link error variable term used where not expected

[parisseb]

While trying to link KhiCAS, I get the error below that I have no idea how to deal with (I did not write assembly code, therefore the FAQ does not apply "You probably have assembly code that was not placed in a linker section, see Assembly Constants").
Source code to reproduce: https://www-fourier.univ-grenoble-alpes.fr/~parisse/ti/khicas.tgz

Thanks for any hint!

========= error log ===========
mkdir -p 'bin'
echo [linking] bin/DEMO.bin
[linking] bin/DEMO.bin
/home/parisse/CEdev/bin/fasmg -v1 '/home/parisse/CEdev/meta/ld.alm' -i 'DEBUG := 0' -i 'HAS_PRINTF := 1' -i 'HAS_LIBC := 1' -i 'HAS_LIBCXX := 1' -i 'PREFER_OS_CRT := 0' -i 'PREFER_OS_LIBC := 1' -i 'include "/home/parisse/CEdev/examples/khicas/app_tools/linker_script"' -i 'range .bss $D052C6 : $D13FD8' -i 'provide __stack = $D1A87E' -i 'locate .header at $0'  -i 'source "obj/icon.src", "/home/parisse/CEdev/lib/crt/crt0.src", "obj/lto.src"' -i 'library "/home/parisse/CEdev/lib/libload/fatdrvce.lib", "/home/parisse/CEdev/lib/libload/fileioc.lib", "/home/parisse/CEdev/lib/libload/fontlibc.lib", "/home/parisse/CEdev/lib/libload/graphx.lib", "/home/parisse/CEdev/lib/libload/keypadc.lib", "/home/parisse/CEdev/lib/libload/msddrvce.lib", "/home/parisse/CEdev/lib/libload/srldrvce.lib", "/home/parisse/CEdev/lib/libload/usbdrvce.lib"' -i 'provide __app_name = "DEMO"' -i 'provide __app_version = "1"' -i 'provide __app_desc = "shell"'  bin/DEMO.bin
flat assembler  version g.jmhx
/home/parisse/CEdev/meta/ld.alm [1570]:
	
macro ? [1297]:
	read source
:read? [14] (CALM)
assemble :temp [1]:
	include?! 'app_tools/app.src'
app_tools/app.src [64]:
	dl relocation
dl? [19] (CALM)
assemble :temp [1]:
	@ez80.word long?, current
macro @ez80.word [16]:
	@ez80.word size, value
@ez80.word [8] (CALM)
Error: variable term used where not expected.
make: *** [/home/parisse/CEdev/meta/makefile.mk:291 : bin/DEMO.bin] Erreur 2
/bin/cp: impossible d'évaluer 'bin/DEMO.8*': Aucun fichier ou dossier de ce type

[adriweb]

repro'd on latest toolchain, and it happens only when using app_tools.

[adriweb]

To give more context to this:

• We were able to repro a (this?) linking issue with a build of KhiCAS using the latest compiler which fixed other things.
  ... which currently (without app_tools) outputs a 8.9 MB lto.bc, and 4.7 MB .bin. :

/Users/adriweb/CEdev/meta/ld.alm [1568] macro ? [1295] :read? [14] app_tools/app.src [64]:
    dl relocation
dl? [19] macro @ez80.word [16] @ez80.word [1] (CALM)
Error: variable term used where not expected.
make: *** [bin/DEMO.bin] Error 2

• This issue doesn't happen with a sample code that has a very large dummy array (referenced somewhere so it doesn't get optimized out) like :
  static const uint8_t filler[2500000] = {0};, so I guess the issue is only happening when there's actual code all over the place.
  (For the record, I generated and successfully tested (on CEmu anyway) a 2.9 MB app and it ran fine, so the OS doesn't have a hardcoded limitation somewhere, or at least it's higher)

[commandblockguy]

Poking around a bit, the first symbol that's causing this issue is __GLOBAL__sub_I_first.cc, so I'm guessing this is an issue with sections related to static object initialization/destruction. They're kinda a weird edge case and the toolchain's linker script handles them differently than other sections, and I haven't actually tested them yet. I'll look into how to make that work.

