AArch64 build fails surprisingly close to the end (building sys0.o) #10791

pao · 2015-04-10T22:31:43Z

Nexus 9, in a Vivid Vervet (Ubuntu 15.04 beta) chroot.

    JULIA usr/lib/julia/sys0.o
exports.jl
base.jl
reflection.jl
build_h.jl
version_git.jl
julia: /home/android/julia/deps/llvm-3.5.1/lib/ExecutionEngine/RuntimeDyld/RuntimeDyldELF.cpp:384: void llvm::RuntimeDyldELF::resolveAArch64Relocation(const
llvm::SectionEntry&, uint64_t, uint64_t, uint32_t, int64_t): Assertion `static_cast<int64_t>(Result) >= (-2147483647-1) && static_cast<int64_t>(Result) <= (4
294967295U)' failed.

signal (6): Aborted
gsignal at /lib/aarch64-linux-gnu/libc.so.6 (unknown line)
Aborted
Makefile:168: recipe for target '/home/android/julia/usr/lib/julia/sys0.o' failed

Make.user

override LLVM_ASSERTIONS=1
LLVM_FLAGS+="--enable-targets=arm --enable-optimized --enable-assertions"

override OPENBLAS_DYNAMIC_ARCH=0
override OPENBLAS_TARGET_ARCH=ARMV8
override USE_BLAS64=0

override LLVM_VER=3.5.1

override USE_SYSTEM_LIBM=1
override USE_SYSTEM_FFTW=1
override USE_SYSTEM_GMP=1
override USE_SYSTEM_MPFR=1
override USE_SYSTEM_LIBUNWIND=1

override OPENBLAS_VER=develop

JCFLAGS += -fsigned-char

Needed system libunwind to pick up aarch64 patches post-1.1 release but applied to the Ubuntu repository. Needed OpenBLAS-develop to pick up this build fix.

Anyone know what LLVM is telling me we've done wrong, here?

The text was updated successfully, but these errors were encountered:

pao · 2015-04-10T23:43:04Z

Imma try this with a newer LLVM and see what happens.

pao · 2015-04-11T01:13:25Z

Same error with LLVM 3.6.0.

I missed this earlier, but:

In file included from intrinsics.cpp:50:0,
from codegen.cp:1171:
ccall.cpp:239:4: warning: #warning "ccall is defaulting to llvm ABI, since no platform ABI has been defined for this CPU/OS combination"

I also missed mentioning that you do need a small patch to Make.inc to define aarch64 as "not x86":

diff --git a/Make.inc b/Make.inc
index 7e6310a..22e9bfb 100644
--- a/Make.inc
+++ b/Make.inc
@@ -495,6 +495,8 @@ else ifneq (,$(findstring powerpc,$(ARCH)))
 ISX86=0
 else ifneq (,$(findstring ppc,$(ARCH)))
 ISX86=0
+else ifneq (,$(findstring aarch64,$(ARCH)))
+ISX86=0
 else
 $(error "unknown word-size for arch: $(ARCH)")
 endif

vtjnash · 2015-05-12T16:17:30Z

what's in /proc/cpuinfo on this machine?

pao · 2015-05-26T15:34:26Z

(taken from http://newandroidbook.com/Articles/Nexus9.html)

root@flounder:/ # cat /proc/cpuinfo
Processor   : NVIDIA Denver 1.0 rev 0 (aarch64)
processor   : 0
processor   : 1
Features    : fp asimd aes pmull sha1 sha2 crc32 
CPU implementer : 0x4e
CPU architecture: AArch64
CPU variant : 0x0
CPU part    : 0x000
CPU revision    : 0

Hardware    : Flounder
Revision    : 0000
Serial      : 0000000000000000
MTS version : 33410787

vtjnash · 2015-05-26T22:35:51Z

interesting, this seems to be the first ARM chip that has been attempted that does not have VFP. that means calling convention is probably correct.

pao · 2015-05-27T01:35:40Z

VFPv3/v4 is apparently pulled into the main instruction set, and NEON is replaced by AdvancedSIMD. I'm not sure why it isn't listed in cpuinfo, though.

Quoting http://people.linaro.org/~rikuvoipio/aarch64-talk/#/17: "VFP/SIMD mandatory - no soft float ABI"

ViralBShah · 2015-07-19T04:02:27Z

With the new updates on using the VFP2 calling convention by default, I wonder if this works now.

pao · 2015-07-19T14:11:32Z

Good question; I'll see what I can do.

