| Message ID | 4D9E1369.5090205@redhat.com |
|---|---|
| State | New |
| Headers | show |
-----BEGIN PGP SIGNED MESSAGE----- Hash: SHA1 On 04/07/11 13:41, Vladimir Makarov wrote: > The following patch should solve problem > http://gcc.gnu.org/bugzilla/show_bug.cgi?id=48435. It might solve other > latest IRA problems too including performance related ones. The patch > is for targets which require some values to be placed in hard registers > starting with an even (odd) hard registers. > > I already addressed an analogous problem recently and the patch I sent > that time although solved some problems it created even more new > problems. I should acknowledge this. > > The problem was in that profitable hard regs were used for colorability > criterion and finding hard registers where allocno values can be resided > but *also* as starting allocno hard registers. It resulted in spilling > allocnos which should be placed in multi-registers starting on a > specific border because profitable hard registers were only starting > registers and when we calculated number of available hard registers > multi-register allocnos can not fit only in their starting hard registers. > > The following patch was successfully bootstrapped on x86/x86-64 and on > i686 with H.J.'s autotester options. > > OK to commit? > > 2011-04-07 Vladimir Makarov <vmakarov@redhat.com> > > PR 4435 > * ira-color.c (setup_profitable_hard_regs): Add comments. > Don't take prohibited hard regs into account. > (setup_conflict_profitable_regs): Rename to > get_conflict_profitable_regs. > (check_hard_reg_p): Check prohibited hard regs. OK. jeff -----BEGIN PGP SIGNATURE----- Version: GnuPG v1.4.11 (GNU/Linux) Comment: Using GnuPG with Fedora - http://enigmail.mozdev.org/ iQEcBAEBAgAGBQJNnycrAAoJEBRtltQi2kC70DsH/A7fnYlawxDvYPeULGb1ctQD 6LYWBWhJY4UDTiUNjLEzwALpq2XccVzubz1KEMI3wOndIHHYM2ykbsgTqCKmXqhf ZthBjzNUtE/i7jSxK8JK3SEShqs2j7vaAvlkUKbdF7O61rhQap8MRnD2umHfrbet PFea7MedpaC96mazzoyUWlzwWCt2CQhOlMFGlC44bxB1FrKy0nUjeIM7kWqjqG5W Y9lV8RlC4O2cLST5qHpqtssLSO4omPlKsEUBEKW4E/87UCoFAxP0n8NOZ1EFgMsc 3SlhSv+Dpinfh79gjHXt5+CH33bXTecf8Jfyd89Rn6huzL9OhAXJy+MjMi6cBm0= =KSaH -----END PGP SIGNATURE-----
Vladimir, The wrong PR 4435 was in referrenced the commit --- On Fri, 8/4/11, Jeff Law <law@redhat.com> wrote: > From: Jeff Law <law@redhat.com> > Subject: Re: patch to fix PR 48435 > To: "Vladimir Makarov" <vmakarov@redhat.com> > Cc: "gcc-patches" <gcc-patches@gcc.gnu.org> > Date: Friday, 8 April, 2011, 16:18 > -----BEGIN PGP SIGNED MESSAGE----- > Hash: SHA1 > > On 04/07/11 13:41, Vladimir Makarov wrote: > > The following patch should solve problem > > http://gcc.gnu.org/bugzilla/show_bug.cgi?id=48435. > It might solve other > > latest IRA problems too including performance related > ones. The patch > > is for targets which require some values to be placed > in hard registers > > starting with an even (odd) hard registers. > > > > I already addressed an analogous problem recently and > the patch I sent > > that time although solved some problems it created > even more new > > problems. I should acknowledge this. > > > > The problem was in that profitable hard regs were used > for colorability > > criterion and finding hard registers where allocno > values can