Aarch64 support

Author: abrandemuehl-rai

conda-forge.yml

@@ -13,5 +13,4 @@ github:
   tooling_branch_name: main
 provider:
   linux_aarch64: default
-  linux_ppc64le: false
 test: native_and_emulated

recipe/build.sh

@@ -32,6 +32,11 @@ if [[ "${target_platform}" == linux* ]]; then
 
   if [[ "${target_platform}" != "${build_platform}" ]]; then
     cmake_extra_args+=("-DACPP_HOST_FORCE_MCPU_TARGET=generic")
+    # Cross build: device libkernel bitcode is target-independent but is
+    # compiled/linked on the build host, so use build-native clang/llvm-link.
+    # CLANG_EXECUTABLE_PATH stays on the target toolchain for the runtime config.
+    cmake_extra_args+=("-DACPP_BITCODE_CLANG=${BUILD_PREFIX}/bin/clang++-${llvm_version}")
+    cmake_extra_args+=("-DACPP_BITCODE_LLVM_LINK=${BUILD_PREFIX}/bin/llvm-link")
   fi
 
   test -f "${PREFIX}/include/spirv/unified1/spirv.hpp"

recipe/patches/0004-Cross-compile-native-bitcode-tools.patch

@@ -0,0 +1,55 @@
+From 7ae15bf526f06b5e6c18e9f908b6c58236f7c6a4 Mon Sep 17 00:00:00 2001
+From: conda-forge <noreply@conda-forge.org>
+Date: Fri, 24 Jul 2026 10:00:23 +0200
+Subject: [PATCH] Use build-native tools for device bitcode when
+ cross-compiling
+
+The SSCP libkernel bitcode is target-independent (the device triple is
+passed explicitly) but is generated on the build host. Route the clang and
+llvm-link invocations through ACPP_BITCODE_CLANG / ACPP_BITCODE_LLVM_LINK,
+which default to the target tools for native builds and can be overridden
+with build-native tools when cross-compiling (e.g. linux-aarch64 on x86_64).
+---
+ src/libkernel/sscp/CMakeLists.txt | 15 +++++++++++++--
+ 1 file changed, 13 insertions(+), 2 deletions(-)
+
+diff --git a/src/libkernel/sscp/CMakeLists.txt b/src/libkernel/sscp/CMakeLists.txt
+index 672553a..7ab74f4 100644
+--- a/src/libkernel/sscp/CMakeLists.txt
++++ b/src/libkernel/sscp/CMakeLists.txt
+@@ -1,3 +1,14 @@
++# Cross-compilation: the device bitcode below is target-independent (the
++# device triple is set explicitly), but it is generated on the *build* host.
++# Use build-native clang/llvm-link so they can execute when cross-compiling,
++# while CLANG_EXECUTABLE_PATH stays pointed at the target toolchain for the
++# runtime configuration. Both default to the target tools for native builds.
++if(NOT ACPP_BITCODE_CLANG)
++  set(ACPP_BITCODE_CLANG "${CLANG_EXECUTABLE_PATH}")
++endif()
++if(NOT ACPP_BITCODE_LLVM_LINK)
++  set(ACPP_BITCODE_LLVM_LINK "${LLVM_TOOLS_BINARY_DIR}/llvm-link")
++endif()
+ 
+ function(libkernel_generate_bitcode_library)
+   set(options)
+@@ -32,7 +43,7 @@ function(libkernel_generate_bitcode_library)
+ 
+   add_custom_command(
+       OUTPUT ${target_output_file}
+-      COMMAND ${CLANG_EXECUTABLE_PATH} -target ${triple} -fno-exceptions -fno-rtti -O3 -emit-llvm -std=c++17 -DHIPSYCL_SSCP_LIBKERNEL_LIBRARY
++      COMMAND ${ACPP_BITCODE_CLANG} -target ${triple} -fno-exceptions -fno-rtti -O3 -emit-llvm -std=c++17 -DHIPSYCL_SSCP_LIBKERNEL_LIBRARY
+                     -I ${HIPSYCL_SOURCE_DIR}/include ${additional} ${platform_specific} -o ${target_output_file} -c ${CMAKE_CURRENT_SOURCE_DIR}/${source}
+       DEPENDS ${CMAKE_CURRENT_SOURCE_DIR}/${source} ${target_depends}
+       VERBATIM)
+@@ -77,7 +88,7 @@ function(libkernel_generate_bitcode_target)
+   set(linked_output ${CMAKE_CURRENT_BINARY_DIR}/libkernel-sscp-${target}-full.bc)
+   add_custom_command(
+     OUTPUT ${linked_output}
+-    COMMAND ${LLVM_TOOLS_BINARY_DIR}/llvm-link -o=${linked_output} ${output_files}
++    COMMAND ${ACPP_BITCODE_LLVM_LINK} -o=${linked_output} ${output_files}
+     DEPENDS ${output_files} ${target_depends}
+   )
+ 
+-- 
+2.53.0
+

recipe/recipe.yaml

@@ -19,6 +19,8 @@ source:
         then: patches/0002-Fall-back-missing-functions.patch
       - if: linux
         then: patches/0003-Use-packaged-dependency-sources.patch
+      - if: linux
+        then: patches/0004-Cross-compile-native-bitcode-tools.patch
   - if: linux and llvm_version == "16"
     then:
       url: https://github.com/AdaptiveCpp/SPIRV-LLVM-Translator/archive/d678e6a2a95849c9985c8d209f67b318999083c6.tar.gz
@@ -80,6 +82,12 @@ requirements:
       then: ${{ compiler("cuda") }}
     - if: cuda_compiler_version != "None" and build_platform != target_platform
       then: cuda-cudart-dev
+    # Cross builds compile the device bitcode with build-native clang/llvm-link
+    # (see ACPP_BITCODE_* in build.sh).
+    - if: build_platform != target_platform and linux
+      then: clangdev ${{ llvm_version }}.*
+    - if: build_platform != target_platform and linux
+      then: llvm-tools ${{ llvm_version }}.*
   host:
     - clangdev
     - if: linux and x86_64