orc @0.4.30_1: ../orc-0.4.30/orc/orccpu-x86.c:91:7: error: instruction requires: Not 64-bit mode

[kencu (Ken)]

On a system running 10.6.8, building orc at 64 bit:

:info:build FAILED: orc/8524582@@orc-0.4@sha/orccpu-x86.c.o 
:info:build /opt/local/bin/clang-mp-9.0 -Iorc/8524582@@orc-0.4@sha -Iorc -I../orc-0.4.30/orc -I. -I../orc-0.4.30/ -I/opt/local/include -Xclang -fcolor-diagnostics -pipe -Wall -Winvalid-pch -std=gnu99 -O2 -g -DHAVE_CONFIG_H -fvisibility=hidden -Os -arch x86_64 -DORC_ENABLE_UNSTABLE_API -D_GNU_SOURCE -DBUILDING_ORC -MD -MQ 'orc/8524582@@orc-0.4@sha/orccpu-x86.c.o' -MF 'orc/8524582@@orc-0.4@sha/orccpu-x86.c.o.d' -o 'orc/8524582@@orc-0.4@sha/orccpu-x86.c.o' -c ../orc-0.4.30/orc/orccpu-x86.c
:info:build ../orc-0.4.30/orc/orccpu-x86.c:91:7: error: instruction requires: Not 64-bit mode
:info:build       "  pushl %%ebx\n"
:info:build       ^
:info:build <inline asm>:1:4: note: instantiated into assembly here
:info:build           pushl %ebx
:info:build           ^
:info:build ../orc-0.4.30/orc/orccpu-x86.c:94:8: error: instruction requires: Not 64-bit mode
:info:build       "  popl %%ebx\n"
:info:build        ^
:info:build <inline asm>:4:3: note: instantiated into assembly here
:info:build   popl %ebx
:info:build   ^
:info:build ../orc-0.4.30/orc/orccpu-x86.c:91:7: error: instruction requires: Not 64-bit mode
:info:build       "  pushl %%ebx\n"
:info:build       ^
:info:build <inline asm>:1:4: note: instantiated into assembly here
:info:build           pushl %ebx
:info:build           ^
:info:build ../orc-0.4.30/orc/orccpu-x86.c:94:8: error: instruction requires: Not 64-bit mode
:info:build       "  popl %%ebx\n"
:info:build        ^
:info:build <inline asm>:4:3: note: instantiated into assembly here
:info:build   popl %ebx
:info:build   ^
:info:build ../orc-0.4.30/orc/orccpu-x86.c:91:7: error: instruction requires: Not 64-bit mode
:info:build       "  pushl %%ebx\n"
:info:build       ^
:info:build <inline asm>:1:4: note: instantiated into assembly here
:info:build           pushl %ebx
:info:build           ^
:info:build ../orc-0.4.30/orc/orccpu-x86.c:94:8: error: instruction requires: Not 64-bit mode
:info:build       "  popl %%ebx\n"
:info:build        ^
:info:build <inline asm>:4:3: note: instantiated into assembly here
:info:build   popl %ebx
:info:build   ^
:info:build ../orc-0.4.30/orc/orccpu-x86.c:91:7: error: instruction requires: Not 64-bit mode
:info:build       "  pushl %%ebx\n"
:info:build       ^
:info:build <inline asm>:1:4: note: instantiated into assembly here
:info:build           pushl %ebx
:info:build           ^
:info:build ../orc-0.4.30/orc/orccpu-x86.c:94:8: error: instruction requires: Not 64-bit mode
:info:build       "  popl %%ebx\n"
:info:build        ^
:info:build <inline asm>:4:3: note: instantiated into assembly here
:info:build   popl %ebx
:info:build   ^
:info:build ../orc-0.4.30/orc/orccpu-x86.c:91:7: error: instruction requires: Not 64-bit mode
