| Message ID | 4E406F1F.8030508@redhat.com |
|---|---|
| State | New |
| Headers | show |
On Mon, Aug 8, 2011 at 4:19 PM, Vladimir Makarov <vmakarov@redhat.com> wrote: > The following patch fixes PR49990. The problem is described on > > http://gcc.gnu.org/bugzilla/show_bug.cgi?id=49936 > > Reg classes which can not change modes for some pseudo were excluded from > the consideration. As I wrote recently, they should not. Instead the > correct cost for changing mode (by moving through memory or other class > register) should be taken into the account. I believe the cost is already > calculated rightly for this case, fortunately. > > The patch was successfully bootstrapped on x86-64 and ppc64. Actually I did > not find a difference in generated code on variety tests on x86/x86-64 and > arm (that is what I tried). > > Committed as rev. 177575. > > > > 2011-08-08 Vladimir Makarov <vmakarov@redhat.com> > > PR rtl-optimization/49990 > * ira-costs.c (print_allocno_costs, print_pseudo_costs): Don't > ignore classes which can not change mode. > (find_costs_and_classes): Ditto. > > It breaks GCC on Linux/x86: http://gcc.gnu.org/bugzilla/show_bug.cgi?id=50026
On 08/08/2011 10:35 PM, H.J. Lu wrote: > On Mon, Aug 8, 2011 at 4:19 PM, Vladimir Makarov<vmakarov@redhat.com> wrote: >> The following patch fixes PR49990. The problem is described on >> >> http://gcc.gnu.org/bugzilla/show_bug.cgi?id=49936 >> >> Reg classes which can not change modes for some pseudo were excluded from >> the consideration. As I wrote recently, they should not. Instead the >> correct cost for changing mode (by moving through memory or other class >> register) should be taken into the account. I believe the cost is already >> calculated rightly for this case, fortunately. >> >> The patch was successfully bootstrapped on x86-64 and ppc64. Actually I did >> not find a difference in generated code on variety tests on x86/x86-64 and >> arm (that is what I tried). >> >> Committed as rev. 177575. >> >> >> >> 2011-08-08 Vladimir Makarov<vmakarov@redhat.com> >> >> PR rtl-optimization/49990 >> * ira-costs.c (print_allocno_costs, print_pseudo_costs): Don't >> ignore classes which can not change mode. >> (find_costs_and_classes): Ditto. >> >> > It breaks GCC on Linux/x86: > > http://gcc.gnu.org/bugzilla/show_bug.cgi?id=50026 > > Thanks for reporting. I reverted the patch although I can not still reproduce it on my machine.
On Tue, Aug 9, 2011 at 7:37 AM, Vladimir Makarov <vmakarov@redhat.com> wrote: > On 08/08/2011 10:35 PM, H.J. Lu wrote: >> >> On Mon, Aug 8, 2011 at 4:19 PM, Vladimir Makarov<vmakarov@redhat.com> >> wrote: >>> >>> The following patch fixes PR49990. The problem is described on >>> >>> http://gcc.gnu.org/bugzilla/show_bug.cgi?id=49936 >>> >>> Reg classes which can not change modes for some pseudo were excluded from >>> the consideration. As I wrote recently, they should not. Instead the >>> correct cost for changing mode (by moving through memory or other class >>> register) should be taken into the account. I believe the cost is >>> already >>> calculated rightly for this case, fortunately. >>> >>> The patch was successfully bootstrapped on x86-64 and ppc64. Actually I >>> did >>> not find a difference in generated code on variety tests on x86/x86-64 >>> and >>> arm (that is what I tried). >>> >>> Committed as rev. 177575. >>> >>> >>> >>> 2011-08-08 Vladimir Makarov<vmakarov@redhat.com> >>> >>> PR rtl-optimization/49990 >>> * ira-costs.c (print_allocno_costs, print_pseudo_costs): Don't >>> ignore classes which can not change mode. >>> (find_costs_and_classes): Ditto. >>> >>> >> It breaks GCC on Linux/x86: >> >> http://gcc.gnu.org/bugzilla/show_bug.cgi?id=50026 >> >> > Thanks for reporting. > > I reverted the patch although I can not still reproduce it on my machine. > My configurations are --enable-clocale=gnu --with-system-zlib --enable-shared --with-demangler-in-ld --enable-cloog-backend=isl --with-ppl-include=/opt/gnu/include --with-ppl-lib=/opt/gnu/lib64 --with-cloog-include=/opt/gnu/include --with-cloog-lib=/opt/gnu/lib64 --with-fpmath=sse --enable-languages=c,c++,fortran,java,lto,objc and --with-arch=corei7 --with-cpu=corei7 --enable-clocale=gnu --with-system-zlib --enable-shared --with-demangler-in-ld --enable-cloog-backend=isl --with-ppl-include=/opt/gnu/include --with-ppl-lib=/opt/gnu/lib64 --with-cloog-include=/opt/gnu/include --with-cloog-lib=/opt/gnu/lib64 --with-fpmath=sse --enable-languages=c,c++,fortran,java,lto,objc on Fedora 15/Intel Core i7. Thanks.
