From d9b8283e7c089566eccc8d1f46de0aa4f7f1cd3c Mon Sep 17 00:00:00 2001
From: yanght24 <yanght24@gmail.com>
Date: Fri, 21 Aug 2026 23:58:28 +0800
Subject: [PATCH] Fix tileiras SIGSEGV with occupancy=2 by retrying without
 occupancy hint

Signed-off-by: yanght24 <yanght24@gmail.com>
---
 changelog.d/fix-occupancy-sigsegv-fallback.md |  1 +
 src/cuda/tile/_compile.py                     | 25 ++++++-
 src/cuda/tile/_exception.py                   |  1 +
 test/test_compiler_sigsegv_fallback.py        | 69 +++++++++++++++++++
 4 files changed, 95 insertions(+), 1 deletion(-)
 create mode 100644 changelog.d/fix-occupancy-sigsegv-fallback.md
 create mode 100644 test/test_compiler_sigsegv_fallback.py

diff --git a/changelog.d/fix-occupancy-sigsegv-fallback.md b/changelog.d/fix-occupancy-sigsegv-fallback.md
new file mode 100644
index 00000000..48ccf9d6
--- /dev/null
+++ b/changelog.d/fix-occupancy-sigsegv-fallback.md
@@ -0,0 +1 @@
+Work around a `tileiras` SIGSEGV triggered by `occupancy=2` on some fused tile kernels by retrying compilation without the occupancy hint.
diff --git a/src/cuda/tile/_compile.py b/src/cuda/tile/_compile.py
index 5b091689..09dedaa5 100644
--- a/src/cuda/tile/_compile.py
+++ b/src/cuda/tile/_compile.py
@@ -6,7 +6,7 @@
 import re
 import warnings
 import contextlib
-from dataclasses import dataclass
+from dataclasses import dataclass, replace
 import datetime
 import functools
 from functools import cache
@@ -607,6 +607,29 @@ def compile_tile(ann_func: AnnotatedFunction | FunctionType,
                                        timeout_sec=context.config.compiler_timeout_sec,
                                        remarks_output_file=remarks_file)
         except TileCompilerError as e:
+            if (isinstance(e, TileCompilerExecutionError)
+                    and getattr(e, "return_code", None) == -11
+                    and compiler_options.occupancy is not None):
+                fallback_options = replace(compiler_options, occupancy=None)
+                warnings.warn(
+                    "tileiras terminated with SIGSEGV while processing the requested "
+                    f"occupancy hint {compiler_options.occupancy!r}; retrying without an "
+                    "occupancy hint.",
+                    UserWarning,
+                    stacklevel=3,
+                )
+                return compile_tile(
+                    ann_func,
+                    signatures,
+                    sm_arch=sm_arch,
+                    compiler_options=fallback_options,
+                    context=context,
+                    bytecode_version=bytecode_version,
+                    return_final_ir=return_final_ir,
+                    return_bytecode=return_bytecode,
+                    return_cubin=return_cubin,
+                )
+
             if context.config.enable_crash_dump:
                 anonymized_bytecode = _get_bytecode(ir_keeper, compiler_options,
                                                     anonymize_debug_info=True)
diff --git a/src/cuda/tile/_exception.py b/src/cuda/tile/_exception.py
index e67bd575..fe636193 100644
--- a/src/cuda/tile/_exception.py
+++ b/src/cuda/tile/_exception.py
@@ -269,6 +269,7 @@ def __init__(self,
                  stderr: str,
                  compiler_flags: str,
                  compiler_version: Optional[str]):
+        self.return_code = return_code
         message, loc = _parse_tileir_stderr(stderr)
         if loc is None:
             loc = _unknown_loc
diff --git a/test/test_compiler_sigsegv_fallback.py b/test/test_compiler_sigsegv_fallback.py
new file mode 100644
index 00000000..41f8a243
--- /dev/null
+++ b/test/test_compiler_sigsegv_fallback.py
@@ -0,0 +1,69 @@
+# SPDX-FileCopyrightText: Copyright (c) <2026> NVIDIA CORPORATION & AFFILIATES. All rights reserved.
+#
+# SPDX-License-Identifier: Apache-2.0
+
+from pathlib import Path
+
+import cuda.tile as ct
+from cuda.tile._bytecode.version import BytecodeVersion
+from cuda.tile._cext import TileContext
+from cuda.tile._compile import compile_tile
+from cuda.tile._context import TileContextConfig
+from cuda.tile._exception import TileCompilerExecutionError
+from cuda.tile.compilation import ArrayConstraint, CallingConvention, KernelSignature
+
+
+def test_sigsegv_with_occupancy_falls_back_to_auto(monkeypatch, tmp_path):
+    @ct.kernel(occupancy=2)
+    def kernel(x, y):
+        t = ct.load(x, (0,), shape=(32,))
+        ct.store(y, (0,), tile=t)
+
+    constraint = ArrayConstraint(
+        ct.float32,
+        1,
+        index_dtype=ct.int32,
+        stride_lower_bound_incl=0,
+        alias_groups=(),
+        may_alias_internally=False,
+        stride_constant=(1,),
+        base_addr_divisible_by=16,
+    )
+    sig = KernelSignature(
+        parameters=[constraint, constraint],
+        calling_convention=CallingConvention.cutile_python_v1(),
+    )
+
+    seen = []
+
+    def fake_compile_cubin(fname_bytecode, compiler_options, sm_arch, timeout_sec,
+                           remarks_output_file=None):
+        seen.append(compiler_options.occupancy)
+        if compiler_options.occupancy is not None:
+            raise TileCompilerExecutionError(-11, "", "--gpu-name sm_120 -O2 --lineinfo",
+                                             "13.3")
+        cubin_path = Path(fname_bytecode).with_suffix(".cubin")
+        cubin_path.write_bytes(b"FAKE_CUBIN")
+        return cubin_path
+
+    monkeypatch.setattr("cuda.tile._compile.compile_cubin", fake_compile_cubin)
+    context = TileContext(config=TileContextConfig(
+        temp_dir=str(tmp_path),
+        compiler_timeout_sec=None,
+        enable_crash_dump=False,
+        cache_dir=None,
+        cache_size_limit=0,
+    ))
+
+    result = compile_tile(
+        kernel._annotated_function,
+        [sig],
+        sm_arch="sm_120",
+        compiler_options=kernel._compiler_options,
+        context=context,
+        bytecode_version=BytecodeVersion.V_13_3,
+        return_cubin=True,
+    )
+
+    assert result.cubin == b"FAKE_CUBIN"
+    assert seen == [2, None]