([0x02, 0xdc, 0x22, 0x2e], "fmul v2.2s, v0.2s, v2.2s"), + ([0x04, 0xdc, 0x22, 0x2e], "fmul v4.2s, v0.2s, v2.2s"), + ([0x05, 0xdc, 0x23, 0x2e], "fmul v5.2s, v0.2s, v3.2s"), + ([0x81, 0xd4, 0x24, 0x2e], "faddp v1.2s, v4.2s, v4.2s"), + ([0x00, 0xd4, 0x20, 0x6e], "faddp v0.4s, v0.4s, v0.4s"), + ([0x21, 0xdc, 0x20, 0x6e], "fmul v1.4s, v1.4s, v0.4s"), + ([0x42, 0xdc, 0x20, 0x6e], "fmul v2.4s, v2.4s, v0.4s"), + ([0x21, 0xd4, 0x21, 0x6e], "faddp v1.4s, v1.4s, v1.4s"), + ([0x04, 0xdc, 0x22, 0x6e], "fmul v4.4s, v0.4s, v2.4s"), + ([0x06, 0xdc, 0x22, 0x6e], "fmul v6.4s, v0.4s, v2.4s"), + ([0x42, 0xd4, 0x23, 0x6e], "faddp v2.4s, v2.4s, v3.4s"), + ([0x05, 0xdc, 0x23, 0x6e], "fmul v5.4s, v0.4s, v3.4s"), + ([0x07, 0xdc, 0x23, 0x6e], "fmul v7.4s, v0.4s, v3.4s"), + ([0x10, 0xdc, 0x24, 0x6e], "fmul v16.4s, v0.4s, v4.4s"), + ([0x83, 0xd4, 0x25, 0x6e], "faddp v3.4s, v4.4s, v5.4s"), + ([0x84, 0xd4, 0x25, 0x6e], "faddp v4.4s, v4.4s, v5.4s"), + ([0x11, 0xdc, 0x25, 0x6e], "fmul v17.4s, v0.4s, v5.4s"), + ([0xc6, 0xd4, 0x27, 0x6e], "faddp v6.4s, v6.4s, v7.4s"), + ([0x44, 0xdc, 0x30, 0x6e], "fmul v4.4s, v2.4s, v16.4s"), + ([0x66, 0xdc, 0x30, 0x6e], "fmul v6.4s, v3.4s, v16.4s"), + ([0x00, 0xf8, 0x30, 0x6e], "fmaxv s0, v0.4s"), + ([0x21, 0xf8, 0x30, 0x6e], "fmaxv s1, v1.4s"), + ([0x41, 0xf8, 0x30, 0x6e], "fmaxv s1, v2.4s"), + ([0x10, 0xd6, 0x31, 0x6e], "faddp v16.4s, v16.4s, v17.4s"), + ([0x45, 0xdc, 0x31, 0x6e], "fmul v5.4s, v2.4s, v17.4s"), + ([0x66, 0xdc, 0x31, 0x6e], "fmul v6.4s, v3.4s, v17.4s"), + ([0x52, 0xd6, 0x33, 0x6e], "faddp v18.4s, v18.4s, v19.4s"), + ([0x94, 0xd6, 0x35, 0x6e], "faddp v20.4s, v20.4s, v21.4s"), + ([0xd6, 0xd6, 0x37, 0x6e], "faddp v22.4s, v22.4s, v23.4s"), + ([0x18, 0xd7, 0x39, 0x6e], "faddp v24.4s, v24.4s, v25.4s"), + ([0xc7, 0x20, 0x06, 0x2e], "ext v7.8b, v6.8b, v6.8b, 0x4"), + ([0x27, 0x1c, 0x20, 0x2e], "eor v7.8b, v1.8b, v0.8b"), + ([0x22, 0x1c, 0x22, 0x2e], "eor v2.8b, v1.8b, v2.8b"), + ([0x42, 0x1c, 0x32, 0x2e], "eor v2.8b, v2.8b, v18.8b"), + ([0xc2, 0x1c, 0x70, 0x2e], "bsl v2.8b, v6.8b, v16.8b"), + ([0xd2, 0x1d, 0xa1, 0x2e], "bit v18.8b, v14.8b, v1.8b"), + ([0x01, 0x1c, 0xa6, 0x2e], "bit v1.8b, v0.8b, v6.8b"), + ([0xa6, 0x1c, 0xe0, 0x2e], "bif v6.8b, v5.8b, v0.8b"), + ([0x15, 0x1e, 