Re: Build failure with clang on older CPU :Illegal instruction

Hi,

On 11. 11. 19 09:46, Luboš Luňák wrote:
>   The guilty commit is actually
> https://cgit.freedesktop.org/libreoffice/core/commit/?id=f43f9b99603736a4d54f550052509eb5f4d04b45 .
>
>   The INTRINSICS_CXXFLAGS variable is misdesigned or misused. What happens is
> that a source file gets unconditionally compiled with whatever configure
> finds to be the most advanced instruction set, and then the code from the
> compiled source gets executed unconditionally, without a runtime check. A
> runtime check is necessary, as flags like -mavx2 allow the compiler to use
> newer instructions even for source code that doesn't explicitly use them.
>
>   Looking at tools/qa/cppunit/test_cpuid.cxx, it seems that it doesn't need any
> special CXXFLAGS, as it doesn't use them. And if it does need them, then
> those parts need to be split to one source file per instruction set, each
> compiled with its matching CXXFLAGS, and then another source file (compiled
> normally) needs to call them only after doing a runtime check. See the
> description of -mavx2 etc. in 'man gcc' and see 'git grep SSE2 sc/' for an
> example.

Right, the test is quite naive and use of INTRINSICS_CXXFLAGS doesn't 
make sense here as it tests the runtime detection only, which is not 
dependent on the compiler flags (it just reports what the CPU supports). 
Probably this was needed in the past because compile and runtime 
detection was mixed with compile time detection, so if the compiler 
flags didn't return support for an instruction set, then runtime 
detection wasn't even performed, but this is not the case with the 
latest code anymore.

I fixed this problem with the test with [1] patch and I'll make a proper 
test at a later date to demonstrate how this is supposed to work.

[1] https://gerrit.libreoffice.org/#/c/82422/

Regards, Tomaž