ViralBShah · 2015-07-25T15:41:31Z

Bump. Closing, but reopen if still breaking.

pao · 2015-07-25T20:10:35Z

What? The close is obviously premature.

pao · 2015-07-25T20:44:44Z

OpenBLAS doesn't build anymore. Yay.

pao · 2015-07-25T20:51:30Z

Switching to system BLAS/LAPACK, looks like the same assertion:

    JULIA usr/lib/julia/inference0.ji
WARNING: unable to determine host cpu name.
essentials.jl
julia: /home/android/julia/deps/llvm-3.6.0/lib/ExecutionEngine/RuntimeDyld/RuntimeDyldELF.cpp:346: void llvm::RuntimeDyldELF::resolveAArch64Relocation(const
llvm::SectionEntry&, uint64_t, uint64_t, uint32_t, int64_t): Assertion `static_cast<int64_t>(Result) >= (-2147483647-1) && static_cast<int64_t>(Result) <= (4
294967295U)' failed.

signal (6): Aborted
gsignal at /lib/aarch64-linux-gnu/libc.so.6 (unknown line)
Aborted
Makefile:175: recipe for target '/home/android/julia/usr/lib/julia/inference0.ji' failed
make[1]: *** [/home/android/julia/usr/lib/julia/inference0.ji] Error 134
Makefile:64: recipe for target 'julia-inference' failed
make: *** [julia-inference] Error 2

Keno · 2015-07-25T22:09:53Z

Can you try with llvm-svn, this may potentially be fixed there as part of my effort to improve GOT relocations. If not, I'll need to take a look at some point.

pao · 2015-07-25T23:32:17Z

Sure thing.

pao · 2015-07-26T01:47:47Z

Do we build at all on SVN right now?

    CC src/codegen.o
    PERL base/build_h.jl.phony
In file included from intrinsics.cpp:52:0,
                 from codegen.cpp:1255:
ccall.cpp:245:4: warning: #warning "ccall is defaulting to llvm ABI, since no platform ABI has been defined for this CPU/OS combination" [-Wcpp]
 #  warning "ccall is defaulting to llvm ABI, since no platform ABI has been defined for this CPU/OS combination"
    ^
    CC src/disasm.o
In file included from codegen.cpp:516:0:
cgutils.cpp: In function ‘void jl_dump_shadow(char*, int, const char*, size_t, bool)’:
cgutils.cpp:495:30: error: ‘class llvm::TargetMachine’ has no member named ‘getDataLayout’
     clone->setDataLayout(TM->getDataLayout()->getStringRepresentation());
                              ^
disasm.cpp: In function ‘void jl_dump_asm_internal(uintptr_t, size_t, size_t, const llvm::object::ObjectFile*, llvm::raw_ostream&)’:
disasm.cpp:314:83: warning: ‘void llvm::MCObjectFileInfo::InitMCObjectFileInfo(llvm::StringRef, llvm::Reloc::Model, llvm::CodeModel::Model, llvm::MCContext&)
’ is deprecated (declared at /home/android/julia/usr/include/llvm/MC/MCObjectFileInfo.h:192) [-Wdeprecated-declarations]
     MOFI->InitMCObjectFileInfo(TripleName, Reloc::Default, CodeModel::Default, Ctx);
                                                                                   ^
codegen.cpp: In function ‘void jl_init_codegen()’:
codegen.cpp:5542:52: error: ‘class llvm::TargetMachine’ has no member named ‘getDataLayout’
     engine_module->setDataLayout(jl_TargetMachine->getDataLayout()->getStringRepresentation());
                                                    ^
Makefile:80: recipe for target 'codegen.o' failed
make[1]: *** [codegen.o] Error 1
make[1]: *** Waiting for unfinished jobs....
Makefile:58: recipe for target 'julia-src-release' failed
make: *** [julia-src-release] Error 2

Keno · 2015-07-26T01:49:55Z

Not anymore apparently. LLVM APIs changed again. Should be straightforward fix, but I'm not currently in a position to do & test it.

pao · 2015-07-26T02:09:46Z

No problem.

tkelman · 2015-07-26T04:09:39Z

You can try the release_37 branch of LLVM which should have fewer API changes on it.

ViralBShah · 2015-07-26T05:18:43Z

@pao Glad that my premature closing helped revisit the current status.

pao · 2015-07-26T14:13:47Z

A bump would have done the same.

