Bump LLVM version to 6.0.0 #26398

vchuravy · 2018-03-09T20:19:56Z

Viral and Jeff cornered me and asked me how hard it would be to switch to 6.0.0.

This is intended as a smoke-screen test to see what CI and more importantly Nanosoldier says.

6.0.0

Base  ────────────────── 76.464808 seconds                 
Base64  ────────────────  3.127560 seconds                 
CRC32c  ────────────────  0.051004 seconds                 
SHA  ───────────────────  0.365380 seconds                 
FileWatching  ──────────  0.268810 seconds                 
Unicode  ───────────────  0.025183 seconds                 
Mmap  ──────────────────  0.183732 seconds                 
Serialization  ─────────  1.652668 seconds                 
Libdl  ─────────────────  0.160410 seconds                 
Markdown  ──────────────  3.160268 seconds                 
LibGit2  ─────────────── 14.052500 seconds                 
Logging  ───────────────  3.085846 seconds                 
Sockets  ───────────────  3.448210 seconds                 
Printf  ────────────────  1.850275 seconds                 
Profile  ───────────────  0.693780 seconds                 
Dates  ─────────────────  5.775388 seconds                 
DelimitedFiles  ────────  3.969369 seconds                 
Random  ──────────────── 10.706404 seconds                 
UUIDs  ─────────────────  1.889006 seconds                 
Future  ────────────────  0.022085 seconds                 
Pkg  ───────────────────  8.691972 seconds                 
LinearAlgebra  ───────── 39.905994 seconds                 
IterativeEigensolvers  ─  3.755822 seconds                 
SparseArrays  ────────── 14.649278 seconds                 
SuiteSparse  ───────────  3.585028 seconds                 
SharedArrays  ────────── 12.941465 seconds                 
Distributed  ───────────  1.774925 seconds                 
Test  ──────────────────  6.042602 seconds                 
REPL  ────────────────── 11.350251 seconds                 
Pkg3  ────────────────── 18.842404 seconds                 
Stdlibs total  ─────────177.743514 seconds                 
WARNING: SparseArrays.blkdiag is deprecated, use blockdiag instead.                                                   
  likely near sysimg.jl:715                                
/home/vchuravy/julia/base/precompile.jl                    
Sysimage built. Summary:                                   
Total ─────── 279.479692 seconds                           
Base: ───────  76.464808 seconds 27.3597%                  
Stdlibs: ──── 177.743514 seconds 63.598%                   
Precompile: ─  25.267449 seconds 9.04089%

`time -v`:

User time (seconds): 531.63
        System time (seconds): 5.29
        Percent of CPU this job got: 100%
        Elapsed (wall clock) time (h:mm:ss or m:ss): 8:55.20
        Average shared text size (kbytes): 0
        Average unshared data size (kbytes): 0
        Average stack size (kbytes): 0
        Average total size (kbytes): 0
        Maximum resident set size (kbytes): 2292672
        Average resident set size (kbytes): 0
        Major (requiring I/O) page faults: 0
        Minor (reclaiming a frame) page faults: 2233141
        Voluntary context switches: 42
        Involuntary context switches: 594
        Swaps: 0
        File system inputs: 0
        File system outputs: 340656
        Socket messages sent: 0
        Socket messages received: 0
        Signals delivered: 0
        Page size (bytes): 4096
        Exit status: 0

3.9.1

Base  ────────────────── 72.322026 seconds
Base64  ────────────────  2.729186 seconds
CRC32c  ────────────────  0.034614 seconds
SHA  ───────────────────  0.363587 seconds
FileWatching  ──────────  0.267717 seconds
Unicode  ───────────────  0.025211 seconds
Mmap  ──────────────────  0.180100 seconds
Serialization  ─────────  1.660259 seconds
Libdl  ─────────────────  0.148583 seconds
Markdown  ──────────────  2.894347 seconds
LibGit2  ─────────────── 12.889084 seconds
Logging  ───────────────  2.859578 seconds
Sockets  ───────────────  3.200110 seconds
Printf  ────────────────  1.714964 seconds
Profile  ───────────────  0.658392 seconds
Dates  ─────────────────  5.509282 seconds
DelimitedFiles  ────────  3.658211 seconds
Random  ──────────────── 10.228097 seconds
UUIDs  ─────────────────  1.660729 seconds
Future  ────────────────  0.022349 seconds
Pkg  ───────────────────  8.261382 seconds
LinearAlgebra  ───────── 39.721122 seconds
IterativeEigensolvers  ─  3.532457 seconds
SparseArrays  ────────── 14.157760 seconds
SuiteSparse  ───────────  3.436138 seconds
SharedArrays  ────────── 12.632652 seconds
Distributed  ───────────  1.634865 seconds
Test  ──────────────────  5.761900 seconds
REPL  ────────────────── 10.515702 seconds
Pkg3  ────────────────── 17.505437 seconds
Stdlibs total  ─────────169.579216 seconds
/home/vchuravy/julia/base/precompile.jl
Sysimage built. Summary:
Total ─────── 267.523895 seconds 
Base: ───────  72.322026 seconds 27.0339%
Stdlibs: ──── 169.579216 seconds 63.3884%
Precompile: ─  25.618957 seconds 9.57632%

`time -v`:

        User time (seconds): 528.65
        System time (seconds): 5.23
        Percent of CPU this job got: 100%
        Elapsed (wall clock) time (h:mm:ss or m:ss): 8:52.16
        Average shared text size (kbytes): 0
        Average unshared data size (kbytes): 0
        Average stack size (kbytes): 0
        Average total size (kbytes): 0
        Maximum resident set size (kbytes): 2225288
        Average resident set size (kbytes): 0
        Major (requiring I/O) page faults: 0
        Minor (reclaiming a frame) page faults: 2340496
        Voluntary context switches: 42
        Involuntary context switches: 567
        Swaps: 0
        File system inputs: 0
        File system outputs: 331216
        Socket messages sent: 0
        Socket messages received: 0
        Signals delivered: 0
        Page size (bytes): 4096
        Exit status: 0

Edit: GPU testing to the list above by @ViralBShah

ViralBShah · 2018-03-09T20:22:44Z

Looks like about 5% slower, which is quite decent, and perhaps even recoverable quickly - should we decide to do something like this.

