Get GMP building on Apple Silicon

[BytesGuy]

• Have you followed the guidelines for contributing?
• Have you checked that there aren't other open pull requests for the same formula update/change?
• Have you built your formula locally with brew install --build-from-source <formula>, where <formula> is the name of the formula you're submitting?
• Is your test running fine brew test <formula>, where <formula> is the name of the formula you're submitting?
• Does your build pass brew audit --strict <formula> (after doing brew install <formula>)?

---

This is a bit of a halfway fix to get GMP building and working on Apple Silicon. The architecture is not yet supported officially, so while we wait this will produce a generic C build - according to the documentation this may run slowly, but at least it gets it working for now!

[claui]

This might be really helpful, given that lots of popular formulae depend on gmp.

[BytesGuy]

This might be really helpful, given that lots of popular formulae depend on gmp.

Yeah, it's probably about as good as it gets for now 😄 I started working on this because I was trying to install GCC which depends on it - working on that next!

[fxcoudert]

Alternatively there is a patch for this on the gmp-bugs mailing list: https://gmplib.org/list-archives/gmp-bugs/2020-July/004837.html
It would enable the use of aarch64 assembly, making GMP more efficient

[BytesGuy]

Alternatively there is a patch for this on the gmp-bugs mailing list: https://gmplib.org/list-archives/gmp-bugs/2020-July/004837.html
It would enable the use of aarch64 assembly, making GMP more efficient

I have just been trying to get this to work with little success for some reason. Are you doing anything noteworthy that I might be missing here? Fairly new to this so apologies if it's something obvious!

Applying the patch and having no --build defined for configure provides:

==> Patching
==> Applying attachment.bin
patching file configure.ac
patching file mpn/arm64/arm64-defs.m4
patching file mpn/arm64/bdiv_q_1.asm
patching file mpn/arm64/darwin.m4
patching file mpn/arm64/invert_limb.asm
==> ./configure
Last 15 lines from /Users/adam/Library/Logs/Homebrew/gmp/01.configure:
checking if globals are prefixed by underscore... yes
checking how to switch to read-only data section... 	.section	__TEXT,__const
checking for assembler .type directive... 
checking for assembler .size directive... 
checking for assembler local label prefix... L
checking for assembler byte directive... .byte
checking how to define a 32-bit word... .long
checking if .align assembly directive is logarithmic... yes
checking size of void *... 8
checking size of unsigned short... 2
checking size of unsigned... 4
checking size of unsigned long... 8
checking size of mp_limb_t... 8
configure: error: Oops, mp_limb_t is 64 bits, but the assembler code
in this configuration expects 32 bits.
/usr/bin/curl --version
/usr/libexec/java_home --xml --failfast
/usr/local/Homebrew/Library/Homebrew/debrew.rb:13:in `raise'

Adding just --build=aarch64-apple-darwin20 to configure provides:

==> Patching
==> Applying attachment.bin
patching file configure.ac
patching file mpn/arm64/arm64-defs.m4
patching file mpn/arm64/bdiv_q_1.asm
patching file mpn/arm64/darwin.m4
patching file mpn/arm64/invert_limb.asm
==> ./configure --build=aarch64-apple-darwin20
==> make
Last 15 lines from /Users/adam/Library/Logs/Homebrew/gmp/02.make:
       ^
tmp-bdiv_q_1.s:76:12: error: expected ')' in parentheses expression
 LEA_LO( x7, ___gmp_binvert_limb_table)
           ^
tmp-bdiv_q_1.s:76:8: error: invalid operand
 LEA_LO( x7, ___gmp_binvert_limb_table)
       ^
make[2]: *** [bdiv_q_1.lo] Error 1
make[2]: *** Waiting for unfinished jobs....
libtool: compile:  clang -DHAVE_CONFIG_H -I. -I.. -D__GMP_WITHIN_GMP -I.. -DOPERATION_dcpi1_bdiv_qr -O2 -pedantic -march=armv8-a -c dcpi1_bdiv_qr.c  -fno-common -DPIC -o .libs/dcpi1_bdiv_qr.o
libtool: compile:  clang -DHAVE_CONFIG_H -I. -I.. -D__GMP_WITHIN_GMP -I.. -DOPERATION_sbpi1_bdiv_r -O2 -pedantic -march=armv8-a -c sbpi1_bdiv_r.c -o sbpi1_bdiv_r.o >/dev/null 2>&1
libtool: compile:  clang -DHAVE_CONFIG_H -I. -I.. -D__GMP_WITHIN_GMP -I.. -DOPERATION_dcpi1_bdiv_q -O2 -pedantic -march=armv8-a -c dcpi1_bdiv_q.c -o dcpi1_bdiv_q.o >/dev/null 2>&1
libtool: compile:  clang -DHAVE_CONFIG_H -I. -I.. -D__GMP_WITHIN_GMP -I.. -DOPERATION_dcpi1_bdiv_qr -O2 -pedantic -march=armv8-a -c dcpi1_bdiv_qr.c -o dcpi1_bdiv_qr.o >/dev/null 2>&1
make[1]: *** [all-recursive] Error 1
make: *** [all] Error 2
/usr/bin/curl --version
/usr/libexec/java_home --xml --failfast
/usr/local/Homebrew/Library/Homebrew/debrew.rb:13:in `raise'

