Python 3.11 loses the ability to set PYTHON_DECIMAL_WITH_MACHINE

[Bo98]

Bug report

Python 3.10 had the ability to set PYTHON_DECIMAL_WITH_MACHINE to override the choice of configuration for the _decimal module:

cpython/setup.py

Lines 2388 to 2397 in dcb342b

	machine = os.environ.get('PYTHON_DECIMAL_WITH_MACHINE')
	if machine:
	# Override automatic configuration to facilitate testing.
	define_macros = config[machine]
	elif MACOS:
	# Universal here means: build with the same options Python
	# was built with.
	define_macros = config['universal']
	elif sizeof_size_t == 8:

Since Python 3.11, this is no longer possible. This feature was necessary, on macOS particularly, with the --with-system-libmpdec option if that system libmpdec is configured differently to the default Python config. On macOS, the default Python config forces universal, while setting PYTHON_DECIMAL_WITH_MACHINE allowed it to be single-arch.

libmpdec produces different headers depending on how it was built, which is why the setting is important to be able to override. Without it, the _decimal module will fail to compile if the default does not match how system libmpdec was built.

Homebrew's Python currently depends on this feature.

A test within CPython also seems to depend on this feature:

cpython/Modules/_decimal/tests/runall-memorydebugger.sh

Lines 63 to 79 in f4c0348

	unset PYTHON_DECIMAL_WITH_MACHINE
	libmpdec_config=$config
	if [ X"$config" != X"auto" ]; then
	PYTHON_DECIMAL_WITH_MACHINE=$config
	export PYTHON_DECIMAL_WITH_MACHINE
	else
	libmpdec_config=""
	fi
	############ refleak tests ###########
	print_config "refleak tests: config=$config" $args
	printf "\nbuilding python ...\n\n"
	cd ../../
	$GMAKE distclean > /dev/null 2>&1
	./configure CFLAGS="$ADD_CFLAGS" LDFLAGS="$ADD_LDFLAGS" --with-pydebug $args > /dev/null 2>&1
	$GMAKE | grep _decimal

Your environment

• CPython versions tested on: 3.11.0rc2
• Operating system and architecture: macOS 12 (x86_64 and arm64)

[AlexWaygood]

Cc. @rhettinger for decimal module expertise.

[erlend-aasland]

FTR, this was removed in GH-29541 (cc. @tiran as PR author and @mdickinson as PR reviewer)

[erlend-aasland]

AFAICS from GH-29541, it should be possible to set PYTHON_DECIMAL_WITH_MACHINE via CFLAGS (for example LIBMPDEC_CFLAGS). @Bo98, can you please test this?

[Bo98]

What reads PYTHON_DECIMAL_WITH_MACHINE? It's not a define to be passed to C but something that used to change what item is selected from:

cpython/configure.ac

Lines 3882 to 3891 in 75a6fad

	AS_CASE([$libmpdec_machine],
	[x64], [AS_VAR_APPEND([LIBMPDEC_CFLAGS], [" -DCONFIG_64=1 -DASM=1"])],
	[uint128], [AS_VAR_APPEND([LIBMPDEC_CFLAGS], [" -DCONFIG_64=1 -DANSI=1 -DHAVE_UINT128_T=1"])],
	[ansi64], [AS_VAR_APPEND([LIBMPDEC_CFLAGS], [" -DCONFIG_64=1 -DANSI=1"])],
	[ppro], [AS_VAR_APPEND([LIBMPDEC_CFLAGS], [" -DCONFIG_32=1 -DANSI=1 -DASM=1 -Wno-unknown-pragmas"])],
	[ansi32], [AS_VAR_APPEND([LIBMPDEC_CFLAGS], [" -DCONFIG_32=1 -DANSI=1"])],
	[ansi-legacy], [AS_VAR_APPEND([LIBMPDEC_CFLAGS], [" -DCONFIG_32=1 -DANSI=1 -DLEGACY_COMPILER=1"])],
	[universal], [AS_VAR_APPEND([LIBMPDEC_CFLAGS], [" -DUNIVERSAL=1"])],
	[AC_MSG_ERROR([_decimal: unsupported architecture])]
	)

Unless you suggest that I instead pass LIBMPDEC_CFLAGS=-DCONFIG_64=1 -DASM=1 on x86_64 and LIBMPDEC_CFLAGS=-DCONFIG_64=1 -DANSI=1 -DHAVE_UINT128_T=1 on arm64? (Assuming this list is never changing.)

[erlend-aasland]

Ah, sorry, you are right. I misread the diff from GH-29541; it is used as an index, not a CFLAG.

[erlend-aasland]

Yes, it seems this ability is now lost; there is no way to manually override that AS_CASE. We'd either have to reintroduce PYTHON_DECIMAL_WITH_MACHINE, or provide a new configure switch for this, if we were to reintroduce this.