ViralBShah · 2015-07-26T14:22:27Z

Sure. My bad.

pao · 2015-07-26T16:30:04Z

Different set of issues on release_37:

In file included from intrinsics.cpp:52:0,
                 from codegen.cpp:1255:
ccall.cpp:245:4: warning: #warning "ccall is defaulting to llvm ABI, since no platform ABI has been defined for this CPU/OS combination" [-Wcpp]
 #  warning "ccall is defaulting to llvm ABI, since no platform ABI has been defined for this CPU/OS combination"
    ^
disasm.cpp: In function ‘void jl_dump_asm_internal(uintptr_t, size_t, size_t, const llvm::object::ObjectFile*, llvm::raw_ostream&)’:
disasm.cpp:314:83: warning: ‘void llvm::MCObjectFileInfo::InitMCObjectFileInfo(llvm::StringRef, llvm::Reloc::Model, llvm::CodeModel::Model, llvm::MCContext&)
’ is deprecated (declared at /home/android/julia/usr/include/llvm/MC/MCObjectFileInfo.h:192) [-Wdeprecated-declarations]
     MOFI->InitMCObjectFileInfo(TripleName, Reloc::Default, CodeModel::Default, Ctx);
                                                                                   ^
codegen.cpp: In function ‘void jl_setup_module(llvm::Module*, bool)’:
codegen.cpp:695:62: error: request for member ‘getStringRepresentation’ in ‘jl_ExecutionEngine->llvm::ExecutionEngine::getDataLayout()’, which is of pointer
type ‘const llvm::DataLayout*’ (maybe you meant to use ‘->’ ?)
         m->setDataLayout(jl_ExecutionEngine->getDataLayout().getStringRepresentation());
                                                              ^
codegen.cpp: In function ‘void jl_init_codegen()’:
codegen.cpp:5558:58: error: request for member ‘getStringRepresentation’ in ‘jl_ExecutionEngine->llvm::ExecutionEngine::getDataLayout()’, which is of pointer
 type ‘const llvm::DataLayout*’ (maybe you meant to use ‘->’ ?)
     m->setDataLayout(jl_ExecutionEngine->getDataLayout().getStringRepresentation());
                                                          ^
codegen.cpp:5559:70: error: request for member ‘getStringRepresentation’ in ‘jl_ExecutionEngine->llvm::ExecutionEngine::getDataLayout()’, which is of pointer
 type ‘const llvm::DataLayout*’ (maybe you meant to use ‘->’ ?)
     engine_module->setDataLayout(jl_ExecutionEngine->getDataLayout().getStringRepresentation());
                                                                      ^
Makefile:80: recipe for target 'codegen.o' failed
make[1]: *** [codegen.o] Error 1
make[1]: *** Waiting for unfinished jobs....
Makefile:58: recipe for target 'julia-src-release' failed
make: *** [julia-src-release] Error 2

pao · 2015-07-26T16:35:11Z

And making the suggested changes in codegen.cpp, we get back to:

    LINK usr/lib/libjulia.so
    JULIA usr/lib/julia/inference0.ji
WARNING: unable to determine host cpu name.
essentials.jl
julia: /home/android/julia/deps/llvm-svn/lib/ExecutionEngine/RuntimeDyld/RuntimeDyldELF.cpp:315: void llvm::RuntimeDyldELF::resolveAArch64Relocation(const ll
vm::SectionEntry&, uint64_t, uint64_t, uint32_t, int64_t): Assertion `static_cast<int64_t>(Result) >= (-2147483647-1) && static_cast<int64_t>(Result) <= (429
4967295U)' failed.

signal (6): Aborted
gsignal at /lib/aarch64-linux-gnu/libc.so.6 (unknown line)
Aborted
Makefile:175: recipe for target '/home/android/julia/usr/lib/julia/inference0.ji' failed
make[1]: *** [/home/android/julia/usr/lib/julia/inference0.ji] Error 134
Makefile:64: recipe for target 'julia-inference' failed
make: *** [julia-inference] Error 2

This allows the build to proceed for AArch64 (mentioned in #10791) and others. (cherry picked from commit f244fb6) ref #13558

yuyichao · 2015-12-30T19:22:47Z

Is this an issue on LLVM 3.7. I'm able to get a REPL on AArch64 (ArchLinuxARM) with system LLVM 3.7.0 after applying #14511 and #14500 .

yuyichao · 2016-01-06T01:44:10Z

@pao could you please confirm if this is still an issue for you on the LLVM 3.7.0 or 3.7.1 release? Given that I'm able to get all tests pass on AArch64 apart from libunwind and blas bugs, it's likely that this is fixed before LLVM 3.7.0 is released.

Will close in a few days.

pao · 2016-01-06T19:29:18Z

Quite possible that it would work. I don't have my tablet rooted at the moment, and don't think I'll get around to it in the "a few days" timeframe--but I think your success is probably sufficient to close this.

tomaklutfu · 2016-01-06T19:51:36Z

Hi. I may have similar issue detailed in the link using LLVM 3.7.0. I added a Make.user file with LLVM=3.7.1 but it still builds LLVM 3.7.0. If not related sorry for the noise. https://groups.google.com/forum/m/#!topic/julia-users/NO6Lox7r_Xc

tkelman · 2016-01-06T19:56:59Z

probably need to add override to fix the arm setting

yuyichao · 2016-01-06T19:58:16Z

Your screenshot suggest that you are building for a 32bit ARM so it's unrelated to this issue. In additional to the override, you also need to clear the old build so that it builds a new one.

eschnett · 2016-01-06T21:58:46Z

It's LLVM_VER=3.7.1 (with a _VER suffix).