JeffBezanson · 2018-03-09T20:22:46Z

👍 I think we should try to upgrade LLVM before 1.0, so that the 0.7-1.0 timeframe is the least stable period, and we can steadily increase reliability and performance over the 1.0.x and 1.x series.

ViralBShah · 2018-03-09T20:24:31Z

@vchuravy Just wondering if you have a comparison of the total time for running tests.

tkelman · 2018-03-09T21:54:33Z

deps/llvm.mk

 $(eval $(call LLVM_PATCH,llvm-PPC-addrspaces)) # PPC
 $(eval $(call LLVM_PATCH,llvm-PR36292-5.0)) # PPC fixes #26249, remove for 6.0
+else ifeq ($(LLVM_VER_SHORT),6.0)
+#$(eval $(call LLVM_PATCH,llvm-5.0.0_threads))


just fyi the comment on this one "Cygwin and openSUSE still use win32-threads mingw" hasn't been true for over a year, ref https://build.opensuse.org/request/show/444383 and https://sourceware.org/ml/cygwin/2016-11/msg00113.html. should be safe to drop this patch nowadays.

Cheers! That is good news.

ViralBShah · 2018-03-10T16:50:25Z

@maleadt Is LLVM 6.0 good from the GPU perspective? Just pinging you here in case you haven't seen it.

maleadt · 2018-03-10T16:54:05Z

Yeah, I was waiting for the AS patch before testing this (if Valentin doesn't beat me to it, I'll do it next week).

vchuravy · 2018-03-10T17:29:59Z

already ported just not pushed yet, vchuravy@2c19817

vchuravy · 2018-03-10T20:15:10Z

On Nanosoldier4 (with -j8):

6.0.0

    SUCCESS
        Command being timed: "make -j8 testall"
        User time (seconds): 6719.85
        System time (seconds): 96.32
        Percent of CPU this job got: 573%
        Elapsed (wall clock) time (h:mm:ss or m:ss): 19:49.26
        Average shared text size (kbytes): 0
        Average unshared data size (kbytes): 0
        Average stack size (kbytes): 0
        Average total size (kbytes): 0
        Maximum resident set size (kbytes): 2258056
        Average resident set size (kbytes): 0
        Major (requiring I/O) page faults: 21
        Minor (reclaiming a frame) page faults: 12547419
        Voluntary context switches: 48900
        Involuntary context switches: 55660107
        Swaps: 0
        File system inputs: 768
        File system outputs: 1013232
        Socket messages sent: 0
        Socket messages received: 0
        Signals delivered: 0
        Page size (bytes): 4096
        Exit status: 0

3.9.1

        Command being timed: "make -j8 testall"
        User time (seconds): 6809.86
        System time (seconds): 89.12
        Percent of CPU this job got: 575%
        Elapsed (wall clock) time (h:mm:ss or m:ss): 19:58.05
        Average shared text size (kbytes): 0
        Average unshared data size (kbytes): 0
        Average stack size (kbytes): 0
        Average total size (kbytes): 0
        Maximum resident set size (kbytes): 2247944
        Average resident set size (kbytes): 0
        Major (requiring I/O) page faults: 4
        Minor (reclaiming a frame) page faults: 11719777
        Voluntary context switches: 48016
        Involuntary context switches: 57011270
        Swaps: 0
        File system inputs: 456
        File system outputs: 997504
        Socket messages sent: 0
        Socket messages received: 0
        Signals delivered: 0
        Page size (bytes): 4096
        Exit status: 0

Builds

vchuravy · 2018-03-10T20:19:04Z

@nanosoldier runbenchmarks(ALL, vs=":master")

KristofferC · 2018-03-10T20:22:47Z

It seems nanosoldier lately only listens to its true master @ararslan

ararslan · 2018-03-10T20:39:49Z

@nanosoldier runbenchmarks(ALL, vs=":master")

ararslan · 2018-03-10T20:42:52Z

He still ignores me too, like a toddler acting out. @nanosoldier runbenchmarks(ALL, vs=":master")

ararslan · 2018-03-10T20:49:55Z

Could be confused by the branch target changing, as it seems both Travis and AppVeyor are. I'd restart the server but there's another build running.

vchuravy · 2018-03-10T21:05:27Z

@nanosoldier runbenchmarks(ALL, vs=":master")

vchuravy · 2018-03-10T21:20:16Z

CircleCI is encountering a Codegen issue.

julia-debug: /home/circleci/project/deps/srccache/llvm-6.0.0/lib/Target/X86/X86FastISel.cpp:1793: bool {anonymous}::X86FastISel::X86SelectShift(const llvm::Instruction*): Assertion `!I->getType()->isIntegerTy(8) && "i8 shifts should be handled by autogenerated table"' failed.

