cutile-compiler 0.1.0

Crate for compiling kernels authored in cuTile Rust to executable kernels.
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
435
436
437
438
439
440
441
442
443
444
445
446
447
448
449
450
451
452
453
454
455
456
457
458
459
460
461
462
463
464
465
466
467
468
469
470
471
472
473
474
/*
 * SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved.
 * SPDX-License-Identifier: Apache-2.0
 */

//! Binary operation compilation for compiler2.
//!
//! Mechanical port of `compiler/compile_binary_op.rs` — translates binary
//! arithmetic, comparison, and bitwise operations into tile-ir operations.
//! Only type and IR-emission changes; the dispatch logic and bounds
//! propagation are identical.

use quote::ToTokens;
use syn::spanned::Spanned;

use super::_function::CUDATileFunctionCompiler;
use super::_value::{CompilerContext, TileRustValue};
use super::shared_types::Kind;
use super::shared_utils::{get_tile_bop_from_rust_bop, TileBinaryOp};
use super::tile_rust_type::TileRustType;
use super::utils::{
    cmp_ordering_attr, cmp_pred_attr, flag_attr, rounding_mode_attr, signedness_attr, NamedAttr,
};
use crate::bounds::bounds_from_bop;
use crate::error::JITError;
use crate::generics::GenericVars;

use cutile_ir::builder::{append_op, OpBuilder};
use cutile_ir::bytecode::Opcode;
use cutile_ir::ir::{Attribute, BlockId, Module, ScalarType, TileElementType, TileType, Type};

use std::collections::HashMap;
use syn::ExprBinary;

/// Port of `get_cmp_predicate_attr` from `compiler/utils.rs`.
///
/// Returns a comparison-predicate named attribute for comparison binary ops,
/// or `None` for non-comparison ops.
fn get_cmp_predicate_attr_ir(expr: &TileBinaryOp) -> Result<Option<NamedAttr>, JITError> {
    match expr {
        TileBinaryOp::Eq => Ok(Some(cmp_pred_attr("equal"))),
        TileBinaryOp::Ne => Ok(Some(cmp_pred_attr("not_equal"))),
        TileBinaryOp::Lt => Ok(Some(cmp_pred_attr("less_than"))),
        TileBinaryOp::Le => Ok(Some(cmp_pred_attr("less_than_or_equal"))),
        TileBinaryOp::Gt => Ok(Some(cmp_pred_attr("greater_than"))),
        TileBinaryOp::Ge => Ok(Some(cmp_pred_attr("greater_than_or_equal"))),
        _ => Ok(None),
    }
}

/// Construct a tile-ir bool (i1) result type that mirrors the shape of `lhs_type`.
///
/// If `lhs_type` is a `Tile`, the result is a tile with the same shape but `I1`
/// element type. If it's a scalar, the result is `Scalar(I1)`.
fn make_bool_result_type(lhs_type: &Type) -> Type {
    match lhs_type {
        Type::Tile(tile_ty) => Type::Tile(TileType {
            shape: tile_ty.shape.clone(),
            element_type: TileElementType::Scalar(ScalarType::I1),
        }),
        _ => Type::Scalar(ScalarType::I1),
    }
}

impl<'m> CUDATileFunctionCompiler<'m> {
    pub fn compile_binary_op(
        &self,
        module: &mut Module,
        block_id: BlockId,
        bin_expr: &ExprBinary,
        generic_vars: &GenericVars,
        ctx: &mut CompilerContext,
        return_type: Option<TileRustType>,
    ) -> Result<Option<TileRustValue>, JITError> {
        let tile_binary_op = get_tile_bop_from_rust_bop(&bin_expr.op)?;
        let is_comparison = matches!(
            tile_binary_op,
            TileBinaryOp::Eq
                | TileBinaryOp::Ne
                | TileBinaryOp::Lt
                | TileBinaryOp::Le
                | TileBinaryOp::Gt
                | TileBinaryOp::Ge
        );
        let is_logical = matches!(bin_expr.op, syn::BinOp::And(_) | syn::BinOp::Or(_));
        let lhs_return_type = if is_comparison || is_logical {
            None
        } else {
            return_type.clone()
        };
        let lhs = self.compile_expression(
            module,
            block_id,
            &bin_expr.left,
            generic_vars,
            ctx,
            lhs_return_type,
        )?;
        if lhs.is_none() {
            return self.jit_error_result(
                &bin_expr.left.span(),
                "failed to compile the left-hand side of this binary operation",
            );
        }
        let lhs = lhs.unwrap();
        let rhs_return_type = if is_logical {
            None
        } else if is_comparison {
            Some(lhs.ty.clone())
        } else {
            return_type.clone().or_else(|| Some(lhs.ty.clone()))
        };
        let rhs = self.compile_expression(
            module,
            block_id,
            &bin_expr.right,
            generic_vars,
            ctx,
            rhs_return_type,
        )?;
        if rhs.is_none() {
            return self.jit_error_result(
                &bin_expr.right.span(),
                "failed to compile the right-hand side of this binary operation",
            );
        }
        let rhs = rhs.unwrap();
        Ok(Some(self.compile_binary_op_from_values(
            module,
            block_id,
            lhs,
            rhs,
            &tile_binary_op,
            generic_vars,
            ctx,
            return_type,
            &bin_expr.span(),
        )?))
    }

