llvm dev - Dec 2018 - Restrict global constructors to base ISA

I'm testing on older OS X 10.8 with older SSE4 hardware from about
2010. I've got updated gear from MacPorts and it includes GCC and
Clang. GCC is the compiler, and Clang is the assembler.

We perform a compile/link on a test file to ensure an ISA is supported
by the toolchain. If an ISA is available then we compile a source file
to the ISA as needed. Then, we guard the higher ISAs at runtime to
avoid SIGILLs. It worked well until we added AVX2.

For AVX2 we see this as expected:

    $ CXX=/opt/local/bin/clang++-mp-5.0 make
    /opt/local/bin/clang++-mp-5.0 ... -c chacha.cpp
    /opt/local/bin/clang++-mp-5.0 ... -mavx2 -c chacha_avx.cpp
    /opt/local/bin/clang++-mp-5.0 ... -msse2 -c chacha_simd.cpp
    ...

At runtime we catch a SIGILL due to chacha_avx.cpp as shown below. It
looks like global constructors are using instructions from AVX
(vxorps), which is beyond what the machine supports.

How do we tell Clang to use the base ISA for global constructors?

Thanks in advance.

=========
Here's the full command line used for a typical file:

/opt/local/bin/clang++-mp-5.0 -DNDEBUG -g2 -O3 -fPIC -pthread -pipe -c
cryptlib.cpp

Here's the source file. I don't believe it has global data provided by
us: https://github.com/weidai11/cryptopp/blob/master/chacha_avx.cpp

=========
(lldb) r v
...
* thread #1: tid = 0x19f83aa, 0x000000010016ec49
cryptest.exe`_GLOBAL__sub_I_chacha_avx.cpp at string:1322, queue
'com.apple.main-thread', stop reason = EXC_BAD_INSTRUCTION
(code=EXC_I386_INVOP, subcode=0x0)
    frame #0: 0x000000010016ec49
cryptest.exe`_GLOBAL__sub_I_chacha_avx.cpp at string:1322
   1319         {
   1320             size_type (&__a)[__n_words] = __r_.first().__r.__words;
   1321             for (unsigned __i = 0; __i < __n_words; ++__i)
-> 1322                 __a[__i] = 0;
   1323         }
   1324
   1325     template <size_type __a> static

And:

(lldb) disass
cryptest.exe`_GLOBAL__sub_I_chacha_avx.cpp at chacha_avx.cpp:
   0x10016ec30:  pushq  %rbp
   0x10016ec31:  movq   %rsp, %rbp
   0x10016ec34:  pushq  %r14
   0x10016ec36:  pushq  %rbx
   0x10016ec37:  movq   0x24cf8a(%rip), %rax      ; (void
*)0x00000001003e69f0: vtable for CryptoPP::NullNameValuePairs
   0x10016ec3e:  addq   $0x10, %rax
   0x10016ec42:  movq   %rax, 0x2dfb6f(%rip)      ;
CryptoPP::s_nullNameValuePairs
-> 0x10016ec49:  vxorps %xmm0, %xmm0, %xmm0
   0x10016ec4d:  vmovups %xmm0, 0x2dfb6b(%rip)     ; CryptoPP::DEFAULT_CHANNEL
   0x10016ec55:  movq   $0x0, 0x2dfb70(%rip)      ;
CryptoPP::DEFAULT_CHANNEL + 12
   0x10016ec60:  leaq   0x2dfb59(%rip), %rsi      ; CryptoPP::DEFAULT_CHANNEL
   0x10016ec67:  movq   0x24ba52(%rip), %rbx      ; (void
*)0x00007fff8f09321e: std::__1::basic_string<char,
std::__1::char_traits<char>,
std::__1::allocator<char>
>::~basic_string()
   0x10016ec6e:  leaq   -0x16ec75(%rip), %r14     ; _mh_execute_header
   ...

On 12/1/2018 2:17 AM, Jeffrey Walton via llvm-dev wrote:
> I'm testing on older OS X 10.8 with older SSE4 hardware from about
> 2010. I've got updated gear from MacPorts and it includes GCC and
> Clang. GCC is the compiler, and Clang is the assembler.
>
> We perform a compile/link on a test file to ensure an ISA is supported
> by the toolchain. If an ISA is available then we compile a source file
> to the ISA as needed. Then, we guard the higher ISAs at runtime to
> avoid SIGILLs. It worked well until we added AVX2.
>
> For AVX2 we see this as expected:
>
>      $ CXX=/opt/local/bin/clang++-mp-5.0 make
>      /opt/local/bin/clang++-mp-5.0 ... -c chacha.cpp
>      /opt/local/bin/clang++-mp-5.0 ... -mavx2 -c chacha_avx.cpp
>      /opt/local/bin/clang++-mp-5.0 ... -msse2 -c chacha_simd.cpp
>      ...
>
> At runtime we catch a SIGILL due to chacha_avx.cpp as shown below. It
> looks like global constructors are using instructions from AVX
> (vxorps), which is beyond what the machine supports.
>
> How do we tell Clang to use the base ISA for global constructors?
There isn't any way to specifically restrict the ISA for global 
constructors/inline functions/etc.  The inverse works, though: you can 
specify the base ISA for the whole file, then mark specific functions 
using __attribute__((target("avx2"))).

-Eli

-- 
Employee of Qualcomm Innovation Center, Inc.
Qualcomm Innovation Center, Inc. is a member of Code Aurora Forum, a Linux
Foundation Collaborative Project

On Mon, Dec 3, 2018 at 2:30 PM Friedman, Eli <efriedma at codeaurora.org>
wrote:
> ...
> > How do we tell Clang to use the base ISA for global constructors?
>
> There isn't any way to specifically restrict the ISA for global
> constructors/inline functions/etc.  The inverse works, though: you can
> specify the base ISA for the whole file, then mark specific functions
> using __attribute__((target("avx2"))).
It looks like this is becoming more of a problem as CPU advance and
folks try to add multiple implementations.
https://stackoverflow.com/a/24136523/608639 .

The problem with attributes is, it is too new. They did not appear
until GCC 5 for x86_64, and GCC 6 for ARM. They also seem to be
missing for some platforms, like MIPS and PowerPC. We support back to
GCC 3 and Visual Studio 2002 for our sources so we need something more
more available.

Jeff