| Message ID | 19480.43134.61768.905494@pilspetsen.it.uu.se |
|---|---|
| State | New |
| Headers | show |
On Wed, Jun 16, 2010 at 12:33 PM, Mikael Pettersson <mikpe@it.uu.se> wrote: > This fixes a powerpc testsuite regression on trunk and 4.5. > The fix for PR43949 improved VRP, causing initialization loops > in two vectorization test cases (gcc.dg/vect/slp-perm-{5,6}.c) > to also be vectorized. The scan-tree-dump-times check then sees > too many "vectorized 1 loops" messages, and fails. > > Fixed by placing asm (""); markers in the initialization loops, > as suggested by Richard Guenther in the PR entry. > > Tested on powerpc-linux where it eliminated the new FAILs for > gcc.dg/vect/slp-perm-{5,6}.c. There was no change in testsuite > results on x86_64-linux; the patch does have the same effect on > x86_64 with regard to the number of vectorized loops, but x86_64 > is not a vect_perm target so the number of vectorized loops is > not actually checked there. > > Ok for trunk and 4.5? (I don't have svn write access.) Ok with the if () removed (which was supposed to prevent vectorization). I'll take care of that change and the commit. Thanks, Richard. > /Mikael > > gcc/testsuite/ > > 2010-06-16 Mikael Pettersson <mikpe@it.uu.se> > > PR testsuite/44538 > * gcc.dg/vect/slp-perm-5.c (main): Prevent > initialization loop from being vectorized. > * gcc.dg/vect/slp-perm-6.c (main): Likewise. > > --- gcc-4.6-20100612/gcc/testsuite/gcc.dg/vect/slp-perm-5.c.~1~ 2008-09-04 12:41:50.000000000 +0200 > +++ gcc-4.6-20100612/gcc/testsuite/gcc.dg/vect/slp-perm-5.c 2010-06-15 22:08:10.000000000 +0200 > @@ -63,6 +63,7 @@ int main (int argc, const char* argv[]) > output2[i] = 0; > if (input[i] > 256) > abort (); > + __asm__ volatile (""); > } > > foo (input, output, input2, output2); > --- gcc-4.6-20100612/gcc/testsuite/gcc.dg/vect/slp-perm-6.c.~1~ 2008-09-04 12:41:50.000000000 +0200 > +++ gcc-4.6-20100612/gcc/testsuite/gcc.dg/vect/slp-perm-6.c 2010-06-15 22:08:10.000000000 +0200 > @@ -62,6 +62,7 @@ int main (int argc, const char* argv[]) > output2[i] = 0; > if (input[i] > 256) > abort (); > + __asm__ volatile (""); > } > > foo (input, output, input2, output2); >
--- gcc-4.6-20100612/gcc/testsuite/gcc.dg/vect/slp-perm-5.c.~1~ 2008-09-04 12:41:50.000000000 +0200 +++ gcc-4.6-20100612/gcc/testsuite/gcc.dg/vect/slp-perm-5.c 2010-06-15 22:08:10.000000000 +0200 @@ -63,6 +63,7 @@ int main (int argc, const char* argv[]) output2[i] = 0; if (input[i] > 256) abort (); + __asm__ volatile (""); } foo (input, output, input2, output2); --- gcc-4.6-20100612/gcc/testsuite/gcc.dg/vect/slp-perm-6.c.~1~ 2008-09-04 12:41:50.000000000 +0200 +++ gcc-4.6-20100612/gcc/testsuite/gcc.dg/vect/slp-perm-6.c 2010-06-15 22:08:10.000000000 +0200 @@ -62,6 +62,7 @@ int main (int argc, const char* argv[]) output2[i] = 0; if (input[i] > 256) abort (); + __asm__ volatile (""); } foo (input, output, input2, output2);
This fixes a powerpc testsuite regression on trunk and 4.5. The fix for PR43949 improved VRP, causing initialization loops in two vectorization test cases (gcc.dg/vect/slp-perm-{5,6}.c) to also be vectorized. The scan-tree-dump-times check then sees too many "vectorized 1 loops" messages, and fails. Fixed by placing asm (""); markers in the initialization loops, as suggested by Richard Guenther in the PR entry. Tested on powerpc-linux where it eliminated the new FAILs for gcc.dg/vect/slp-perm-{5,6}.c. There was no change in testsuite results on x86_64-linux; the patch does have the same effect on x86_64 with regard to the number of vectorized loops, but x86_64 is not a vect_perm target so the number of vectorized loops is not actually checked there. Ok for trunk and 4.5? (I don't have svn write access.) /Mikael gcc/testsuite/ 2010-06-16 Mikael Pettersson <mikpe@it.uu.se> PR testsuite/44538 * gcc.dg/vect/slp-perm-5.c (main): Prevent initialization loop from being vectorized. * gcc.dg/vect/slp-perm-6.c (main): Likewise.