signal (6): Aborted
in expression starting at mpfr.jl:927
gsignal at /lib/x86_64-linux-gnu/libc.so.6 (unknown line)
abort at /lib/x86_64-linux-gnu/libc.so.6 (unknown line)
unknown function (ip: 0x7fed84874265)
__assert_fail at /lib/x86_64-linux-gnu/libc.so.6 (unknown line)
_ZN12_GLOBAL__N_111X86FastISel21fastSelectInstructionEPKN4llvm11InstructionE at /home/circleci/project/usr/bin/../lib/libLLVM-6.0.so (unknown line)
_ZN4llvm8FastISel17selectInstructionEPKNS_11InstructionE at /home/circleci/project/usr/bin/../lib/libLLVM-6.0.so (unknown line)
_ZN4llvm16SelectionDAGISel20SelectAllBasicBlocksERKNS_8FunctionE at /home/circleci/project/usr/bin/../lib/libLLVM-6.0.so (unknown line)
_ZN4llvm16SelectionDAGISel20runOnMachineFunctionERNS_15MachineFunctionE.part.1097 at /home/circleci/project/usr/bin/../lib/libLLVM-6.0.so (unknown line)
_ZN12_GLOBAL__N_115X86DAGToDAGISel20runOnMachineFunctionERN4llvm15MachineFunctionE at /home/circleci/project/usr/bin/../lib/libLLVM-6.0.so (unknown line)
_ZN4llvm19MachineFunctionPass13runOnFunctionERNS_8FunctionE at /home/circleci/project/usr/bin/../lib/libLLVM-6.0.so (unknown line)
_ZN4llvm13FPPassManager13runOnFunctionERNS_8FunctionE at /home/circleci/project/usr/bin/../lib/libLLVM-6.0.so (unknown line)
_ZN4llvm13FPPassManager11runOnModuleERNS_6ModuleE at /home/circleci/project/usr/bin/../lib/libLLVM-6.0.so (unknown line)
_ZN4llvm6legacy15PassManagerImpl3runERNS_6ModuleE at /home/circleci/project/usr/bin/../lib/libLLVM-6.0.so (unknown line)
operator() at /home/circleci/project/src/jitlayers.cpp:466
addModule at /home/circleci/project/usr/include/llvm/ExecutionEngine/Orc/IRCompileLayer.h:57
addModule at /home/circleci/project/src/jitlayers.cpp:593
jl_add_to_ee at /home/circleci/project/src/jitlayers.cpp:831
jl_finalize_function at /home/circleci/project/src/jitlayers.cpp:839
getAddressForFunction at /home/circleci/project/src/codegen.cpp:1283
jl_generate_fptr at /home/circleci/project/src/codegen.cpp:1386
jl_compile_method_internal at /home/circleci/project/src/julia_internal.h:370
jl_call_method_internal at /home/circleci/project/src/julia_internal.h:398
jl_apply_generic at /home/circleci/project/src/gf.c:2098
jl_apply at /home/circleci/project/src/julia.h:1528
jl_invoke at /home/circleci/project/src/gf.c:55

Win32 is failing during building LLVM

/usr/bin/ar: ../libLLVMSupport.a: Error reading CMakeFiles/LLVMSupport.dir/Watchdog.cpp.o: File truncated
make[8]: *** [lib/Support/CMakeFiles/LLVMSupport.dir/build.make:2878: lib/libLLVMSupport.a] Error 1
make[8]: *** Deleting file 'lib/libLLVMSupport.a'
make[7]: *** [CMakeFiles/Makefile2:619: lib/Support/CMakeFiles/LLVMSupport.dir/all] Error 2
make[6]: *** [CMakeFiles/Makefile2:10680: tools/llvm-config/CMakeFiles/llvm-config.dir/rule] Error 2
make[5]: *** [Makefile:3356: llvm-config] Error 2

Wind64 is having a link time issue:

/usr/bin/x86_64-w64-mingw32-g++.exe  -m64    -D__USING_SJLJ_EXCEPTIONS__ -D__CRT__NO_INLINE -Wa,-mbig-obj -Werror=date-time -std=gnu++11 -Wall -W -Wno-unused-parameter -Wwrite-strings -Wcast-qual -Wno-missing-field-initializers -pedantic -Wno-long-long -Wno-maybe-uninitialized -Wdelete-non-virtual-dtor -Wno-comment  -O2 -DNDEBUG    -shared -o ../../bin/LLVM.dll -Wl,--out-implib,../../lib/libLLVM.dll.a -Wl,--major-image-version,0,--minor-image-version,0 -Wl,--whole-archive CMakeFiles/LLVM.dir/objects.a -Wl,--no-whole-archive @CMakeFiles/LLVM.dir/linklibs.rsp
../../lib/libLLVMNVPTXCodeGen.a(NVPTXInstrInfo.cpp.obj):NVPTXInstrInfo.cpp:(.text$_ZNK4llvm15TargetInstrInfo22insertOutlinerPrologueERNS_17MachineBasicBlockERNS_15MachineFunctionERKNS0_19MachineOutlinerInfoE+0x0): multiple definition of `llvm::TargetInstrInfo::insertOutlinerPrologue(llvm::MachineBasicBlock&, llvm::MachineFunction&, llvm::TargetInstrInfo::MachineOutlinerInfo const&) const'
../../lib/libLLVMCodeGen.a(TargetInstrInfo.cpp.obj):TargetInstrInfo.cpp:(.text$_ZNK4llvm15TargetInstrInfo22insertOutlinerPrologueERNS_17MachineBasicBlockERNS_15MachineFunctionERKNS0_19MachineOutlinerInfoE+0x0): first defined here
../../lib/libLLVMNVPTXCodeGen.a(NVPTXInstrInfo.cpp.obj):NVPTXInstrInfo.cpp:(.text$_ZNK4llvm15TargetInstrInfo18insertOutlinedCallERNS_6ModuleERNS_17MachineBasicBlockERNS_26MachineInstrBundleIteratorINS_12MachineInstrELb0EEERNS_15MachineFunctionERKNS0_19MachineOutlinerInfoE+0x0): multiple definition of `llvm::TargetInstrInfo::insertOutlinedCall(llvm::Module&, llvm::MachineBasicBlock&, llvm::MachineInstrBundleIterator<llvm::MachineInstr, false>&, llvm::MachineFunction&, llvm::TargetInstrInfo::MachineOutlinerInfo const&) const'
../../lib/libLLVMCodeGen.a(TargetInstrInfo.cpp.obj):TargetInstrInfo.cpp:(.text$_ZNK4llvm15TargetInstrInfo18insertOutlinedCallERNS_6ModuleERNS_17MachineBasicBlockERNS_26MachineInstrBundleIteratorINS_12MachineInstrELb0EEERNS_15MachineFunctionERKNS0_19MachineOutlinerInfoE+0x0): first defined here
../../lib/libLLVMNVPTXCodeGen.a(NVPTXInstrInfo.cpp.obj):NVPTXInstrInfo.cpp:(.text$_ZNK4llvm15TargetInstrInfo22insertOutlinerEpilogueERNS_17MachineBasicBlockERNS_15MachineFunctionERKNS0_19MachineOutlinerInfoE+0x0): multiple definition of `llvm::TargetInstrInfo::insertOutlinerEpilogue(llvm::MachineBasicBlock&, llvm::MachineFunction&, llvm::TargetInstrInfo::MachineOutlinerInfo const&) const'
../../lib/libLLVMCodeGen.a(TargetInstrInfo.cpp.obj):TargetInstrInfo.cpp:(.text$_ZNK4llvm15TargetInstrInfo22insertOutlinerEpilogueERNS_17MachineBasicBlockERNS_15MachineFunctionERKNS0_19MachineOutlinerInfoE+0x0): first defined here
../../lib/libLLVMNVPTXCodeGen.a(NVPTXInstrInfo.cpp.obj):NVPTXInstrInfo.cpp:(.text$_ZNK4llvm15TargetInstrInfo16getOutliningTypeERNS_12MachineInstrE+0x0): multiple definition of `llvm::TargetInstrInfo::getOutliningType(llvm::MachineInstr&) const'
../../lib/libLLVMCodeGen.a(TargetInstrInfo.cpp.obj):TargetInstrInfo.cpp:(.text$_ZNK4llvm15TargetInstrInfo16getOutliningTypeERNS_12MachineInstrE+0x0): first defined here
../../lib/libLLVMNVPTXCodeGen.a(NVPTXInstrInfo.cpp.obj):NVPTXInstrInfo.cpp:(.text$_ZNK4llvm15TargetInstrInfo27getOutlininingCandidateInfoERSt6vectorISt4pairINS_26MachineInstrBundleIteratorINS_12MachineInstrELb0EEES5_ESaIS6_EE+0x0): multiple definition of `llvm::TargetInstrInfo::getOutlininingCandidateInfo(std::vector<std::pair<llvm::MachineInstrBundleIterator<llvm::MachineInstr, false>, llvm::MachineInstrBundleIterator<llvm::MachineInstr, false> >, std::allocator<std::pair<llvm::MachineInstrBundleIterator<llvm::MachineInstr, false>, llvm::MachineInstrBundleIterator<llvm::MachineInstr, false> > > >&) const'
../../lib/libLLVMCodeGen.a(TargetInstrInfo.cpp.obj):TargetInstrInfo.cpp:(.text$_ZNK4llvm15TargetInstrInfo27getOutlininingCandidateInfoERSt6vectorISt4pairINS_26MachineInstrBundleIteratorINS_12MachineInstrELb0EEES5_ESaIS6_EE+0x0): first defined here
../../lib/libLLVMNVPTXCodeGen.a(NVPTXInstrInfo.cpp.obj):NVPTXInstrInfo.cpp:(.text$_ZNK4llvm15TargetInstrInfo25replaceBranchWithTailCallERNS_17MachineBasicBlockERNS_15SmallVectorImplINS_14MachineOperandEEERKNS_12MachineInstrE+0x0): multiple definition of `llvm::TargetInstrInfo::replaceBranchWithTailCall(llvm::MachineBasicBlock&, llvm::SmallVectorImpl<llvm::MachineOperand>&, llvm::MachineInstr const&) const'
../../lib/libLLVMCodeGen.a(TargetInstrInfo.cpp.obj):TargetInstrInfo.cpp:(.text$_ZNK4llvm15TargetInstrInfo25replaceBranchWithTailCallERNS_17MachineBasicBlockERNS_15SmallVectorImplINS_14MachineOperandEEERKNS_12MachineInstrE+0x0): first defined here
../../lib/libLLVMNVPTXCodeGen.a(NVPTXInstrInfo.cpp.obj):NVPTXInstrInfo.cpp:(.text$_ZNK4llvm15TargetInstrInfo19shouldClusterMemOpsERNS_12MachineInstrEjS2_jj+0x0): multiple definition of `llvm::TargetInstrInfo::shouldClusterMemOps(llvm::MachineInstr&, unsigned int, llvm::MachineInstr&, unsigned int, unsigned int) const'
../../lib/libLLVMCodeGen.a(TargetInstrInfo.cpp.obj):TargetInstrInfo.cpp:(.text$_ZNK4llvm15TargetInstrInfo19shouldClusterMemOpsERNS_12MachineInstrEjS2_jj+0x0): first defined here
../../lib/libLLVMNVPTXCodeGen.a(NVPTXInstrInfo.cpp.obj):NVPTXInstrInfo.cpp:(.text$_ZNK4llvm15TargetInstrInfo20loadRegFromStackSlotERNS_17MachineBasicBlockENS_26MachineInstrBundleIteratorINS_12MachineInstrELb0EEEjiPKNS_19TargetRegisterClassEPKNS_18TargetRegisterInfoE+0x0): multiple definition of `llvm::TargetInstrInfo::loadRegFromStackSlot(llvm::MachineBasicBlock&, llvm::MachineInstrBundleIterator<llvm::MachineInstr, false>, unsigned int, int, llvm::TargetRegisterClass const*, llvm::TargetRegisterInfo const*) const'
../../lib/libLLVMCodeGen.a(TargetInstrInfo.cpp.obj):TargetInstrInfo.cpp:(.text$_ZNK4llvm15TargetInstrInfo20loadRegFromStackSlotERNS_17MachineBasicBlockENS_26MachineInstrBundleIteratorINS_12MachineInstrELb0EEEjiPKNS_19TargetRegisterClassEPKNS_18TargetRegisterInfoE+0x0): first defined here