    pub fn compile_binary_op_from_values(
        &self,
        module: &mut Module,
        block_id: BlockId,
        lhs: TileRustValue,
        rhs: TileRustValue,
        tile_rust_arithmetic_op: &TileBinaryOp,
        generic_vars: &GenericVars,
        _ctx: &mut CompilerContext,
        return_type: Option<TileRustType>,
        span: &proc_macro2::Span,
    ) -> Result<TileRustValue, JITError> {
        if lhs.ty.rust_ty != rhs.ty.rust_ty {
            return self.jit_error_result(
                span,
                &format!(
                    "binary `{:?}` requires operands of the same type, but got `{}` and `{}`",
                    tile_rust_arithmetic_op,
                    lhs.ty.rust_ty.to_token_stream().to_string(),
                    rhs.ty.rust_ty.to_token_stream().to_string()
                ),
            );
        }
        let lhs_value = lhs.value;
        if lhs_value.is_none() {
            return self.jit_error_result(
                span,
                "left-hand side of binary operation did not produce a value",
            );
        }
        let lhs_value = lhs_value.unwrap();
        let rhs_value = rhs.value;
        if rhs_value.is_none() {
            return self.jit_error_result(
                span,
                "right-hand side of binary operation did not produce a value",
            );
        }
        let rhs_value = rhs_value.unwrap();
        let operand_type = lhs.ty.clone();
        let operand_rust_ty = &operand_type.rust_ty;
        let Some(operand_rust_element_type) =
            operand_type.get_instantiated_rust_element_type(&self.modules.primitives())
        else {
            return self.jit_error_result(
                span,
                &format!(
                    "unable to determine element type for `{:?}` on `{}`",
                    tile_rust_arithmetic_op,
                    operand_type.rust_ty.to_token_stream().to_string()
                ),
            );
        };
        let Some(_operand_tile_ir_ty) = &operand_type.tile_ir_ty else {
            return self.jit_error_result(
                span,
                &format!(
                    "type `{}` cannot be used with binary `{:?}`",
                    operand_type.rust_ty.to_token_stream().to_string(),
                    tile_rust_arithmetic_op
                ),
            );
        };
        // For tile-ir, the result type for same-type operations comes from the
        // lhs value's type in the module (preserves tile shape).
        let operand_result_ty = module.value_type(lhs_value).clone();

