Skip to content

fix(callconv): align alloca slots to CUDA vector type ABI requirements #321

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Open
isVoid wants to merge 2 commits into
base: main
Choose a base branch
from
Open

fix(callconv): align alloca slots to CUDA vector type ABI requirements #321

Changes from all commits
Commits
Show all changes
2 commits
Select commit Hold shift + click to select a range
4fc0c38
fix(callconv): align alloca slots to CUDA vector type ABI requirements
isVoid Apr 15, 2026
f32141d
refactor(callconv): use alignof_ from Numba type instead of sizeof he…
isVoid Apr 15, 2026
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
36 changes: 34 additions & 2 deletions numbast/src/numbast/callconv.py
View file
Open in desktop
Original file line number Diff line number Diff line change
@@ -1,6 +1,7 @@
# SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved.
# SPDX-License-Identifier: Apache-2.0

# NUMBAST_RETVAL_ALIGN_FIX_APPLIED
Copy link
Copy Markdown
Contributor

@coderabbitai coderabbitai bot Apr 15, 2026

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

⚠️ Potential issue | 🟡 Minor

Remove debugging marker comment.

This marker comment appears to be a development/tracking artifact (possibly from Claude Code generation) and should not be committed to the production codebase.

Proposed fix 
-# NUMBAST_RETVAL_ALIGN_FIX_APPLIED
 from numbast.args import prepare_ir_types
🤖 Prompt for AI Agents 
Verify each finding against the current code and only fix it if needed.

In `@numbast/src/numbast/callconv.py` at line 4, Remove the development/tracking
marker comment "NUMBAST_RETVAL_ALIGN_FIX_APPLIED" from the top of
numbast/src/numbast/callconv.py; simply delete that standalone comment line so
the file contains only production code/comments, and ensure no leftover blank
line or stray artifact remains.

from numbast.args import prepare_ir_types
from numbast.intent import IntentPlan

Expand Down Expand Up @@ -107,6 +108,28 @@ def _lower_impl(self, builder, context, sig, args):
else:
retval_ty = context.get_value_type(cxx_return_type)
retval_ptr = builder.alloca(retval_ty, name="retval")
# Use the Numba type's alignof_ to set the alloca alignment.
#
# LLVM computes struct ABI alignment as the max alignment of its
# members. For CUDA vector types (float2, float4, uchar4, …)
# declared with __align__(N) in the CUDA headers, N can exceed
# the member alignment: float2 is {float,float} with member
# alignment 4 B but __align__(8). LLVM therefore assigns a 4 B
# alloca, while the NVRTC shim uses a vector instruction
# (ld/st.v2.f32) that requires 8 B alignment, causing
# cudaErrorMisalignedAddress at runtime.
#
# Numbast's struct binder already sets alignof_ on user-defined
# bound structs (propagated from ast_canopy). For built-in CUDA
# vector types, callers must set alignof_ on the Numba type when
# registering it in CTYPE_MAPS. When alignof_ is present it is
# used here, matching the convention already applied to loads and
# stores (getattr(argty, "alignof_", None)). When absent, LLVM's
# default ABI alignment is used, which is correct for scalars and
# structs without an explicit __align__ attribute.
_nb_align = getattr(cxx_return_type, "alignof_", None)
if _nb_align is not None:
retval_ptr.align = _nb_align

# 2. Prepare arguments
if self._intent_plan is None:
Expand Down Expand Up @@ -154,8 +177,11 @@ def _lower_impl(self, builder, context, sig, args):
ptrs.append(arg)
else:
ptr = cgutils.alloca_once(builder, vty)
_nb_align = getattr(argty, "alignof_", None)
if _nb_align is not None:
ptr.align = _nb_align # see retval_ptr comment above
builder.store(
arg, ptr, align=getattr(argty, "alignof_", None)
arg, ptr, align=_nb_align
)
ptrs.append(ptr)
else:
Expand All @@ -175,6 +201,9 @@ def _lower_impl(self, builder, context, sig, args):
out_nbty = self._out_return_types[out_pos]
vty = context.get_value_type(out_nbty)
ptr = cgutils.alloca_once(builder, vty)
_nb_align = getattr(out_nbty, "alignof_", None)
if _nb_align is not None:
ptr.align = _nb_align # see retval_ptr comment above
ptrs.append(ptr)
arg_pointer_types.append(ir.PointerType(vty))
out_return_ptrs.append((out_nbty, ptr))
Expand All @@ -194,8 +223,11 @@ def _lower_impl(self, builder, context, sig, args):
arg_pointer_types.append(vty)
else:
ptr = cgutils.alloca_once(builder, vty)
_nb_align = getattr(argty, "alignof_", None)
if _nb_align is not None:
ptr.align = _nb_align # see retval_ptr comment above
builder.store(
arg, ptr, align=getattr(argty, "alignof_", None)
arg, ptr, align=_nb_align
)
ptrs.append(ptr)
arg_pointer_types.append(ir.PointerType(vty))
Expand Down