[RFC,IRA] Fix PR87507, IRA unnecessarily uses non-volatile registers during register assignment

Message ID c72e8e4d-2a2b-fd42-01ad-1f7c4c8fba52@linux.ibm.com
State New
Headers show
Series
  • [RFC,IRA] Fix PR87507, IRA unnecessarily uses non-volatile registers during register assignment
Related show

Commit Message

Peter Bergner Oct. 8, 2018, 7:36 p.m.
PR87507 shows a problem where IRA assigns a non-volatile TImode reg pair to
a pseudo when there is a volatile reg pair available to use.  This then
causes us to emit save/restore code for the non-volatile reg usage.

The problem here is that the only volatile reg pair that is available is an
odd/even reg pair (r7,r8) and ira-costs.c:ira_tune_allocno_costs() disparages
odd/even reg pairs by increasing their cost.  That's fine, but an even/odd
non-volatile reg pair should still be more expensive than an odd/even reg
pair given the save/restore that we'd need to use it.  However, the costs
used in assign_hard_reg() show that non-volatile reg pair (r30,r31) is
cheaper than odd/even reg pair (r7,r8) (15 versus 1000).  That's a huge
disparity in costs, so looking at where the non-volatile cost comes from,
it comes from the code below in ira-color.c:assign_hard_reg():

       if (!HONOR_REG_ALLOC_ORDER)
	{
	  if ((saved_nregs = calculate_saved_nregs (hard_regno, mode)) != 0)
	  /* We need to save/restore the hard register in
	     epilogue/prologue.  Therefore we increase the cost.  */
	  {
	    rclass = REGNO_REG_CLASS (hard_regno);
	    add_cost = ((ira_memory_move_cost[mode][rclass][0]
			 + ira_memory_move_cost[mode][rclass][1])
			* saved_nregs / hard_regno_nregs (hard_regno,
							  mode) - 1);
	    cost += add_cost;
	    full_cost += add_cost;
	  }
	}

I'm not sure I understand the "* saved_nregs / h_r_n (h_r, m) - 1" part
of the calculation.  If saved_nregs is the number of hard regs that
need to be saved for hard_regno in mode MODE (ie, we don't need to
save a hard reg if it's already been saved, etc.), then why aren't we
just multiplying by saved_nregs?  The other problem I see here is
that we're not scaling the cost by the basic block frequency of the
prologue/epilogue, which is what is causing the non-volatile reg
cost to be so low compared to the odd/even volatile reg use, which
is scaled.  However, even if I fix that, improve_allocation() comes
along and undoes it, because it too does not correctly compute the
cost of non-volatiles, so that seems to me that it needs fixing too.

I have the following work in progress patch I'd like some comments on.
Am I on the right track here?  I noticed that assign_hard_reg() tracks
min_cost and min_full_cost, but min_cost is actually never used for
anything other than setting min_cost, so I removed it.  I also don't
understand why we don't charge non-volatile usage for targets that define
HONOR_REG_ALLOC_ORDER.  Why shouldn't we always account for save/restore
of non-volatile reg usage?  I'll note I did not change that behavior.

Thoughts on the patch below?  Vlad, can you comment on some of my
questions above?

Peter


gcc/
	PR rtl-optimization/87507
	* ira-color.c (calculate_saved_nregs): Rename from this...
	(calculate_saved_nregs_cost): ...to this.  Return cost of saving NREGS.
	(assign_hard_reg):
	(improve_allocation):

gcc/testsuite/
	PR rtl-optimization/87507
	* gcc.dg/pr10474.c: Don't XFAIL for powerpc*-*-*.
	* gcc.target/powerpc/vsx-vector-6.p8.c: Update expected output.

