On Wed, Nov 18, 2015 at 8:35 PM, Siarhei Siamashka <siarhei.siamas...@gmail.com> wrote: > On Wed, 18 Nov 2015 14:22:09 -0800 > Matt Turner <matts...@gmail.com> wrote: > >> On Sun, Oct 11, 2015 at 10:34 AM, Andrea Canciani <ranm...@gmail.com> wrote: >> > On Sun, Oct 11, 2015 at 5:30 AM, Siarhei Siamashka >> > <siarhei.siamas...@gmail.com> wrote: >> >> >> >> On Sun, 11 Oct 2015 04:53:08 +0300 >> >> Siarhei Siamashka <siarhei.siamas...@gmail.com> wrote: >> >> >> >> > On Sat, 10 Oct 2015 16:03:53 -0700 >> >> > Jeremy Huddleston Sequoia <jerem...@freedesktop.org> wrote: >> >> > >> >> > > > On Oct 10, 2015, at 13:48, Andrea Canciani <ranm...@gmail.com> >> >> > > > wrote: >> >> > > > The attached hack gets the code to compile on modern clang, but I >> >> > > > believe first of all we should improve the configure.ac detection >> >> > > > code >> >> > > > so that pixman can actually build both on old and on new clang >> >> > > > versions (possibly with mmx disabled, if the asm constraints we need >> >> > > > are not implemented). >> >> > >> >> > This workaround looks reasonable to me. We should probably just drop >> >> > the whole "ifdef __OPTIMIZE__" part in >> >> > >> >> > http://cgit.freedesktop.org/pixman/tree/pixman/pixman-mmx.c?id=pixman-0.32.8#n92 >> >> > >> >> > I don't quite like the fact that this way of returning results from >> >> > a macro is a GNU C specific extension. But as you said, the configure >> >> > test can be updated to better match the code and also check if the >> >> > compiler supports this particular construct. >> >> > >> >> > Could you please submit the final variant of your patch in a >> >> > "git format-patch" format with a commit message and your >> >> > Signed-off-by tag? >> >> >> >> After looking at this issue a bit more, I realized that we are >> >> about to add a second layer of workarounds on top of the existing >> >> old workarounds :-) >> > >> > >> > The attached patch should fix the issue with only minor changes. >> > It keeps the workarounds :( but somewhat it simplifies them :) >> > I followed your suggestion of checking&using block expressions. >> > Given that the _mm_shuffle_pi16() function is always used in a "return" >> > statement, if needed we could avoid the usage of block expressions by >> > defining a macro "_return_mm_shuffle_pi16()" (which would return the result >> > of the operation instead of making it available as an expression) both for >> > the xmmintrin branch and for the hand-coded one. >> > >> >> The original problem is that certain compilers (just GCC?) did not >> >> support some intrinsics when compiling MMX code (_mm_movemask_pi8, >> >> _mm_mulhi_pu16, _mm_shuffle_pi16) and we got the following code: >> >> >> >> http://cgit.freedesktop.org/pixman/tree/pixman/pixman-mmx.c?id=pixman-0.32.8#n66 >> >> >> >> In fact, these instructions were not available as part of the original >> >> MMX, but only got introduced later with AMD Extended 3DNow! and Intel >> >> SSE1. This is mentioned in the commit messages: >> >> >> >> http://cgit.freedesktop.org/pixman/commit/?id=84221f4c1687b8ea14e9cbdc78b2ba7258e62c9e >> >> >> >> http://cgit.freedesktop.org/pixman/commit/?id=14208344964f341a7b4a704b05cf4804c23792e9 >> >> >> >> These extra instructions are unofficially known as MMX2. But GCC does >> >> not have a separate option for "-mmmx2". Instead the GCC manual says >> >> that these intrinsics are available when either "-msse" or a >> >> combination of "-m3dnow -march=athlon" is used: >> >> >> >> https://gcc.gnu.org/onlinedocs/gcc-5.2.0/gcc/x86-Built-in-Functions.html#x86-Built-in-Functions >> >> >> >> >> >> Now I wonder if the comment "We have to compile with -msse to use >> >> xmmintrin.h" is still valid. I tried to tweak the following ifdef to >> >> use the part of code, which includes <xmmintrin.h> and the it compiled >> >> fine for me with CFLAGS="-O2 -m32" using recent versions of GCC and >> >> Clang: >> >> >> >> http://cgit.freedesktop.org/pixman/tree/pixman/pixman-mmx.c?id=pixman-0.32.8#n63 >> >> >> >> I believe that this might be somehow related to the new __ALL_ISA__ >> >> define, which had been mentioned in 2013: >> >> https://gcc.gnu.org/ml/gcc-patches/2013-04/txts5M0c0uU9y.txt >> >> >> >> So what about just dropping this ugly stuff and adding a configure >> >> check, which would verify if the MMX code can include <xmmintrin.h>? >> > >> > >> > I would love getting rid of the workarounds, but I'm somewhat worried about >> > the possibility of regressions. >> > If you believe is a valid option, we might definitely try to pursue it. >> > >> > What is the best way forward? >> >> I've now reverted my commit and pushed yours. > > Oh, thanks. Though it is always a good idea to give a short notice > before landing patches to git, especially controversial ones. > > What we had in pixman-0.33.4 was not exactly bad. Sure, GCC did not > detect MMX support, but the other compilers had no problems with it. > In a way, this was a good thing because GCC is known to have a broken > _mm_empty() intrinsic handling and it is probably not going to be > ever fixed: > > https://gcc.gnu.org/PR47759 > > In the case of pixman, it happened to cause problems on more than > one occasion. We do have a test in the test suite, which can detect > x87 registers corruption caused by a misplaced EMMS instruction due > to compiler optimizations. But not everyone runs the test suite for > 32-bit pixman builds on x86 with PIXMAN_DISABLE="sse2 ssse3" > regularly enough. > > It is not like we want to retire the MMX support just because it > is old and out of fashion. The reason is that it is a source of > recurring troubles. Sigh. Let's see how long will it take until > the MMX code breaks again with one compiler version or another.
I had a look at the GCC code for handling -mmmx/-m3dnow flags. It doesn't look difficult to add a flag for -munion-of-sse-and-3dnowa (-mmmx2?). I think I'll give that a try, but the most difficult part might finding a good name. :) _______________________________________________ Pixman mailing list Pixman@lists.freedesktop.org http://lists.freedesktop.org/mailman/listinfo/pixman