llvm-3.3 can't link on Leopard/PPC

[khepler]

I've tried all permutations I can think of on my G4 PowerBook, can't get past:

llvm[1]: Linking Release+Debug+Asserts Shared Library libLLVM-3.3.dylib
ccache /usr/bin/g++-4.2 -I/opt/local/var/macports/build/_opt_local_var_macports_sources_rsync.macports.org_release_tarballs_ports_lang_llvm-3.3/llvm-3.3/work/llvm-3.3.src/include -I/opt/local/var/macports/build/_opt_local_var_macports_sources_rsync.macports.org_release_tarballs_ports_lang_llvm-3.3/llvm-3.3/work/llvm-3.3.src/tools/llvm-shlib  -D_DEBUG -D_GNU_SOURCE -D__STDC_CONSTANT_MACROS -D__STDC_FORMAT_MACROS -D__STDC_LIMIT_MACROS -Os  -g -fvisibility-inlines-hidden -fno-exceptions -fno-common -Woverloaded-virtual -Wcast-qual -pipe -O2 -arch ppc -pipe -O2 -arch ppc -Os -g -Wl,-rpath -Wl,@executable_path/../lib -L/opt/local/var/macports/build/_opt_local_var_macports_sources_rsync.macports.org_release_tarballs_ports_lang_llvm-3.3/llvm-3.3/work/llvm-3.3.src/Release+Debug+Asserts/lib -L/opt/local/var/macports/build/_opt_local_var_macports_sources_rsync.macports.org_release_tarballs_ports_lang_llvm-3.3/llvm-3.3/work/llvm-3.3.src/Release+Debug+Asserts/lib -arch ppc -arch ppc -mmacosx-version-min=10.5  -pedantic -Wno-long-long -Wall -W -Wno-unused-parameter -Wwrite-strings     -Wno-missing-field-initializers  -dynamiclib -mmacosx-version-min=10.5 -o /opt/local/var/macports/build/_opt_local_var_macports_sources_rsync.macports.org_release_tarballs_ports_lang_llvm-3.3/llvm-3.3/work/llvm-3.3.src/Release+Debug+Asserts/lib/libLLVM-3.3.dylib  \
	   -lLLVMAArch64AsmParser -lLLVMAArch64AsmPrinter -lLLVMAArch64CodeGen -lLLVMAArch64Desc -lLLVMAArch64Disassembler -lLLVMAArch64Info -lLLVMAArch64Utils -lLLVMARMAsmParser -lLLVMARMAsmPrinter -lLLVMARMCodeGen -lLLVMARMDesc -lLLVMARMDisassembler -lLLVMARMInfo -lLLVMAnalysis -lLLVMArchive -lLLVMAsmParser -lLLVMAsmPrinter -lLLVMBitReader -lLLVMBitWriter -lLLVMCodeGen -lLLVMCore -lLLVMCppBackendCodeGen -lLLVMCppBackendInfo -lLLVMDebugInfo -lLLVMExecutionEngine -lLLVMHexagonAsmPrinter -lLLVMHexagonCodeGen -lLLVMHexagonDesc -lLLVMHexagonInfo -lLLVMIRReader -lLLVMInstCombine -lLLVMInstrumentation -lLLVMInterpreter -lLLVMJIT -lLLVMLinker -lLLVMMBlazeAsmParser -lLLVMMBlazeAsmPrinter -lLLVMMBlazeCodeGen -lLLVMMBlazeDesc -lLLVMMBlazeDisassembler -lLLVMMBlazeInfo -lLLVMMC -lLLVMMCDisassembler -lLLVMMCJIT -lLLVMMCParser -lLLVMMSP430AsmPrinter -lLLVMMSP430CodeGen -lLLVMMSP430Desc -lLLVMMSP430Info -lLLVMMipsAsmParser -lLLVMMipsAsmPrinter -lLLVMMipsCodeGen -lLLVMMipsDesc -lLLVMMipsDisassembler -lLLVMMipsInfo -lLLVMNVPTXAsmPrinter -lLLVMNVPTXCodeGen -lLLVMNVPTXDesc -lLLVMNVPTXInfo -lLLVMObjCARCOpts -lLLVMObject -lLLVMOption -lLLVMPowerPCAsmParser -lLLVMPowerPCAsmPrinter -lLLVMPowerPCCodeGen -lLLVMPowerPCDesc -lLLVMPowerPCInfo -lLLVMRuntimeDyld -lLLVMScalarOpts -lLLVMSelectionDAG -lLLVMSparcCodeGen -lLLVMSparcDesc -lLLVMSparcInfo -lLLVMSupport -lLLVMSystemZAsmParser -lLLVMSystemZAsmPrinter -lLLVMSystemZCodeGen -lLLVMSystemZDesc -lLLVMSystemZInfo -lLLVMTarget -lLLVMTransformUtils -lLLVMVectorize -lLLVMX86AsmParser -lLLVMX86AsmPrinter -lLLVMX86CodeGen -lLLVMX86Desc -lLLVMX86Disassembler -lLLVMX86Info -lLLVMX86Utils -lLLVMXCoreAsmPrinter -lLLVMXCoreCodeGen -lLLVMXCoreDesc -lLLVMXCoreDisassembler -lLLVMXCoreInfo -lLLVMipa -lLLVMipo -all_load -Wl,-dead_strip -Wl,-seg1addr -Wl,0xE0000000  -Wl,-install_name -Wl,"@executable_path/../lib/libLLVM-3.3.dylib" -lz -lpthread -lffi -lm 
ld: bl out of range (-16932856 max is +/-16M) from llvm::EmitStrLen(llvm::Value*, llvm::IRBuilder<true, llvm::ConstantFolder, llvm::IRBuilderDefaultInserter<true> >&, llvm::DataLayout const*, llvm::TargetLibraryInfo const*)at 0xE103A560 in __text of /opt/local/var/macports/build/_opt_local_var_macports_sources_rsync.macports.org_release_tarballs_ports_lang_llvm-3.3/llvm-3.3/work/llvm-3.3.src/Release+Debug+Asserts/lib/libLLVMTransformUtils.a(BuildLibCalls.o) to llvm::Twine::Twine(char const*)at 0xE00148F0 in __text of  /opt/local/var/macports/build/_opt_local_var_macports_sources_rsync.macports.org_release_tarballs_ports_lang_llvm-3.3/llvm-3.3/work/llvm-3.3.src/Release+Debug+Asserts/lib/libLLVMAArch64AsmParser.a(AArch64AsmParser.o) in llvm::EmitStrLen(llvm::Value*, llvm::IRBuilder<true, llvm::ConstantFolder, llvm::IRBuilderDefaultInserter<true> >&, llvm::DataLayout const*, llvm::TargetLibraryInfo const*)from /opt/local/var/macports/build/_opt_local_var_macports_sources_rsync.macports.org_release_tarballs_ports_lang_llvm-3.3/llvm-3.3/work/llvm-3.3.src/Release+Debug+Asserts/lib/libLLVMTransformUtils.a(BuildLibCalls.o)
collect2: ld returned 1 exit status

