Regression: meson 0.53.0 cross compilation "Unable to determine dynamic linker"

[asavah]

Describe the bug
I get "Unable to determine dynamic linker" when trying to cross compile for example dav1d.

To Reproduce
Try cross compiling dav1d https://code.videolan.org/videolan/dav1d/blob/master/meson.build

Expected behavior
The build succeeds as it it did with 0.52.x
Adding ld = 'ld-binary" did not help
Disregard that this is x86_64 to x86_64 cross compilation, toolchain is different and cross is needed to keep host and target environments apart.

Build started at 2020-01-08T17:11:17.925584
Main binary: /home/asavah/kross/host/bin/python3
Build Options: -Db_ndebug=true -Db_lto=true -Dbitdepths=8,16 -Denable_asm=true -Denable_tools=true -Denable_examples=false -Denable_tests=false -Dlogging=false -Dprefix=/usr -Dbindir=/usr/bin -Ddatadir=/usr/share -Dincludedir=/usr/include -Dinfodir=/usr/share/info -Dlibdir=/usr/lib -Dlibexecdir=/usr/lib -Dlocalstatedir=/var -Dmandir=/usr/share/man -Dsbindir=/usr/sbin -Dsharedstatedir=/var -Dsysconfdir=/etc -Dbackend=ninja -Dbuildtype=plain '--cross-file /home/asavah/kross/tmp/asusb450eg/toolchain.meson'
Python system: Linux
The Meson build system
Version: 0.53.0
Source dir: /home/asavah/kross/src/dav1d
Build dir: /home/asavah/kross/tmp/asusb450eg/dav1d
Build type: cross build
Project name: dav1d
Project version: 0.5.2
No CFLAGS in the environment, not changing global flags.
No LDFLAGS in the environment, not changing global flags.
Appending CPPFLAGS from environment: '-DPIC'

meson.build:25:0: ERROR: Unable to determine dynamic linker

[binaries]
c = ['ccache','x86_64-asusb450eg-linux-gnu-gcc']
cpp = ['ccache','x86_64-asusb450eg-linux-gnu-g++']
ar = 'x86_64-asusb450eg-linux-gnu-ar'
strip = 'x86_64-asusb450eg-linux-gnu-strip'
nm = 'x86_64-asusb450eg-linux-gnu-nm'
ld = 'x86_64-asusb450eg-linux-gnu-ld'
objdump = 'x86_64-asusb450eg-linux-gnu-objdump'
pkgconfig = 'x86_64-asusb450eg-linux-gnu-pkg-config'
exe_wrapper = 'kqemu-wrapper'
llvm-config = '/home/asavah/kross/host/asusb450eg/bin-config/llvm-config'
cmake = '/home/asavah/kross/host/bin/cmake'

[host_machine]
system = 'linux'
cpu_family = 'x86_64'
cpu = 'znver1'
endian = 'little'

[properties]
needs_exe_wrapper = true
has_function_printf = true

system parameters

• cross build
• ubuntu 19.10
• python 3.8.1
• 0.53.0
• 1.9.0.git

Edit:
Same happens when building systemd or any other project using meson.
Reverted to 0.52.1.
Suggest reverting ld nonsense and test cross compilation more thoroughly.

[asavah]

This check is bogus https://github.com/mesonbuild/meson/blob/master/mesonbuild/environment.py#L820
This is what actually runs with popen with or withoud ld= removed from the toolchain hint file.

cc -Wl,--version -DPIC -fuse-ld=x86_64-asusb450eg-linux-gnu-ld
cc: error: unrecognized command line option ‘-fuse-ld=x86_64-asusb450eg-linux-gnu-ld’

Which is not valid for gcc,
-fuse-ld can be either bfd or gold or lld without host prefix
I suggest whoever coded this mess reads gcc manual first https://gcc.gnu.org/onlinedocs/gcc/Link-Options.html

setting
ld='bfd'
does not help either, "machine-prefixed-ld" is still passed to fuse-ld=.

This is what should run for such a test

machine-prefixed-cc -Wl,--version -DPIC -fuse-ld=bfd #or gold

[ryanwoodsmall]

Seconded. This broke a musl cross-compile on a glibc-based OS and I couldn't find a suitable workaround. Things I tried:

• manually setting "ld = '...'" in a cross file
• manually setting "ld = '...'" in a native file
• setting LD environment variable
• symlinking toolchain bins to a temp dir at the front of PATH

I would assume LD in the environment would override the linker guessing logic, but it doesn't appear to do so. This is somewhat ironic as I was actually attempting to compile C code to a static binary. Another assumption would be that --default-library static (or -Wl,-static in CFLAGS or -static in LDFLAGS) would also short-circuit checks here, or at least make the failure non-fatal, but that's not the case.

Meson figured that I was using a cross-compiler as "native" and warned such, but still couldn't resolve a dynamic linker. None of this was workable on x86_64 -> x86_64 cross with musl compilers on CentOS 7. For giggles, I tried on ARM 32- and 64-bit glibc dists with musl compilers as well with the same results. Native (glibc or musl) compiles seem to work fine, or at least better, for a quick set of tests.

