| 1 | //! Lowering of Fortran intrinsic functions to IR. |
| 2 | //! |
| 3 | //! Extracted from `core.rs` in Sprint 11 Stage B.1. Pure mechanical |
| 4 | //! move — behavior unchanged. Helpers consulted via `core::*`. |
| 5 | |
| 6 | use crate::ir::builder::FuncBuilder; |
| 7 | use crate::ir::inst::*; |
| 8 | use crate::ir::types::*; |
| 9 | |
| 10 | use super::core::*; |
| 11 | use super::helpers::{coerce_to_type, storage_size_bits_for_ir_type}; |
| 12 | |
| 13 | /// Lower a Fortran intrinsic function call to IR instructions. |
| 14 | /// Returns Some(ValueId) if recognized, None for external functions. |
| 15 | pub(crate) fn lower_intrinsic(b: &mut FuncBuilder, name: &str, args: &[ValueId]) -> Option<ValueId> { |
| 16 | match name { |
| 17 | "cmplx" => { |
| 18 | if let Some(real_arg) = args.first() { |
| 19 | let kind = args |
| 20 | .get(2) |
| 21 | .and_then(|arg| extract_const_int_from_value(b, *arg)) |
| 22 | .unwrap_or_else(|| { |
| 23 | if args.iter().any(|arg| { |
| 24 | let ty = b.func().value_type(*arg); |
| 25 | matches!(ty, Some(IrType::Float(FloatWidth::F64))) |
| 26 | || ty.as_ref().is_some_and(|ty| { |
| 27 | is_complex_ty(ty) && complex_float_width(ty) == FloatWidth::F64 |
| 28 | }) |
| 29 | }) { |
| 30 | 8 |
| 31 | } else { |
| 32 | 4 |
| 33 | } |
| 34 | }); |
| 35 | let fw = if kind == 8 { |
| 36 | FloatWidth::F64 |
| 37 | } else { |
| 38 | FloatWidth::F32 |
| 39 | }; |
| 40 | let elem_ty = IrType::Float(fw); |
| 41 | let buf = b.alloca(IrType::Array(Box::new(elem_ty.clone()), 2)); |
| 42 | let zero = b.const_i64(0); |
| 43 | let imag_offset = b.const_i64(if fw == FloatWidth::F64 { 8 } else { 4 }); |
| 44 | let real_val = coerce_to_type(b, *real_arg, &elem_ty); |
| 45 | let imag_val = if let Some(imag_arg) = args.get(1) { |
| 46 | coerce_to_type(b, *imag_arg, &elem_ty) |
| 47 | } else { |
| 48 | match fw { |
| 49 | FloatWidth::F64 => b.const_f64(0.0), |
| 50 | FloatWidth::F32 => b.const_f32(0.0), |
| 51 | } |
| 52 | }; |
| 53 | let real_ptr = b.gep(buf, vec![zero], IrType::Int(IntWidth::I8)); |
| 54 | b.store(real_val, real_ptr); |
| 55 | let imag_ptr = b.gep(buf, vec![imag_offset], IrType::Int(IntWidth::I8)); |
| 56 | b.store(imag_val, imag_ptr); |
| 57 | Some(buf) |
| 58 | } else { |
| 59 | None |
| 60 | } |
| 61 | } |
| 62 | "conjg" => { |
| 63 | if let Some(arg) = args.first() { |
| 64 | let ty = b |
| 65 | .func() |
| 66 | .value_type(*arg) |
| 67 | .unwrap_or(IrType::Int(IntWidth::I32)); |
| 68 | if is_complex_ty(&ty) { |
| 69 | let fw = complex_float_width(&ty); |
| 70 | let elem_ty = IrType::Float(fw); |
| 71 | let buf = b.alloca(IrType::Array(Box::new(elem_ty.clone()), 2)); |
| 72 | let zero = b.const_i64(0); |
| 73 | let imag_offset = b.const_i64(if fw == FloatWidth::F64 { 8 } else { 4 }); |
| 74 | let real_ptr = b.gep(*arg, vec![zero], IrType::Int(IntWidth::I8)); |
| 75 | let imag_ptr = b.gep(*arg, vec![imag_offset], IrType::Int(IntWidth::I8)); |
| 76 | let real_val = b.load_typed(real_ptr, elem_ty.clone()); |
| 77 | let imag_val = b.load_typed(imag_ptr, elem_ty.clone()); |
| 78 | let neg_imag = b.fneg(imag_val); |
| 79 | let out_real_ptr = b.gep(buf, vec![zero], IrType::Int(IntWidth::I8)); |
| 80 | b.store(real_val, out_real_ptr); |
| 81 | let out_imag_ptr = b.gep(buf, vec![imag_offset], IrType::Int(IntWidth::I8)); |
| 82 | b.store(neg_imag, out_imag_ptr); |
| 83 | Some(buf) |
| 84 | } else { |
| 85 | None |
| 86 | } |
| 87 | } else { |
| 88 | None |
| 89 | } |
| 90 | } |
| 91 | "merge" => { |
| 92 | if args.len() >= 3 { |
| 93 | let mut ty = b |
| 94 | .func() |
| 95 | .value_type(args[0]) |
| 96 | .unwrap_or(IrType::Int(IntWidth::I32)); |
| 97 | if ty.is_float() { |
| 98 | if matches!( |
| 99 | b.func().value_type(args[1]), |
| 100 | Some(IrType::Float(FloatWidth::F64)) |
| 101 | ) { |
| 102 | ty = IrType::Float(FloatWidth::F64); |
| 103 | } |
| 104 | } else if ty.is_int() { |
| 105 | let width = [args[0], args[1]] |
| 106 | .iter() |
| 107 | .filter_map(|arg| b.func().value_type(*arg).and_then(|ty| ty.int_width())) |
| 108 | .max_by_key(|width| width.bits()) |
| 109 | .unwrap_or(IntWidth::I32); |
| 110 | ty = IrType::Int(width); |
| 111 | } |
| 112 | let true_val = coerce_to_type(b, args[0], &ty); |
| 113 | let false_val = coerce_to_type(b, args[1], &ty); |
| 114 | let mask = coerce_to_type(b, args[2], &IrType::Bool); |
| 115 | Some(b.select(mask, true_val, false_val)) |
| 116 | } else { |
| 117 | None |
| 118 | } |
| 119 | } |
| 120 | "transfer" => { |
| 121 | if args.len() >= 2 { |
| 122 | let mold_ty = b |
| 123 | .func() |
| 124 | .value_type(args[1]) |
| 125 | .unwrap_or(IrType::Int(IntWidth::I32)); |
| 126 | Some(coerce_to_type(b, args[0], &mold_ty)) |
| 127 | } else { |
| 128 | None |
| 129 | } |
| 130 | } |
| 131 | "mod" => { |
| 132 | // MOD(a, p) = a - INT(a/p) * p (sign of dividend) |
| 133 | if args.len() >= 2 { |
| 134 | let (lhs, rhs) = unify_int_widths(b, args[0], args[1]); |
| 135 | let ty = b |
| 136 | .func() |
| 137 | .value_type(lhs) |
| 138 | .unwrap_or(IrType::Int(IntWidth::I32)); |
| 139 | if ty.is_float() { |
| 140 | Some(b.call(FuncRef::External("fmod".into()), vec![lhs, rhs], ty)) |
| 141 | } else { |
| 142 | Some(b.imod(lhs, rhs)) |
| 143 | } |
| 144 | } else { |
| 145 | None |
| 146 | } |
| 147 | } |
| 148 | "modulo" => { |
| 149 | // MODULO(a, p) = a - FLOOR(a/p) * p (sign of divisor, result in [0, |p|)) |
| 150 | // For integers: if result has opposite sign to p, add p. |
| 151 | if args.len() >= 2 { |
| 152 | let (lhs, rhs) = unify_int_widths(b, args[0], args[1]); |
| 153 | let ty = b |
| 154 | .func() |
| 155 | .value_type(lhs) |
| 156 | .unwrap_or(IrType::Int(IntWidth::I32)); |
| 157 | if ty.is_float() { |
| 158 | // Float modulo: use fmod then adjust. |
| 159 | let rem = b.call(FuncRef::External("fmod".into()), vec![lhs, rhs], ty.clone()); |
| 160 | let sum = b.fadd(rem, rhs); |
| 161 | let rem2 = b.call(FuncRef::External("fmod".into()), vec![sum, rhs], ty); |
| 162 | Some(rem2) |
| 163 | } else { |
| 164 | // Integer modulo: rem = a % p; if (rem != 0 && (rem ^ p) < 0) rem += p |
| 165 | let rem = b.imod(lhs, rhs); |
| 166 | let zero = match &ty { |
| 167 | IrType::Int(IntWidth::I64) => b.const_i64(0), |
| 168 | _ => b.const_i32(0), |
| 169 | }; |
| 170 | let rem_ne_zero = b.icmp(CmpOp::Ne, rem, zero); |
| 171 | let rem_xor_p = b.bit_xor(rem, rhs); |
| 172 | let sign_differs = b.icmp(CmpOp::Lt, rem_xor_p, zero); |
| 173 | let needs_adjust = b.and(rem_ne_zero, sign_differs); |
| 174 | let adjusted = b.iadd(rem, rhs); |
| 175 | Some(b.select(needs_adjust, adjusted, rem)) |
| 176 | } |
| 177 | } else { |
| 178 | None |
| 179 | } |
