Re: x86 - weird cross-compile build problem with try-run next-20210602

[Valdis Klētnieks]

On Sat, 05 Jun 2021 17:19:30 +0900, Masahiro Yamada said:

> > Anybody have a clue why $(x32_ld_ok)  is null rather than 'y' or 'n'?
>
>
> What command did you run?
>
> I see this warning message for 'make install' for example.
>
> $ make install
> arch/x86/Makefile:148: CONFIG_X86_X32 enabled but no binutils support

> Please add one more debug line:
>   $(warning need-compiler is +$(need-compiler)+)
> and what will you get?
>
Bingo.

arch/x86/Makefile:143: need-compiler is ++ x32_ld_ok is ++ with CC=(.... same as before)

And it was hitting on a 'make kernelrelease'

805b2e1d427aa (Masahiro Yamada          2021-02-28 15:10:28 +0900  275) # is an exception where build artifacts may be updated. This must be fixed.
805b2e1d427aa (Masahiro Yamada          2021-02-28 15:10:28 +0900  276) no-compiler-targets := $(no-dot-config-targets) install dtbs_install \
805b2e1d427aa (Masahiro Yamada          2021-02-28 15:10:28 +0900  277)                         headers_install modules_install kernelrelease image_name
993bdde945478 (Masahiro Yamada          2021-02-28 15:10:25 +0900  278) no-sync-config-targets := $(no-dot-config-targets) %install kernelrelease \

I suspect it's this:

 # Include this also for config targets because some architectures need
 # cc-cross-prefix to determine CROSS_COMPILE.
+ifdef need-compiler
 include $(srctree)/scripts/Makefile.compiler
+endif

and as a result, try-run isn't defined when we get into the ifdef CONFIG_X86_32
chunk in arch/x86/Makefile, which silently fails and returns a null.   I added
more debugging to double-check...

ifdef CONFIG_X86_X32
        x32_ld_ok := $(call try-run,\
                        /bin/echo -e '1: .quad 1b' | \ 
                        $(CC) $(KBUILD_AFLAGS) -c -x assembler -o "$$TMP" - && \
                        $(OBJCOPY) -O elf32-x86-64 "$$TMP" "$$TMP.o" && \
                        $(LD) -m elf32_x86_64 "$$TMP.o" -o "$$TMP",y,n)
 $(warning need-compiler is +$(need-compiler)+ x32_ld_ok is +$(x32_ld_ok)+ with CC=$(CC) $(KBUILD_AFLAGS) OBJ=$(OBJCOPY) LD=$(LD) )
 foo := $(call wombats-r-us )
 $(warning foo is +$(foo)+ )
        ifeq ($(x32_ld_ok),y)

and that gets:

/usr/src/linux-next/arch/x86/Makefile:143: need-compiler is ++ x32_ld_ok is ++ with CC=x86_64-unknown-linux-gnu-gcc -D__ASSEMBLY__ -fno-PIE -m64 OBJ=x86_64-unknown-linux-gnu-objcopy LD=x86_64-unknown-linux-gnu-ld 
/usr/src/linux-next/arch/x86/Makefile:145: foo is ++ 
/usr/src/linux-next/arch/x86/Makefile:151: CONFIG_X86_X32 enabled but no binutils support

So the call to the undefined 'wombats-r-us' fails silently and returns a null..
and try-run is also failing silently the same way because it's not defined either.

There's only a few uses of try-run outside Makefile.compiler, and it looks like
x32_ld_ok is the only place where Makefile logic changes based on what try-run
returns (the rest just change compiler flags).

Havings said that, I'm not sure what the proper fix is. Move try-run out of Makefile.compiler?