On Sun, Dec 11, 2011 at 8:46 AM, Marek Olšák <mar...@gmail.com> wrote: > On Mon, Nov 21, 2011 at 6:52 PM, Eric Anholt <e...@anholt.net> wrote: >> On Sun, 20 Nov 2011 15:08:55 +0100, Marek Olšák <mar...@gmail.com> wrote: >>> unpack_float_z_Z24_X8 fails with -O2 in: >>> - fbo-blit-d24s8 >>> - fbo-depth-sample-compare >>> - fbo-readpixels-depth-formats >>> - glean/depthStencil >>> >>> With -O0, it works fine. >>> >>> I am removing all the assertions. There's not much point in having them, >>> is there? >> >> Is -ffast-math involved at all? > > Yes, -fno-fast-math makes the problem go away. > >> >> At a guess, replacing "* scale" with "/ (float)0xffffff" makes the >> failure happen either way? > > Yes. Only using GLdouble fixes it. > > It makes sense. The mantissa without the sign has 23 bits, but 24 bits > are required to exactly represent 0xffffff if I am not mistaken. > > I have found a couple more cases where it could fail in the same way > and fixed them in the attached patch. Please review.
Looks good. Thanks for fixing this. I guess it would have saved you some work if we'd have #defined some Z24SCALE and Z32SCALE constants at the top of the file in the first place... Reviewed-by: Brian Paul <bri...@vmware.com> _______________________________________________ mesa-dev mailing list mesa-dev@lists.freedesktop.org http://lists.freedesktop.org/mailman/listinfo/mesa-dev
