Arm backend: Add Clang bare-metal toolchain - #24831

Summary

This PR adds Arm Toolchain for Embedded (Clang + Picolibc) as an alternative bare-metal toolchain for Corstone-300/320 runners. There is one blocker (the -Oz scope) and one major linker-script issue (.tbss reserves no space). Several of the linker-script changes also alter GCC builds without being guarded.

Code Quality

• .tbss is missing a space reservation, so TLS overlaps the next section.
  • Where: backends/arm/cmake/linker_scripts/Corstone-300.ld:250-256 and Corstone-320.ld:173-180.
  • Cause: .tbss is a TLS NOBITS section. Neither ld.lld nor GNU ld moves the location counter past it, so the next section in the same region starts at ADDR(.tbss). On C300 that section is .sram.rodata/.ddr.rodata. On C320 it is .ARM.extab/.ARM.exidx/the copy and zero tables in BRAM.
  • Effect: the new zero-table entry LONG(ADDR(.tbss)) LONG(SIZEOF(.tbss)/4) (C300:183, C320:212) wipes the start of that section. Every later write to TLS, such as picolibc's TLS errno, corrupts it too. On C320 this can hit the zero table while it is still being walked. The FVP runs may pass only because .tbss was empty or the clobbered bytes were never read.
  • Fix: picolibc's own picolibc.ld handles this with .tbss_space (NOLOAD) : { . = ADDR(.tbss); . = . + SIZEOF(.tbss); } right after .tbss.
  • To confirm: check llvm-readelf -S on a Clang ELF with a non-empty .tbss. (Raised by all three reviewers; inferred from the linker semantics, not linked here.)
• Linker-script changes outside #ifdef __clang__ also change GCC images.
  • Corstone-300:
    • .got moved from .ddr into .text/ITCM (Corstone-300.ld:123-124).
    • The copy-table rodata source changed to LOADADDR(.rodata).
    • A new ALIGN(4) before __sram_data_end__.
  • Corstone-320:
    • A new rom_sram PT_LOAD (Corstone-320.ld:53).
    • .sram moved from :rom_dram to :rom_sram (:258).
  • Most of these look like fixes or no-ops, but only C300 U55 was tested with GCC. Please run C320/U85 with GCC as well, or say in the commit message that these GCC changes are intended.
• std::random_device shim (arm-none-eabi-clang-random-device.h, force-included at arm-none-eabi-clang.cmake:49).
  • Problem: when libc++ has no random_device, the shim defines a class inside namespace std, which is undefined behaviour. It is force-included into every C++ file of every target that uses the toolchain.
  • Effect: it always returns 0x9e3779b9. Without any warning, op_rand, op_randn and op_native_dropout in kernels/portable/cpu/ produce identical output on every run, and extension/tensor/tensor_ptr_maker.cpp is affected the same way.
  • Suggestion: handle the missing random_device at those call sites (or behind an ExecuTorch macro) instead of patching std globally.
• ArmRunnerUtilsInternal.cmake:338-348 link-option rewrite.
  • If INTERFACE_LINK_OPTIONS is unset, get_target_property returns _ethosu_target_link_options-NOTFOUND, and that string gets added as a link option. A guard is needed.
  • REMOVE_ITEM --entry Reset_Handler only works if core_platform stores those as two separate list items.
  • This check uses STREQUAL "Clang" while the root CMakeLists uses MATCHES "Clang".
  • (Low confidence: core_platform isn't in the checkout, so I couldn't check how it stores these.)
• Tool checks resolve whatever is first on PATH, not the ATfE install.
  • examples/arm/run.sh:202 checks toolchain_compiler="clang" with command -v/hash. Because setup_path_toolchain appends ATfE to the end of PATH, a host /usr/bin/clang passes the check even when ATfE isn't installed. The build then fails later in the toolchain file.
  • build_executor_runner.sh:104 sets size_tool=llvm-size and has the same problem: it picks up the host tool if there is one. If there isn't, the size lines come out empty without an error.
  • Suggestion: use ${ATFE_TOOLCHAIN_ROOT}/bin/....

Infrastructure

• CMakeLists.txt:316-322: the -Oz workaround applies to every Clang compiler. Blocker.
  • Problem: CMAKE_CXX_COMPILER_ID MATCHES "Clang" also matches host clang, AppleClang, the Android NDK and iOS. Every EXECUTORCH_OPTIMIZE_SIZE=ON build on those compilers switches from -Os to -Oz. That is a performance change for users who aren't on Arm.
  • CI impact: the failing job, test-binary-size-linux, runs on the clang12 image (.github/workflows/pull.yml:686); the GCC variant is a separate job. It builds a host size_test with -DEXECUTORCH_OPTIMIZE_SIZE=ON and never touches the Arm linker scripts. Its log shows Fail 48312 > 45000, against a previous size of about 44160. This flag is the only change in the diff that reaches that build, so the linker-script reshuffle is not the cause.
  • Fix: move the -Oz override into examples/arm/ethos-u-setup/arm-none-eabi-clang.cmake, or guard it on the bare-metal target, e.g. CMAKE_SYSTEM_NAME STREQUAL "Generic".
• No CI coverage for the Clang path. The description says Clang is wired through the Cortex-M workflows, but no .ci/ or .github/ changes exercise --toolchain=arm-none-eabi-clang. The new download URL and SHA-256 values in toolchain_utils.sh are therefore unverified by CI. Consider adding at least one Corstone FVP job that uses Clang.

Recommendation

Request Changes. Limit the -Oz workaround to the Arm Clang toolchain; this should also fix the size-test failure. Add the .tbss space reservation. The remaining items are minor.