make: Use ld -r for intermediate linking of each module #8711

jnohlgard · 2018-03-01T13:26:53Z

Contribution description

As discussed in #5757, using ld -r --unique and creating intermediate object files can be preferable to creating archives (using ar rcs, current practice). This eliminates hidden issues such as a weak symbol may override a strong symbol even though both are included in the final link.
This means that each module will first be linked into a single object file module.o instead of a static archive module.a.

New build system environment variables are introduced:

LD: target low level linker, default $(PREFIX)ld, must be ld, gcc does not work here.
ILINK: command used for intermediate linking (incremental linking), default $(LD).
ILINKFLAGS: flags for incremental linking, default -r. Certain platforms may need to specify the CPU type or other flags here, e.g. AVR, MIPS

This PR eliminates the variable UNDEF entirely because it is no longer needed.

Issues/PRs references

Initial suggestion in #5757

jnohlgard · 2018-03-01T13:27:46Z

TODO:

verify that binaries are still correct.
Check if any pkgs need updating
Check if the special case for MacOS native builds is still necessary in Makefile.base

cladmi · 2018-03-01T15:09:59Z

In #5757 PR, @kaspar030 needed to handle certain libraries without the '--whole-archive' and so with the lazy linking algorithm.

Even if some/(all?) of the packages could be fixed to produce an object file, there may be reasons for an external libraries or a module to need to be built as an archive, or also we could be lazy to migrate external build scripts all at once.
So starting with a mechanism that allows some code to be compiled and used as library could help migrate in a more smoother way. It could be adapted from the EXTRAL_LIBS of #5757 PR.

Also, if moving away from 'libraries' archives, would it make sense to rename BASELIBS?

I found some references to the '.a' output in bindist:

examples/bindist/Makefile:# and that the resulting .a should be saved when doing "make bindist"
examples/bindist/README.md:compiled and linked binary, bindist.a, abc.a and Makefiles.
makefiles/bindist.inc.mk:DIST_FILES += $(BIN_USEMODULE:%=bin/$(BOARD)/%.a)

And the bindist example does not build for the moment because it does not find main.

Kijewski · 2018-03-01T16:12:42Z

Makefile.base

+# An empty object file is included in the link to avoid situations where modules
+# do not provide any code (e.g. sys, some boards), causing the build to fail
+# with a missing object file.
+EMPTYOBJ = $(BINDIR)/$(MODULE)/.empty.o


I am not sure that using a generic empty.o will work. If I'm not mistaken (I did not test it), the linker will reject an object file if it's not the right architecture.

I guess something like that would be more robust:

ifeq $(,$(OBJ)) OBJ := $(BINDIR)/$(MODULE)/empty.o $(BINDIR)/$(MODULE)/empty.o: $(RIOTBASE)/empty.c $(CC) $(CFLAGS) -w -c -o $@ $< ifeq

static void __attribute__((unused)) _placeholder(void) {}

It works, try it. It is the correct architecture because of using the real CFLAGS when generating it.
Not sure the ifeq suggestion is robust because of lazy evaluation of recursive variables, does OBJ always have its final value when interpreting this makefile?

kaspar030 · 2018-03-14T16:33:35Z

I'm very looking forward to this one!

jnohlgard · 2018-03-14T17:39:05Z

Feel free to join in if you feel like working on this. Currently some pkgs have problems linking because of multiple inclusions of the same object file in the final link

cladmi · 2018-03-14T18:58:16Z

I really like the concept of packages registering themselves as building an .a but I do not really like the name USEARCHIVE. For me it does not match what we have currently in USEMODULE USEPKG used in dependencies. Its more something like MODULES_PRODUCING_ARCHIVES_INSTEAD_OF_OBJECTS but should have a better name than this.

I will create sub pull requests to fix what is currently broken in master.

jnohlgard · 2018-03-22T13:52:25Z

rebased to fix conflicts

jnohlgard · 2018-03-22T13:53:45Z

@cladmi I take any suggestions for a name better than USEARCHIVE. I think it's a bit too cryptic right now.

kaspar030 · 2018-03-22T14:02:39Z

I take any suggestions for a name better than USEARCHIVE. I think it's a bit too cryptic right now.