be resided > > but *also* as starting allocno hard registers. > It resulted in spilling > > allocnos which should be placed in multi-registers > starting on a > > specific border because profitable hard registers were > only starting > > registers and when we calculated number of available > hard registers > > multi-register allocnos can not fit only in their > starting hard registers. > > > > The following patch was successfully bootstrapped on > x86/x86-64 and on > > i686 with H.J.'s autotester options. > > > > OK to commit? > > > > 2011-04-07 Vladimir Makarov <vmakarov@redhat.com> > > > > PR 4435 > > * ira-color.c > (setup_profitable_hard_regs): Add comments. > > Don't take prohibited hard > regs into account. > > > (setup_conflict_profitable_regs): Rename > to > > get_conflict_profitable_regs. > > (check_hard_reg_p): Check > prohibited hard regs. > OK. > jeff > -----BEGIN PGP SIGNATURE----- > Version: GnuPG v1.4.11 (GNU/Linux) > Comment: Using GnuPG with Fedora - http://enigmail.mozdev.org/ > > iQEcBAEBAgAGBQJNnycrAAoJEBRtltQi2kC70DsH/A7fnYlawxDvYPeULGb1ctQD > 6LYWBWhJY4UDTiUNjLEzwALpq2XccVzubz1KEMI3wOndIHHYM2ykbsgTqCKmXqhf > ZthBjzNUtE/i7jSxK8JK3SEShqs2j7vaAvlkUKbdF7O61rhQap8MRnD2umHfrbet > PFea7MedpaC96mazzoyUWlzwWCt2CQhOlMFGlC44bxB1FrKy0nUjeIM7kWqjqG5W > Y9lV8RlC4O2cLST5qHpqtssLSO4omPlKsEUBEKW4E/87UCoFAxP0n8NOZ1EFgMsc > 3SlhSv+Dpinfh79gjHXt5+CH33bXTecf8Jfyd89Rn6huzL9OhAXJy+MjMi6cBm0= > =KSaH > -----END PGP SIGNATURE----- >
On 04/08/2011 01:28 PM, Graham Stott wrote: > Vladimir, > > The wrong PR 4435 was in referrenced the commit > Thanks, Graham. I've just fixed it. I should change my keyboard. This is not the first time when some symbols are missed.
Index: ira-color.c =================================================================== --- ira-color.c (revision 172107) +++ ira-color.c (working copy) @@ -1057,6 +1057,8 @@ setup_profitable_hard_regs (void) enum reg_class aclass; enum machine_mode mode; + /* Initial set up from allocno classes and explicitly conflicting + hard regs. */ EXECUTE_IF_SET_IN_BITMAP (coloring_allocno_bitmap, 0, i, bi) { a = ira_allocnos[i]; @@ -1076,9 +1078,6 @@ setup_profitable_hard_regs (void) { COPY_HARD_REG_SET (obj_data->profitable_hard_regs, reg_class_contents[aclass]); - AND_COMPL_HARD_REG_SET - (obj_data->profitable_hard_regs, - ira_prohibited_class_mode_regs[aclass][mode]); AND_COMPL_HARD_REG_SET (obj_data->profitable_hard_regs, ira_no_alloc_regs); AND_COMPL_HARD_REG_SET (obj_data->profitable_hard_regs, @@ -1086,6 +1085,7 @@ setup_profitable_hard_regs (void) } } } + /* Exclude hard regs already assigned for conflicting objects. */ EXECUTE_IF_SET_IN_BITMAP (consideration_allocno_bitmap, 0, i, bi) { a = ira_allocnos[i]; @@ -1124,6 +1124,7 @@ setup_profitable_hard_regs (void) } } } + /* Exclude too costly hard regs. */ EXECUTE_IF_SET_IN_BITMAP (coloring_allocno_bitmap, 0, i, bi) { int min_cost = INT_MAX; @@ -1451,9 +1452,9 @@ update_conflict_hard_regno_costs (int *c profitable regs exclude hard regs which can not hold value of mode of allocno A. */ static inline void -setup_conflict_profitable_regs (ira_allocno_t a, bool retry_p, - HARD_REG_SET *conflict_regs, - HARD_REG_SET *profitable_regs) +get_conflict_profitable_regs (ira_allocno_t a, bool