Reverted to 0.52.1 as well until this is baked.

[asavah]

@ryanwoodsmall thanks for the pointer,

I thought I did not export LD in environment when cross compiling with meson,
I do unexport CC/CXX vars for meson because it freaks meson out.
I rechecked and yes, LD still was exported.
Unexporting it lets meson go one step further and it correctly guesses build machine linker

The Meson build system
Version: 0.53.0
Source dir: /home/asavah/kross/src/dav1d
Build dir: /home/asavah/kross/tmp/asusb450eg/dav1d
Build type: cross build
Project name: dav1d
Project version: 0.5.2
Appending CPPFLAGS from environment: '-DPIC'
C compiler for the build machine: ccache cc (gcc 9.2.1 "cc (Ubuntu 9.2.1-9ubuntu2) 9.2.1 20191008")
C linker for the build machine: cc GNU ld.bfd 2.33

meson.build:25:0: ERROR: Unable to determine dynamic linker

A full log can be found at /home/asavah/kross/tmp/asusb450eg/dav1d/meson-logs/meson-log.txt

Ok, lets see what meson does by adding a simple print(compiler + check_args) just above https://github.com/mesonbuild/meson/blob/master/mesonbuild/environment.py#L820

This is with `ld = 'x86_64-asusb450eg-linux-gnu-ld' in toolchain file, LD unexported:

The Meson build system
Version: 0.53.0
Source dir: /home/asavah/kross/src/dav1d
Build dir: /home/asavah/kross/tmp/asusb450eg/dav1d
Build type: cross build
Project name: dav1d
Project version: 0.5.2
Appending CPPFLAGS from environment: '-DPIC'
['cc', '-Wl,--version', '-DPIC']
C compiler for the build machine: ccache cc (gcc 9.2.1 "cc (Ubuntu 9.2.1-9ubuntu2) 9.2.1 20191008")
C linker for the build machine: cc GNU ld.bfd 2.33
['x86_64-asusb450eg-linux-gnu-gcc', '-Wl,--version', '-fuse-ld=x86_64-asusb450eg-linux-gnu-ld']

meson.build:25:0: ERROR: Unable to determine dynamic linker

A full log can be found at /home/asavah/kross/tmp/asusb450eg/dav1d/meson-logs/meson-log.txt

This is with without ld = and LD unexported:

The Meson build system
Version: 0.53.0
Source dir: /home/asavah/kross/src/dav1d
Build dir: /home/asavah/kross/tmp/asusb450eg/dav1d
Build type: cross build
Project name: dav1d
Project version: 0.5.2
Appending CPPFLAGS from environment: '-DPIC'
['cc', '-Wl,--version', '-DPIC']
C compiler for the build machine: ccache cc (gcc 9.2.1 "cc (Ubuntu 9.2.1-9ubuntu2) 9.2.1 20191008")
C linker for the build machine: cc GNU ld.bfd 2.33
['x86_64-asusb450eg-linux-gnu-gcc', '-Wl,--version']
C compiler for the host machine: ccache x86_64-asusb450eg-linux-gnu-gcc (gcc 9.2.1 "x86_64-asusb450eg-linux-gnu-gcc (GCC) 9.2.1 20200102")
C linker for the host machine: x86_64-asusb450eg-linux-gnu-gcc GNU ld.bfd 2.33.1.20191206
Build machine cpu family: x86_64
Build machine cpu: x86_64
Host machine cpu family: x86_64
Host machine cpu: znver1
Target machine cpu family: x86_64
Target machine cpu: znver1
Run-time dependency threads found: YES
Checking for function "clock_gettime" : YES
Library dl found: YES
Checking for function "dlsym" with dependency -ldl: YES
Check usable header "stdatomic.h" : YES
Check usable header "unistd.h" : YES
Check usable header "io.h" : NO
Checking for function "getopt_long" : YES
Checking for function "posix_memalign" : YES
Compiler for C supports arguments -fvisibility=hidden: YES
Compiler for C supports arguments -Wundef: YES
Compiler for C supports arguments -Werror=vla: YES
Compiler for C supports arguments -Wno-maybe-uninitialized -Wmaybe-uninitialized: YES
Compiler for C supports arguments -Wno-missing-field-initializers -Wmissing-field-initializers: YES
Compiler for C supports arguments -Wno-unused-parameter -Wunused-parameter: YES
Compiler for C supports arguments -Werror=missing-prototypes: YES
Compiler for C supports arguments -Wshorten-64-to-32: NO
Compiler for C supports arguments -mpreferred-stack-boundary=5: YES
Configuring config.h using configuration
Configuring config.asm using configuration
Program nasm found: YES (/home/asavah/kross/host/bin/nasm)
Configuring version.h using configuration
Program doxygen found: NO
Configuring cli_config.h using configuration
Build targets in project: 10

Option buildtype is: plain [default: release]
Option b_ndebug is: true [default: if-release]
Found ninja-1.9.0.git at /home/asavah/kross/host/bin/ninja
dav1d: configure DONE.