0xf7, 0x2e], "bif v21.8b, v16.8b, v23.8b"), + ([0x08, 0x85, 0xa0, 0x0e], "add v8.2s, v8.2s, v0.2s"), + ([0x21, 0x1c, 0x21, 0x6e], "eor v1.16b, v1.16b, v1.16b"), + ([0x42, 0x1c, 0x22, 0x6e], "eor v2.16b, v2.16b, v2.16b"), + ([0x84, 0x1c, 0x24, 0x6e], "eor v4.16b, v4.16b, v4.16b"), + ([0x10, 0x1e, 0x30, 0x6e], "eor v16.16b, v16.16b, v16.16b"), + ([0x31, 0x1e, 0x31, 0x6e], "eor v17.16b, v17.16b, v17.16b"), + ([0x73, 0x1e, 0x33, 0x6e], "eor v19.16b, v19.16b, v19.16b"), + ([0xb5, 0x1e, 0x35, 0x6e], "eor v21.16b, v21.16b, v21.16b"), + ([0xf7, 0x1e, 0x37, 0x6e], "eor v23.16b, v23.16b, v23.16b"), + ([0x39, 0x1f, 0x39, 0x6e], "eor v25.16b, v25.16b, v25.16b"), + ([0x7b, 0x1f, 0x3b, 0x6e], "eor v27.16b, v27.16b, v27.16b"), + ([0xbd, 0x1f, 0x3d, 0x6e], "eor v29.16b, v29.16b, v29.16b"), + ([0xff, 0x1f, 0x3f, 0x6e], "eor v31.16b, v31.16b, v31.16b"), + ([0xa5, 0x40, 0x05, 0x6e], "ext v5.16b, v5.16b, v5.16b, 8"), + ([0xc7, 0x40, 0x06, 0x6e], "ext v7.16b, v6.16b, v6.16b, 8"), + ([0x02, 0x1d, 0x21, 0x0e], "and v2.8b, v8.8b, v1.8b"), + ([0x21, 0x1c, 0x22, 0x0e], "and v1.8b, v1.8b, v2.8b"), + ([0x21, 0x1c, 0x32, 0x0e], "and v1.8b, v1.8b, v18.8b"), + ([0x08, 0x54, 0x22, 0x0f], "shl v8.2s, v0.2s, 2"), + ]; + + let errs = run_tests(TESTS); + + for err in errs.iter() { + println!("{}", err); + } + + assert!(errs.is_empty()); +} + +#[test] +fn test_openblas_simd_movs() { + const TESTS: &[([u8; 4], &'static str)] = &[ + ([0x55, 0x04, 0x04, 0x4e], "dup v21.4s, v2.s[0]"), + ([0x38, 0x07, 0x08, 0x4e], "dup v24.2d, v25.d[0]"), + ([0x64, 0x64, 0x04, 0x6e], "mov v4.s[0], v3.s[3]"), + ([0x67, 0x64, 0x04, 0x6e], "mov v7.s[0], v3.s[3]"), + ([0x4b, 0x64, 0x14, 0x6e], "mov v11.s[2], v2.s[3]"), + ([0x00, 0x04, 0x18, 0x6e], "mov v0.d[1], v0.d[0]"), + ([0x21, 0x04, 0x18, 0x6e], "mov v1.d[1], v1.d[0]"), + ([0x40, 0x04, 0x18, 0x6e], "mov v0.d[1], v2.d[0]"), + ([0x41, 0x04, 0x18, 0x6e], "mov v1.d[1], v2.d[0]"), + ([0x42, 0x04, 0x18, 0x6e], "mov v2.d[1], v2.d[0]"), + ([0x48, 0x04, 0x18, 0x6e], "mov v8.d[1], v2.d[0]"), + ([0x50, 0x04, 0x18, 0x6e], "mov v16.d[1], v2.d[0]"), + ([0x63, 0x04, 0x18, 0x6e], "mov v3.d[1], v3.d[0]"), + ([0x6b, 0x44, 0x18, 0x6e], "mov v11.d[1], v3.d[1]"), + ([0xff, 0x45, 0x18, 0x6e], "mov v31.d[1], v15.d[1]"), + ([0x5a, 