| 180 | } |
| 181 | "abs" | "iabs" | "dabs" => { |
| 182 | if let Some(arg) = args.first() { |
| 183 | let ty = b |
| 184 | .func() |
| 185 | .value_type(*arg) |
| 186 | .unwrap_or(IrType::Int(IntWidth::I32)); |
| 187 | match &ty { |
| 188 | IrType::Int(w) => { |
| 189 | let zero = match w { |
| 190 | IntWidth::I64 => b.const_i64(0), |
| 191 | _ => b.const_i32(0), |
| 192 | }; |
| 193 | let is_pos = b.icmp(CmpOp::Ge, *arg, zero); |
| 194 | let neg = b.ineg(*arg); |
| 195 | Some(b.select(is_pos, *arg, neg)) |
| 196 | } |
| 197 | IrType::Float(_) => Some(b.fabs(*arg)), |
| 198 | _ => None, |
| 199 | } |
| 200 | } else { |
| 201 | None |
| 202 | } |
| 203 | } |
| 204 | "int" | "idint" | "ifix" => { |
| 205 | if let Some(arg) = args.first() { |
| 206 | let ty = b |
| 207 | .func() |
| 208 | .value_type(*arg) |
| 209 | .unwrap_or(IrType::Int(IntWidth::I32)); |
| 210 | let requested_width = args |
| 211 | .get(1) |
| 212 | .and_then(|kind| extract_const_int_from_value(b, *kind)) |
| 213 | .and_then(int_width_from_kind_value) |
| 214 | .unwrap_or(IntWidth::I32); |
| 215 | if ty.is_float() { |
| 216 | Some(b.float_to_int(*arg, requested_width)) |
| 217 | } else { |
| 218 | Some(coerce_to_type(b, *arg, &IrType::Int(requested_width))) |
| 219 | } |
| 220 | } else { |
| 221 | None |
| 222 | } |
| 223 | } |
| 224 | "nint" | "idnint" => { |
| 225 | // NINT: round to nearest integer (not truncate). |
| 226 | // Round via libm round(), then convert to int. |
| 227 | if let Some(arg) = args.first() { |
| 228 | let ty = b |
| 229 | .func() |
| 230 | .value_type(*arg) |
| 231 | .unwrap_or(IrType::Float(FloatWidth::F64)); |
| 232 | let requested_width = args |
| 233 | .get(1) |
| 234 | .and_then(|kind| extract_const_int_from_value(b, *kind)) |
| 235 | .and_then(int_width_from_kind_value) |
| 236 | .unwrap_or(IntWidth::I32); |
| 237 | if ty.is_float() { |
| 238 | let func = if matches!(ty, IrType::Float(FloatWidth::F32)) { |
| 239 | "roundf" |
| 240 | } else { |
| 241 | "round" |
| 242 | }; |
| 243 | let rounded = b.call(FuncRef::External(func.into()), vec![*arg], ty.clone()); |
| 244 | Some(b.float_to_int(rounded, requested_width)) |
| 245 | } else { |
| 246 | Some(coerce_to_type(b, *arg, &IrType::Int(requested_width))) |
| 247 | } |
| 248 | } else { |
| 249 | None |
| 250 | } |
| 251 | } |
| 252 | "anint" | "dnint" => { |
| 253 | // ANINT: round to nearest whole number, return as real. |
| 254 | if let Some(arg) = args.first() { |
| 255 | let ty = b |
| 256 | .func() |
| 257 | .value_type(*arg) |
| 258 | .unwrap_or(IrType::Float(FloatWidth::F64)); |
| 259 | let func = if matches!(ty, IrType::Float(FloatWidth::F32)) { |
| 260 | "roundf" |
| 261 | } else { |
| 262 | "round" |
| 263 | }; |
| 264 | Some(b.call(FuncRef::External(func.into()), vec![*arg], ty)) |
| 265 | } else { |
| 266 | None |
| 267 | } |
| 268 | } |
| 269 | "real" | "float" | "sngl" => { |
| 270 | if let Some(arg) = args.first() { |
| 271 | let ty = b |
| 272 | .func() |
| 273 | .value_type(*arg) |
| 274 | .unwrap_or(IrType::Int(IntWidth::I32)); |
| 275 | let requested_fw = args |
| 276 | .get(1) |
| 277 | .and_then(|kind| extract_const_int_from_value(b, *kind)) |
| 278 | .map(|kind| { |
| 279 | if kind == 8 { |
| 280 | FloatWidth::F64 |
| 281 | } else { |
| 282 | FloatWidth::F32 |
| 283 | } |
| 284 | }); |
| 285 | let default_fw = if matches!(ty, IrType::Float(FloatWidth::F64)) |
| 286 | || (is_complex_ty(&ty) && complex_float_width(&ty) == FloatWidth::F64) |
| 287 | { |
| 288 | FloatWidth::F64 |
| 289 | } else { |
| 290 | FloatWidth::F32 |
| 291 | }; |
| 292 | let target_ty = IrType::Float(requested_fw.unwrap_or(default_fw)); |
| 293 | if ty.is_int() || ty.is_float() { |
| 294 | Some(coerce_to_type(b, *arg, &target_ty)) |
| 295 | } else if is_complex_ty(&ty) { |
| 296 | // real(z) extracts the real component of a complex number. |
| 297 | // Complex values live as ptr<[f32/f64 x 2]>; load element 0. |
| 298 | let fw = complex_float_width(&ty); |
| 299 | let zero = b.const_i64(0); |
| 300 | let re_ptr = b.gep(*arg, vec![zero], IrType::Int(IntWidth::I8)); |
| 301 | let real_part = b.load_typed(re_ptr, IrType::Float(fw)); |
| 302 | Some(coerce_to_type(b, real_part, &target_ty)) |
| 303 | } else { |
| 304 | Some(*arg) |
| 305 | } |
| 306 | } else { |
| 307 | None |
| 308 | } |
| 309 | } |
| 310 | "logical" => { |
| 311 | if let Some(arg) = args.first() { |
| 312 | let requested_ty = args |
| 313 | .get(1) |
| 314 | .and_then(|kind| extract_const_int_from_value(b, *kind)) |
| 315 | .and_then(|kind| match kind { |
| 316 | 1 => Some(IrType::Int(IntWidth::I8)), |
| 317 | 2 => Some(IrType::Int(IntWidth::I16)), |
| 318 | 4 => Some(IrType::Bool), |
| 319 | 8 => Some(IrType::Int(IntWidth::I64)), |
| 320 | _ => None, |
| 321 | }) |
| 322 | .unwrap_or(IrType::Bool); |
| 323 | let ty = b.func().value_type(*arg).unwrap_or(IrType::Bool); |
| 324 | match ty { |
| 325 | IrType::Bool | IrType::Int(_) => Some(coerce_to_type(b, *arg, &requested_ty)), |
| 326 | _ => None, |
| 327 | } |
| 328 | } else { |
| 329 | None |
| 330 | } |
| 331 | } |
| 332 | "aimag" | "dimag" => { |
| 333 | // aimag(z) extracts the imaginary component of a complex number. |
| 334 | // Complex values live as ptr<[f32/f64 x 2]>; load element 1 at |
| 335 | // byte offset 4 (f32) or 8 (f64). |
| 336 | if let Some(arg) = args.first() { |
| 337 | let ty = b |
| 338 | .func() |
| 339 | .value_type(*arg) |
| 340 | .unwrap_or(IrType::Int(IntWidth::I32)); |
| 341 | if is_complex_ty(&ty) { |
| 342 | let fw = complex_float_width(&ty); |
| 343 | let offset = b.const_i64(if fw == FloatWidth::F64 { 8 } else { 4 }); |
| 344 | let im_ptr = b.gep(*arg, vec![offset], IrType::Int(IntWidth::I8)); |
| 345 | Some(b.load_typed(im_ptr, IrType::Float(fw))) |
| 346 | } else { |
| 347 | None |
| 348 | } |
| 349 | } else { |
| 350 | None |
| 351 | } |
| 352 | } |
| 353 | "dble" | "dfloat" => { |
| 354 | if let Some(arg) = args.first() { |
| 355 | let ty = b |
| 356 | .func() |
| 357 | .value_type(*arg) |
| 358 | .unwrap_or(IrType::Int(IntWidth::I32)); |
| 359 | if ty.is_int() { |
| 360 | Some(b.int_to_float(*arg, FloatWidth::F64)) |
| 361 | } else if matches!(ty, IrType::Float(FloatWidth::F32)) { |
| 362 | Some(b.float_extend(*arg, FloatWidth::F64)) |
| 363 | } else { |
| 364 | Some(*arg) |
| 365 | } |
| 366 | } else { |
| 367 | None |
| 368 | } |
| 369 | } |
| 370 | "max" | "max0" | "amax1" | "dmax1" => { |
| 371 | if args.len() >= 2 { |
| 372 | let mut ty = b |
| 373 | .func() |
| 374 | .value_type(args[0]) |
| 375 | .unwrap_or(IrType::Int(IntWidth::I32)); |
| 376 | if ty.is_float() { |
| 377 | if args.iter().any(|arg| { |