:info:build       "  pushl %%ebx\n"
:info:build       ^
:info:build <inline asm>:1:4: note: instantiated into assembly here
:info:build           pushl %ebx
:info:build           ^
:info:build ../orc-0.4.30/orc/orccpu-x86.c:94:8: error: instruction requires: Not 64-bit mode
:info:build       "  popl %%ebx\n"
:info:build        ^
:info:build <inline asm>:4:3: note: instantiated into assembly here
:info:build   popl %ebx
:info:build   ^
:info:build ../orc-0.4.30/orc/orccpu-x86.c:91:7: error: instruction requires: Not 64-bit mode
:info:build       "  pushl %%ebx\n"
:info:build       ^
:info:build <inline asm>:1:4: note: instantiated into assembly here
:info:build           pushl %ebx
:info:build           ^
:info:build ../orc-0.4.30/orc/orccpu-x86.c:94:8: error: instruction requires: Not 64-bit mode
:info:build       "  popl %%ebx\n"
:info:build        ^
:info:build <inline asm>:4:3: note: instantiated into assembly here
:info:build   popl %ebx
:info:build   ^
:info:build ../orc-0.4.30/orc/orccpu-x86.c:91:7: error: instruction requires: Not 64-bit mode
:info:build       "  pushl %%ebx\n"
:info:build       ^
:info:build <inline asm>:1:4: note: instantiated into assembly here
:info:build           pushl %ebx
:info:build           ^
:info:build ../orc-0.4.30/orc/orccpu-x86.c:94:8: error: instruction requires: Not 64-bit mode
:info:build       "  popl %%ebx\n"
:info:build        ^
:info:build <inline asm>:4:3: note: instantiated into assembly here
:info:build   popl %ebx
:info:build   ^
:info:build ../orc-0.4.30/orc/orccpu-x86.c:91:7: error: instruction requires: Not 64-bit mode
:info:build       "  pushl %%ebx\n"
:info:build       ^
:info:build <inline asm>:1:4: note: instantiated into assembly here
:info:build           pushl %ebx
:info:build           ^
:info:build ../orc-0.4.30/orc/orccpu-x86.c:94:8: error: instruction requires: Not 64-bit mode
:info:build       "  popl %%ebx\n"
:info:build        ^
:info:build <inline asm>:4:3: note: instantiated into assembly here
:info:build   popl %ebx
:info:build   ^
:info:build ../orc-0.4.30/orc/orccpu-x86.c:91:7: error: instruction requires: Not 64-bit mode
:info:build       "  pushl %%ebx\n"
:info:build       ^
:info:build <inline asm>:1:4: note: instantiated into assembly here
:info:build           pushl %ebx
:info:build           ^
:info:build ../orc-0.4.30/orc/orccpu-x86.c:94:8: error: instruction requires: Not 64-bit mode
:info:build       "  popl %%ebx\n"
:info:build        ^
:info:build <inline asm>:4:3: note: instantiated into assembly here
:info:build   popl %ebx
:info:build   ^
:info:build ../orc-0.4.30/orc/orccpu-x86.c:91:7: error: instruction requires: Not 64-bit mode
:info:build       "  pushl %%ebx\n"
:info:build       ^
:info:build <inline asm>:1:4: note: instantiated into assembly here
:info:build           pushl %ebx
:info:build           ^
:info:build fatal error: too many errors emitted, stopping now [-ferror-limit=]
:info:build 20 errors generated.