0x46, 0x18, 0x6e], "mov v26.d[1], v18.d[1]"), + ([0x00, 0x04, 0x1c, 0x6e], "mov v0.s[3], v0.s[0]"), + ([0x21, 0x04, 0x1c, 0x6e], "mov v1.s[3], v1.s[0]"), + ([0x68, 0x04, 0x1c, 0x6e], "mov v8.s[3], v3.s[0]"), + ([0x69, 0x24, 0x1c, 0x6e], "mov v9.s[3], v3.s[1]"), + ([0x6a, 0x44, 0x1c, 0x6e], "mov v10.s[3], v3.s[2]"), + ([0x6b, 0x64, 0x1c, 0x6e], "mov v11.s[3], v3.s[3]"), + ([0x43, 0x1c, 0xa2, 0x4e], "mov v3.16b, v2.16b"), + ([0x1c, 0x04, 0x00, 0x0f], "movi v28.2s, 0x0"), + ([0x28, 0x04, 0x00, 0x0f], "movi v8.2s, 0x1"), + ([0x80, 0x66, 0x01, 0x0f], "movi v0.2s, 0x34, lsl 24"), + ([0x17, 0x64, 0x04, 0x0f], "movi v23.2s, 0x80, lsl 24"), + ([0x01, 0xe4, 0x00, 0x6f], "movi v1.2d, 0x0"), + ([0x02, 0xe4, 0x00, 0x6f], "movi v2.2d, 0x0"), + ]; + + let errs = run_tests(TESTS); + + for err in errs.iter() { + println!("{}", err); + } + + assert!(errs.is_empty()); +} + +#[test] +fn test_openblas_misc_ops() { + const TESTS: &[([u8; 4], &'static str)] = &[ + ([0x1f, 0x20, 0x03, 0xd5], "nop"), + ([0xbf, 0x3a, 0x03, 0xd5], "dmb ishst"), + ([0xbf, 0x3b, 0x03, 0xd5], "dmb ish"), + ([0xdf, 0x3f, 0x03, 0xd5], "isb"), + ([0x01, 0x00, 0x38, 0xd5], "mrs x1, midr_el1"), + ([0x54, 0xd0, 0x3b, 0xd5], "mrs x20, tpidr_el0"), + ([0x41, 0xe0, 0x3b, 0xd5], "mrs x1, cntvct_el0"), + ([0x55, 0xe0, 0x3b, 0xd5], "mrs x21, cntvct_el0"), + ([0xe2, 0x69, 0xb6, 0xf8], "prfm pldl2keep, [x15, x22]"), + ([0x00, 0x6a, 0xb7, 0xf8], "prfm pldl1keep, [x16, x23]"), + ([0x80, 0x00, 0x80, 0xf9], "prfm pldl1keep, [x4]"), + ([0x81, 0x00, 0x80, 0xf9], "prfm pldl1strm, [x4]"), + ([0x00, 0x02, 0x80, 0xf9], "prfm pldl1keep, [x16]"), + ([0x21, 0x20, 0x80, 0xf9], "prfm pldl1strm, [x1, 0x40]"), + ([0x00, 0x22, 0x80, 0xf9], "prfm pldl1keep, [x16, 0x40]"), + ([0x61, 0x30, 0x80, 0xf9], "prfm pldl1strm, [x3, 0x60]"), + ([0xa2, 0x51, 0x80, 0xf9], "prfm pldl2keep, [x13, 0xa0]"), + ([0xc2, 0x51, 0x80, 0xf9], "prfm pldl2keep, [x14, 0xa0]"), + ([0xe2, 0x51, 0x80, 0xf9], "prfm pldl2keep, [x15, 0xa0]"), + ([0xa1, 0x00, 0x82, 0xf9], "prfm pldl1strm, [x5, 0x400]"), + ([0x00, 0x82, 0x87, 0xf9], "prfm pldl1keep, [x16, 0xf00]"), + ]; + + let errs = run_tests(TESTS); + + for err in errs.iter() { + println!("{}", err); + } + + assert!(errs.is_empty()); +} |