[commandblockguy]

It seems like it would take some changes to the crt to get .init_array working properly. I'm inclined to just write this off as a "known limitation" for the time being, considering how rare usage of .init_array seems to be. For the code provided, it should be fairly easy to work around this considering there's only a single .init_array item, and all it does is register a no-op destructor for init_gmp_memory_instance, by just moving the init_gmp_memory_instance inside of the #ifdef HAVE_LIBGC block.

[parisseb]

Thanks, I modified first.cc to avoid this linking error (with #ifndef USE_GMP_REPLACEMENTS since GMP is not used). However I still get thousends of errors at link time

flat assembler  version g.k4v8
/home/parisse/CEdev/meta/ld.alm [1568]:
	
macro ? [1583]:
	output.hex section.base?
:output.hex [7] (CALM)
Error: cannot subdivide variable term.
/home/parisse/CEdev/meta/ld.alm [1568]:
	
macro ? [1583]:
	output.hex section.base?
:output.hex [7] (CALM)
Error: variable term used where not expected.
/home/parisse/CEdev/meta/ld.alm [1568]:
	
...

[commandblockguy]

That's strange, I'm able to compile it successfully. Do you have the debug map enabled in the makefile? If so, try turning that off and see if it still gives the same error.

[parisseb]

I'm unsure, but it seems not, I see OUTPUT_MAP = NO in app_tools/makefile. If you have a successfull compile, maybe the easiest is that you send me a working archive.

[adriweb]

I'm also able to build it successfully now (I'm running the latest toolchain and app_tools, maybe that's why?).
(the linking step does take a while, 7min 07s here, so be warned)

However it's a bit too big for now: 3 857 172 bytes .8ek file, so it's not testable as-is. It would need to be 850k lighter... at least.

[parisseb]

I have updated my linux version to the latest nightly update (https://github.com/CE-Programming/toolchain/releases), also updated app.src in app_tools, I still get all these errors. I'm using -Oz (not -Os). Can you describe the precise OS/version you are using?

[adriweb]

I've built it on my mac running the latest macOS but that doesn't matter anyway. I used -Oz too.
I realize I have a slightly older fasmg ("version g.jya9") but I don't think this matters either.

Did you make clean and make app -j ?

[parisseb]

Still no luck.

ls -l ~/CEdev/bin/
total 226268
-rwxr-xr-x 1 parisse parisse   979152 14 déc.  22:29 cedev-config
-rwxr-xr-x 1 parisse parisse  6081048 14 déc.  22:28 cemu-autotester
-rwxr-xr-x 1 parisse parisse    40152 14 déc.  22:29 convbin
-rwxr-xr-x 1 parisse parisse    62848 14 déc.  22:29 convfont
-rwxr-xr-x 1 parisse parisse   417920 14 déc.  22:29 convimg
-rwxr-xr-x 1 parisse parisse 78772240 13 déc.  20:01 ez80-clang
-rwxr-xr-x 1 parisse parisse 78763696 29 oct.  14:43 ez80-clang.old
-rwxr-xr-x 1 parisse parisse  4376976 13 déc.  18:54 ez80-link
-rwxr-xr-x 1 parisse parisse 26193976 15 févr.  2023 ez80-lto
-rwxr-xr-x 1 parisse parisse 35830104 15 févr.  2023 ez80-lto2
-rwxr-xr-x 1 parisse parisse    70710 14 déc.  22:28 fasmg

ez80-lto seems outdated, where can I update it?

[adriweb]

ez80-lto/lto2 aren't actually needed or used.

[parisseb]

I also tried on windows, same error.

[commandblockguy]

Here's a tarball of my working compile: khicas.tar.gz
I'm using the latest toolchain version.

[parisseb]

Thanks, it now works! The issue was in the makefile where #include $(shell cedev-config --makefile) was not commented.
I will now try to shrink the code size, by removing the plotting/geometry commands, since the output is too large for emulator testing. Once this is done, there will still remain some work to support loading large app on the real calculator...