More information in ocamlc -config

[vicuna]

Original bug ID: 7172
Reporter: @dbuenzli
Assigned to: @dra27
Status: assigned (set by @mshinwell on 2016-12-14T11:38:40Z)
Resolution: open
Priority: normal
Severity: feature
Version: 4.02.3
Target version: 4.07.0+dev/beta2/rc1/rc2
Category: configure and build/install
Monitored by: @gasche @diml @dbuenzli

Bug description

It would be nice if ocamlc -config could report whether:

1. Native compilation is available
2. Native dynlinking is available.

[vicuna]

Comment author: @mshinwell

I think we may need to fix matters for 4.03 such that there is a means of discovering whether Flambda is enabled or not. Perhaps we could just fix these all at once.

[vicuna]

Comment author: @dbuenzli

@shinwell I'd be delighted if this could go into 4.03 and I may be tempted to hide the fact that ocamlc -config already reports if flambda is enabled or not.

[vicuna]

Comment author: @mshinwell

Oh, so it does. Well it will probably have to wait then :p

[vicuna]

Comment author: @dbuenzli

Grrrrr.

[vicuna]

Comment author: @dbuenzli

Somehow I had hoped that I could fill the missing slots with:

$(ocamlc -where)/Makefile.config

but in fact this file seems to reflect the output of the configure process, not of what is actually installed, e.g. on a bytecode only switch like:

https://github.com/ocaml/opam-repository/blob/master/compilers/4.02.3/4.02.3%2Bbytecode-only/4.02.3%2Bbytecode-only.comp

NATDYNLINK will be true even though there no actual dynlink.cmxa. In fact reading the INSTALL file it seems there's no way to configure ocaml to be bytecode only.

So I guess I'm left with discovery hacks.

I'm thinking about:

• native code compilation is available if $(ocamlc -where)/libasmrun.$(EXT_DLL) exists
  (looking up ocamlopt in the path can be a bit unreliable, for example in a bytecode only switch ocamlopt may actually be in the path, but it's the ocamlopt of the system compiler...)

• native dynlinking is available if $(ocamlc -where)/dynlink.cmxa exists

Does it sound reasonable ?

[vicuna]

Comment author: @dbuenzli

That is libasmrun.$(EXT_LIB) rather than $(EXT_DLL).

[vicuna]

Comment author: @gasche

I needed a -no-native-compiler configure option to test the bytecode-only build of ocamlbuild, and added it to the script in #387. The relevant variables should be set to a meaningful value in this case, and the fact that this is not done for NATDYNLINK comes from an (independent) omission in the relevant logic:

https://github.com/ocaml/ocaml/blob/d3d30f3/configure#L874-L878

I need to document the -no-native-compiler switch in the INSTALL file, and I think I will fix the configure script as well.

[vicuna]

Comment author: @dbuenzli

Also it would be nice to give us ext_exe, since the logic is a little bit twisted and requires to consult both ccomp_type and os_type. There is the EXE variable in Makefile.config but I now decided I'd rely only on the output of ocamlc -config.

[vicuna]

Comment author: @damiendoligez

@dbuenzli your "discovery hacks" look quite reasonable to me.

It's not clear to me how ocamlc -config could know whether the native compiler was built or anything about the build process. Currently, it relies only on information from configure.

I agree with ext_exe.

[vicuna]

Comment author: @dbuenzli

@doligez The solution is to be able to specify whether native compilers should be built in the configure of course.

[vicuna]

Comment author: @gasche

I proposed
#513
to solve this issue.

[vicuna]

Comment author: @gasche

#513 has just been merged in 4.03 and trunk. ./configure -no-native-compiler now sets NATDYNLINK=false in Makefile.config.

[vicuna]

Comment author: @dbuenzli

Another bit that is useful for build systems to know an that should be reported by ocamlc -config is the value of Sys.word_size on the target platform. This is useful even for OCaml programs as it affects the size of ocaml integers.

Currently workarounds include:

1. Grepping for SIZEOF_PTR in $(ocamlc -where)/caml/config.h
2. Trying to pattern match on the values of architecture: and model: of ocamlc -config and try to infer the size based on the meaning of these keys according to