Seems similar to #33987, but the patch is already applied to the Portfile in tree. Perhaps llvm 3.3 has simply grown too large for relative jumps on ppc32? The "Linking Release+Debug+Asserts Shared Library" bit seems suspect, as I'm not requesting assertions variant, and I don't expect a debug version.

The attached log file is under the default compiler, in this case apple-gcc-4.2 from Macports. Same result with many combinations of the following toolchain components. Heck, I even tried the stock Apple ld for laughs.

The following ports are currently installed:
  apple-gcc42 @5666.3_11
  apple-gcc42 @5666.3_11+gpl3 (active)
  cctools @806_2+llvm31
  cctools @806_2+llvm32 (active)
  clang-3.2 @3.2_1+analyzer+assertions+python27 (active)
  gcc44 @4.4.7_6 (active)
  gcc45 @4.5.4_7 (active)
  ld64 @97.17_2+llvm31
  ld64 @97.17_2+llvm32
  llvm-3.1 @3.1_7 (active)
  llvm-3.2 @3.2_1+assertions (active)

[khepler]

build log

[mf2k (Frank Schima)]

In the future, please Cc the port maintainer(s).

[jeremyhu (Jeremy Huddleston Sequoia)]

Yep ... we were bound to crash into that ceiling eventually... there's only so much -Os can do.

[jeremyhu (Jeremy Huddleston Sequoia)]

You may be able to squeeze out some more room if you turn off assertions:

sudo port -v -f uninstall llvm-3.3
sudo port -v -s install llvm-3.3 -assertions

[jeremyhu (Jeremy Huddleston Sequoia)]

Limiting our targets will also help...

[khepler]

As I understood it, the assertions variant was disabled by default. Guess I can try it though.

Thoughts on building with -mlongcall on ppc32? Seems like a likely solution for GCC; I don't know an equivalent for clang/llvm.

[jeremyhu (Jeremy Huddleston Sequoia)]

I'd rather just cut out half the targets, but please feel free to try that as well.

[udbraumann]

Replying to jeremyhu@…:

You may be able to squeeze out some more room if you turn off assertions:

To explicitly turning off assertions has also led to the same linker error as was reported initially on my G4 Powerbook running 10.5.8 (I made a clean install, of course). Ironically, a few days before clang-3.3 seemingly has become mandatory to build boost which in turn is required for inkscape. It would be really sad if llvm-3.3 could not be build to due to exceeding 32bit limitations. - Is anybody aware how to apply the -mlongcall switch as was proposed above? And what is meant by cutting out targets?

[udbraumann]

Cc Me!

[jeremyhu (Jeremy Huddleston Sequoia)]

Replying to braumann@…:

Is anybody aware how to apply the -mlongcall switch as was proposed above? And what is meant by cutting out targets?

Yes, add -mlongcall along with -Os (search for -Os in the Portfile).

By removing targets, I mean disabling support for other targets that you probably don't need (eg: sparc, mips, etc). My guess is that the addition of AArch64 in llvm-3.3 is what put this over the top.

[cooljeanius (Eric Gallager)]

Cc Me!

[udbraumann]

Replying to jeremyhu@…:

Yes, add -mlongcall along with -Os (search for -Os in the Portfile).

I have added this to the Portfile

	configure.args-append --with-extra-options=-mlongcall

just next to the existing entry

        # http://trac.macports.org/ticket/33987
        configure.args-append --with-optimize-option=-Os

However, it basically ends up in the same final error

ld: bl out of range (-16991880 max is +/-16M)

Without -mlongcall the overflow message twice was

ld: bl out of range (-16936552 max is +/-16M)

So, there was some effect, but seemingly too little. What is new is a series complaints which might indicate that the -mlongcall idea is not really appropriate:

PPC_RELOC_JBSR should not be using an external relocation

This is my full error output concerning the linking phase, if you want to see more details:

llvm[1]: Linking Release+Debug+Asserts Shared Library libLLVM-3.3.dylib
/usr/bin/g++-4.2 -I/opt/local/var/macports/build/_opt_local_var_macports_sources_rsync.macports.org_release_ports_lang_llvm-3.3/llvm-3.3/work/llvm-3.3.src/include -I/opt/local/var/macports/build/_opt_local_var_macports_sources_rsync.macports.org_release_ports_lang_llvm-3.3/llvm-3.3/work/llvm-3.3.src/tools/llvm-shlib  -D_DEBUG -D_GNU_SOURCE -D__STDC_CONSTANT_MACROS -D__STDC_FORMAT_MACROS -D__STDC_LIMIT_MACROS -Os  -g -fvisibility-inlines-hidden -fno-exceptions -fno-common -Woverloaded-virtual -Wcast-qual -O2 -arch ppc -O2 -arch ppc -Os -g -Wl,-rpath -Wl,@executable_path/../lib -L/opt/local/var/macports/build/_opt_local_var_macports_sources_rsync.macports.org_release_ports_lang_llvm-3.3/llvm-3.3/work/llvm-3.3.src/Release+Debug+Asserts/lib -L/opt/local/var/macports/build/_opt_local_var_macports_sources_rsync.macports.org_release_ports_lang_llvm-3.3/llvm-3.3/work/llvm-3.3.src/Release+Debug+Asserts/lib -arch ppc -arch ppc -mmacosx-version-min=10.5  -pedantic -Wno-long-long -Wall -W -Wno-unused-parameter -Wwrite-strings -mlongcall    -Wno-missing-field-initializers  -dynamiclib -mmacosx-version-min=10.5 -o /opt/local/var/macports/build/_opt_local_var_macports_sources_rsync.macports.org_release_ports_lang_llvm-3.3/llvm-3.3/work/llvm-3.3.src/Release+Debug+Asserts/lib/libLLVM-3.3.dylib  \
	   -lLLVMAArch64AsmParser -lLLVMAArch64AsmPrinter -lLLVMAArch64CodeGen -lLLVMAArch64Desc -lLLVMAArch64Disassembler -lLLVMAArch64Info -lLLVMAArch64Utils -lLLVMARMAsmParser -lLLVMARMAsmPrinter -lLLVMARMCodeGen -lLLVMARMDesc -lLLVMARMDisassembler -lLLVMARMInfo -lLLVMAnalysis -lLLVMArchive -lLLVMAsmParser -lLLVMAsmPrinter -lLLVMBitReader -lLLVMBitWriter -lLLVMCodeGen -lLLVMCore -lLLVMCppBackendCodeGen -lLLVMCppBackendInfo -lLLVMDebugInfo -lLLVMExecutionEngine -lLLVMHexagonAsmPrinter -lLLVMHexagonCodeGen -lLLVMHexagonDesc -lLLVMHexagonInfo -lLLVMIRReader -lLLVMInstCombine -lLLVMInstrumentation -lLLVMInterpreter -lLLVMJIT -lLLVMLinker -lLLVMMBlazeAsmParser -lLLVMMBlazeAsmPrinter -lLLVMMBlazeCodeGen -lLLVMMBlazeDesc -lLLVMMBlazeDisassembler -lLLVMMBlazeInfo -lLLVMMC -lLLVMMCDisassembler -lLLVMMCJIT -lLLVMMCParser -lLLVMMSP430AsmPrinter -lLLVMMSP430CodeGen -lLLVMMSP430Desc -lLLVMMSP430Info -lLLVMMipsAsmParser -lLLVMMipsAsmPrinter -lLLVMMipsCodeGen -lLLVMMipsDesc -lLLVMMipsDisassembler -lLLVMMipsInfo -lLLVMNVPTXAsmPrinter -lLLVMNVPTXCodeGen -lLLVMNVPTXDesc -lLLVMNVPTXInfo -lLLVMObjCARCOpts -lLLVMObject -lLLVMOption -lLLVMPowerPCAsmParser -lLLVMPowerPCAsmPrinter -lLLVMPowerPCCodeGen -lLLVMPowerPCDesc -lLLVMPowerPCInfo -lLLVMRuntimeDyld -lLLVMScalarOpts -lLLVMSelectionDAG -lLLVMSparcCodeGen -lLLVMSparcDesc -lLLVMSparcInfo -lLLVMSupport -lLLVMSystemZAsmParser -lLLVMSystemZAsmPrinter -lLLVMSystemZCodeGen -lLLVMSystemZDesc -lLLVMSystemZInfo -lLLVMTarget -lLLVMTransformUtils -lLLVMVectorize -lLLVMX86AsmParser -lLLVMX86AsmPrinter -lLLVMX86CodeGen -lLLVMX86Desc -lLLVMX86Disassembler -lLLVMX86Info -lLLVMX86Utils -lLLVMXCoreAsmPrinter -lLLVMXCoreCodeGen -lLLVMXCoreDesc -lLLVMXCoreDisassembler -lLLVMXCoreInfo -lLLVMipa -lLLVMipo -all_load -Wl,-dead_strip -Wl,-seg1addr -Wl,0xE0000000  -Wl,-install_name -Wl,"@executable_path/../lib/libLLVM-3.3.dylib" -lz -lpthread -lffi -lm 