        let Some(operand_cuda_tile_element_type) =
            operand_type.get_cuda_tile_element_type(&self.modules.primitives())?
        else {
            return self.jit_error_result(
                span,
                &format!(
                    "unable to determine compiled element type for `{:?}`",
                    tile_rust_arithmetic_op
                ),
            );
        };
        let mut is_cmp = false;
        let signedness_str = match operand_rust_element_type.as_str() {
            "bool" | "u32" | "u64" => "unsigned",
            _ => "signed",
        };
        let sign_attr = signedness_attr("signedness", signedness_str);
        // Build the operation (allocates in module) but do NOT append to the
        // block yet. The old compiler defers `.build()` + `append_operation`
        // until after the exact-bounds early-return check, so we replicate that:
        // build now, check bounds, append only if we actually need the op.
        let (op_id, results) = match operand_cuda_tile_element_type.as_ref() {
            "i1" | "i4" | "i8" | "i32" | "i64" => {
                // TODO (hme): Add i4, i8, i16 support, as needed.
                if let Some(comparison_predicate) =
                    get_cmp_predicate_attr_ir(tile_rust_arithmetic_op)?
                {
                    is_cmp = true;
                    let bool_result_ty = make_bool_result_type(&operand_result_ty);
                    OpBuilder::new(Opcode::CmpI, self.ir_location(span))
                        .attr(comparison_predicate.0, comparison_predicate.1)
                        .attr(sign_attr.0, sign_attr.1)
                        .operand(lhs_value)
                        .operand(rhs_value)
                        .result(bool_result_ty)
                        .build(module)
                } else {
                    // If both operands have bounds, we can generate bounds on the output.
                    match tile_rust_arithmetic_op {
                        TileBinaryOp::Min => OpBuilder::new(Opcode::MinI, self.ir_location(span))
                            .operand(lhs_value)
                            .operand(rhs_value)
                            .attr(sign_attr.0, sign_attr.1)
                            .result(operand_result_ty.clone())
                            .build(module),
                        TileBinaryOp::Max => OpBuilder::new(Opcode::MaxI, self.ir_location(span))
                            .operand(lhs_value)
                            .operand(rhs_value)
                            .attr(sign_attr.0, sign_attr.1)
                            .result(operand_result_ty.clone())
                            .build(module),
                        TileBinaryOp::Add => OpBuilder::new(Opcode::AddI, self.ir_location(span))
                            .operand(lhs_value)
                            .operand(rhs_value)
                            .attr("overflow", Attribute::i32(0))
                            .result(operand_result_ty.clone())
                            .build(module),
                        TileBinaryOp::Sub => OpBuilder::new(Opcode::SubI, self.ir_location(span))
                            .operand(lhs_value)
                            .operand(rhs_value)
                            .attr("overflow", Attribute::i32(0))
                            .result(operand_result_ty.clone())
                            .build(module),
                        TileBinaryOp::Mul => OpBuilder::new(Opcode::MulI, self.ir_location(span))
                            .operand(lhs_value)
                            .operand(rhs_value)
                            .attr("overflow", Attribute::i32(0))
                            .result(operand_result_ty.clone())
                            .build(module),
                        TileBinaryOp::Rem => OpBuilder::new(Opcode::RemI, self.ir_location(span))
                            .operand(lhs_value)
                            .operand(rhs_value)
                            .attr(sign_attr.0, sign_attr.1)
                            .result(operand_result_ty.clone())
                            .build(module),
                        TileBinaryOp::Div => {
                            // DivI uses "rounding" (not "rounding_mode") in bytecode
                            OpBuilder::new(Opcode::DivI, self.ir_location(span))
                                .operand(lhs_value)
                                .operand(rhs_value)
                                .result(operand_result_ty.clone())
                                .attr(sign_attr.0, sign_attr.1)
                                .attr("rounding", Attribute::i32(2)) // negative_inf
                                .build(module)
                        }
                        TileBinaryOp::CeilDiv => {
                            // DivI uses "rounding" (not "rounding_mode") in bytecode
                            OpBuilder::new(Opcode::DivI, self.ir_location(span))
                                .operand(lhs_value)
                                .operand(rhs_value)
                                .result(operand_result_ty.clone())
                                .attr(sign_attr.0, sign_attr.1)
                                .attr("rounding", Attribute::i32(3)) // positive_inf
                                .build(module)
                        }
                        TileBinaryOp::BitAnd => {
                            OpBuilder::new(Opcode::AndI, self.ir_location(span))
                                .operand(lhs_value)
                                .operand(rhs_value)
                                .result(operand_result_ty.clone())
                                .build(module)
                        }
                        TileBinaryOp::BitOr => OpBuilder::new(Opcode::OrI, self.ir_location(span))
                            .operand(lhs_value)
                            .operand(rhs_value)
                            .result(operand_result_ty.clone())
                            .build(module),
                        TileBinaryOp::BitXor => {
                            OpBuilder::new(Opcode::XOrI, self.ir_location(span))
                                .operand(lhs_value)
                                .operand(rhs_value)
                                .result(operand_result_ty.clone())
                                .build(module)
                        }
                        _ => {
                            return self.jit_error_result(
                                span,
                                &format!("Unimplemented binary op {tile_rust_arithmetic_op:#?}"),
                            );
                        }
                    }
                }
            }
            "bf16" | "f16" | "f32" | "f64" => {
                if let Some(comparison_predicate) =
                    get_cmp_predicate_attr_ir(tile_rust_arithmetic_op)?
                {
                    let comparison_ordering = cmp_ordering_attr("ordered");
                    is_cmp = true;
                    let bool_result_ty = make_bool_result_type(&operand_result_ty);
                    OpBuilder::new(Opcode::CmpF, self.ir_location(span))
                        .attr(comparison_predicate.0, comparison_predicate.1)
                        .attr(comparison_ordering.0, comparison_ordering.1)
                        .operand(lhs_value)
                        .operand(rhs_value)
                        .result(bool_result_ty)
                        .build(module)
                } else {
                    let default_rm_attr = rounding_mode_attr("nearest_even");
                    match tile_rust_arithmetic_op {
                        TileBinaryOp::Min => OpBuilder::new(Opcode::MinF, self.ir_location(span))
                            .attr(default_rm_attr.0, default_rm_attr.1)
                            .operand(lhs_value)
                            .operand(rhs_value)
                            .result(operand_result_ty.clone())
                            .build(module),
                        TileBinaryOp::Max => OpBuilder::new(Opcode::MaxF, self.ir_location(span))
                            .attr(default_rm_attr.0, default_rm_attr.1)
                            .operand(lhs_value)
                            .operand(rhs_value)
                            .result(operand_result_ty.clone())
                            .build(module),
                        TileBinaryOp::Add => OpBuilder::new(Opcode::AddF, self.ir_location(span))
                            .attr(default_rm_attr.0, default_rm_attr.1)
                            .operand(lhs_value)
                            .operand(rhs_value)
                            .result(operand_result_ty.clone())
                            .build(module),
                        TileBinaryOp::Sub => OpBuilder::new(Opcode::SubF, self.ir_location(span))
                            .attr(default_rm_attr.0, default_rm_attr.1)
                            .operand(lhs_value)
                            .operand(rhs_value)
                            .result(operand_result_ty.clone())
                            .build(module),
                        TileBinaryOp::Mul => OpBuilder::new(Opcode::MulF, self.ir_location(span))
                            .attr(default_rm_attr.0, default_rm_attr.1)
                            .operand(lhs_value)
                            .operand(rhs_value)
                            .result(operand_result_ty.clone())
                            .build(module),
                        TileBinaryOp::Rem => OpBuilder::new(Opcode::RemF, self.ir_location(span))
                            .operand(lhs_value)
                            .operand(rhs_value)
                            .result(operand_result_ty.clone())
                            .build(module),
                        TileBinaryOp::Div => OpBuilder::new(Opcode::DivF, self.ir_location(span))
                            .attr(default_rm_attr.0, default_rm_attr.1)
                            .operand(lhs_value)
                            .operand(rhs_value)
                            .result(operand_result_ty.clone())
                            .build(module),
                        TileBinaryOp::TrueDiv => {
                            let approx_rm_attr = rounding_mode_attr("approx");
                            let mut builder = OpBuilder::new(Opcode::DivF, self.ir_location(span))
                                .attr(approx_rm_attr.0, approx_rm_attr.1);
                            if operand_cuda_tile_element_type.as_str() == "f32" {
                                let ftz = flag_attr("flush_to_zero");
                                builder = builder.attr(ftz.0, ftz.1);
                            }
                            builder
                                .operand(lhs_value)
                                .operand(rhs_value)
                                .result(operand_result_ty.clone())
                                .build(module)
                        }
                        _ => {
                            return self.jit_error_result(
                                span,
                                &format!("Unimplemented binary op {tile_rust_arithmetic_op:#?}"),
                            );
                        }
                    }
                }
            }
            _ => {
                return self.jit_error_result(
                    span,
                    &format!(
                        "Binary operation is not implemented for {}",
                        operand_rust_ty.to_token_stream().to_string()
                    ),
                );
            }
        };