And it builds!

So the key for cross compilation is to make sure LD is NOT exported and ld = is NOT specified as target-triplet-ld (as we do for c||cpp||etc) but as linker type (bfd||gold||lld) in toolchain file or NOT specified at all.
This should be properly documented and as far as I understand it contradicts the purpose of [binaries] where one should specify the real executable substite. Eg doing ld = 'bfd' does not specify the binary for cross environment, but specifies linker type.

And AFAIK the fact that you can't have CC/CXX/LD variables exported while cross compiling with meson is not documented (if someone points me to the sentence where this is documented I'll happily stfu :) )
This is more important for OS builders/maintainers than for software developers.

[ryanwoodsmall]

Seconded to this as well. Have something working now, but feel this is still somewhat underdocumented. Furthermore, precedence seems underspecified, i.e., environment vs. cross/native file vs command line vs... Tables at https://mesonbuild.com/Reference-tables.html#linker-ids mention ld.bfd but that didn't work for either ld = ... or LD environment variable. Also attempted with a custom CC and LD=bfd env with no luck. Had to unset my entire cross environment, i.e. all of:

• AR
• AS
• CC
• CFLAGS
• CPPFLAGS
• CPP
• CXXFLAGS
• CXX
• LDFLAGS
• LD
• STRIP

Working cross file looks like this:

[binaries]
ld = 'bfd'
c = 'x86_64-linux-musl-gcc'
ar = 'x86_64-linux-musl-ar'
as = 'x86_64-linux-musl-as'
cpp = 'x86_64-linux-musl-cpp'
strip = 'x86_64-linux-musl-strip'
[properties]
c_args = '-I/usr/local/crosware/software/muslfts/current/include'
c_link_args = '-L/usr/local/crosware/software/muslfts/current/lib -lfts -static'

Thanks for working through this and for pointers. Looks like I can now build ksh93 (from https://github.com/att/ast) with compilers built from musl-cross-make (https://github.com/richfelker/musl-cross-make) w/Meson 0.53.0.

[asavah]

Yep I had to unexport everything too:

ifeq ($(MESON_HACKS),1)
export HINT_CC := $(CC)
export HINT_CXX := $(CXX)
export HINT_LD := $(LD)
export HINT_CFLAGS := $(CFLAGS)
export HINT_CXXFLAGS := $(CXXFLAGS)
export HINT_LDFLAGS := $(LDFLAGS)
ifeq ($(BUILD_FOR),target)
export CC =
unexport CC
export CXX =
unexport CXX
export LD =
unexport LD
export CFLAGS =
unexport CFLAGS
export HOST_CFLAGS =
unexport HOST_CFLAGS
export CXXFLAGS =
unexport CXXFLAGS
export HOST_CXXFLAGS =
unexport HOST_CXXFLAGS
export LDFLAGS =
unexport LDFLAGS
export HOST_LDFLAGS =
unexport HOST_LDFLAGS
export STRIP =
unexport STRIP
export CMAKE =
unexport CMAKE
endif
endif

Quote from https://mesonbuild.com/Overview.html

The main design goals of this language has been simplicity, clarity and conciseness.

Yep, totally

[jameshilliard]

See if #6436 fixes this.

[FlyingBBQ]

I experienced the same problem when updating to Meson 0.53.0 and I agree that documentation for this can be better.

In my project I cross compile for ARM where I have the following linker binaries on my Linux machine:

arm-none-eabi-ld
arm-none-eabi-ld.bfd
arm-none-eabi-ld.gold

Setting any of these binaries in my cross file resulted in the "Unable to determine dynamic linker" error. Eventually it was by trial and error that I found the correct setting should be ld = 'bfd', but as @asavah mentioned, this sort of contradicts the [binaries] section in the cross file.

To me it would make sense that I can explicitly set the linker binary e.g.

[binaries]
ld = 'arm-none-eabi-ld.bfd'
c = 'arm-none-eabi-gcc'
cpp = 'arm-none-eabi-g++'

[codyps]

@jameshilliard that will work around the issue, but it does seem like we might also want a way to set the -fuse-ld parameter separately from the ld program being passed in the cross file spec. It also could try to guess at the ld type (for example, someone might be using ld = 'arm-none-eabi-ld.gold', etc)

To clarify, right now I use ld = find_program('ld') in my meson.build file, and invoke ld directly in a custom_target to generate certain binary files.

[jameshilliard]

@jmesmon Yeah, might be a good idea to do that as well(I'll let someone else figure that out if it's important for them), the -fuse-ld whitelist in my PR probably still makes sense to have as a last resort sanity check to prevent passing invalid args to gcc.

[dcbaker]

Closing in favor of #6442, which is mean to find a solution to all of the linker stuff.

[marc-h38]

Furthermore, precedence seems underspecified, i.e., environment vs. cross/native file vs command line vs...

ld_args and c_args precedence leaves me confused too, see #6362