[llvm-commits] [llvm] r61030 - in /llvm/trunk: include/llvm/Attributes.h lib/AsmParser/llvmAsmParser.y lib/Bitcode/Writer/BitcodeWriter.cpp lib/VMCore/Attributes.cpp

[Bill Wendling]

On Tue, Dec 16, 2008 at 10:53 AM, Bill Wendling <isanbard at gmail.com> wrote:
> On Tue, Dec 16, 2008 at 12:57 AM, Nick Lewycky <nicholas at mxc.ca> wrote:
>> Bill Wendling wrote:
>>> Nick,
>>>
>>> You patch r61019 is causing a bootstrap failure on PPC. Here's the
>>> minimal testcase and the error:
>>>
>>> $ cat testcase.i
>>> struct S {
>>>    unsigned long long f1;
>>>    unsigned long long f2;
>>> };
>>> static void func2(struct S);
>>> void func1() {
>>>    struct S a;
>>>    func2(a);
>>> }
>>>
>>> $ Volumes/SandBox/NightlyTest/test/llvm-gcc.obj/./prev-gcc/xgcc -B/
>>> Volumes/SandBox/NightlyTest/test/llvm\
>>> -gcc.obj/./prev-gcc/ -B/Volumes/SandBox/NightlyTest/test/llvm-
>>> gcc.install/powerpc-apple-darwin9.5.0/bin\
>>> / -c testcase.i
>>> bash-3.2$ bash-3.2$ Assertion failed: (isPowerOf2_32(i) && "Alignment
>>> must be a power of two."), functi\
>>> on constructAlignmentFromInt, file /Volumes/SandBox/NightlyTest/test/
>>> llvm.src/include/llvm/Attributes.h\
>>> , line 81.
>>> testcase.i: In function 'func1':
>>> testcase.i:6: internal compiler error: Abort trap
>>> Please submit a full bug report,
>>> with preprocessed source if appropriate.
>>> See <URL:http://developer.apple.com/bugreporter> for instructions.
>>>
>>> Unfortunately, this occurs in the front-end, so I can't create a .bc
>>> file. Here is the backtrace:
>>>
>>> (gdb) bt
>>> #0  0x95bfd278 in abort ()
>>> #1  0x95bec904 in __assert_rtn ()
>>> #2  0x0093396c in llvm::Attribute::constructAlignmentFromInt ()
>>> #3  0x00862a6c in DefaultABI<(anonymous
>>> namespace)::FunctionTypeConversion>::HandleArgument ()
>>> #4  0x0084b4ec in TypeConverter::ConvertFunctionType ()
>>> #5  0x007cf6fc in make_decl_llvm ()
>>> #6  0x007e3790 in llvm_get_decl ()
>>> #7  0x007e6fe0 in TreeToLLVM::EmitLV_DECL ()
>>> #8  0x00816dcc in TreeToLLVM::EmitLV ()
>>> #9  0x0081e7fc in TreeToLLVM::EmitADDR_EXPR ()
>>> #10 0x00802734 in TreeToLLVM::Emit ()
>>> #11 0x00801988 in TreeToLLVM::EmitCALL_EXPR ()
>>> #12 0x00802750 in TreeToLLVM::Emit ()
>>> #13 0x0081f558 in TreeToLLVM::EmitFunction ()
>>> #14 0x007d22a4 in llvm_emit_code_for_current_function ()
>>> #15 0x0018c3a8 in tree_rest_of_compilation ()
>>> #16 0x0002d9bc in c_expand_body ()
>>> #17 0x008df4dc in cgraph_expand_function ()
>>> #18 0x008dba94 in cgraph_assemble_pending_functions ()
>>> #19 0x008dc01c in cgraph_finalize_function ()
>>> #20 0x0002d854 in finish_function ()
>>> #21 0x000d88a0 in c_parser_declaration_or_fndef ()
>>> #22 0x000d7afc in c_parser_external_declaration ()
>>> #23 0x000d7968 in c_parser_translation_unit ()
>>> #24 0x000eb214 in c_parse_file ()
>>> #25 0x000c35dc in c_common_parse_file ()
>>> #26 0x0070688c in compile_file ()
>>> #27 0x007095e0 in do_compile ()
>>> #28 0x00709680 in toplev_main ()
>>> #29 0x0010b308 in main ()
>>>
>>> Could you back-out your changes and investigate?
>>
>> I'll do that if we can't resolve this quickly, but it's complicated by
>> the fact that other patches have already gone in which depend on this
>> one. I think it's just r61019 r61030 r61040 so far.
>>
>> The crash you pasted should only happen if there's a mismatch between
>> LLVM and llvm-gcc versions. To be exact, it looks like llvm-gcc was
>> picking up the newer Attribute.h from LLVM, while still linking against
>> an older version of the LLVM libraries (perhaps installed in the system?).
>>
>> Could you please check to make sure that's not the case? If you're sure,
>> I'll be happy to go ahead and back this change out (or you can do it if
>> you like), but I want to be sure firts.
>>
> Hi Nick,
>
> It failed again last night, which was r61086. It was using the correct
> header files.
>
Now it's failing on X86 in release mode. I'm going to go ahead and
revert those patches.

-bw