[LLVMdev] Makefile.rules Changes / automake update (IMPORTANT!)

[Michael McCracken]

Hi Reid, just a quick note while you're doing this, a while ago I ran
into a problem that the standard makefiles weren't building libraries
(like, say, a new pass) correctly on OS X - they were being built as
shared libraries and not bundles, so they couldn't be loaded with
dlopen. The discussion is in the archives if you want more details...

I fixed it locally by doing the following hack:

% diff Makefile.rules.old Makefile.rules
326c326
< Link     := $(LIBTOOL) --mode=link $(CXX)
---
> Link     := $(LIBTOOL) --mode=link $(CXX) -module

(Note that libtool on OS X is not quite gnu libtool)

If there's a saner way to incorporate this fix, now might be the time,
since you're working on it already.

Hope this isn't piling too much work on, but it'd be nice to not have
to worry about keeping this local change.

Thanks!
-mike


On Fri, 22 Oct 2004 14:40:42 -0700, Reid Spencer <reid at x10sys.com> wrote:
> Hello,
> 
> I've closed PR106 (use automake) as WONTFIX. I've already delineated the
> problems with automake in previous posts but as of now, all the automake
> related stuff has been removed from the repository.
> 
> In an effort to start making our makefile system better, I've committed changes
> to Makefile.rules and a few library Makefiles that nearly double the speed of
> our compilations. I can now build all of LLVM (lib, utils, tools, runtime,
> examples) in just over 17 minutes on my system. This used to take over 30
> minutes. There are two main speed ups: (a) we do automatic dependency checking
> in one pass instead of two, and (b) we use "ar" to archive libraries, not
> libtool. The first change makes the build generally go faster in all
> directories. The second change speeds up links slightly because libraries are
> properly searched for symbols. I have also removed redundant variable
> definitions and sub-shell invocations (replaced with GNU Make features).
> 
> I have tested these changes quite extensively including the Stacker, sample,
> llvm-test, Java and reopt projects. If you have another project, you MIGHT be
> impacted because of a few variable name changes.
> 
> All changes will (soon) be documented in a new "LLVM Makefiles" document that
> I'll write, but quickly, here's what changed from a USER's perspective:
> 
> * ExtraSource variable became BUILT_SOURCES (standard variable name)
> * SourceDir variable eliminated (redundant with BUILD_SRC_DIR, few uses)
> * Properly allow standard variables to be used (C,CFLAGS,CXX,CXXFLAGS,CPP,
>    CPPFLAGS, LD, LDFLAGS, AR, ARFLAGS, etc) to alter compilation settings on a
>    per directory basis.
> * Build BUILT_SOURCES first (ensures *.td -> *.inc translation occurs)
> * Dependency information is now, by default, generated on every compile. This
>    keeps it constantly up to date. System headers are not tracked so as to
>    reduce make's load time. There is no separate dependency generation pass.
> * Implement auto-reconfiguration (if the configure script changes, you get
>    an automatic reconfiguration instead of a message saying that you need to
>    do so.)
> * Implement auto update of Makefile.config (if you change Makefile.config.in
>    it automatically gets regenerated to Makefile.config)
> * Added tblgen build rules to Makefile.rules and removed from the X86 and
>    PPC target Makefiles. There was some overlap. Standardizing these will
>    encourage future targets to use them. These rules are available if you set
>    the TARGET variable.
> * $(DEST*) variables are gone unless they specifically refer to the
>    installation destination. These conflicted with standard makefile usage
> * $(BUILD_OBJ_DIR)/$(CONFIGURATION) became $(OBJDIR) to simplify it. This is
>    where object files should go.
> * $(LIBDIR) is where libraries should go
> * $(TOOLDIR) is where tools (executables) should go
> * VERBOSE=1 mode is now even quieter (a few rules were missing $(VERB)
> * Build messages have been corrected and made more indicative
> * The "prdirs" target has been replaced with the "printvars" target which
>    prints values for interesting make variables, not just the directories.
> 
> Brian: I've emailed your one Makefile change that you need to use in
> WholeReoptimizer library after you take the Makefile.rules update.
> 
> Alkis: Some very minor changes were needed in your Java project, but since I
> don't have access to your cvs repository any more, I can't tell you what
> changed (my file timestamps are all the same for some reason). So, try building
> it, if it isn't obvious what's wrong, let me know.
> 
> Misha/Brian: I didn't try compiling LLVM-TV with this, but I doubt you'll run
> into any significant problems, the user changes are few
> 
> Everyone: llvm-test has successfully compiled and run ONCE with these changes.
> However, please note that llvm-test no longer depends on llvm/Makefile.rules. I
> copied the OLD Makefile.rules to llvm-test/Makefile.rules.  The ONLY remaining
> dependency between llvm-test and llvm proper is the Makefile.confg(.in) file.
> This was the only "safe/quick" way to ensure I didn't break llvm-test.  As time
> goes on we need to straighten out the Makefile mess in llvm-test but that's
> another project.
> 
> Reid.
> 
> _______________________________________________
> LLVM Developers mailing list
> LLVMdev at cs.uiuc.edu         http://llvm.cs.uiuc.edu
> http://mail.cs.uiuc.edu/mailman/listinfo/llvmdev
> 


-- 
Michael McCracken
UCSD CSE PhD Student
San Diego Supercomputer Center
http://www.cse.ucsd.edu/~mmccrack/