[erlend-aasland]

You mentioned Homebrew? Does this prevent Homebrew from building and distributing Python 3.11? Is there a Homebrew ticket for this?

[erlend-aasland]

From #98557 (comment):

Unless you suggest that I instead pass LIBMPDEC_CFLAGS=-DCONFIG_64=1 -DASM=1 on x86_64 and LIBMPDEC_CFLAGS=-DCONFIG_64=1 -DANSI=1 -DHAVE_UINT128_T=1 on arm64? (Assuming this list is never changing.)

Then from the OP:

libmpdec produces different headers depending on how it was built, which is why the setting is important to be able to override. Without it, the _decimal module will fail to compile if the default does not match how system libmpdec was built.

I suggest you use LIBMPDEC_CFLAGS to set these flags depending on how your libmpdec was built. That seems to me to be a more robust approach than letting configure chose flags which might not reflect how your libmpdec was built.

ISTM we don't need a way to override the libmpdec_machine switch in configure.

[Bo98]

Is there a Homebrew ticket for this?

Not yet considering we don't ship betas, but what we'll probably do for 3.11.0 (and what I've done in a local test rc2 build) is a s/libmpdec_machine=universal/libmpdec_machine=x64/ (uint128 on arm64) on configure. Not a desirable long-term solution but should be ok for a short-term fix.

I suggest you use LIBMPDEC_CFLAGS to set these flags depending on how your libmpdec was built. That seems to me to be a more robust approach than letting configure chose flags which might not reflect how your libmpdec was built.

PYTHON_DECIMAL_WITH_MACHINE did directly match the option you would pass to libmpdec configure. If you passed MACHINE=x64 to libmpdec while building then you would pass the same to Python.

It made sense and is what Stefan Krah recommended we did.

Might I propose an alternative way forward here however. I propose looking into not setting any CFLAGS for the system libmpdec and checking MPD_CONFIG_64 and MPD_CONFIG_32 in _decimal source code, given mpdecimal.h should tell you how it was built by setting those.

FWIW, this is what the relevent section of mpdecimal.h looks like on a different configurations of system libmpdec:

Universal

/* Mac OS X: support for building universal binaries */
#if defined(MPD_CONFIG_64) || defined(MPD_CONFIG_32)
  #error "cannot use MPD_CONFIG_64 or MPD_CONFIG_32 with universal header."
#endif

#if defined(CONFIG_64) || defined(CONFIG_32)
  #error "cannot use CONFIG_64 or CONFIG_32 with universal header."
#endif

#if defined(__ppc__)
  #define MPD_CONFIG_32 1
  #ifdef UNIVERSAL
    #define CONFIG_32
    #define ANSI
  #endif
#elif defined(__ppc64__)
  #define MPD_CONFIG_64 1
  #ifdef UNIVERSAL
    #define CONFIG_64
    #define ANSI
  #endif
#elif defined(__i386__)
  #define MPD_CONFIG_32 1
  #ifdef UNIVERSAL
    #define CONFIG_32
    #define ANSI
  #endif
#elif defined(__x86_64__)
  #define MPD_CONFIG_64 1
  #ifdef UNIVERSAL
    #define CONFIG_64
    #define ASM
  #endif
#elif defined(__arm64__)
  #define MPD_CONFIG_64 1
  #ifdef UNIVERSAL
    #define CONFIG_64
    #define ANSI
  #endif
#else
  #error "unknown architecture for universal build."
#endif

x64 and uint128

/* ABI: 64-bit */
#define MPD_CONFIG_64 1

#ifdef MPD_CONFIG_32
  #error "cannot use MPD_CONFIG_32 with 64-bit header."
#endif

#ifdef CONFIG_32
  #error "cannot use CONFIG_32 with 64-bit header."
#endif

[erlend-aasland]

Might I propose an alternative way forward here however. I propose looking into not setting any CFLAGS for the system libmpdec and checking MPD_CONFIG_64 and MPD_CONFIG_32 in _decimal source code, given mpdecimal.h should tell you how it was built by setting those.

That sounds like a better way forward indeed. I've already pinged Christian on this issue; let's wait and see if he also agrees.

[carlocab]

It's been a while here. Can we find a way to move this forward?

[erlend-aasland]

It's been a while here. Can we find a way to move this forward?

For 3.11? Unfortunately not; 3.11 only received security fixes. For 3.12 and newer, the environment variables LIBMPDEC_CFLAGS and LIBMPDEC_LIBS are available if you want to customise how _decimal is built.

See also #115119.

[Bo98]

Looks like #115406 might have fixed this by moving the assumptions into a AS_VAR_IF([with_system_libmpdec], [no] in 3.13. Specifically, the flag and decimal.c changes rather than the actual pkg-config part, so the fix works for pre-4.0 too.

For 3.11 and 3.12 it was previously unfixable without patching the configure file (LIBMPDEC_CFLAGS won't override it), unless a backport of that new patch is applied.