On 3/21/06, Daniel Berlin <[EMAIL PROTECTED]> wrote: > On Tue, 2006-03-21 at 17:30 -0500, Andrew MacLeod wrote: > > > > I seem to have narrowed it down to this patch: > > > > http://gcc.gnu.org/ml/gcc-patches/2006-01/msg00908.html > > > > That's quite a while ago :). > > > > > > > Dan, this appear to *not* be compile time neutral: > > > > Timings on this patch show that it is not faster or slower than > > what we > > do now (even with the removal of the call clobbering patch). This > > is > > true even on fortran tests i had that clobber a lot of stuff. > > > > > > > running cpgram.ii shows a regression: > > > > before patch: > > > > tree alias analysis : 2.49 ( 7%) usr 0.25 ( 5%) sys 6.13 ( 5%) > > wall 4971 kB ( 1%) ggc > > TOTAL : 36.90 4.72 130.34 > > 467341 kB > > > > after patch: > > > > tree alias analysis : 59.00 (63%) usr 0.40 ( 7%) sys 70.43 (36%) > > wall 4957 kB ( 1%) ggc > > TOTAL : 94.13 5.43 193.85 > > 468339 kB > > > > > on a 386 linux machine bootstrapped with checking disabled. > > Can you send me cpgram.ii, so i can look into it? > > i will note the patch is pretty much required for correctness. We were > getting seriously wrong answers before in some cases.
Maybe someone can have a look at the attribute((pointer_no_escape)) patch I posted a while ago. With some IPA machinery we could possibly trim down the clobber lists quite a bit. Richard.