../../lib/libLLVMNVPTXCodeGen.a(NVPTXInstrInfo.cpp.obj):NVPTXInstrInfo.cpp:(.text$_ZNK4llvm15TargetInstrInfo19storeRegToStackSlotERNS_17MachineBasicBlockENS_26MachineInstrBundleIteratorINS_12MachineInstrELb0EEEjbiPKNS_19TargetRegisterClassEPKNS_18TargetRegisterInfoE+0x0): multiple definition of `llvm::TargetInstrInfo::storeRegToStackSlot(llvm::MachineBasicBlock&, llvm::MachineInstrBundleIterator<llvm::MachineInstr, false>, unsigned int, bool, int, llvm::TargetRegisterClass const*, llvm::TargetRegisterInfo const*) const'
../../lib/libLLVMCodeGen.a(TargetInstrInfo.cpp.obj):TargetInstrInfo.cpp:(.text$_ZNK4llvm15TargetInstrInfo19storeRegToStackSlotERNS_17MachineBasicBlockENS_26MachineInstrBundleIteratorINS_12MachineInstrELb0EEEjbiPKNS_19TargetRegisterClassEPKNS_18TargetRegisterInfoE+0x0): first defined here
../../lib/libLLVMNVPTXCodeGen.a(NVPTXInstrInfo.cpp.obj):NVPTXInstrInfo.cpp:(.text$_ZNK4llvm15TargetInstrInfo11copyPhysRegERNS_17MachineBasicBlockENS_26MachineInstrBundleIteratorINS_12MachineInstrELb0EEERKNS_8DebugLocEjjb+0x0): multiple definition of `llvm::TargetInstrInfo::copyPhysReg(llvm::MachineBasicBlock&, llvm::MachineInstrBundleIterator<llvm::MachineInstr, false>, llvm::DebugLoc const&, unsigned int, unsigned int, bool) const'
../../lib/libLLVMCodeGen.a(TargetInstrInfo.cpp.obj):TargetInstrInfo.cpp:(.text$_ZNK4llvm15TargetInstrInfo11copyPhysRegERNS_17MachineBasicBlockENS_26MachineInstrBundleIteratorINS_12MachineInstrELb0EEERKNS_8DebugLocEjjb+0x0): first defined here
../../lib/libLLVMNVPTXCodeGen.a(NVPTXInstrInfo.cpp.obj):NVPTXInstrInfo.cpp:(.text$_ZNK4llvm15TargetInstrInfo14optimizeSelectERNS_12MachineInstrERNS_15SmallPtrSetImplIPS1_EEb+0x0): multiple definition of `llvm::TargetInstrInfo::optimizeSelect(llvm::MachineInstr&, llvm::SmallPtrSetImpl<llvm::MachineInstr*>&, bool) const'
../../lib/libLLVMCodeGen.a(TargetInstrInfo.cpp.obj):TargetInstrInfo.cpp:(.text$_ZNK4llvm15TargetInstrInfo14optimizeSelectERNS_12MachineInstrERNS_15SmallPtrSetImplIPS1_EEb+0x0): first defined here
../../lib/libLLVMNVPTXCodeGen.a(NVPTXInstrInfo.cpp.obj):NVPTXInstrInfo.cpp:(.text$_ZNK4llvm15TargetInstrInfo12insertSelectERNS_17MachineBasicBlockENS_26MachineInstrBundleIteratorINS_12MachineInstrELb0EEERKNS_8DebugLocEjNS_8ArrayRefINS_14MachineOperandEEEjj+0x0): multiple definition of `llvm::TargetInstrInfo::insertSelect(llvm::MachineBasicBlock&, llvm::MachineInstrBundleIterator<llvm::MachineInstr, false>, llvm::DebugLoc const&, unsigned int, llvm::ArrayRef<llvm::MachineOperand>, unsigned int, unsigned int) const'
../../lib/libLLVMCodeGen.a(TargetInstrInfo.cpp.obj):TargetInstrInfo.cpp:(.text$_ZNK4llvm15TargetInstrInfo12insertSelectERNS_17MachineBasicBlockENS_26MachineInstrBundleIteratorINS_12MachineInstrELb0EEERKNS_8DebugLocEjNS_8ArrayRefINS_14MachineOperandEEEjj+0x0): first defined here
../../lib/libLLVMNVPTXCodeGen.a(NVPTXInstrInfo.cpp.obj):NVPTXInstrInfo.cpp:(.text$_ZNK4llvm15TargetInstrInfo15reduceLoopCountERNS_17MachineBasicBlockEPNS_12MachineInstrERS3_RNS_15SmallVectorImplINS_14MachineOperandEEERNS6_IS4_EEjj+0x0): multiple definition of `llvm::TargetInstrInfo::reduceLoopCount(llvm::MachineBasicBlock&, llvm::MachineInstr*, llvm::MachineInstr&, llvm::SmallVectorImpl<llvm::MachineOperand>&, llvm::SmallVectorImpl<llvm::MachineInstr*>&, unsigned int, unsigned int) const'
../../lib/libLLVMCodeGen.a(TargetInstrInfo.cpp.obj):TargetInstrInfo.cpp:(.text$_ZNK4llvm15TargetInstrInfo15reduceLoopCountERNS_17MachineBasicBlockEPNS_12MachineInstrERS3_RNS_15SmallVectorImplINS_14MachineOperandEEERNS6_IS4_EEjj+0x0): first defined here
../../lib/libLLVMNVPTXCodeGen.a(NVPTXInstrInfo.cpp.obj):NVPTXInstrInfo.cpp:(.text$_ZNK4llvm15TargetInstrInfo12insertBranchERNS_17MachineBasicBlockEPS1_S3_NS_8ArrayRefINS_14MachineOperandEEERKNS_8DebugLocEPi+0x0): multiple definition of `llvm::TargetInstrInfo::insertBranch(llvm::MachineBasicBlock&, llvm::MachineBasicBlock*, llvm::MachineBasicBlock*, llvm::ArrayRef<llvm::MachineOperand>, llvm::DebugLoc const&, int*) const'
../../lib/libLLVMCodeGen.a(TargetInstrInfo.cpp.obj):TargetInstrInfo.cpp:(.text$_ZNK4llvm15TargetInstrInfo12insertBranchERNS_17MachineBasicBlockEPS1_S3_NS_8ArrayRefINS_14MachineOperandEEERKNS_8DebugLocEPi+0x0): first defined here
../../lib/libLLVMNVPTXCodeGen.a(NVPTXInstrInfo.cpp.obj):NVPTXInstrInfo.cpp:(.text$_ZNK4llvm15TargetInstrInfo12removeBranchERNS_17MachineBasicBlockEPi+0x0): multiple definition of `llvm::TargetInstrInfo::removeBranch(llvm::MachineBasicBlock&, int*) const'
../../lib/libLLVMCodeGen.a(TargetInstrInfo.cpp.obj):TargetInstrInfo.cpp:(.text$_ZNK4llvm15TargetInstrInfo12removeBranchERNS_17MachineBasicBlockEPi+0x0): first defined here
../../lib/libLLVMNVPTXCodeGen.a(NVPTXInstrInfo.cpp.obj):NVPTXInstrInfo.cpp:(.text$_ZNK4llvm15TargetInstrInfo20insertIndirectBranchERNS_17MachineBasicBlockES2_RKNS_8DebugLocExPNS_12RegScavengerE+0x0): multiple definition of `llvm::TargetInstrInfo::insertIndirectBranch(llvm::MachineBasicBlock&, llvm::MachineBasicBlock&, llvm::DebugLoc const&, long long, llvm::RegScavenger*) const'
../../lib/libLLVMCodeGen.a(TargetInstrInfo.cpp.obj):TargetInstrInfo.cpp:(.text$_ZNK4llvm15TargetInstrInfo20insertIndirectBranchERNS_17MachineBasicBlockES2_RKNS_8DebugLocExPNS_12RegScavengerE+0x0): first defined here
../../lib/libLLVMNVPTXCodeGen.a(NVPTXInstrInfo.cpp.obj):NVPTXInstrInfo.cpp:(.text$_ZNK4llvm15TargetInstrInfo18getBranchDestBlockERKNS_12MachineInstrE+0x0): multiple definition of `llvm::TargetInstrInfo::getBranchDestBlock(llvm::MachineInstr const&) const'
../../lib/libLLVMCodeGen.a(TargetInstrInfo.cpp.obj):TargetInstrInfo.cpp:(.text$_ZNK4llvm15TargetInstrInfo18getBranchDestBlockERKNS_12MachineInstrE+0x0): first defined here
../../lib/libLLVMNVPTXCodeGen.a(NVPTXInstrInfo.cpp.obj):NVPTXInstrInfo.cpp:(.text$_ZNK4llvm15TargetInstrInfo21isBranchOffsetInRangeEjx+0x0): multiple definition of `llvm::TargetInstrInfo::isBranchOffsetInRange(unsigned int, long long) const'
../../lib/libLLVMCodeGen.a(TargetInstrInfo.cpp.obj):TargetInstrInfo.cpp:(.text$_ZNK4llvm15TargetInstrInfo21isBranchOffsetInRangeEjx+0x0): first defined here
collect2: error: ld returned 1 exit status