| 378 | matches!( |
| 379 | b.func().value_type(*arg), |
| 380 | Some(IrType::Float(FloatWidth::F64)) |
| 381 | ) |
| 382 | }) { |
| 383 | ty = IrType::Float(FloatWidth::F64); |
| 384 | } |
| 385 | } else if ty.is_int() { |
| 386 | let width = args |
| 387 | .iter() |
| 388 | .filter_map(|arg| b.func().value_type(*arg).and_then(|ty| ty.int_width())) |
| 389 | .max_by_key(|width| width.bits()) |
| 390 | .unwrap_or(IntWidth::I32); |
| 391 | ty = IrType::Int(width); |
| 392 | } |
| 393 | let coerced: Vec<ValueId> = args |
| 394 | .iter() |
| 395 | .map(|arg| coerce_to_type(b, *arg, &ty)) |
| 396 | .collect(); |
| 397 | let cmp = if ty.is_float() { |
| 398 | b.fcmp(CmpOp::Ge, coerced[0], coerced[1]) |
| 399 | } else { |
| 400 | b.icmp(CmpOp::Ge, coerced[0], coerced[1]) |
| 401 | }; |
| 402 | let mut result = b.select(cmp, coerced[0], coerced[1]); |
| 403 | // Variadic: max(a, b, c, ...) chains. |
| 404 | for arg in &coerced[2..] { |
| 405 | let cmp = if ty.is_float() { |
| 406 | b.fcmp(CmpOp::Ge, result, *arg) |
| 407 | } else { |
| 408 | b.icmp(CmpOp::Ge, result, *arg) |
| 409 | }; |
| 410 | result = b.select(cmp, result, *arg); |
| 411 | } |
| 412 | Some(result) |
| 413 | } else { |
| 414 | None |
| 415 | } |
| 416 | } |
| 417 | "min" | "min0" | "amin1" | "dmin1" => { |
| 418 | if args.len() >= 2 { |
| 419 | let mut ty = b |
| 420 | .func() |
| 421 | .value_type(args[0]) |
| 422 | .unwrap_or(IrType::Int(IntWidth::I32)); |
| 423 | if ty.is_float() { |
| 424 | if args.iter().any(|arg| { |
| 425 | matches!( |
| 426 | b.func().value_type(*arg), |
| 427 | Some(IrType::Float(FloatWidth::F64)) |
| 428 | ) |
| 429 | }) { |
| 430 | ty = IrType::Float(FloatWidth::F64); |
| 431 | } |
| 432 | } else if ty.is_int() { |
| 433 | let width = args |
| 434 | .iter() |
| 435 | .filter_map(|arg| b.func().value_type(*arg).and_then(|ty| ty.int_width())) |
| 436 | .max_by_key(|width| width.bits()) |
| 437 | .unwrap_or(IntWidth::I32); |
| 438 | ty = IrType::Int(width); |
| 439 | } |
| 440 | let coerced: Vec<ValueId> = args |
| 441 | .iter() |
| 442 | .map(|arg| coerce_to_type(b, *arg, &ty)) |
| 443 | .collect(); |
| 444 | let cmp = if ty.is_float() { |
| 445 | b.fcmp(CmpOp::Le, coerced[0], coerced[1]) |
| 446 | } else { |
| 447 | b.icmp(CmpOp::Le, coerced[0], coerced[1]) |
| 448 | }; |
| 449 | let mut result = b.select(cmp, coerced[0], coerced[1]); |
| 450 | for arg in &coerced[2..] { |
| 451 | let cmp = if ty.is_float() { |
| 452 | b.fcmp(CmpOp::Le, result, *arg) |
| 453 | } else { |
| 454 | b.icmp(CmpOp::Le, result, *arg) |
| 455 | }; |
| 456 | result = b.select(cmp, result, *arg); |
| 457 | } |
| 458 | Some(result) |
| 459 | } else { |
| 460 | None |
| 461 | } |
| 462 | } |
| 463 | "sign" | "dsign" | "isign" => { |
| 464 | // sign(a, b) = abs(a) * sign_of(b) = b >= 0 ? abs(a) : -abs(a) |
| 465 | if args.len() >= 2 { |
| 466 | let ty = b |
| 467 | .func() |
| 468 | .value_type(args[0]) |
| 469 | .unwrap_or(IrType::Int(IntWidth::I32)); |
| 470 | let abs_a = if ty.is_float() { |
| 471 | b.fabs(args[0]) |
| 472 | } else { |
| 473 | let zero = match &ty { |
| 474 | IrType::Int(IntWidth::I64) => b.const_i64(0), |
| 475 | _ => b.const_i32(0), |
| 476 | }; |
| 477 | let is_pos = b.icmp(CmpOp::Ge, args[0], zero); |
| 478 | let neg = b.ineg(args[0]); |
| 479 | b.select(is_pos, args[0], neg) |
| 480 | }; |
| 481 | let neg_abs = if ty.is_float() { |
| 482 | b.fneg(abs_a) |
| 483 | } else { |
| 484 | b.ineg(abs_a) |
| 485 | }; |
| 486 | let zero = match &ty { |
| 487 | IrType::Float(FloatWidth::F32) => b.const_f32(0.0), |
| 488 | IrType::Float(_) => b.const_f64(0.0), |
| 489 | IrType::Int(IntWidth::I64) => b.const_i64(0), |
| 490 | _ => b.const_i32(0), |
| 491 | }; |
| 492 | let b_pos = if ty.is_float() { |
| 493 | b.fcmp(CmpOp::Ge, args[1], zero) |
| 494 | } else { |
| 495 | b.icmp(CmpOp::Ge, args[1], zero) |
| 496 | }; |
| 497 | Some(b.select(b_pos, abs_a, neg_abs)) |
| 498 | } else { |
| 499 | None |
| 500 | } |
| 501 | } |
| 502 | "sqrt" | "dsqrt" => args.first().map(|a| b.fsqrt(*a)), |
| 503 | // ---- Bit manipulation (inline) ---- |
| 504 | // Mixed-kind bit ops (e.g. iand(c_long, c_int)) must unify |
| 505 | // widths to the wider operand before the IR-level bit_and, |
| 506 | // or the verifier rejects "operand width mismatch". F2018 |
| 507 | // §16.9.104 doesn't require same kinds; gfortran silently |
| 508 | // promotes. Audit31 Finding 14. |
| 509 | "iand" => { |
| 510 | if args.len() >= 2 { |
| 511 | let (l, r) = unify_int_widths(b, args[0], args[1]); |
| 512 | Some(b.bit_and(l, r)) |
| 513 | } else { |
| 514 | None |
| 515 | } |
| 516 | } |
| 517 | "ior" => { |
| 518 | if args.len() >= 2 { |
| 519 | let (l, r) = unify_int_widths(b, args[0], args[1]); |
| 520 | Some(b.bit_or(l, r)) |
| 521 | } else { |
| 522 | None |
| 523 | } |
| 524 | } |
| 525 | "ieor" => { |
| 526 | if args.len() >= 2 { |
| 527 | let (l, r) = unify_int_widths(b, args[0], args[1]); |
| 528 | Some(b.bit_xor(l, r)) |
| 529 | } else { |
| 530 | None |
| 531 | } |
| 532 | } |
| 533 | "not" => args.first().map(|a| b.bit_not(*a)), |
| 534 | "leadz" => args.first().map(|a| b.clz(*a)), |
| 535 | "trailz" => args.first().map(|a| b.ctz(*a)), |
| 536 | "popcount" | "popcnt" => { |
| 537 | // Use __builtin_popcountll via runtime call since ARM64 NEON popcount |
| 538 | // requires a complex instruction sequence. |
| 539 | args.first().map(|a| { |
| 540 | let widened = b.int_extend(*a, IntWidth::I64, false); |
| 541 | b.call( |
| 542 | FuncRef::External("afs_popcount".into()), |
| 543 | vec![widened], |
| 544 | IrType::Int(IntWidth::I32), |
| 545 | ) |
| 546 | }) |
| 547 | } |
| 548 | "ishft" => { |
| 549 | // F2018 §16.9.95: ISHFT does a *logical* shift on the bit |
| 550 | // representation of the integer. For int8/int16 values |
| 551 | // already sign-extended into the 32-bit AArch64 register |
| 552 | // (e.g. -32767_int16 = 0x8001 lives as 0xFFFF8001 in |
| 553 | // w-reg), `lshr` would shift the upper sign-fill bits in |
| 554 | // alongside, producing 0x00FFFF80 instead of 0x0080. |
| 555 | // Mask args[0] to the kind's width before the shift so |
| 556 | // the logical-right shift sees the unsigned bit pattern. |
| 557 | if args.len() >= 2 { |
| 558 | let shift_cmp_width = int_width_of_value(b, args[1]).unwrap_or(IntWidth::I32); |
| 559 | let zero = int_const_for_width(b, shift_cmp_width, 0); |
| 560 | let is_left = b.icmp(CmpOp::Ge, args[1], zero); |
| 561 | let neg_shift = b.ineg(args[1]); |
| 562 | let value_width = int_width_of_value(b, args[0]).unwrap_or(IntWidth::I32); |
| 563 | let shift = coerce_int_like_to_width(b, args[1], value_width); |