Traditionally statically linked libraries are linked as .a. Thus I'd suggest LINKLIBS or just LIBS.

cladmi · 2018-04-13T12:46:42Z

I would like to have a precise name, even maybe a strange name so that it cannot be understood directly as something else, and I would prefer that people look it up in the documentation as it should be an exception when not possible to do otherwise.

LIBSMODULES, ARCHIVEMODULES, ARCHIVELIBS ?

As it's meant to be used in modules.inc.mk when setting BASELIBS, with the following I would understand it

_OBJECTMODULES = $(filter-out $(ARCHIVEMODULES), $(REALMODULES))
BASELIBS += $(_OBJECTMODULES:%=$(BINDIR)/%.o)
BASELIBS += $(ARCHIVEMODULES:%=$(BINDIR)/%.a)

kaspar030 · 2018-04-13T12:50:15Z

As it's meant to be used in modules.inc.mk when setting BASELIBS

The name BASELIBS is not a set-in-stone one either, more something from the past that can be re-thought when oit makes sense.

makefiles/modules.inc.mk must be cleaned. There is no need for special 'PKGS' handling. Just find all real modules and define both Add variables names for '$(sort $(USEMODULE) $(USEPKG))' and build from this. Or just add "USEPKG" to USEMODULE at that moment. Should be removed later anyway. REALLIBS and something to describe modules that are not 'LIBS' This needs proper names and refactoring. I will do it outside.

This will allow special handling as a library.

… pkg/oonf_api: Update build to use ld -r for submodules Also using '$(BINDIR)/oonf_*.a as a dependency' is wrong… This commit should be this at the end with a commit with prerequisites to fix issues. pkg/oonf_api: define 'oonf_api' as LIBS This will allow special handling as a library.

It is only needed to overwrite it if `LIBS` is exported in the build system. But resetting it anyway does not hurt. Could be removed if `LIBS` is not exported.

Add an exception to not depend on periph_uart when using MIPS UHI syscalls for newlib syscalls.

Do not make 'oonf_api' group all oonf_* objects. I find it strange to merge modules there.

kaspar030 · 2019-08-29T11:10:05Z

Seems like "ld -r" is incompatible with LTO: https://lwn.net/Articles/744507/

jcarrano · 2019-08-29T11:44:00Z

Seems like "ld -r" is incompatible with LTO:

If that's the case, we may have to go with --whole-archive, which is the route I prefer, as it is less invasive.

stale · 2020-03-01T12:00:32Z

This issue has been automatically marked as stale because it has not had recent activity. It will be closed if no further activity occurs. If you want me to ignore this issue, please mark it with the "State: don't stale" label. Thank you for your contributions.

jnohlgard assigned cladmi Mar 1, 2018

jnohlgard requested review from cladmi, aabadie and kaspar030 March 1, 2018 13:26

jnohlgard force-pushed the pr/ld-r branch from fdab390 to 2e0d00d Compare March 1, 2018 13:30

jnohlgard mentioned this pull request Mar 1, 2018

make: fix linking #5757

Closed

Kijewski reviewed Mar 1, 2018

View reviewed changes

This was referenced Mar 1, 2018

cpu/efm32/periph/timer: remove timer_set duplicate #8718

Merged

cpu/mips32r2_common: use periph_common timer_set #8720

Closed

jnohlgard force-pushed the pr/ld-r branch 3 times, most recently from 472383d to e5a7376 Compare March 13, 2018 13:38

This was referenced Mar 22, 2018

mips32r2_common: timer fixes #8817

Merged

bindist: remove APPLICATION_MODULE duplicate #8818

Merged

jnohlgard force-pushed the pr/ld-r branch from fc047fc to 8631915 Compare March 22, 2018 13:49

kaspar030 mentioned this pull request May 9, 2018

core: introduce crossfile arrays (xfa), refactored #9105

Closed

Joakim Nohlgård and others added 16 commits March 5, 2019 15:13

pkg/nordic_softdevice_ble: define 'ble_6lowpan' as LIBS

c0c9288

This will allow special handling as a library.