   ocaml/configure

   Lines 810 to 862 in 4ca0ca4

   	case "$target" in
   	sparc*-*-solaris2.*) arch=sparc; system=solaris;;
   	sparc*-*-*bsd*) arch=sparc; system=bsd;;
   	sparc*-*-linux*) arch=sparc; system=linux;;
   	sparc*-*-gnu*) arch=sparc; system=gnu;;
   	i[3456]86-*-linux*) arch=i386; system=linux_`sh ./runtest elf.c`;;
   	i[3456]86-*-*bsd*) arch=i386; system=bsd_`sh ./runtest elf.c`;;
   	i[3456]86-*-nextstep*) arch=i386; system=nextstep;;
   	i[3456]86-*-solaris*) if $arch64; then
   	arch=amd64; system=solaris
   	else
   	arch=i386; system=solaris
   	fi;;
   	i[3456]86-*-haiku*) arch=i386; system=beos;;
   	i[3456]86-*-beos*) arch=i386; system=beos;;
   	i[3456]86-*-cygwin*) arch=i386; system=cygwin;;
   	i[3456]86-*-darwin*) if $arch64; then
   	arch=amd64; system=macosx
   	else
   	arch=i386; system=macosx
   	fi;;
   	i[3456]86-*-gnu*) arch=i386; system=gnu;;
   	i[3456]86-*-mingw*) arch=i386; system=mingw;;
   	powerpc64le*-*-linux*) arch=power; model=ppc64le; system=elf;;
   	powerpc*-*-linux*) arch=power;
   	if $arch64; then model=ppc64; else model=ppc; fi
   	system=elf;;
   	powerpc-*-netbsd*) arch=power; model=ppc; system=elf;;
   	powerpc-*-openbsd*) arch=power; model=ppc; system=bsd_elf;;
   	s390x*-*-linux*) arch=s390x; model=z10; system=elf;;
   	armv6*-*-linux-gnueabihf) arch=arm; model=armv6; system=linux_eabihf;;
   	arm*-*-linux-gnueabihf) arch=arm; system=linux_eabihf;;
   	armv7*-*-linux-gnueabi) arch=arm; model=armv7; system=linux_eabi;;
   	armv6t2*-*-linux-gnueabi) arch=arm; model=armv6t2; system=linux_eabi;;
   	armv6*-*-linux-gnueabi) arch=arm; model=armv6; system=linux_eabi;;
   	armv6*-*-freebsd*) arch=arm; model=armv6; system=freebsd;;
   	earmv6*-*-netbsd*) arch=arm; model=armv6; system=netbsd;;
   	earmv7*-*-netbsd*) arch=arm; model=armv7; system=netbsd;;
   	armv5te*-*-linux-gnueabi) arch=arm; model=armv5te; system=linux_eabi;;
   	armv5*-*-linux-gnueabi) arch=arm; model=armv5; system=linux_eabi;;
   	arm*-*-linux-gnueabi) arch=arm; system=linux_eabi;;
   	arm*-*-openbsd*) arch=arm; system=bsd;;
   	zaurus*-*-openbsd*) arch=arm; system=bsd;;
   	x86_64-*-linux*) arch=amd64; system=linux;;
   	x86_64-*-gnu*) arch=amd64; system=gnu;;
   	x86_64-*-freebsd*) arch=amd64; system=freebsd;;
   	x86_64-*-netbsd*) arch=amd64; system=netbsd;;
   	x86_64-*-openbsd*) arch=amd64; system=openbsd;;
   	x86_64-*-darwin*) arch=amd64; system=macosx;;
   	x86_64-*-mingw*) arch=amd64; system=mingw;;
   	aarch64-*-linux*) arch=arm64; system=linux;;
   	x86_64-*-cygwin*) arch=amd64; system=cygwin;;
   	esac

[vicuna]

Comment author: @dbuenzli

shinwell asked me to update on the status of this bug. So in 4.04, ocamlc -config still does not report the following useful tidbits:

• ext_exe
• word_size
• natdynlink
• nativecomp

[vicuna]

Comment author: @gasche

#970

[vicuna]

Comment author: @dra27

#970 adds ext_exe, word_size and possibly natdynlink.

A suggestion for nativecomp (and really natdynlink) would to be change the logic for make install. At present, if native code has been built then it is installed, but we could instead make it that if native code has been configured then make install fails if it has not also been built (same could sensibly be done at the same time for opt.opt).

At present, Windows should be configured so that opt and opt.opt are assumed.

[vicuna]

Comment author: @gasche

So after discussion with dra, #970 only includes int_size, word_size and ext_exe. I won't work on natdynlink and nativecomp myself in the 4.05 timeline, so I just un-assigned myself.

[vicuna]

Comment author: @dra27

So int_size, word_size and ext_exe are merged and will be in 4.05.0

I've assigned the issue and hope to have a GPR for natdynlink and nativecomp ready for the 4.05 freeze.

[dbuenzli]

• natdynlink
• nativecomp

Are still missing (at least in 4.10) and would be nice to have.

[dbuenzli]

I think I can do without nativecomp. But natdynlink is needed to understand the output of ocamlmklib for C stubs (if not available only a .a is generated).

Is natdynlink implied by supports_shared_libraries by any chance ?

[dbuenzli]

But natdynlink is needed to understand the output of ocamlmklib for C stubs (if not available only a .a is generated).

That formulation is a bit inexact I'm confusing two things here:

1. There is support for C dynamic linking (e.g. used by bytecode for loading primitives of C stubs)
2. The OCaml native dynlink API, (i.e. the ability to produce .cmxs files)

I highly suspect supports_shared_libraries answers at least 1. I'm unsure whether it answers 2.

[shindere]

This has now been fixed by #11819, which should be available in OCaml 5.1.