PPC_RELOC_JBSR should not be using an external relocationPPC_RELOC_JBSR should not be using an external relocationPPC_RELOC_JBSR should not be using an external relocationPPC_RELOC_JBSR should not be using an external relocationPPC_RELOC_JBSR should not be using an external relocationPPC_RELOC_JBSR should not be using an external relocationPPC_RELOC_JBSR should not be using an external relocationPPC_RELOC_JBSR should not be using an external relocationPPC_RELOC_JBSR should not be using an external relocationPPC_RELOC_JBSR should not be using an external relocationPPC_RELOC_JBSR should not be using an external relocationPPC_RELOC_JBSR should not be using an external relocationPPC_RELOC_JBSR should not be using an external relocationPPC_RELOC_JBSR should not be using an external relocationPPC_RELOC_JBSR should not be using an external relocationPPC_RELOC_JBSR should not be using an external relocationPPC_RELOC_JBSR should not be using an external relocationPPC_RELOC_JBSR should not be using an external relocationPPC_RELOC_JBSR should not be using an external relocationPPC_RELOC_JBSR should not be using an external relocationPPC_RELOC_JBSR should not be using an external relocationPPC_RELOC_JBSR should not be using an external relocationPPC_RELOC_JBSR should not be using an external relocationPPC_RELOC_JBSR should not be using an external relocationPPC_RELOC_JBSR should not be using an external relocationPPC_RELOC_JBSR should not be using an external relocationPPC_RELOC_JBSR should not be using an external relocationPPC_RELOC_JBSR should not be using an external relocationPPC_RELOC_JBSR should not be using an external relocationPPC_RELOC_JBSR should not be using an external relocationPPC_RELOC_JBSR should not be using an external relocationPPC_RELOC_JBSR should not be using an external relocationPPC_RELOC_JBSR should not be using an external relocationPPC_RELOC_JBSR should not be using an external relocationPPC_RELOC_JBSR should not be using an external relocationPPC_RELOC_JBSR should not be using an external relocationPPC_RELOC_JBSR should not be using an external relocationPPC_RELOC_JBSR should not be using an external relocationPPC_RELOC_JBSR should not be using an external relocationPPC_RELOC_JBSR should not be using an external relocationPPC_RELOC_JBSR should not be using an external relocationPPC_RELOC_JBSR should not be using an external relocationld: bl out of range (-16991880 max is +/-16M) from llvm::EmitStrLen(llvm::Value*, llvm::IRBuilder<true, llvm::ConstantFolder, llvm::IRBuilderDefaultInserter<true> >&, llvm::DataLayout const*, llvm::TargetLibraryInfo const*)at 0xE1048A00 in __text of /opt/local/var/macports/build/_opt_local_var_macports_sources_rsync.macports.org_release_ports_lang_llvm-3.3/llvm-3.3/work/llvm-3.3.src/Release+Debug+Asserts/lib/libLLVMTransformUtils.a(BuildLibCalls.o) to llvm::Twine::Twine(char const*)at 0xE0014700 in __text of  /opt/local/var/macports/build/_opt_local_var_macports_sources_rsync.macports.org_release_ports_lang_llvm-3.3/llvm-3.3/work/llvm-3.3.src/Release+Debug+Asserts/lib/libLLVMAArch64AsmParser.a(AArch64AsmParser.o) in llvm::EmitStrLen(llvm::Value*, llvm::IRBuilder<true, llvm::ConstantFolder, llvm::IRBuilderDefaultInserter<true> >&, llvm::DataLayout const*, llvm::TargetLibraryInfo const*)from /opt/local/var/macports/build/_opt_local_var_macports_sources_rsync.macports.org_release_ports_lang_llvm-3.3/llvm-3.3/work/llvm-3.3.src/Release+Debug+Asserts/lib/libLLVMTransformUtils.a(BuildLibCalls.o)
collect2: ld returned 1 exit status

