HOSTCC always uses its default linker

[nathanchance]

Currently, there is no way to specify the host linker on a kernel wide scale. HOSTLD exists in a couple of Makefiles but much like the initial premise of #342, when $(HOSTCC) is used to compile then link executables, the default ld gets invoked.

See the failure here when ld is not available in PATH: https://travis-ci.com/nathanchance/continuous-integration/jobs/194766150

We want to have completely hermetic builds where no GNU tools are invoked, meaning that we need to provide -fuse-ld= to $(HOSTCC) somehow (or have $(HOSTLD) be exclusively used for linking). We can currently do this with the variable HOSTLDFLAGS (commit) but it would be nice for HOSTLD to work like LD currently does.

[nickdesaulniers]

cc @masahir0y

[nathanchance]

Along the same lines, objtool fails to compile even when HOSTLDFLAGS=-fuse-ld=lld is provided:

error: Cannot generate ORC metadata for CONFIG_UNWINDER_ORC=y, please install libelf-dev, libelf-devel or elfutils-libelf-devel
make: *** [Makefile:1141: prepare-objtool] Error 1

This can be fixed with the following diff:

diff --git a/Makefile b/Makefile
index 95421654a16d..f6b878084f36 100644
--- a/Makefile
+++ b/Makefile
@@ -979,7 +979,7 @@ HOST_LIBELF_LIBS = $(shell pkg-config libelf --libs 2>/dev/null || echo -lelf)
 
 ifdef CONFIG_STACK_VALIDATION
   has_libelf := $(call try-run,\
-               echo "int main() {}" | $(HOSTCC) -xc -o /dev/null $(HOST_LIBELF_LIBS) -,1,0)
+               echo "int main() {}" | $(HOSTCC) -xc -o /dev/null $(HOSTLDFLAGS) $(HOST_LIBELF_LIBS) -,1,0)
   ifeq ($(has_libelf),1)
     objtool_target := tools/objtool FORCE
   else

[nathanchance]

Recently, I found out this can be worked around by setting CLANG_DEFAULT_LINKER=lld during LLVM’s cmake’s invocation.

[tpimh]

This is what I use to set HOSTLDFLAGS:

--- a/Makefile
+++ b/Makefile
@@ -401,6 +401,9 @@ KBUILD_HOSTCFLAGS   := -Wall -Wmissing-prototypes -Wstrict-prototypes -O2 \
 		$(HOSTCFLAGS)
 KBUILD_HOSTCXXFLAGS := -O2 $(HOST_LFS_CFLAGS) $(HOSTCXXFLAGS)
 KBUILD_HOSTLDFLAGS  := $(HOST_LFS_LDFLAGS) $(HOSTLDFLAGS)
+ifeq ($(shell $(HOSTLD) -v | grep -o ^LLD),LLD)
+KBUILD_HOSTLDFLAGS += -fuse-ld=lld
+endif
 KBUILD_HOSTLDLIBS   := $(HOST_LFS_LIBS) $(HOSTLDLIBS)
 
 # Make variables (CC, etc...)
@@ -996,7 +999,7 @@ HOST_LIBELF_LIBS = $(shell pkg-config libelf --libs 2>/dev/null || echo -lelf)
 
 ifdef CONFIG_STACK_VALIDATION
   has_libelf := $(call try-run,\
-		echo "int main() {}" | $(HOSTCC) -xc -o /dev/null $(HOST_LIBELF_LIBS) -,1,0)
+		echo "int main() {}" | $(HOSTCC) -xc -o /dev/null $(KBUILD_HOSTLDFLAGS) $(HOST_LIBELF_LIBS) -,1,0)
   ifeq ($(has_libelf),1)
     objtool_target := tools/objtool FORCE
   else

Still not perfect, would fail if lld is not found in path and HOSTLD variable contains absolute path to it. Using $(which ${LD}) would solve it, but would also fail in other uses.

[nickdesaulniers]

I wonder what hack I put in pixel 4 kernel?
https://android.googlesource.com/kernel/msm/+/refs/heads/android-msm-coral-4.14-android10-c2f2/Makefile#508

Yeah, I should probably revisit upstreaming that series; IIRC I did send that as a patch.

[nickdesaulniers]

oh, nvm, this bug is about HOSTCC