        let return_type = match return_type {
            Some(rt) => rt,
            None => {
                // Try to infer from lhs/rhs.
                if is_cmp {
                    let bool_ty = syn::parse2::<syn::Type>("bool".parse().unwrap()).unwrap();
                    self.compile_type(&bool_ty, &generic_vars, &HashMap::new())?
                        .unwrap()
                } else {
                    operand_type
                }
            }
        };

        let op_bounds = if let (Some(a), Some(b)) = (lhs.bounds, rhs.bounds) {
            if !(lhs.kind == Kind::PrimitiveType && rhs.kind == Kind::PrimitiveType) {
                return self.jit_error_result(
                    span,
                    &format!(
                        "Expected PrimitiveType for binary op bounds, got lhs={:#?}, rhs={:#?}",
                        lhs.kind, rhs.kind
                    ),
                );
            }
            bounds_from_bop(tile_rust_arithmetic_op, &a, &b)
        } else {
            None
        };
        if let Some(bounds) = &op_bounds {
            if bounds.is_exact() {
                // The lower/upper bounds are equivalent — emit a constant
                // instead. The op allocated above becomes dead (not appended
                // to any block).
                return Ok(self.compile_constant_from_exact_bounds(
                    module,
                    block_id,
                    bounds.clone(),
                    return_type,
                )?);
            }
        }

        // Only now append the binary op to the block (mirrors the old
        // compiler which only calls `builder.append_operation` after the
        // bounds check).
        append_op(module, block_id, op_id);
        let value = results[0];
        let mut tr_value = TileRustValue::new_value_kind_like(value, return_type.clone());
        tr_value.bounds = op_bounds;
        Ok(tr_value)
    }
}