pkg/jerryscript: define jerryscript modules as LIBS

ddb5d87

This will allow special handling as a library.

pkg/ccn-lite: define 'ccn-lite' as LIBS

2ee9d0b

This will allow special handling as a library.

pkg/relic: define 'relic' module as LIBS

4eaf7cd

This will allow special handling as a library.

pkg/openthread: define openthread modules as LIBS

b4afd66

This will allow special handling as a library.

pkg/openthread: do not pass RIOT 'LIBS' to openthread build

43f9a43

It is only needed to overwrite it if `LIBS` is exported in the build system. But resetting it anyway does not hurt. Could be removed if `LIBS` is not exported.

make: Use incremental link for modules

95e8559

make: Remove UNDEF

b61ecae

squash introduce ILINK, ILINKFLAGS

b305094

squash mips fix

abb1374

squash MIPS newlib UHI fix

79c666f

Makefile.dep: Special case newlib_syscalls_mips_uhi

cc4e6dd

Add an exception to not depend on periph_uart when using MIPS UHI syscalls for newlib syscalls.

squash riscv fix

1886161

SQUASHME pkg/oonf_api: Update build to use ld -r for submodules

976b014

Do not make 'oonf_api' group all oonf_* objects. I find it strange to merge modules there.

cladmi force-pushed the pr/ld-r branch from ca2edb6 to 976b014 Compare March 5, 2019 14:26

This was referenced Mar 5, 2019

makefiles/modules.inc.mk: refactoring #11109

Merged

makefiles: introduce 'LIBS' variable #11111

Closed

miri64 modified the milestones: Release 2019.04, Release 2019.07 May 7, 2019

MrKevinWeiss modified the milestones: Release 2019.07, Release 2019.10 Jul 16, 2019

stale bot added the State: stale State: The issue / PR has no activity for >185 days label Mar 1, 2020

stale bot closed this Apr 1, 2020

leandrolanzieri mentioned this pull request Aug 13, 2020

build system: link object files #14754

Merged

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

make: Use ld -r for intermediate linking of each module #8711

make: Use ld -r for intermediate linking of each module #8711

jnohlgard commented Mar 1, 2018 •

edited

Loading

jnohlgard commented Mar 1, 2018 •

edited

Loading

cladmi commented Mar 1, 2018

Kijewski Mar 1, 2018 •

edited

Loading

jnohlgard Mar 1, 2018

kaspar030 commented Mar 14, 2018

jnohlgard commented Mar 14, 2018

cladmi commented Mar 14, 2018

jnohlgard commented Mar 22, 2018

jnohlgard commented Mar 22, 2018

kaspar030 commented Mar 22, 2018

cladmi commented Apr 13, 2018

kaspar030 commented Apr 13, 2018

kaspar030 commented Aug 29, 2019

jcarrano commented Aug 29, 2019

stale bot commented Mar 1, 2020

make: Use ld -r for intermediate linking of each module #8711

make: Use ld -r for intermediate linking of each module #8711

Conversation

jnohlgard commented Mar 1, 2018 • edited Loading

Contribution description

Issues/PRs references

jnohlgard commented Mar 1, 2018 • edited Loading

cladmi commented Mar 1, 2018

Kijewski Mar 1, 2018 • edited Loading

Choose a reason for hiding this comment

jnohlgard Mar 1, 2018

Choose a reason for hiding this comment

kaspar030 commented Mar 14, 2018

jnohlgard commented Mar 14, 2018

cladmi commented Mar 14, 2018

jnohlgard commented Mar 22, 2018

jnohlgard commented Mar 22, 2018

kaspar030 commented Mar 22, 2018

cladmi commented Apr 13, 2018

kaspar030 commented Apr 13, 2018

kaspar030 commented Aug 29, 2019

jcarrano commented Aug 29, 2019

stale bot commented Mar 1, 2020

jnohlgard commented Mar 1, 2018 •

edited

Loading

jnohlgard commented Mar 1, 2018 •

edited

Loading

Kijewski Mar 1, 2018 •

edited

Loading