Keno · 2018-03-10T21:38:15Z

Please add https://reviews.llvm.org/D42260 to the patch list.

ararslan · 2018-03-10T21:47:49Z

@nanosoldier runbenchmarks(ALL, vs=":master")

vchuravy · 2018-03-10T21:47:51Z

Sure, You haven't landed it yet, right?

Keno · 2018-03-10T21:49:09Z

No, I should though.

vchuravy · 2018-03-11T02:09:05Z

Quick update on the codegen failure. It is not triggered by the debug build, but rather by O0 and O1 so it might be more likely related to pass order.

maleadt · 2018-03-11T08:53:10Z

Could you include D44140? With it, LLVM.jl and CUDAnative.jl test without issues.

vchuravy · 2018-03-11T13:38:04Z

@nanosoldier runbenchmarks(ALL, vs=":master")

ararslan · 2018-03-11T19:15:45Z

Man, Nanosoldier really hates this PR. @nanosoldier runbenchmarks(ALL, vs=":master")

ararslan · 2018-03-11T20:42:02Z

@nanosoldier runbenchmarks(ALL, vs=":master")

vchuravy · 2018-03-17T21:32:30Z

Good news. After updating a copy of the buildbots from mingw 5.4 to 6.2 and including the last three patches I got Julia to build and pass tests on Win64!
So while the commit history needs to be cleaned up and 4436b90 needs to be submitted upstream this is now at a point where I would want to encourage other people to take this for a spin.

staticfloat · 2018-03-17T23:10:42Z

@vchuravy What do you need in order to check off the armv7l and ppc64le boxes on your checklist?

vchuravy · 2018-03-17T23:19:03Z

armv7l finished the builds, but I don't think the buildbot managed to run the tests due to the qemu syscall issue? ppc64le is more complicated. Right now we are triggering an assertion and I am trying to reduce the issue so that I can bisect it and report it upstream.

…

On Sat, 17 Mar 2018 at 19:10 Elliot Saba ***@***.***> wrote: @vchuravy <https://github.com/vchuravy> What do you need in order to check off the armv7l and ppc64le boxes on your checklist? — You are receiving this because you were mentioned. Reply to this email directly, view it on GitHub <#26398 (comment)>, or mute the thread <https://github.com/notifications/unsubscribe-auth/AAI3agorDIPad3l0eT3Ng138Cbvudanfks5tfZf0gaJpZM4Sk0LP> .

tshort · 2018-03-18T16:48:17Z

+1 from me. It compiled fine on Arch Linux, and tests for the LLVM.jl package ran fine. I also tried recompiling with the WebAssembly backend included, and that worked fine, too. (Once this is in, I'll probably submit a PR to include the wasm backend.)

vchuravy · 2018-03-19T03:42:49Z

Thanks Tom! When you submit a PR for the WebAssembly backend just make sure to measure increase in compile time for LLVM and increase in binary size, but as long as it builds on all platforms it shouldn't be a problem to include it.