By removing targets, I mean disabling support for other targets that you probably don't need (eg: sparc, mips, etc). My guess is that the addition of AArch64 in llvm-3.3 is what put this over the top.

Well, how is the best way to let people deselect unwanted architectures (targets) like Sun Sparc etc? From the Portfile level, i.e. as soon as a 32 bit PPC is identified, ancient Silicon Graphics and Sun get dropped?

Anyway, I do hope that we manage to get llvm-3.3 to be built again on 32bit PPC, thanks for any help!

[jeremyhu (Jeremy Huddleston Sequoia)]

I think the easiest way at this point is just to force them off if ppc is a build arch. For example, something like this:

Index: Portfile
===================================================================
--- Portfile	(revision 108504)
+++ Portfile	(working copy)
@@ -105,6 +105,13 @@
     select.file         ${filespath}/mp-${subport}
 }
 
+# https://trac.macports.org/ticket/39761
+# If building to run on ppc, only build "Apple" targets to conserve space
+if {${build_arch} == "ppc" || ([variant_isset universal] && [lsearch ${universal_archs} ppc] != -1)} {
+    configure.args-append \
+        --enable-targets="x86 x86_64 powerpc arm"
+}
+
 # g++-4.0 fails to build some of the newer C++ for ppc
 # Intel looks ok, but I prefer using gcc-4.2 for consistency
 compiler.blacklist gcc-4.0

[udbraumann]

after applying the patch to build only "Apple" targets

[udbraumann]

Replying to jeremyhu@…:

I think the easiest way at this point is just to force them off if ppc is a build arch.

Unfortunately, after applying your patch I basically get the same linker error:

ld: bl out of range (-16777524 max is +/-16M)

See the full build log I have uploaded before as main.2.log.

I have no further idea what is going wrong. I mean, a considerable portion of the code is skipped now, isn't it? Should I again add the -mlongcall option?

Under ​http://www.paraview.org/pipermail/paraview/2008-March/007359.html I have found this dialogue:

What you want to avoid is the *error* "ld: bl out of range".
Quoting from a message on this list 10 days ago:

> The problem was fixed by turning BUILD_SHARED_LIBS on.

Would this be a solution to turn this variable on? If yes, again from the Portfile level?

Further, in the log file I found two strange lines:

:info:configure ./configure: line 1990: -O2: command not found
:info:configure ./configure: line 1991: -O2: command not found

Maybe this is also affecting our builds?

[jeremyhu (Jeremy Huddleston Sequoia)]

The 1990/1991 lines look like a bug in their configure script. The lines are just:

${CFLAGS=}
${CXXFLAGS=}

[jeremyhu (Jeremy Huddleston Sequoia)]

I think I see the issue:

-Os  -g -fvisibility-inlines-hidden -fno-exceptions -fno-common -Woverloaded-virtual -Wcast-qual -O2

The -Os is getting overridden by -O2. I'm guessing maybe they intended those lines to unset CFLAGS and CXXFLAGS or something

[jeremyhu (Jeremy Huddleston Sequoia)]

Try doing this (possibly in addition to the "Apple" targets if it fails alone:

--- Portfile	(revision 108504)
+++ Portfile	(working copy)
@@ -173,6 +173,7 @@
     if {[string match "*ppc*" [get_canonical_archs]]} {
         # http://trac.macports.org/ticket/33987
         configure.args-append --with-optimize-option=-Os
+        configure.optflags    -Os
     }
 
     if {${subport} == "clang-${llvm_version}" && ${os.major} <= 9} {

[udbraumann]

Replying to jeremyhu@…:

Try doing this (possibly in addition to the "Apple" targets if it fails alone:

Thanks, I have applied your -Os patch in addition to the previous patch reducing the targets, and violà, this solved the linker problem and llvm-3.3 was successfully built on my PowerBook G4 running 10.5.8.

These lines 1990 and 1991 in the configure script should be removed, shouldn't they? BTW, the same strange lines occur in the clang-3.3 configure script.

Shall I try out if building works again with full targets? Well, at the moment I will continue building clang-3.3...

[jeremyhu (Jeremy Huddleston Sequoia)]

Ok, I'll push the -Os change for now which will still give full targets. Please let me know if that still fails.

I've file a bug with upstream about the configure script issue: ​http://llvm.org/bugs/show_bug.cgi?id=16724

[udbraumann]

Replying to jeremyhu@…:

Ok, I'll push the -Os change for now which will still give full targets. Please let me know if that still fails.

Thanks, llvm-3.3 builds now on ppc as it comes from the repository (with full targets), but gets huge!

[jeremyhu (Jeremy Huddleston Sequoia)]

Ok, thanks. I'm marking this fixed then.