| 564 | let neg_shift = coerce_int_like_to_width(b, neg_shift, value_width); |
| 565 | let masked_value = match value_width { |
| 566 | IntWidth::I8 => { |
| 567 | let mask = int_const_for_width(b, value_width, 0xFF); |
| 568 | b.bit_and(args[0], mask) |
| 569 | } |
| 570 | IntWidth::I16 => { |
| 571 | let mask = int_const_for_width(b, value_width, 0xFFFF); |
| 572 | b.bit_and(args[0], mask) |
| 573 | } |
| 574 | _ => args[0], |
| 575 | }; |
| 576 | let left = b.shl(args[0], shift); |
| 577 | let right = b.lshr(masked_value, neg_shift); |
| 578 | Some(b.select(is_left, left, right)) |
| 579 | } else { |
| 580 | None |
| 581 | } |
| 582 | } |
| 583 | "shiftl" => { |
| 584 | // F2008 §13.7.150: logical left shift, shift>=0. |
| 585 | if args.len() >= 2 { |
| 586 | let value_width = int_width_of_value(b, args[0]).unwrap_or(IntWidth::I32); |
| 587 | let shift = coerce_int_like_to_width(b, args[1], value_width); |
| 588 | Some(b.shl(args[0], shift)) |
| 589 | } else { |
| 590 | None |
| 591 | } |
| 592 | } |
| 593 | "shiftr" => { |
| 594 | // F2008 §13.7.151: logical right shift (zero fill). |
| 595 | if args.len() >= 2 { |
| 596 | let value_width = int_width_of_value(b, args[0]).unwrap_or(IntWidth::I32); |
| 597 | let shift = coerce_int_like_to_width(b, args[1], value_width); |
| 598 | Some(b.lshr(args[0], shift)) |
| 599 | } else { |
| 600 | None |
| 601 | } |
| 602 | } |
| 603 | "shifta" => { |
| 604 | // F2008 §13.7.149: arithmetic right shift (sign-extending). |
| 605 | // No native ashr in the IR yet — synthesize as |
| 606 | // shifta(x, n) = lshr(x, n) | (sign_mask << (width - n)) |
| 607 | // where sign_mask is all-ones if MSB(x) is set, else 0. |
| 608 | if args.len() >= 2 { |
| 609 | let value_width = int_width_of_value(b, args[0]).unwrap_or(IntWidth::I32); |
| 610 | let shift = coerce_int_like_to_width(b, args[1], value_width); |
| 611 | let logical = b.lshr(args[0], shift); |
| 612 | let bits = int_const_for_width(b, value_width, value_width.bits() as i64); |
| 613 | let pos_top = b.isub(bits, shift); |
| 614 | // Pre-fill: -1 if top bit of x is 1, else 0. |
| 615 | let one = int_const_for_width(b, value_width, 1); |
| 616 | let top_bit_pos = int_const_for_width(b, value_width, (value_width.bits() - 1) as i64); |
| 617 | let top_bit = b.lshr(args[0], top_bit_pos); |
| 618 | let sign = b.bit_and(top_bit, one); |
| 619 | let neg_one = int_const_for_width(b, value_width, -1); |
| 620 | let zero = int_const_for_width(b, value_width, 0); |
| 621 | let is_neg = b.icmp(CmpOp::Ne, sign, zero); |
| 622 | let mask_full = b.select(is_neg, neg_one, zero); |
| 623 | let high_mask = b.shl(mask_full, pos_top); |
| 624 | Some(b.bit_or(logical, high_mask)) |
| 625 | } else { |
| 626 | None |
| 627 | } |
| 628 | } |
| 629 | "btest" => { |
| 630 | // btest(a, pos) = (a >> pos) & 1 /= 0 |
| 631 | if args.len() >= 2 { |
| 632 | let value_width = int_width_of_value(b, args[0]).unwrap_or(IntWidth::I32); |
| 633 | let pos = coerce_int_like_to_width(b, args[1], value_width); |
| 634 | let shifted = b.lshr(args[0], pos); |
| 635 | let one = int_const_for_width(b, value_width, 1); |
| 636 | let masked = b.bit_and(shifted, one); |
| 637 | let zero = int_const_for_width(b, value_width, 0); |
| 638 | Some(b.icmp(CmpOp::Ne, masked, zero)) |
| 639 | } else { |
| 640 | None |
| 641 | } |
| 642 | } |
| 643 | "ibset" => { |
| 644 | // ibset(a, pos) = a | (1 << pos) |
| 645 | if args.len() >= 2 { |
| 646 | let value_width = int_width_of_value(b, args[0]).unwrap_or(IntWidth::I32); |
| 647 | let one = int_const_for_width(b, value_width, 1); |
| 648 | let pos = coerce_int_like_to_width(b, args[1], value_width); |
| 649 | let mask = b.shl(one, pos); |
| 650 | Some(b.bit_or(args[0], mask)) |
| 651 | } else { |
| 652 | None |
| 653 | } |
| 654 | } |
| 655 | "ibclr" => { |
| 656 | // ibclr(a, pos) = a & ~(1 << pos) |
| 657 | if args.len() >= 2 { |
| 658 | let value_width = int_width_of_value(b, args[0]).unwrap_or(IntWidth::I32); |
| 659 | let one = int_const_for_width(b, value_width, 1); |
| 660 | let pos = coerce_int_like_to_width(b, args[1], value_width); |
| 661 | let mask = b.shl(one, pos); |
| 662 | let inv = b.bit_not(mask); |
| 663 | Some(b.bit_and(args[0], inv)) |
| 664 | } else { |
| 665 | None |
| 666 | } |
| 667 | } |
| 668 | "ibits" => { |
| 669 | // ibits(i, pos, len) = (i >> pos) & ((1 << len) - 1) |
| 670 | if args.len() >= 3 { |
| 671 | let value_width = int_width_of_value(b, args[0]).unwrap_or(IntWidth::I32); |
| 672 | let pos = coerce_int_like_to_width(b, args[1], value_width); |
| 673 | let len = coerce_int_like_to_width(b, args[2], value_width); |
| 674 | let shifted = b.lshr(args[0], pos); |
| 675 | let one = int_const_for_width(b, value_width, 1); |
| 676 | let mask_hi = b.shl(one, len); |
| 677 | let one2 = int_const_for_width(b, value_width, 1); |
| 678 | let mask = b.isub(mask_hi, one2); |
| 679 | Some(b.bit_and(shifted, mask)) |
| 680 | } else { |
| 681 | None |
| 682 | } |
| 683 | } |
| 684 | // F2018 §16.9.21-24 — unsigned bitwise comparisons. Implemented |
| 685 | // via the "flip the sign bit, then signed compare" trick so the |
| 686 | // existing signed icmp ops produce unsigned ordering. |
| 687 | "bge" | "bgt" | "ble" | "blt" => { |
| 688 | if args.len() >= 2 { |
| 689 | let (l, r) = unify_int_widths(b, args[0], args[1]); |
| 690 | let value_width = int_width_of_value(b, l).unwrap_or(IntWidth::I32); |
| 691 | let sign_bit_pos = |
| 692 | int_const_for_width(b, value_width, (value_width.bits() - 1) as i64); |
| 693 | let one = int_const_for_width(b, value_width, 1); |
| 694 | let sign_mask = b.shl(one, sign_bit_pos); |
| 695 | let l_flipped = b.bit_xor(l, sign_mask); |
| 696 | let r_flipped = b.bit_xor(r, sign_mask); |
| 697 | let op = match name { |
| 698 | "bge" => CmpOp::Ge, |
| 699 | "bgt" => CmpOp::Gt, |
| 700 | "ble" => CmpOp::Le, |
| 701 | "blt" => CmpOp::Lt, |
| 702 | _ => unreachable!(), |
| 703 | }; |
| 704 | Some(b.icmp(op, l_flipped, r_flipped)) |
| 705 | } else { |
| 706 | None |
| 707 | } |
| 708 | } |
| 709 | "poppar" => { |
| 710 | // F2018 §16.9.179: 1 if popcount(i) is odd, 0 otherwise. |
| 711 | args.first().map(|a| { |
| 712 | let widened = b.int_extend(*a, IntWidth::I64, false); |
| 713 | let cnt = b.call( |
| 714 | FuncRef::External("afs_popcount".into()), |
| 715 | vec![widened], |
| 716 | IrType::Int(IntWidth::I32), |
| 717 | ); |
| 718 | let one = b.const_i32(1); |
| 719 | b.bit_and(cnt, one) |
| 720 | }) |
| 721 | } |
| 722 | "merge_bits" => { |
| 723 | // F2018 §16.9.150: (i AND mask) IOR (j AND NOT mask) |
| 724 | if args.len() >= 3 { |
| 725 | let (i, j) = unify_int_widths(b, args[0], args[1]); |
