use $CC if it's set externally

[catawbasam]

Getting the following after checking out master just now and running make clean all on 64-bit red hat linux:

CC src/llvm-simdloop.o
CC src/gc.o
gc.c:40: error: flexible array member in otherwise empty struct
make[2]: *** [gc.o] Error 1
make[1]: *** [julia-release] Error 2
make: *** [release] Error 2

[tkelman]

Previously reported at #9984 - the fix there is kind of ugly, and we're going to be increasing the minimum GCC version requirement on 0.4 before too much longer. I'd recommend you install and start using the devtoolset to get a modern version of GCC.

[catawbasam]

[keithc@giada julia]$ echo $CC
/usr/depot/gcc-4.8.2

This is too old?

[tkelman]

No, not at all, 4.8.2 is fine. Are you sure that's the gcc that is getting used to build julia?

[nalimilan]

I guess you ended up using the system gcc for some reason. I did the same mistake before I filed #9984.

Can you run make VERBOSE=1 to see the actual command that is run? Maybe our handling of CC is broken. FWIW, for the RPMs I'm using devtoolset-2 and I just run this command before make to set the required environment variables: /opt/rh/devtoolset-2/enable.

[catawbasam]

If the Makefile is using $CC, then it is 4.8.2. If Makefile ever references gcc directly then it will pick up gcc 4.4.

The release 0.3 branch builds fine. Only master has the problem.

Thank you for the tip re devtoolset. Unfortunately, on my fairly locked down system:
/opt/rh/devtoolset-2/enable
bash: /opt/rh/devtoolset-2/enable: No such file or directory

alias gcc="/usr/depot/gcc-4.8.2" Does not help

[nalimilan]

@catawbasam Please post the output of make VERBOSE=1.

[catawbasam]

[keithc@giada julia]$ git branch
* master
  release-0.3
[keithc@giada julia]$ make VERBOSE=1
make[1]: Entering directory `/devel/asias/keithc/giada/julia'
make[2]: Entering directory `/devel/asias/keithc/giada/julia/deps'
make[2]: Leaving directory `/devel/asias/keithc/giada/julia/deps'
make[2]: Entering directory `/devel/asias/keithc/giada/julia/src'
 gcc -m64  "-DJL_SYSTEM_IMAGE_PATH=\"../lib/julia/sys.ji\"" -DJULIA_TARGET_ARCH=native -std=gnu99 -pipe -fPIC -fno-strict-aliasing -D_FILE_OFFSET_BITS=64 -DCOPY_STACKS -O3 -ggdb3 -falign-functions -momit-leaf-frame-pointer -D_GNU_SOURCE -Iflisp -Isupport -I/devel/asias/keithc/giada/julia/usr/include -I/devel/asias/keithc/giada/julia/deps/libuv/include -I/devel/asias/keithc/giada/julia/usr/include -DLIBRARY_EXPORTS -I. -I/devel/asias/keithc/giada/julia/deps/valgrind -Wall -Wno-strict-aliasing -fno-omit-frame-pointer -fvisibility=hidden -fno-common -DNDEBUG -c gc.c -o gc.o
gc.c:40: error: flexible array member in otherwise empty struct
make[2]: *** [gc.o] Error 1
make[2]: Leaving directory `/devel/asias/keithc/giada/julia/src'
make[1]: *** [julia-release] Error 2
make[1]: Leaving directory `/devel/asias/keithc/giada/julia'
make: *** [release] Error 2
[keithc@giada julia]$ 

[nalimilan]

Clearly we're calling the system gcc here. And yet src/Makefile looks correct:

$(BUILDDIR)/%.o: %.c $(HEADERS) | $(BUILDDIR)
        @$(call PRINT_CC, $(CC) $(CPPFLAGS) $(CFLAGS) $(SHIPFLAGS) -DNDEBUG -c $< -o $@)

Could try replacing all occurrences of gcc in Make.inc with the path to your gcc and see if that works?

[catawbasam]

Yes, that did work.

[catawbasam]

I'd like to have it use my CC environment variable instead.
However that led to a circular reference error Make.inc lines 278-281:

CC_BASE = $(CROSS_COMPILE)/usr/depot/gcc-4.8.2/bin/gcc
CXX_BASE = $(CROSS_COMPILE)g++
CC = $(CC_BASE)
CXX = $(CXX_BASE)

[nalimilan]

Can you identify the precise place where replacing gcc with the correct path makes a difference? Is changing CC_BASE enough?

[catawbasam]

Yes, changing CC_BASE is enough. Looks like several of the other gcc references are in if-then branches that may not be exercised when I make.

[nalimilan]

OK, thanks. That CC_BASE = $(CROSS_COMPILE)gcc line indeed hardcodes the gcc command, which doesn't look right at all. I think we should always respect CC when it is set, and only use that line when it's missing.

[catawbasam]

Sounds great. Thanks!

[tkelman]

That line is a workaround for cmake not playing well with ccache. But yeah we should probably check whether CC is set before hard-coding that variable.

[nalimilan]

@tkelman What's the issue with CMake exactly? Can we pass it an absolute path to the compiler when CC is defined?

[tkelman]

See the libgit2 PR's. The issue is Julia's makefiles do some really ugly bad-idea things with appending flags (-m32 for 32-bit compilation from 64-bit Linux, and a set of OSX-specific flags) to CC. I proposed to fix that, because cmake requires the compiler argument to be a path, not including any flags. But rather than add a new variable to hold the flags, which was a more invasive change and there was concern that I didn't catch every usage (all the dependencies, etc), I came up with this simpler workaround using CC_BASE. We could try an alternate approach, sending only the first word of CC to cmake, but be careful that it doesn't break USE_CCACHE.