-erik

On Wed, Jan 6, 2016 at 2:51 PM, tomaklutfu [email protected] wrote:

Hi. I may have similar issue detailed in the link using LLVM 3.7.0. I
added a Make.user file with LLVM=3.7.1 but it still builds LLVM 3.7.0. If
not related sorry for the noise.
https://groups.google.com/forum/m/#!topic/julia-users/NO6Lox7r_Xc

—
Reply to this email directly or view it on GitHub
#10791 (comment).

Erik Schnetter [email protected]
http://www.perimeterinstitute.ca/personal/eschnetter/

tomaklutfu · 2016-01-06T22:09:04Z

Thanks all. Yes, I mean LLVM_VER. Issue was 'override' . I'll try building again.

pao added building Build system, or building Julia or its dependencies system:arm ARMv7 and AArch64 labels Apr 10, 2015

ViralBShah closed this as completed Jul 25, 2015

pao reopened this Jul 25, 2015

pao mentioned this issue Jul 29, 2015

ccall test failure on arm #12343

Closed

pao mentioned this issue Oct 10, 2015

baseline 300% CPU usage with ARMv7 binaries on quad-core AArch64 ARM #13522

Closed

ginggs mentioned this issue Oct 12, 2015

Define AArch64 as "not x86" #13558

Merged

tkelman pushed a commit that referenced this issue Oct 31, 2015

Define all other architectures as "not x86"

57ab2f6

This allows the build to proceed for AArch64 (mentioned in #10791) and others. (cherry picked from commit f244fb6) ref #13558

pao closed this as completed Jan 6, 2016

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

AArch64 build fails surprisingly close to the end (building sys0.o) #10791

AArch64 build fails surprisingly close to the end (building sys0.o) #10791

pao commented Apr 10, 2015

pao commented Apr 10, 2015

pao commented Apr 11, 2015

vtjnash commented May 12, 2015

pao commented May 26, 2015

vtjnash commented May 26, 2015

pao commented May 27, 2015

ViralBShah commented Jul 19, 2015

pao commented Jul 19, 2015

ViralBShah commented Jul 25, 2015

pao commented Jul 25, 2015

pao commented Jul 25, 2015

pao commented Jul 25, 2015

Keno commented Jul 25, 2015

pao commented Jul 25, 2015

pao commented Jul 26, 2015

Keno commented Jul 26, 2015

pao commented Jul 26, 2015

tkelman commented Jul 26, 2015

ViralBShah commented Jul 26, 2015

pao commented Jul 26, 2015

ViralBShah commented Jul 26, 2015

pao commented Jul 26, 2015

pao commented Jul 26, 2015

yuyichao commented Dec 30, 2015

yuyichao commented Jan 6, 2016

pao commented Jan 6, 2016

tomaklutfu commented Jan 6, 2016

tkelman commented Jan 6, 2016

yuyichao commented Jan 6, 2016

eschnett commented Jan 6, 2016

tomaklutfu commented Jan 6, 2016

AArch64 build fails surprisingly close to the end (building sys0.o) #10791

AArch64 build fails surprisingly close to the end (building sys0.o) #10791

Comments

pao commented Apr 10, 2015

pao commented Apr 10, 2015

pao commented Apr 11, 2015

vtjnash commented May 12, 2015

pao commented May 26, 2015

vtjnash commented May 26, 2015

pao commented May 27, 2015

ViralBShah commented Jul 19, 2015

pao commented Jul 19, 2015

ViralBShah commented Jul 25, 2015

pao commented Jul 25, 2015

pao commented Jul 25, 2015

pao commented Jul 25, 2015

Keno commented Jul 25, 2015

pao commented Jul 25, 2015

pao commented Jul 26, 2015

Keno commented Jul 26, 2015

pao commented Jul 26, 2015

tkelman commented Jul 26, 2015

ViralBShah commented Jul 26, 2015

pao commented Jul 26, 2015

ViralBShah commented Jul 26, 2015

pao commented Jul 26, 2015

pao commented Jul 26, 2015

yuyichao commented Dec 30, 2015

yuyichao commented Jan 6, 2016

pao commented Jan 6, 2016

tomaklutfu commented Jan 6, 2016

tkelman commented Jan 6, 2016

yuyichao commented Jan 6, 2016

eschnett commented Jan 6, 2016

tomaklutfu commented Jan 6, 2016