| 726 | let value_width = int_width_of_value(b, i).unwrap_or(IntWidth::I32); |
| 727 | let mask = coerce_int_like_to_width(b, args[2], value_width); |
| 728 | let lhs = b.bit_and(i, mask); |
| 729 | let inv_mask = b.bit_not(mask); |
| 730 | let rhs = b.bit_and(j, inv_mask); |
| 731 | Some(b.bit_or(lhs, rhs)) |
| 732 | } else { |
| 733 | None |
| 734 | } |
| 735 | } |
| 736 | "maskl" => { |
| 737 | // F2018 §16.9.139: i leftmost bits set; rest cleared. |
| 738 | // maskl(i) = (-1) << (bits - i) if i > 0, else 0. |
| 739 | args.first().map(|a| { |
| 740 | let value_width = int_width_of_value(b, *a).unwrap_or(IntWidth::I32); |
| 741 | let bits = int_const_for_width(b, value_width, value_width.bits() as i64); |
| 742 | let i_in_w = coerce_int_like_to_width(b, *a, value_width); |
| 743 | let shift = b.isub(bits, i_in_w); |
| 744 | let neg_one = int_const_for_width(b, value_width, -1); |
| 745 | let zero = int_const_for_width(b, value_width, 0); |
| 746 | let shifted = b.shl(neg_one, shift); |
| 747 | let is_zero = b.icmp(CmpOp::Le, i_in_w, zero); |
| 748 | b.select(is_zero, zero, shifted) |
| 749 | }) |
| 750 | } |
| 751 | "maskr" => { |
| 752 | // F2018 §16.9.140: i rightmost bits set; rest cleared. |
| 753 | // maskr(i) = (1 << i) - 1 for 0 < i < bits; 0 for i==0; -1 for i>=bits. |
| 754 | args.first().map(|a| { |
| 755 | let value_width = int_width_of_value(b, *a).unwrap_or(IntWidth::I32); |
| 756 | let one = int_const_for_width(b, value_width, 1); |
| 757 | let i_in_w = coerce_int_like_to_width(b, *a, value_width); |
| 758 | let shifted = b.shl(one, i_in_w); |
| 759 | let one_again = int_const_for_width(b, value_width, 1); |
| 760 | let computed = b.isub(shifted, one_again); |
| 761 | let zero = int_const_for_width(b, value_width, 0); |
| 762 | let bits = int_const_for_width(b, value_width, value_width.bits() as i64); |
| 763 | let neg_one = int_const_for_width(b, value_width, -1); |
| 764 | let too_big = b.icmp(CmpOp::Ge, i_in_w, bits); |
| 765 | let is_zero = b.icmp(CmpOp::Le, i_in_w, zero); |
| 766 | let big_or_normal = b.select(too_big, neg_one, computed); |
| 767 | b.select(is_zero, zero, big_or_normal) |
| 768 | }) |
| 769 | } |
| 770 | "dshiftl" => { |
| 771 | // F2018 §16.9.59: combine i and j as a 2*bits value with |
| 772 | // i on the left, then logical-shift left by `shift` and |
| 773 | // return the leftmost `bits` bits. |
| 774 | // dshiftl(i, j, s) = (i << s) | (j >> (bits - s)) |
| 775 | if args.len() >= 3 { |
| 776 | let (i, j) = unify_int_widths(b, args[0], args[1]); |
| 777 | let value_width = int_width_of_value(b, i).unwrap_or(IntWidth::I32); |
| 778 | let shift = coerce_int_like_to_width(b, args[2], value_width); |
| 779 | let bits = int_const_for_width(b, value_width, value_width.bits() as i64); |
| 780 | let comp = b.isub(bits, shift); |
| 781 | let left = b.shl(i, shift); |
| 782 | let right = b.lshr(j, comp); |
| 783 | Some(b.bit_or(left, right)) |
| 784 | } else { |
| 785 | None |
| 786 | } |
| 787 | } |
| 788 | "dshiftr" => { |
| 789 | // F2018 §16.9.60: combine i and j and shift right by `shift`, |
| 790 | // returning the rightmost `bits` bits. |
| 791 | // dshiftr(i, j, s) = (j >> s) | (i << (bits - s)) |
| 792 | if args.len() >= 3 { |
| 793 | let (i, j) = unify_int_widths(b, args[0], args[1]); |
| 794 | let value_width = int_width_of_value(b, i).unwrap_or(IntWidth::I32); |
| 795 | let shift = coerce_int_like_to_width(b, args[2], value_width); |
| 796 | let bits = int_const_for_width(b, value_width, value_width.bits() as i64); |
| 797 | let comp = b.isub(bits, shift); |
| 798 | let right = b.lshr(j, shift); |
| 799 | let left = b.shl(i, comp); |
| 800 | Some(b.bit_or(left, right)) |
| 801 | } else { |
| 802 | None |
| 803 | } |
| 804 | } |
| 805 | // ---- Math intrinsics → libm calls ---- |
| 806 | // Dispatch to sinf/sin based on argument type for F32/F64 correctness. |
| 807 | "sin" | "dsin" | "cos" | "dcos" | "tan" | "dtan" | "asin" | "dasin" | "acos" | "dacos" |
| 808 | | "atan" | "datan" | "sinh" | "dsinh" | "cosh" | "dcosh" | "tanh" | "dtanh" | "asinh" |
| 809 | | "acosh" | "atanh" | "exp" | "dexp" | "log" | "dlog" | "alog" | "log10" | "dlog10" |
| 810 | | "alog10" | "erf" | "derf" | "erfc" | "derfc" | "ceiling" | "floor" => { |
| 811 | if let Some(arg) = args.first() { |
| 812 | let ty = b |
| 813 | .func() |
| 814 | .value_type(*arg) |
| 815 | .unwrap_or(IrType::Float(FloatWidth::F64)); |
| 816 | let is_f32 = matches!(ty, IrType::Float(FloatWidth::F32)); |
| 817 | let base_name = match name { |
| 818 | "dsin" | "sin" => "sin", |
| 819 | "dcos" | "cos" => "cos", |
| 820 | "dtan" | "tan" => "tan", |
| 821 | "dasin" | "asin" => "asin", |
| 822 | "dacos" | "acos" => "acos", |
| 823 | "datan" | "atan" => "atan", |
| 824 | "dsinh" | "sinh" => "sinh", |
| 825 | "dcosh" | "cosh" => "cosh", |
| 826 | "dtanh" | "tanh" => "tanh", |
| 827 | "asinh" => "asinh", |
| 828 | "acosh" => "acosh", |
| 829 | "atanh" => "atanh", |
| 830 | "dexp" | "exp" => "exp", |
| 831 | "dlog" | "log" | "alog" => "log", |
| 832 | "dlog10" | "log10" | "alog10" => "log10", |
| 833 | "derf" | "erf" => "erf", |
| 834 | "derfc" | "erfc" => "erfc", |
| 835 | "ceiling" => "ceil", |
| 836 | "floor" => "floor", |
| 837 | _ => name, |
| 838 | }; |
| 839 | let func_name = if is_f32 { |
| 840 | format!("{}f", base_name) |
| 841 | } else { |
| 842 | base_name.to_string() |
| 843 | }; |
| 844 | let ret_ty = if is_f32 { |
| 845 | IrType::Float(FloatWidth::F32) |
| 846 | } else { |
| 847 | IrType::Float(FloatWidth::F64) |
| 848 | }; |
| 849 | Some(b.call(FuncRef::External(func_name), vec![*arg], ret_ty)) |
| 850 | } else { |
| 851 | None |
| 852 | } |
| 853 | } |
| 854 | "atan2" | "datan2" => { |
| 855 | if args.len() >= 2 { |
| 856 | let ty = b |
| 857 | .func() |
| 858 | .value_type(args[0]) |
| 859 | .unwrap_or(IrType::Float(FloatWidth::F64)); |
| 860 | let is_f32 = matches!(ty, IrType::Float(FloatWidth::F32)); |
| 861 | let func = if is_f32 { "atan2f" } else { "atan2" }; |
| 862 | let ret_ty = if is_f32 { |
| 863 | IrType::Float(FloatWidth::F32) |
| 864 | } else { |
| 865 | IrType::Float(FloatWidth::F64) |
| 866 | }; |
| 867 | Some(b.call( |
| 868 | FuncRef::External(func.into()), |
| 869 | vec![args[0], args[1]], |
| 870 | ret_ty, |
| 871 | )) |
| 872 | } else { |
| 873 | None |
| 874 | } |
| 875 | } |
| 876 | "gamma" | "dgamma" => args.first().map(|a| { |
| 877 | let ty = b |
| 878 | .func() |
| 879 | .value_type(*a) |
| 880 | .unwrap_or(IrType::Float(FloatWidth::F64)); |
| 881 | let is_f32 = matches!(ty, IrType::Float(FloatWidth::F32)); |
| 882 | let func = if is_f32 { "tgammaf" } else { "tgamma" }; |