retry_p, + HARD_REG_SET *conflict_regs, + HARD_REG_SET *profitable_regs) { int i, nwords; ira_object_t obj; @@ -1485,8 +1486,15 @@ check_hard_reg_p (ira_allocno_t a, int h HARD_REG_SET *conflict_regs, HARD_REG_SET *profitable_regs) { int j, nwords, nregs; + enum reg_class aclass; + enum machine_mode mode; - nregs = hard_regno_nregs[hard_regno][ALLOCNO_MODE (a)]; + aclass = ALLOCNO_CLASS (a); + mode = ALLOCNO_MODE (a); + if (TEST_HARD_REG_BIT (ira_prohibited_class_mode_regs[aclass][mode], + hard_regno)) + return false; + nregs = hard_regno_nregs[hard_regno][mode]; nwords = ALLOCNO_NUM_OBJECTS (a); for (j = 0; j < nregs; j++) { @@ -1554,8 +1562,8 @@ assign_hard_reg (ira_allocno_t a, bool r #endif ira_assert (! ALLOCNO_ASSIGNED_P (a)); - setup_conflict_profitable_regs (a, retry_p, - conflicting_regs, profitable_hard_regs); + get_conflict_profitable_regs (a, retry_p, + conflicting_regs, profitable_hard_regs); aclass = ALLOCNO_CLASS (a); class_size = ira_class_hard_regs_num[aclass]; best_hard_regno = -1; @@ -2233,7 +2241,8 @@ setup_allocno_available_regs_num (ira_al ira_object_t obj = ALLOCNO_OBJECT (a, k); object_color_data_t obj_data = OBJECT_COLOR_DATA (obj); - /* Checking only profitable hard regs. */ + /* Checking only profitable hard regs which exclude + object's conflict hard regs. */ if (TEST_HARD_REG_BIT (OBJECT_TOTAL_CONFLICT_HARD_REGS (obj), hard_regno + j) || ! TEST_HARD_REG_BIT (obj_data->profitable_hard_regs, @@ -2403,8 +2412,8 @@ improve_allocation (void) else base_cost = allocno_costs[ira_class_hard_reg_index[aclass][hregno]]; try_p = false; - setup_conflict_profitable_regs (a, false, - conflicting_regs, profitable_hard_regs); + get_conflict_profitable_regs (a, false, + conflicting_regs, profitable_hard_regs); class_size = ira_class_hard_regs_num[aclass]; /* Set up cost improvement for usage of each profitable hard register for allocno A. */
The following patch should solve problem http://gcc.gnu.org/bugzilla/show_bug.cgi?id=48435. It might solve other latest IRA problems too including performance related ones. The patch is for targets which require some values to be placed in hard registers starting with an even (odd) hard registers. I already addressed an analogous problem recently and the patch I sent that time although solved some problems it created even more new problems. I should acknowledge this. The problem was in that profitable hard regs were used for colorability criterion and finding hard registers where allocno values can be resided but *also* as starting allocno hard registers. It resulted in spilling allocnos which should be placed in multi-registers starting on a specific border because profitable hard registers were only starting registers and when we calculated number of available hard registers multi-register allocnos can not fit only in their starting hard registers. The following patch was successfully bootstrapped on x86/x86-64 and on i686 with H.J.'s autotester options. OK to commit? 2011-04-07 Vladimir Makarov <vmakarov@redhat.com> PR 4435 * ira-color.c (setup_profitable_hard_regs): Add comments. Don't take prohibited hard regs into account. (setup_conflict_profitable_regs): Rename to get_conflict_profitable_regs. (check_hard_reg_p): Check prohibited hard regs.