…

On Sun, Mar 18, 2018, 12:48 Tom Short ***@***.***> wrote: +1 from me. It compiled fine on Arch Linux, and tests for the LLVM.jl package ran fine. I also tried recompiling with the WebAssembly backend included, and that worked fine, too. (Once this is in, I'll probably submit a PR to include the wasm backend.) — You are receiving this because you were mentioned. Reply to this email directly, view it on GitHub <#26398 (comment)>, or mute the thread <https://github.com/notifications/unsubscribe-auth/AAI3ar4zhFq7yJdNI252MGTpyLnBwvYXks5tfo_SgaJpZM4Sk0LP> .

vchuravy · 2018-03-19T23:21:37Z

Once I replace 4436b90 with https://reviews.llvm.org/D44650 this should be good to go, but I will give upstream a couple of days to respond to that. There is also movement on the ssp patch for mingw.

tkelman · 2018-03-19T23:41:17Z

Should delete the old checksums. And Mac and Windows CI are not actually testing LLVM 6 yet: https://travis-ci.org/JuliaLang/julia/jobs/354081449#L6185, https://ci.appveyor.com/project/JuliaLang/julia/build/1.0.25526/job/pdt7a2i4xvn8nvj7#L2596

vchuravy · 2018-03-19T23:45:49Z

I tested Mac and Windows via manually triggering the buildbots. But we will need to make sure to update the cache files for windows and presumably homebrew bottles for Mac.

…

On Mon, 19 Mar 2018 at 19:41 Tony Kelman ***@***.***> wrote: Should delete the old checksums. And Mac and Windows CI are not actually testing LLVM 6 yet: https://travis-ci.org/JuliaLang/julia/jobs/354081449#L6185, https://ci.appveyor.com/project/JuliaLang/julia/build/1.0.25526/job/pdt7a2i4xvn8nvj7#L2596 — You are receiving this because you were mentioned. Reply to this email directly, view it on GitHub <#26398 (comment)>, or mute the thread <https://github.com/notifications/unsubscribe-auth/AAI3anBOi_4du1DV54-zar8t4FwlJLdsks5tgEIegaJpZM4Sk0LP> .

KristofferC · 2018-03-23T11:24:15Z

Some (encouraging?) benchmarks. Defining a fixed size array, defining methods for unrolled matmat and matvec multiplication as in as in https://gist.github.com/KristofferC/07814a1dc4e6dfc07697ca2452b309da, and disabling/enabling the SLP-vectorizer and running the following benchmark:

n = 4
v4 = FixedVector((rand(n)...,))
m4x4 = FixedMatrix{n, n, Float64, n*n}((rand(n*n)...,))

n = 8
v8_32 = FixedVector((rand(Float32, n)...,))
m8x8_32 = FixedMatrix{n, n, Float32, n*n}((rand(Float32, n*n)...,))

using BenchmarkTools

@btime $m4x4 * $m4x4;
@btime $m4x4 * $v4;

@btime $m8x8_32 * $m8x8_32;
@btime $m8x8_32 * $v8_32;

I get:

	0.6-O3	0.7-O3	0.7 LLVM 6-O3	0.6	0.7	0.7 LLVM 6
`m4x4 * m4x4`	24.753 ns	26.277 ns	8.398 ns	26.861 ns	25.693 ns	25.986 ns
`m4x4 * v4`	4.715 ns	4.604 ns	2.732 ns	10.938 ns	4.593 ns	4.615 ns
`m8x8_32 * m8x8_32`	27.090 ns	180.194 ns	26.131 ns	180.476 ns	191.392 ns	181.375 ns
`m8x8_32 * v8_32`	5.396 ns	20.582 ns	5.723 ns	21.074 ns	20.588 ns	20.759 ns

LLVM 6 with SLP enabled does a very good job. 0.7 with current LLVM seems to fail to vectorize in all cases even with SLP.

tkoolen · 2018-03-23T13:05:22Z

Hmm, on my machine with 0.6 -O3:

julia> using StaticArrays, BenchmarkTools

julia> @btime a * b setup = (a = rand(SMatrix{4, 4}); b = rand(SMatrix{4, 4}))
  9.261 ns (0 allocations: 0 bytes)

and with your gist:

@btime $m4x4 * $m4x4;
  20.611 ns

StaticArrays doesn't do any explicit SIMD stuff. Why are these baseline results so different?

KristofferC · 2018-03-23T13:09:08Z

I used the definitions in https://github.com/JuliaCI/BaseBenchmarks.jl/blob/master/src/tuple/TupleBenchmarks.jl. These were originally taken from StaticArrays but perhaps they have been updated?

tkoolen · 2018-03-23T13:15:34Z

With an added Base.@_inline_meta in the expression returned from the generated function on 0.6 with -O3:

julia> @btime $m4x4 * $m4x4;
  9.242 ns (0 allocations: 0 bytes)

- forward-ports the NVPTX address space changes - drops the Windows specific threading changes - add patch for D42260 - add patch for D44140 - backport patch for FastISel bug - final version of ssp patch for Windows - fix CMAKE issue for mingw32 compilers

vchuravy · 2018-03-26T17:17:36Z

Ok, does anybody have any other comments on this. I would like to merge this PR this week and from my side it is ready (modulo the infrastructure work).
I don't think ppc64le should hold this up, since we currently can't claim to support power on 0.7 anyway. I pledge to continue my work on restoring support for ppc64le and hopefully have that all figured out for 1.0 or a patch release.

tkelman · 2018-03-26T17:56:20Z

Don't think LLVM_VER should be changed for source builds until CI is actually testing it.

vchuravy · 2018-03-26T18:38:47Z

That's what I meant by "modulo the infrastructure work": My plan is the following: 1. Switch builds servers to Mingw32 6.2 2. Regenerate the llvm-juliadeps archive (is there documentation for this?) 3. Update homebrew bottle Am I missing anything?