| 883 | let ret_ty = if is_f32 { |
| 884 | IrType::Float(FloatWidth::F32) |
| 885 | } else { |
| 886 | IrType::Float(FloatWidth::F64) |
| 887 | }; |
| 888 | b.call(FuncRef::External(func.into()), vec![*a], ret_ty) |
| 889 | }), |
| 890 | "log_gamma" => args.first().map(|a| { |
| 891 | let ty = b |
| 892 | .func() |
| 893 | .value_type(*a) |
| 894 | .unwrap_or(IrType::Float(FloatWidth::F64)); |
| 895 | let is_f32 = matches!(ty, IrType::Float(FloatWidth::F32)); |
| 896 | let func = if is_f32 { "lgammaf" } else { "lgamma" }; |
| 897 | let ret_ty = if is_f32 { |
| 898 | IrType::Float(FloatWidth::F32) |
| 899 | } else { |
| 900 | IrType::Float(FloatWidth::F64) |
| 901 | }; |
| 902 | b.call(FuncRef::External(func.into()), vec![*a], ret_ty) |
| 903 | }), |
| 904 | "bessel_j0" => args.first().map(|a| { |
| 905 | b.call( |
| 906 | FuncRef::External("j0".into()), |
| 907 | vec![*a], |
| 908 | IrType::Float(FloatWidth::F64), |
| 909 | ) |
| 910 | }), |
| 911 | "bessel_j1" => args.first().map(|a| { |
| 912 | b.call( |
| 913 | FuncRef::External("j1".into()), |
| 914 | vec![*a], |
| 915 | IrType::Float(FloatWidth::F64), |
| 916 | ) |
| 917 | }), |
| 918 | "bessel_y0" => args.first().map(|a| { |
| 919 | b.call( |
| 920 | FuncRef::External("y0".into()), |
| 921 | vec![*a], |
| 922 | IrType::Float(FloatWidth::F64), |
| 923 | ) |
| 924 | }), |
| 925 | "bessel_y1" => args.first().map(|a| { |
| 926 | b.call( |
| 927 | FuncRef::External("y1".into()), |
| 928 | vec![*a], |
| 929 | IrType::Float(FloatWidth::F64), |
| 930 | ) |
| 931 | }), |
| 932 | "hypot" => { |
| 933 | if args.len() >= 2 { |
| 934 | let ty = b |
| 935 | .func() |
| 936 | .value_type(args[0]) |
| 937 | .unwrap_or(IrType::Float(FloatWidth::F64)); |
| 938 | let is_f32 = matches!(ty, IrType::Float(FloatWidth::F32)); |
| 939 | let func = if is_f32 { "hypotf" } else { "hypot" }; |
| 940 | let ret_ty = if is_f32 { |
| 941 | IrType::Float(FloatWidth::F32) |
| 942 | } else { |
| 943 | IrType::Float(FloatWidth::F64) |
| 944 | }; |
| 945 | Some(b.call( |
| 946 | FuncRef::External(func.into()), |
| 947 | vec![args[0], args[1]], |
| 948 | ret_ty, |
| 949 | )) |
| 950 | } else { |
| 951 | None |
| 952 | } |
| 953 | } |
| 954 | "ishftc" => { |
| 955 | // ishftc(a, shift, size): circular shift of the rightmost `size` bits. |
| 956 | // F2018 §16.9.108. The previous implementation used |
| 957 | // mask = (1 << size) - 1 |
| 958 | // which is undefined when size equals the operand bit width |
| 959 | // (AArch64 LSL masks the shift amount mod 64, so 1 << 64 wraps |
| 960 | // back to 1, leaving mask = 0 and erasing the result). |
| 961 | // stdlib_random's xoshiro256ss called ishftc(x, 7) with the |
| 962 | // default size = 64 and got 0 every time. |
| 963 | if args.len() >= 2 { |
| 964 | let value_width = int_width_of_value(b, args[0]).unwrap_or(IntWidth::I32); |
| 965 | let bit_width = match value_width { |
| 966 | IntWidth::I64 => 64, |
| 967 | IntWidth::I16 => 16, |
| 968 | IntWidth::I8 => 8, |
| 969 | _ => 32, |
| 970 | }; |
| 971 | let shift = coerce_int_like_to_width(b, args[1], value_width); |
| 972 | let bw_minus_1 = int_const_for_width(b, value_width, bit_width - 1); |
| 973 | if args.len() >= 3 { |
| 974 | // Explicit size: rotate within the rightmost `size` bits, |
| 975 | // leaving the upper (bit_width - size) bits untouched. |
| 976 | let size = coerce_int_like_to_width(b, args[2], value_width); |
| 977 | // Build mask = ((1 << (size-1)) - 1) << 1 | 1. |
| 978 | // Valid for size in [1, bit_width]; at size == bit_width |
| 979 | // this yields all ones without ever shifting by bit_width. |
| 980 | let one_a = int_const_for_width(b, value_width, 1); |
| 981 | let one_b = int_const_for_width(b, value_width, 1); |
| 982 | let one_c = int_const_for_width(b, value_width, 1); |
| 983 | let one_d = int_const_for_width(b, value_width, 1); |
| 984 | let size_minus_1 = b.isub(size, one_a); |
| 985 | let half = b.shl(one_b, size_minus_1); |
| 986 | let half_minus_1 = b.isub(half, one_c); |
| 987 | let half_minus_1_shifted = b.shl(half_minus_1, one_d); |
| 988 | let one_e = int_const_for_width(b, value_width, 1); |
| 989 | let mask = b.bit_or(half_minus_1_shifted, one_e); |
| 990 | let not_mask_pre = int_const_for_width(b, value_width, -1); |
| 991 | let not_mask = b.bit_xor(not_mask_pre, mask); |
| 992 | // Rotate the low bits, preserve the high bits. |
| 993 | let low = b.bit_and(args[0], mask); |
| 994 | let high = b.bit_and(args[0], not_mask); |
| 995 | // shift_safe = shift mod size (avoid UB when shift == size). |
| 996 | // For the common stdlib usage shift < size, so the modulo is |
| 997 | // a no-op; we still emit isub-based fallback in case. |
| 998 | let left_pre = b.shl(low, shift); |
| 999 | let left = b.bit_and(left_pre, mask); |
| 1000 | let diff = b.isub(size, shift); |
| 1001 | let right = b.lshr(low, diff); |
| 1002 | let rotated = b.bit_or(left, right); |
| 1003 | let rotated_low = b.bit_and(rotated, mask); |
| 1004 | Some(b.bit_or(rotated_low, high)) |
| 1005 | } else { |
| 1006 | // Default size: full-width rotate. Use the standard |
| 1007 | // (a << (s & (BITS-1))) | (a >> ((-s) & (BITS-1))) |
| 1008 | // formula which is well-defined for shift = 0 because |
| 1009 | // both lanes shift by 0 and OR back the same value. |
| 1010 | let s_masked = b.bit_and(shift, bw_minus_1); |
| 1011 | let bw_const = int_const_for_width(b, value_width, bit_width); |
| 1012 | let neg_s = b.isub(bw_const, s_masked); |
| 1013 | let bw_minus_1_b = int_const_for_width(b, value_width, bit_width - 1); |
| 1014 | let neg_s_masked = b.bit_and(neg_s, bw_minus_1_b); |
| 1015 | let left = b.shl(args[0], s_masked); |
| 1016 | let right = b.lshr(args[0], neg_s_masked); |
| 1017 | Some(b.bit_or(left, right)) |
| 1018 | } |
| 1019 | } else { |
| 1020 | None |
| 1021 | } |
| 1022 | } |
| 1023 | |
| 1024 | // ---- Numeric inquiry intrinsics (compile-time constants) ---- |
| 1025 | // These depend on the argument's type, which we determine from the first arg. |
| 1026 | "huge" => { |
| 1027 | if let Some(arg) = args.first() { |
| 1028 | let ty = b |
| 1029 | .func() |
| 1030 | .value_type(*arg) |
| 1031 | .unwrap_or(IrType::Int(IntWidth::I32)); |
| 1032 | match &ty { |
| 1033 | IrType::Int(IntWidth::I8) => Some(b.const_i32(i8::MAX as i64 as i32)), |
| 1034 | IrType::Int(IntWidth::I16) => Some(b.const_i32(i16::MAX as i64 as i32)), |
| 1035 | IrType::Int(IntWidth::I32) => Some(b.const_i32(i32::MAX)), |
| 1036 | IrType::Int(IntWidth::I64) => Some(b.const_i64(i64::MAX)), |
| 1037 | IrType::Float(FloatWidth::F32) => Some(b.const_f32(f32::MAX)), |
