On Fri, Nov 03, 2017 at 05:51:56PM +0000, Mark Rutland wrote:

Hi,

On Fri, Nov 03, 2017 at 10:11:44AM -0700, Sami Tolvanen wrote:

quoted

This series adds build system support for compiling the kernel with clang
Link Time Optimization (LTO), using GNU gold with the LLVMgold plug-in
for linking. Some background for clang's LTO support is available here:

  https://llvm.org/docs/LinkTimeOptimization.html

With -flto, clang produces LLVM bitcode instead of object files, and
the compilation to native code happens at link time. In addition, clang
cannot use an external assembler for inline assembly when LTO is enabled,
which causes further compatibility issues.

The patches in this series remove intermediate linking steps when LTO is
used, postpone processing done on object files until after the LTO link
step, add workarounds for GNU gold incompatibilities, and address inline
assembly incompatibilities for arm64.

These changes allow arm64 defconfig to be compiled with LTO, but other
architectures are not enabled until compatibility issues have been
addressed. In particular, x86 inline assembly doesn't currently compile
with clang's integrated assembler due to this LLVM bug:

  https://bugs.llvm.org/show_bug.cgi?id=24487

Due to recent bug fixes in the toolchain, it's recommended to use clang
5.0 or later, and GNU gold from binutils 2.27 or later

Which tree are you testing with?

I had a go building the arm64 for/next/core branch defconfig using clang
5.0.0 from llvm.org, with the Linaro 17.05 toolchain's binutils
(2.27.0.20161019), and I hit what appears to be a compiler bug quite
quickly.

So as per my reply to patch 7, it looks like this series addresses this
case.

What's the minimum set of patches necessary to work with clang (ignoring
LTO)?

Thanks,
Mark.