Le mardi 6 d?cembre 2011 22:24:49 Pierre-Louis Bossart, vous avez ?crit : > > Writing an autoconf test is OK if you have a fixed target CPU at build- > > time. But that rarely is the case for x86. > > So would it be better to just detect x86, assume at compile time that the > targets supports SSE, Assuming the assembler supports SSE, which is presumably what the autoconf test ensures: (1) If you do use a clobber list and SSE is enabled in CFLAGS, the compiler will generate code that can only run on SSE-capable processors. (2) If you do use a clobber list and SSE is not enabled in CFLAGS, the compiler would fail with "unknown registers". (3) If you do not use a clobber list and SSE is enabled, then weird bugs may appear as the compiler might use the SSE registers for other purposes. (4) If you do not use a clobber list and SSE is not enabled, then it might work. Clearly (2) and (3) are non-starters. (4) is a very bad practice as code breaks silently depending on compiler flags. (1) is hard if you need to retain run-time support for non-SSE. > and then at run time disable the resampling if no > SSE3 is detected when the daemon starts? Yeah sure. But you must somehow ensure that all code compiled with -msse3, NOT ONLY the inline assembly, is only ever run on SSE3-capable processors. Without the target function attribute from GCC 4.4, this becomes a tough requirement for the build system. -- R?mi Denis-Courmont http://www.remlab.net/ http://fi.linkedin.com/in/remidenis