On 08/09/2011 10:50 AM, H.J. Lu wrote: > On Tue, Aug 9, 2011 at 7:37 AM, Vladimir Makarov<vmakarov@redhat.com> wrote: >> On 08/08/2011 10:35 PM, H.J. Lu wrote: >>> On Mon, Aug 8, 2011 at 4:19 PM, Vladimir Makarov<vmakarov@redhat.com> >>> wrote: >>>> The following patch fixes PR49990. The problem is described on >>>> >>>> http://gcc.gnu.org/bugzilla/show_bug.cgi?id=49936 >>>> >>>> Reg classes which can not change modes for some pseudo were excluded from >>>> the consideration. As I wrote recently, they should not. Instead the >>>> correct cost for changing mode (by moving through memory or other class >>>> register) should be taken into the account. I believe the cost is >>>> already >>>> calculated rightly for this case, fortunately. >>>> >>>> The patch was successfully bootstrapped on x86-64 and ppc64. Actually I >>>> did >>>> not find a difference in generated code on variety tests on x86/x86-64 >>>> and >>>> arm (that is what I tried). >>>> >>>> Committed as rev. 177575. >>>> >>>> >>>> >>>> 2011-08-08 Vladimir Makarov<vmakarov@redhat.com> >>>> >>>> PR rtl-optimization/49990 >>>> * ira-costs.c (print_allocno_costs, print_pseudo_costs): Don't >>>> ignore classes which can not change mode. >>>> (find_costs_and_classes): Ditto. >>>> >>>> >>> It breaks GCC on Linux/x86: >>> >>> http://gcc.gnu.org/bugzilla/show_bug.cgi?id=50026 >>> >>> >> Thanks for reporting. >> >> I reverted the patch although I can not still reproduce it on my machine. >> > My configurations are > > --enable-clocale=gnu --with-system-zlib --enable-shared > --with-demangler-in-ld --enable-cloog-backend=isl > --with-ppl-include=/opt/gnu/include --with-ppl-lib=/opt/gnu/lib64 > --with-cloog-include=/opt/gnu/include --with-cloog-lib=/opt/gnu/lib64 > --with-fpmath=sse --enable-languages=c,c++,fortran,java,lto,objc > > and > > --with-arch=corei7 --with-cpu=corei7 --enable-clocale=gnu > --with-system-zlib --enable-shared --with-demangler-in-ld > --enable-cloog-backend=isl --with-ppl-include=/opt/gnu/include > --with-ppl-lib=/opt/gnu/lib64 --with-cloog-include=/opt/gnu/include > --with-cloog-lib=/opt/gnu/lib64 --with-fpmath=sse > --enable-languages=c,c++,fortran,java,lto,objc Thanks, HJ. I reproduced it now.
Index: ira-costs.c =================================================================== --- ira-costs.c (revision 177573) +++ ira-costs.c (working copy) @@ -1367,11 +1367,7 @@ print_allocno_costs (FILE *f) for (k = 0; k < cost_classes_ptr->num; k++) { rclass = cost_classes[k]; - if (contains_reg_of_mode[rclass][PSEUDO_REGNO_MODE (regno)] -#ifdef CANNOT_CHANGE_MODE_CLASS - && ! invalid_mode_change_p (regno, (enum reg_class) rclass) -#endif - ) + if (contains_reg_of_mode[rclass][PSEUDO_REGNO_MODE (regno)]) { fprintf (f, " %s:%d", reg_class_names[rclass], COSTS (costs, i)->cost[k]); @@ -1409,11 +1405,7 @@ print_pseudo_costs (FILE *f) for (k = 0; k < cost_classes_ptr->num; k++) { rclass = cost_classes[k]; - if (contains_reg_of_mode[rclass][PSEUDO_REGNO_MODE (regno)] -#ifdef CANNOT_CHANGE_MODE_CLASS - && ! invalid_mode_change_p (regno, (enum reg_class) rclass) -#endif - ) + if (contains_reg_of_mode[rclass][PSEUDO_REGNO_MODE (regno)]) fprintf (f, " %s:%d", reg_class_names[rclass], COSTS (costs, regno)->cost[k]); } @@ -1650,11 +1642,7 @@ find_costs_and_classes (FILE *dump_file) rclass = cost_classes[k]; /* Ignore classes that are too small or invalid for this operand. */ - if (! contains_reg_of_mode[rclass][PSEUDO_REGNO_MODE (i)] -#ifdef CANNOT_CHANGE_MODE_CLASS - || invalid_mode_change_p (i, (enum reg_class) rclass) -#endif - ) + if (! contains_reg_of_mode[rclass][PSEUDO_REGNO_MODE (i)]) continue; if (i_costs[k] < best_cost) { @@ -1725,11 +1713,7 @@ find_costs_and_classes (FILE *dump_file) continue; /* Ignore classes that are too small or invalid for this operand. */ - if (! contains_reg_of_mode[rclass][PSEUDO_REGNO_MODE (i)] -#ifdef CANNOT_CHANGE_MODE_CLASS - || invalid_mode_change_p (i, (enum reg_class) rclass) -#endif - ) + if (! contains_reg_of_mode[rclass][PSEUDO_REGNO_MODE (i)]) ; else if (total_a_costs[k] < best_cost) {
The following patch fixes PR49990. The problem is described on http://gcc.gnu.org/bugzilla/show_bug.cgi?id=49936 Reg classes which can not change modes for some pseudo were excluded from the consideration. As I wrote recently, they should not. Instead the correct cost for changing mode (by moving through memory or other class register) should be taken into the account. I believe the cost is already calculated rightly for this case, fortunately. The patch was successfully bootstrapped on x86-64 and ppc64. Actually I did not find a difference in generated code on variety tests on x86/x86-64 and arm (that is what I tried). Committed as rev. 177575. 2011-08-08 Vladimir Makarov <vmakarov@redhat.com> PR rtl-optimization/49990 * ira-costs.c (print_allocno_costs, print_pseudo_costs): Don't ignore classes which can not change mode. (find_costs_and_classes): Ditto.