On Mon, 7 Aug 2017 07:52:29 -0700 Rob Tsuk <[email protected]> wrote: > From 662b17d0d691b567a3343af3439f5fbefe989584 Mon Sep 17 00:00:00 2001 > From: Rob Tsuk <[email protected]> > Date: Mon, 7 Aug 2017 07:11:10 -0700 > Subject: [PATCH 1/1] Fix test for GCC vector extensions > > The previous test was getting a false positive for > clang version 4. > --- > configure.ac | 2 +- > 1 file changed, 1 insertion(+), 1 deletion(-) > > diff --git a/configure.ac b/configure.ac > index e833e45..cbebc82 100644 > --- a/configure.ac > +++ b/configure.ac > @@ -1101,7 +1101,7 @@ support_for_gcc_vector_extensions=no > AC_MSG_CHECKING(for GCC vector extensions) > AC_LINK_IFELSE([AC_LANG_SOURCE([[ > unsigned int __attribute__ ((vector_size(16))) e, a, b; > -int main (void) { e = a - ((b << 27) + (b >> (32 - 27))) + 1; return e[0]; } > +int main (void) { __builtin_shuffle(a,b);e = a - ((b << 27) + (b >> (32 - > 27))) + 1; return e[0]; } > ]])], support_for_gcc_vector_extensions=yes) > > if test x$support_for_gcc_vector_extensions = xyes; then
Thanks, the patch looks mostly good. We indeed have to add the use of __builtin_shuffle() to the configure check code snippet because it is used by pixman and Clang implements this functionality in a different and incompatible way. We did not see this problem with older versions of Clang because it did not support the shift vector by scalar operation in its vector extensions. Too bad that GCC and Clang parted their ways and we can't have a single unified implementation. The whole reason to use GCC vector extension was that we could avoid platform-dependent intrinsics in this code. About the patch. It would be great to use the result of __builtin_shuffle() in the intermediate calculations just in case. See https://cgit.freedesktop.org/pixman/commit/?id=a566f627dbd6ea8f2cba70a446e62caaa2ecbd26 -- Best regards, Siarhei Siamashka _______________________________________________ Pixman mailing list [email protected] https://lists.freedesktop.org/mailman/listinfo/pixman
