From: "Cui, Lili" <lili....@intel.com>

Hi,

This patch aims to remove issue code in improve_allocation that was causing 
expensive allocno overflows.

Bootstrapped and regtested on x86_64-pc-linux-gnu, OK for trunk?

Thanks,
Lili.



The original intention of this code was to allow more allocnos
to share the same register, but this led to expensive allocno
overflows. Extracted a small case (a bit large, see Bugzilla
PR117838 for details) from 548.exchange2_r to analyze this
register allocation issue.

Before improve_allocation function:

a537 (cost 1896, reg42)
a20  (cost 270, reg1)
a13  (cost 144, spill)
a551 (cost 70, reg40)
a5   (cost 43, spill)
a493 (cost 30, reg42)
a499 (cost 12, reg40)

------------------------------
Dump info in improve_allocation function:

Base:
Spilling a493r125 for a5r113
Spilling a573r202 for a5r113
Spilling a499r248 for a13r106
Spilling a551r120 for a13r106
Spilling a20r237 for a551r120

With patch:
Spilling a499r248 for a13r106
Spilling a551r120 for a13r106
Spilling a493r125 for a551r120
------------------------------

After assign_hard_reg (at the end of improve_allocation):

Base:
a537 (cost 1896, reg1)
a20  (cost 270, spill) -----> This is unreasonable
a13  (cost 144, reg40)
a551 (cost 70, reg1)
a5   (cost 43, reg42)
a493 (cost 30, spill)
a499 (cost 12, reg1)

With patch:
a537 (cost 1896, reg42)
a20  (cost 270, reg1)
a13  (cost 144, reg40)
a551 (cost 70, reg42)
a5   (cost 43, spill)
a493 (cost 30, spill)
a499 (cost 12, reg42)
-----------------------------

Collected spec2017 performance on Znver3/Graviton4/EMR/SRF for O2 and Ofast.
No performance regression was observed.

FOR multi-copy O2
SRF: 548.exchange2_r increased by 7.5%, 500.perlbench_r increased by 2.0%.
EMR: 548.exchange2_r increased by 4.5%, 500.perlbench_r increased by 1.7%.
Graviton4: 548.exchange2_r Increased by 2.2%, 511.povray_r increased by 2.8%.
Znver3 : 500.perlbench_r increased by 2.0%.

gcc/ChangeLog:

        PR rtl-optimization/117838
        * ira-color.cc (improve_allocation): Remove the issue code.
---
 gcc/ira-color.cc | 41 +++++++++++------------------------------
 1 file changed, 11 insertions(+), 30 deletions(-)

diff --git a/gcc/ira-color.cc b/gcc/ira-color.cc
index 4b9296029cc..fa2ea61cadf 100644
--- a/gcc/ira-color.cc
+++ b/gcc/ira-color.cc
@@ -3304,8 +3304,6 @@ improve_allocation (void)
           assigning hard register to allocno A even without spilling
           conflicting allocnos.  */
        continue;
-      auto_bitmap allocnos_to_spill;
-      HARD_REG_SET soft_conflict_regs = {};
       mode = ALLOCNO_MODE (a);
       nwords = ALLOCNO_NUM_OBJECTS (a);
       /* Process each allocno conflicting with A and update the cost
@@ -3331,40 +3329,24 @@ improve_allocation (void)
              ALLOCNO_COLOR_DATA (conflict_a)->temp = check;
              if ((conflict_hregno = ALLOCNO_HARD_REGNO (conflict_a)) < 0)
                continue;
-             auto spill_a = ira_soft_conflict (a, conflict_a);
-             if (spill_a)
-               {
-                 if (!bitmap_set_bit (allocnos_to_spill,
-                                      ALLOCNO_NUM (spill_a)))
-                   continue;
-                 ira_loop_border_costs border_costs (spill_a);
-                 spill_cost = border_costs.spill_inside_loop_cost ();
-               }
+             spill_cost = ALLOCNO_UPDATED_MEMORY_COST (conflict_a);
+             k = (ira_class_hard_reg_index
+                  [ALLOCNO_CLASS (conflict_a)][conflict_hregno]);
+             ira_assert (k >= 0);
+             if ((allocno_costs = ALLOCNO_HARD_REG_COSTS (conflict_a))
+                 != NULL)
+               spill_cost -= allocno_costs[k];
              else
-               {
-                 spill_cost = ALLOCNO_UPDATED_MEMORY_COST (conflict_a);
-                 k = (ira_class_hard_reg_index
-                      [ALLOCNO_CLASS (conflict_a)][conflict_hregno]);
-                 ira_assert (k >= 0);
-                 if ((allocno_costs = ALLOCNO_HARD_REG_COSTS (conflict_a))
-                     != NULL)
-                   spill_cost -= allocno_costs[k];
-                 else
-                   spill_cost -= ALLOCNO_UPDATED_CLASS_COST (conflict_a);
-                 spill_cost
-                   += allocno_copy_cost_saving (conflict_a, conflict_hregno);
-               }
+               spill_cost -= ALLOCNO_UPDATED_CLASS_COST (conflict_a);
+             spill_cost
+               += allocno_copy_cost_saving (conflict_a, conflict_hregno);
              conflict_nregs = hard_regno_nregs (conflict_hregno,
                                                 ALLOCNO_MODE (conflict_a));
              auto note_conflict = [&](int r)
                {
                  if (check_hard_reg_p (a, r,
                                        conflicting_regs, profitable_hard_regs))
-                   {
-                     if (spill_a)
-                       SET_HARD_REG_BIT (soft_conflict_regs, r);
-                     costs[r] += spill_cost;
-                   }
+                   costs[r] += spill_cost;
                };
              for (r = conflict_hregno;
                   r >= 0 && (int) end_hard_regno (mode, r) > conflict_hregno;
@@ -3396,7 +3378,6 @@ improve_allocation (void)
           by spilling some conflicting allocnos does not improve the
           allocation cost.  */
        continue;
-      spill_soft_conflicts (a, allocnos_to_spill, soft_conflict_regs, best);
       nregs = hard_regno_nregs (best, mode);
       /* Now spill conflicting allocnos which contain a hard register
         of A when we assign the best chosen hard register to it.  */
-- 
2.34.1

Reply via email to