| 1038 | IrType::Float(FloatWidth::F64) => Some(b.const_f64(f64::MAX)), |
| 1039 | _ => None, |
| 1040 | } |
| 1041 | } else { |
| 1042 | None |
| 1043 | } |
| 1044 | } |
| 1045 | "tiny" => { |
| 1046 | if let Some(arg) = args.first() { |
| 1047 | let ty = b |
| 1048 | .func() |
| 1049 | .value_type(*arg) |
| 1050 | .unwrap_or(IrType::Float(FloatWidth::F32)); |
| 1051 | match &ty { |
| 1052 | IrType::Float(FloatWidth::F32) => Some(b.const_f32(f32::MIN_POSITIVE)), |
| 1053 | IrType::Float(FloatWidth::F64) => Some(b.const_f64(f64::MIN_POSITIVE)), |
| 1054 | _ => None, |
| 1055 | } |
| 1056 | } else { |
| 1057 | None |
| 1058 | } |
| 1059 | } |
| 1060 | "epsilon" => { |
| 1061 | if let Some(arg) = args.first() { |
| 1062 | let ty = b |
| 1063 | .func() |
| 1064 | .value_type(*arg) |
| 1065 | .unwrap_or(IrType::Float(FloatWidth::F32)); |
| 1066 | match &ty { |
| 1067 | IrType::Float(FloatWidth::F32) => Some(b.const_f32(f32::EPSILON)), |
| 1068 | IrType::Float(FloatWidth::F64) => Some(b.const_f64(f64::EPSILON)), |
| 1069 | _ => None, |
| 1070 | } |
| 1071 | } else { |
| 1072 | None |
| 1073 | } |
| 1074 | } |
| 1075 | "precision" => { |
| 1076 | if let Some(arg) = args.first() { |
| 1077 | let ty = b |
| 1078 | .func() |
| 1079 | .value_type(*arg) |
| 1080 | .unwrap_or(IrType::Float(FloatWidth::F32)); |
| 1081 | let prec = match &ty { |
| 1082 | IrType::Float(FloatWidth::F32) => 6, // ~7.2 decimal digits → 6 |
| 1083 | IrType::Float(FloatWidth::F64) => 15, // ~15.9 decimal digits → 15 |
| 1084 | _ => 0, |
| 1085 | }; |
| 1086 | Some(b.const_i32(prec)) |
| 1087 | } else { |
| 1088 | None |
| 1089 | } |
| 1090 | } |
| 1091 | "range" => { |
| 1092 | if let Some(arg) = args.first() { |
| 1093 | let ty = b |
| 1094 | .func() |
| 1095 | .value_type(*arg) |
| 1096 | .unwrap_or(IrType::Int(IntWidth::I32)); |
| 1097 | let range = match &ty { |
| 1098 | IrType::Int(IntWidth::I8) => 2, |
| 1099 | IrType::Int(IntWidth::I16) => 4, |
| 1100 | IrType::Int(IntWidth::I32) => 9, |
| 1101 | IrType::Int(IntWidth::I64) => 18, |
| 1102 | IrType::Int(IntWidth::I128) => 38, |
| 1103 | IrType::Float(FloatWidth::F32) => 37, |
| 1104 | IrType::Float(FloatWidth::F64) => 307, |
| 1105 | _ => 0, |
| 1106 | }; |
| 1107 | Some(b.const_i32(range)) |
| 1108 | } else { |
| 1109 | None |
| 1110 | } |
| 1111 | } |
| 1112 | "digits" => { |
| 1113 | if let Some(arg) = args.first() { |
| 1114 | let ty = b |
| 1115 | .func() |
| 1116 | .value_type(*arg) |
| 1117 | .unwrap_or(IrType::Int(IntWidth::I32)); |
| 1118 | let digits = match &ty { |
| 1119 | IrType::Int(IntWidth::I8) => 7, |
| 1120 | IrType::Int(IntWidth::I16) => 15, |
| 1121 | IrType::Int(IntWidth::I32) => 31, |
| 1122 | IrType::Int(IntWidth::I64) => 63, |
| 1123 | IrType::Int(IntWidth::I128) => 127, |
| 1124 | IrType::Float(FloatWidth::F32) => 24, // significand bits |
| 1125 | IrType::Float(FloatWidth::F64) => 53, |
| 1126 | _ => 0, |
| 1127 | }; |
| 1128 | Some(b.const_i32(digits)) |
| 1129 | } else { |
| 1130 | None |
| 1131 | } |
| 1132 | } |
| 1133 | "radix" => { |
| 1134 | // Always 2 for binary machines. |
| 1135 | Some(b.const_i32(2)) |
| 1136 | } |
| 1137 | "bit_size" => { |
| 1138 | if let Some(arg) = args.first() { |
| 1139 | let ty = b |
| 1140 | .func() |
| 1141 | .value_type(*arg) |
| 1142 | .unwrap_or(IrType::Int(IntWidth::I32)); |
| 1143 | let bits = match &ty { |
| 1144 | IrType::Int(IntWidth::I8) => 8, |
| 1145 | IrType::Int(IntWidth::I16) => 16, |
| 1146 | IrType::Int(IntWidth::I32) => 32, |
| 1147 | IrType::Int(IntWidth::I64) => 64, |
| 1148 | IrType::Int(IntWidth::I128) => 128, |
| 1149 | _ => 0, |
| 1150 | }; |
| 1151 | Some(b.const_i32(bits)) |
| 1152 | } else { |
| 1153 | None |
| 1154 | } |
| 1155 | } |
| 1156 | // F2018 §16.9.196: STORAGE_SIZE(A [, KIND]) returns the size in |
| 1157 | // bits a value of the same type as A occupies. For non-polymorphic |
| 1158 | // arguments this is determined entirely by the argument's IR type |
| 1159 | // and can be folded at compile time (8 * sizeof). The optional |
| 1160 | // KIND argument names the kind of the result; we always return I32 |
| 1161 | // and let downstream coerce. |
| 1162 | "storage_size" => { |
| 1163 | if let Some(arg) = args.first() { |
| 1164 | let ty = b |
| 1165 | .func() |
| 1166 | .value_type(*arg) |
| 1167 | .unwrap_or(IrType::Int(IntWidth::I32)); |
| 1168 | let bits = storage_size_bits_for_ir_type(&ty); |
| 1169 | Some(b.const_i32(bits)) |
| 1170 | } else { |
| 1171 | None |
| 1172 | } |
| 1173 | } |
| 1174 | "kind" => { |
| 1175 | if let Some(arg) = args.first() { |
| 1176 | let ty = b |
| 1177 | .func() |
| 1178 | .value_type(*arg) |
| 1179 | .unwrap_or(IrType::Int(IntWidth::I32)); |
| 1180 | let kind = match &ty { |
| 1181 | IrType::Int(IntWidth::I8) => 1, |
| 1182 | IrType::Int(IntWidth::I16) => 2, |
| 1183 | IrType::Int(IntWidth::I32) => 4, |
| 1184 | IrType::Int(IntWidth::I64) => 8, |
| 1185 | IrType::Int(IntWidth::I128) => 16, |
| 1186 | IrType::Float(FloatWidth::F32) => 4, |
| 1187 | IrType::Float(FloatWidth::F64) => 8, |
| 1188 | IrType::Bool => 4, |
| 1189 | _ => 4, |
| 1190 | }; |
| 1191 | Some(b.const_i32(kind)) |
| 1192 | } else { |
| 1193 | None |
| 1194 | } |
| 1195 | } |
| 1196 | // ---- System inquiry functions ---- |
| 1197 | "command_argument_count" => Some(b.call( |
| 1198 | FuncRef::External("afs_command_argument_count".into()), |
| 1199 | vec![], |
| 1200 | IrType::Int(IntWidth::I32), |
| 1201 | )), |
| 1202 | "is_iostat_end" => args.first().map(|status| { |
| 1203 | let zero = match b |
| 1204 | .func() |
| 1205 | .value_type(*status) |
| 1206 | .unwrap_or(IrType::Int(IntWidth::I32)) |
| 1207 | { |
| 1208 | IrType::Int(IntWidth::I64) => b.const_i64(-1), |
| 1209 | _ => b.const_i32(-1), |
| 1210 | }; |
| 1211 | b.icmp(CmpOp::Eq, *status, zero) |
| 1212 | }), |
| 1213 | "is_iostat_eor" => args.first().map(|status| { |
| 1214 | let zero = match b |
| 1215 | .func() |
| 1216 | .value_type(*status) |
| 1217 | .unwrap_or(IrType::Int(IntWidth::I32)) |
| 1218 | { |
| 1219 | IrType::Int(IntWidth::I64) => b.const_i64(-2), |
| 1220 | _ => b.const_i32(-2), |
| 1221 | }; |
| 1222 | b.icmp(CmpOp::Eq, *status, zero) |
| 1223 | }), |
| 1224 | |
| 1225 | // ---- iso_c_binding functions ---- |
| 1226 | "c_loc" => None, |
| 1227 | "c_sizeof" => { |
| 1228 | // c_sizeof(x) — return byte size of x's C representation. |
| 1229 | if let Some(arg) = args.first() { |
| 1230 | let ty = b |
| 1231 | .func() |
| 1232 | .value_type(*arg) |
| 1233 | .unwrap_or(IrType::Int(IntWidth::I32)); |
| 1234 | let size: i64 = match &ty { |
| 1235 | IrType::Int(IntWidth::I8) | IrType::Bool => 1, |
| 1236 | IrType::Int(IntWidth::I16) => 2, |