…

On Mon, Mar 26, 2018, 10:56 Tony Kelman ***@***.***> wrote: Don't think LLVM_VER should be changed for source builds until CI is actually testing it. — You are receiving this because you were mentioned. Reply to this email directly, view it on GitHub <#26398 (comment)>, or mute the thread <https://github.com/notifications/unsubscribe-auth/AAI3atsPeIB6Og2Ix27oPupndpsnl2dXks5tiSvGgaJpZM4Sk0LP> .

tkelman · 2018-03-26T20:19:30Z

Switch builds servers to Mingw32 6.2

Why? That's not what CI will use, or anyone following the README.windows.md instructions.

vchuravy · 2018-03-26T22:26:34Z

On the current version of mingw that the buildbots are using I get a linker failure that makes little sense to me (#26398 (comment)) on 6.2 that failure goes away. I managed to fix a similar linker failure in https://reviews.llvm.org/D44650, but that was prevalent on both 5.4 and 6.2.

vchuravy · 2018-04-30T14:26:28Z

Closed in favour of #26925

ararslan added the external dependencies Involves LLVM, OpenBLAS, or other linked libraries label Mar 9, 2018

tkelman reviewed Mar 9, 2018

View reviewed changes

vchuravy force-pushed the vc/llvm6fixes branch from 1a72568 to 4a1946b Compare March 10, 2018 00:53

vchuravy force-pushed the vc/llvm6 branch from 39b3e80 to 125ebdb Compare March 10, 2018 20:05

vchuravy changed the base branch from vc/llvm6fixes to master March 10, 2018 20:17

vchuravy closed this Mar 10, 2018

vchuravy reopened this Mar 10, 2018

KristofferC mentioned this pull request Mar 23, 2018

enable the SLP Vectorizer optimization pass by default #26594

Merged

vchuravy added 3 commits March 25, 2018 16:44

create patch list for 6.0.0

8a826b1

- forward-ports the NVPTX address space changes - drops the Windows specific threading changes - add patch for D42260 - add patch for D44140 - backport patch for FastISel bug - final version of ssp patch for Windows - fix CMAKE issue for mingw32 compilers

force ELF up to LLVM 7.0

c12922e

bump LLVM to 6.0.0

91baf58

vchuravy force-pushed the vc/llvm6 branch from f008714 to 91baf58 Compare March 25, 2018 20:46

vchuravy mentioned this pull request Mar 27, 2018

Patches for llvm 6.0 #26626

Merged

ihnorton mentioned this pull request Apr 11, 2018

Is it possible to specify that Cxx.jl use a particular version of clang++ or g++? JuliaInterop/Cxx.jl#377

Open

vchuravy mentioned this pull request Apr 28, 2018

Upgrade to LLVM 6 and use BinaryBuilder for LLVM on CI #26925

Merged

Sacha0 mentioned this pull request Apr 28, 2018

build fails on Raspberry Pi 3 #26122

Closed

vchuravy closed this Apr 30, 2018

vchuravy deleted the vc/llvm6 branch April 30, 2018 14:28

vchuravy mentioned this pull request May 6, 2018

build failure on Windows under Cygwin #27003

Closed

Uh oh!

Bump LLVM version to 6.0.0 #26398

Bump LLVM version to 6.0.0 #26398

Uh oh!

Conversation

vchuravy commented Mar 9, 2018 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

6.0.0

time -v:

3.9.1

time -v:

Uh oh!

ViralBShah commented Mar 9, 2018

Uh oh!

JeffBezanson commented Mar 9, 2018

Uh oh!

ViralBShah commented Mar 9, 2018

Uh oh!

tkelman Mar 9, 2018

Choose a reason for hiding this comment

Uh oh!

vchuravy Mar 9, 2018

Choose a reason for hiding this comment

Uh oh!

ViralBShah commented Mar 10, 2018

Uh oh!

maleadt commented Mar 10, 2018 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

vchuravy commented Mar 10, 2018

Uh oh!

vchuravy commented Mar 10, 2018

6.0.0

3.9.1

Builds

Uh oh!

vchuravy commented Mar 10, 2018

Uh oh!

KristofferC commented Mar 10, 2018 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

ararslan commented Mar 10, 2018

Uh oh!

ararslan commented Mar 10, 2018

Uh oh!

ararslan commented Mar 10, 2018

Uh oh!

vchuravy commented Mar 10, 2018

Uh oh!

vchuravy commented Mar 10, 2018

Uh oh!

Keno commented Mar 10, 2018

Uh oh!

ararslan commented Mar 10, 2018

Uh oh!

vchuravy commented Mar 10, 2018

Uh oh!

Keno commented Mar 10, 2018

Uh oh!

vchuravy commented Mar 11, 2018

Uh oh!

maleadt commented Mar 11, 2018

Uh oh!

vchuravy commented Mar 11, 2018

Uh oh!

ararslan commented Mar 11, 2018

Uh oh!

ararslan commented Mar 11, 2018

Uh oh!

vchuravy commented Mar 17, 2018

Uh oh!

staticfloat commented Mar 17, 2018

Uh oh!

vchuravy commented Mar 17, 2018 via email

Uh oh!

tshort commented Mar 18, 2018

Uh oh!

vchuravy commented Mar 19, 2018 via email

Uh oh!

vchuravy commented Mar 19, 2018

vchuravy commented Mar 9, 2018 •

edited

Loading

`time -v`:

`time -v`:

maleadt commented Mar 10, 2018 •

edited

Loading

KristofferC commented Mar 10, 2018 •

edited

Loading

KristofferC commented Mar 23, 2018 •

edited

Loading