[kencu (Ken)]

this hack fixes it, on my system at least:

$ cat patch-orc-meson-build-override-cpu-choice.diff
--- meson.build.orig	2019-10-28 12:47:16.000000000 -0700
+++ meson.build	2019-10-28 12:48:22.000000000 -0700
@@ -54,7 +54,7 @@
 
 cpu_family = host_machine.cpu_family()
 if cpu_family == 'x86'
-  cdata.set('HAVE_I386', true)
+  cdata.set('x86_64', true)
 elif cpu_family == 'x86_64'
   cdata.set('HAVE_AMD64', true)
 elif cpu_family == 'ppc' or cpu_family == 'ppc64'

Edit: -- looking at this now, this doesn't look like what worked. I'll try it again. It should be this instead, I think. I'll verify that later today -- still a hack, though, and not an acceptable fix for an actual port patch:

$ cat patch-orc-meson-build-override-cpu-choice.diff
--- meson.build.orig	2019-10-28 12:47:16.000000000 -0700
+++ meson.build	2019-10-28 12:48:22.000000000 -0700
@@ -54,7 +54,7 @@
 
 cpu_family = host_machine.cpu_family()
 if cpu_family == 'x86'
-  cdata.set('HAVE_I386', true)
+  cdata.set('HAVE_AMD64', true)
 elif cpu_family == 'x86_64'
   cdata.set('HAVE_AMD64', true)
 elif cpu_family == 'ppc' or cpu_family == 'ppc64'

[ryandesign (Ryan Carsten Schmidt)]

That doesn't look accurate. I haven't looked at this build system, but wouldn't the CPU family be "x86" on 32-bit Intel Macs? If so, setting "x86_64" to "true" would be wrong.

Perhaps the build system is misusing the output of uname -m (which returns "i386" when using the 32-bit kernel and "x86_64" when using the 64-bit kernel but even the 32-bit kernel can happily run 64-bit programs on 64-bit systems).

I can't find this cpu_family in the orc source code. Maybe it's a feature of meson.

[ryandesign (Ryan Carsten Schmidt)]

​https://mesonbuild.com/Reference-tables.html#cpu-families

These are returned by the cpu_family method of build_machine, host_machine and target_machine

x86	32 bit x86 processor
x86_64	64 bit x86 processor

[ryandesign (Ryan Carsten Schmidt)]

​https://github.com/mesonbuild/meson/blob/master/mesonbuild/environment.py#L307

[kencu (Ken)]

It's certainly not an accurate fix -- that's why I left it here in this ticket. But it does fix it on my system, as a hack.

I think this is similar to why OpenBLAS is broken on < 10.7. The system reports itself being i386 in uname, etc, but it's really x86_64, and so during the build it tries to build i386 code as x86_64 and fails -- just like OpenBLAS does.

I have a fix worked out for OpenBLAS, by querying sysctl and probing for 64bit-capable, that I haven't sent up to MacPorts because .. well. But anyway, perhaps the same kind of sysctl probing fix would work out here.

[kencu (Ken)]

BTW, I did try a few variations changing the code to look at target machine, etc, but that didn't seem to return 64bit on any variation I tried. I might have missed one variation that works, can't verify I tried every one.

This is probably a very MacOSX-legacy edge case; we might need to fix this ourselves, rather than expect meson to accommodate it...

[kencu (Ken)]

I just had to build orc again and used this trade successfully, as mentioned above, to force the compiler detection on 10.6.8.

cpu_family = host_machine.cpu_family()
 if cpu_family == 'x86'
-  cdata.set('HAVE_I386', true)
+  cdata.set('HAVE_AMD64', true)

[ryandesign (Ryan Carsten Schmidt)]

And again, as far as I can tell, that would be telling the build system "if the system is 32-bit, build for 64-bit" which would not be correct. Instead, you should be fixing whatever it is that is erroneously identifying your system as 32-bit.

The system reports itself being i386 in uname, etc, but it's really x86_64

uname -m indicates the kernel architecture. Anybody using it to decide the userland build architecture is using it wrong.

But I cannot find any place where meson is using uname -m. Rather, meson seems to be deciding that Intel systems are 32-bit if the preprocessor define __i386__ is defined. That should only be the case when compiling for 32-bit (-arch i386), which you're not, so I don't know where this is going wrong.

Meson is a build system. As such it should work on the widest possible range of systems. If this is a meson bug, then there's no reason to assume they would not want to fix it.

[kencu (Ken)]

Yes, of course, just wanted to be clear for those who stumbly by later how to get past the error of the moment and move on.

I'm afraid I am not really sure what the right approach might be for the meson devs to sort this out for the general case.

"You have a 64bit processor running a 32bit kernel supporting a 64bit operating system building +universal for both 32bit and 64bit architectures ... " I'm not sure I would take this upstream -- I can only tolerate so much abuse :>

[kencu (Ken)]

fixed by the mson cross-file implementation I added to meson