| 1237 | IrType::Int(IntWidth::I32) | IrType::Float(FloatWidth::F32) => 4, |
| 1238 | IrType::Int(IntWidth::I64) | IrType::Float(FloatWidth::F64) => 8, |
| 1239 | IrType::Int(IntWidth::I128) => 16, |
| 1240 | IrType::Ptr(_) => 8, // pointers are 8 bytes on ARM64 |
| 1241 | // Arrays use element size * count, but we don't have shape info here. |
| 1242 | // For now, return element size. Proper impl needs descriptor access. |
| 1243 | IrType::Array(elem, count) => { |
| 1244 | let elem_size = ir_scalar_byte_size(elem.as_ref()); |
| 1245 | elem_size * (*count as i64) |
| 1246 | } |
| 1247 | _ => 8, // default to pointer size for unknown types |
| 1248 | }; |
| 1249 | Some(b.const_i64(size)) |
| 1250 | } else { |
| 1251 | None |
| 1252 | } |
| 1253 | } |
| 1254 | "c_associated" => { |
| 1255 | // c_associated(p) → p /= null |
| 1256 | // c_associated(p, q) → p == q |
| 1257 | if args.len() >= 2 { |
| 1258 | Some(b.icmp(CmpOp::Eq, args[0], args[1])) |
| 1259 | } else if let Some(p) = args.first() { |
| 1260 | // Use type-matched zero to avoid register width mismatch. |
| 1261 | let ty = b |
| 1262 | .func() |
| 1263 | .value_type(*p) |
| 1264 | .unwrap_or(IrType::Int(IntWidth::I64)); |
| 1265 | let null = match &ty { |
| 1266 | IrType::Int(IntWidth::I32) => b.const_i32(0), |
| 1267 | _ => b.const_i64(0), |
| 1268 | }; |
| 1269 | Some(b.icmp(CmpOp::Ne, *p, null)) |
| 1270 | } else { |
| 1271 | None |
| 1272 | } |
| 1273 | } |
| 1274 | |
| 1275 | // ---- Kind selection intrinsics ---- |
| 1276 | "selected_int_kind" => { |
| 1277 | // selected_int_kind(r): smallest integer kind whose range covers [-10^r, 10^r]. |
| 1278 | if let Some(arg) = args.first() { |
| 1279 | if let Some(r) = extract_const_int_from_value(b, *arg) { |
| 1280 | let kind: i32 = if r <= 2 { |
| 1281 | 1 |
| 1282 | } |
| 1283 | // i8: ±127 |
| 1284 | else if r <= 4 { |
| 1285 | 2 |
| 1286 | } |
| 1287 | // i16: ±32767 |
| 1288 | else if r <= 9 { |
| 1289 | 4 |
| 1290 | } |
| 1291 | // i32: ±2.1e9 |
| 1292 | else if r <= 18 { |
| 1293 | 8 |
| 1294 | } |
| 1295 | // i64: ±9.2e18 |
| 1296 | else if r <= 38 { |
| 1297 | 16 |
| 1298 | } |
| 1299 | // i128: ±1.7e38 |
| 1300 | else { |
| 1301 | -1 |
| 1302 | }; // no kind available |
| 1303 | Some(b.const_i32(kind)) |
| 1304 | } else { |
| 1305 | Some(b.const_i32(4)) // non-constant: default to 4 |
| 1306 | } |
| 1307 | } else { |
| 1308 | None |
| 1309 | } |
| 1310 | } |
| 1311 | "selected_real_kind" => { |
| 1312 | // selected_real_kind(p[, r]): smallest real kind with ≥p decimal digits. |
| 1313 | if let Some(arg) = args.first() { |
| 1314 | if let Some(p) = extract_const_int_from_value(b, *arg) { |
| 1315 | let kind: i32 = if p <= 6 { |
| 1316 | 4 |
| 1317 | } |
| 1318 | // f32: ~7 digits |
| 1319 | else if p <= 15 { |
| 1320 | 8 |
| 1321 | } |
| 1322 | // f64: ~15 digits |
| 1323 | else { |
| 1324 | -1 |
| 1325 | }; // no kind available |
| 1326 | Some(b.const_i32(kind)) |
| 1327 | } else { |
| 1328 | Some(b.const_i32(8)) // non-constant: default to 8 |
| 1329 | } |
| 1330 | } else { |
| 1331 | None |
| 1332 | } |
| 1333 | } |
| 1334 | |
| 1335 | // ---- IEEE arithmetic intrinsics ---- |
| 1336 | "ieee_is_nan" => { |
| 1337 | // IEEE_IS_NAN(x) → x != x (NaN is the only value that is not equal to itself) |
| 1338 | args.first().map(|arg| b.fcmp(CmpOp::Ne, *arg, *arg)) |
| 1339 | } |
| 1340 | "ieee_is_finite" => { |
| 1341 | // IEEE_IS_FINITE(x) → (x - x) == 0.0 |
| 1342 | // For finite values, x-x is 0.0. For inf, x-x is NaN. For NaN, x-x is NaN. |
| 1343 | if let Some(arg) = args.first() { |
| 1344 | let diff = b.fsub(*arg, *arg); |
| 1345 | let ty = b |
| 1346 | .func() |
| 1347 | .value_type(*arg) |
| 1348 | .unwrap_or(IrType::Float(FloatWidth::F64)); |
| 1349 | let zero = match &ty { |
| 1350 | IrType::Float(FloatWidth::F32) => b.const_f32(0.0), |
| 1351 | _ => b.const_f64(0.0), |
| 1352 | }; |
| 1353 | Some(b.fcmp(CmpOp::Eq, diff, zero)) |
| 1354 | } else { |
| 1355 | None |
| 1356 | } |
| 1357 | } |
| 1358 | "ieee_support_datatype" |
| 1359 | | "ieee_support_denormal" |
| 1360 | | "ieee_support_inf" |
| 1361 | | "ieee_support_nan" |
| 1362 | | "ieee_support_subnormal" => { |
| 1363 | // ARM64 + Apple Silicon supports the full IEEE 754 model. |
| 1364 | Some(b.const_bool(true)) |
| 1365 | } |
| 1366 | "maxexponent" => { |
| 1367 | // F2018 §16.9.124: returns the maximum exponent in the model |
| 1368 | // for the same kind as the argument. For IEEE binary32 = 128, |
| 1369 | // binary64 = 1024. |
| 1370 | let arg_ty = args |
| 1371 | .first() |
| 1372 | .and_then(|a| b.func().value_type(*a)) |
| 1373 | .unwrap_or(IrType::Float(FloatWidth::F32)); |
| 1374 | let val = match arg_ty { |
| 1375 | IrType::Float(FloatWidth::F64) => 1024_i32, |
| 1376 | _ => 128_i32, |
| 1377 | }; |
| 1378 | Some(b.const_i32(val)) |
| 1379 | } |
| 1380 | "minexponent" => { |
| 1381 | // F2018 §16.9.146: minimum exponent in the model. binary32 = -125, |
| 1382 | // binary64 = -1021. |
| 1383 | let arg_ty = args |
| 1384 | .first() |
| 1385 | .and_then(|a| b.func().value_type(*a)) |
| 1386 | .unwrap_or(IrType::Float(FloatWidth::F32)); |
| 1387 | let val = match arg_ty { |
| 1388 | IrType::Float(FloatWidth::F64) => -1021_i32, |
| 1389 | _ => -125_i32, |
| 1390 | }; |
| 1391 | Some(b.const_i32(val)) |
| 1392 | } |
| 1393 | "ieee_value" => { |
| 1394 | if args.len() < 2 { |
| 1395 | None |
| 1396 | } else { |
| 1397 | let ty = b |
| 1398 | .func() |
| 1399 | .value_type(args[0]) |
| 1400 | .unwrap_or(IrType::Float(FloatWidth::F64)); |
| 1401 | let class = extract_const_int_from_value(b, args[1]).unwrap_or(0); |
| 1402 | let zero = match ty { |
| 1403 | IrType::Float(FloatWidth::F32) => b.const_f32(0.0), |
| 1404 | _ => b.const_f64(0.0), |
| 1405 | }; |
| 1406 | let one = match ty { |
| 1407 | IrType::Float(FloatWidth::F32) => b.const_f32(1.0), |
| 1408 | _ => b.const_f64(1.0), |
| 1409 | }; |
| 1410 | match class { |
| 1411 | 1 | 4 => Some(b.fdiv(zero, zero)), |
| 1412 | 2 => Some(b.fdiv(one, zero)), |
| 1413 | 3 => { |
| 1414 | let neg_one = match ty { |
| 1415 | IrType::Float(FloatWidth::F32) => b.const_f32(-1.0), |
| 1416 | _ => b.const_f64(-1.0), |
| 1417 | }; |
| 1418 | Some(b.fdiv(neg_one, zero)) |
| 1419 | } |
| 1420 | 5 => Some(zero), |
| 1421 | 6 => { |
| 1422 | let neg_one = match ty { |
| 1423 | IrType::Float(FloatWidth::F32) => b.const_f32(-1.0), |
| 1424 | _ => b.const_f64(-1.0), |
| 1425 | }; |
| 1426 | Some(b.fmul(neg_one, zero)) |
| 1427 | } |
| 1428 | _ => Some(args[0]), |
| 1429 | } |
| 1430 | } |
| 1431 | } |
| 1432 | |
| 1433 | _ => None, |
| 1434 | } |
| 1435 | } |
| 1436 |