[fxcoudert]

Since it's a configure.ac patch, it needs to regenerate configure before running it, with:

system "autoreconf", "-fiv"

and build-dependencies on autoconf, automake, and libtool.

[claui]

Example:

homebrew-core/Formula/libopendkim.rb

Lines 30 to 31 in 628584b

	# Recreate configure due to patch
	system "autoreconf", "-fvi"

[claui]

With autoreconf in place, the formula builds now but the static linking test fails.

homebrew-core/Formula/gmp.rb

Lines 50 to 52 in dc859fc

	# Test the static library to catch potential linking issues
	system ENV.cc, "test.c", "#{lib}/libgmp.a", "-o", "test"
	system "./test"

Undefined symbols for architecture arm64:
  "__gmpn_pi1_bdiv_q_1", referenced from:
      ___gmpn_bdiv_q_1 in libgmp.a(bdiv_q_1.o)
     (maybe you meant: ___gmpn_pi1_bdiv_q_1)
ld: symbol(s) not found for architecture arm64
clang: error: linker command failed with exit code 1 (use -v to see invocation)

It’s hard to see with some fonts but somehow it seems to be about the number of leading _s, again.

[BytesGuy]

@fxcoudert @claui Thank you for the pointers! 🙂 I really need to brush up on these things again, it's been a long time! Just pushed a new commit which now seems to build ok. Tests fail though, as mentioned. Hopefully this is at least a step in the right direction 🤞

[claui]

I think the branch statement in bdiv_q_1.asm is what’s causing the error:

LEA_HI(	x7, binvert_limb_table)
ubfx	x6, d, 1, 7
LEA_LO(	x7, binvert_limb_table)
ldrb	w6, [x7, x6]
ubfiz	x7, x6, 1, 8
umull	x6, w6, w6
msub	x6, x6, d, x7
lsl	x7, x6, 1
mul	x6, x6, x6
msub	x6, x6, d, x7
lsl	x7, x6, 1
mul	x6, x6, x6
msub	di, x6, d, x7

b	mpn_pi1_bdiv_q_1    ←

Evidence

$ nm /usr/local/lib/libgmp.a | grep gmpn_pi 
                 U ___gmpn_pi1_bdiv_q_1
                 U ___gmpn_pi1_bdiv_q_1
                 U ___gmpn_pi1_bdiv_q_1
                 U ___gmpn_pi1_bdiv_q_1
                 U ___gmpn_pi1_bdiv_q_1
                 U ___gmpn_pi1_bdiv_q_1
0000000000000048 T ___gmpn_pi1_bdiv_q_1
                 U __gmpn_pi1_bdiv_q_1    ←
                 U ___gmpn_pi1_bdiv_q_1
                 U ___gmpn_pi1_bdiv_q_1
$ nm /usr/local/lib/libgmp.a | grep -A 6 bdiv_q_1.o

/usr/local/lib/libgmp.a(bdiv_q_1.o):

U ___gmp_binvert_limb_table

0000000000000000 T ___gmpn_bdiv_q_1

0000000000000048 T ___gmpn_pi1_bdiv_q_1

U __gmpn_pi1_bdiv_q_1    ←

0000000000000000 t ltmp0

[BytesGuy]

Awesome, thank you! 🎉

[claui]

The CI failures in adios2, agda and ncview are obviously unrelated.

[claui]

Pinging @fxcoudert for a second review/opinion.

[fxcoudert]

See comments

[claui]

Double-checked that the consolidated patch works on arm64.

Waiting on CI (style/audit only) to pass, and for Homebrew/formula-patches#288 to be merged.

[claui]

Thanks to all of you! This is a textbook example for excellent collaboration among contributors and with upstream. ❤️

[BytesGuy]

Thanks to all of you! This is a textbook example for excellent collaboration among contributors and with upstream. ❤️

Thank you for working on this too! So glad to see this fixed as it should unblock a number of other packages 🎉