Comments

Vladimir Makarov Oct. 9, 2018, 10:53 p.m. | #1
On 10/08/2018 03:36 PM, Peter Bergner wrote:
> PR87507 shows a problem where IRA assigns a non-volatile TImode reg pair to
> a pseudo when there is a volatile reg pair available to use.  This then
> causes us to emit save/restore code for the non-volatile reg usage.
>
> The problem here is that the only volatile reg pair that is available is an
> odd/even reg pair (r7,r8) and ira-costs.c:ira_tune_allocno_costs() disparages
> odd/even reg pairs by increasing their cost.  That's fine, but an even/odd
> non-volatile reg pair should still be more expensive than an odd/even reg
> pair given the save/restore that we'd need to use it.  However, the costs
> used in assign_hard_reg() show that non-volatile reg pair (r30,r31) is
> cheaper than odd/even reg pair (r7,r8) (15 versus 1000).  That's a huge
> disparity in costs, so looking at where the non-volatile cost comes from,
> it comes from the code below in ira-color.c:assign_hard_reg():
>
>         if (!HONOR_REG_ALLOC_ORDER)
> 	{
> 	  if ((saved_nregs = calculate_saved_nregs (hard_regno, mode)) != 0)
> 	  /* We need to save/restore the hard register in
> 	     epilogue/prologue.  Therefore we increase the cost.  */
> 	  {
> 	    rclass = REGNO_REG_CLASS (hard_regno);
> 	    add_cost = ((ira_memory_move_cost[mode][rclass][0]
> 			 + ira_memory_move_cost[mode][rclass][1])
> 			* saved_nregs / hard_regno_nregs (hard_regno,
> 							  mode) - 1);
> 	    cost += add_cost;
> 	    full_cost += add_cost;
> 	  }
> 	}
>
> I'm not sure I understand the "* saved_nregs / h_r_n (h_r, m) - 1" part
> of the calculation.  If saved_nregs is the number of hard regs that
> need to be saved for hard_regno in mode MODE (ie, we don't need to
> save a hard reg if it's already been saved, etc.), then why aren't we
> just multiplying by saved_nregs?  The other problem I see here is
> that we're not scaling the cost by the basic block frequency of the
> prologue/epilogue, which is what is causing the non-volatile reg
> cost to be so low compared to the odd/even volatile reg use, which
> is scaled.  However, even if I fix that, improve_allocation() comes
> along and undoes it, because it too does not correctly compute the
> cost of non-volatiles, so that seems to me that it needs fixing too.
It would be right if we have only one pseudo using non-volatile regs.  
But at this stage we don't know how many pseudos will use the 
non-volatile reg.  Once the non-volotile reg is used, it is free for all 
other pseudo uses (and they can be in hot regions as loops).

The costs were heuristically tuned on SPEC2000 on x86/x86-64.  As the 
functions are bigger (now we have aggressive inlining), the only small 
increase for non-volatile reg works better.  That is why I don't like 
small test RA PRs, the heuristics can be bad for them but they can work 
better for real programs.  When I change this code, I always check SPEC 
because sometimes the results are opposite to the expected ones.

May be you could find better heuristics which works for small and big 
functions.
> I have the following work in progress patch I'd like some comments on.
> Am I on the right track here?  I noticed that assign_hard_reg() tracks
> min_cost and min_full_cost, but min_cost is actually never used for
> anything other than setting min_cost, so I removed it.  I also don't
> understand why we don't charge non-volatile usage for targets that define
> HONOR_REG_ALLOC_ORDER.  Why shouldn't we always account for save/restore
> of non-volatile reg usage?
Once min_cost was used for different heuristics.  Still it is useful for 
debugging because full_cost calculations is very complicated (it is 
affected by conflict allocnos preferences and allocnos connected 
directly and indirectly by copies).  Min_cost can be useful in debugging 
to see the picture because its calculation is more straightforward.

A few years ago IRA did not use HONOR_REG_ALLOC_ORDER but people filled 
a PR (I don't remember the number) and that is what behaviour they 
exactly wanted.

Patch

Index: gcc/ira-color.c
===================================================================
--- gcc/ira-color.c	(revision 264795)
+++ gcc/ira-color.c	(working copy)
@@ -1648,11 +1648,10 @@  check_hard_reg_p (ira_allocno_t a, int h
   return j == nregs;
 }
 
-/* Return number of registers needed to be saved and restored at
-   function prologue/epilogue if we allocate HARD_REGNO to hold value
-   of MODE.  */
+/* Return the cost of saving and restoring HARD_REGNO in mode MODE at
+   function prologue and epilogue.  */
 static int
-calculate_saved_nregs (int hard_regno, machine_mode mode)
+calculate_saved_nregs_cost (int hard_regno, machine_mode mode)
 {
   int i;
   int nregs = 0;
@@ -1663,7 +1662,14 @@  calculate_saved_nregs (int hard_regno, m
 	&& !TEST_HARD_REG_BIT (call_used_reg_set, hard_regno + i)
 	&& !LOCAL_REGNO (hard_regno + i))
       nregs++;
-  return nregs;
+  if (nregs == 0)
+    return 0;
+
+  enum reg_class rclass = REGNO_REG_CLASS (hard_regno);
+  return (ira_memory_move_cost[mode][rclass][0]
+	  + ira_memory_move_cost[mode][rclass][1])
+	 * nregs
+	 * REG_FREQ_FROM_BB (ENTRY_BLOCK_PTR_FOR_FN (cfun));
 }
 
 /* Choose a hard register for allocno A.  If RETRY_P is TRUE, it means
@@ -1694,14 +1700,11 @@  assign_hard_reg (ira_allocno_t a, bool r
 {
   HARD_REG_SET conflicting_regs[2], profitable_hard_regs;
   int i, j, hard_regno, best_hard_regno, class_size;
-  int cost, mem_cost, min_cost, full_cost, min_full_cost, nwords, word;
+  int cost, mem_cost, full_cost, min_full_cost, nwords, word;
   int *a_costs;
   enum reg_class aclass;
   machine_mode mode;
-  static int costs[FIRST_PSEUDO_REGISTER], full_costs[FIRST_PSEUDO_REGISTER];
-  int saved_nregs;
-  enum reg_class rclass;
-  int add_cost;
+  static int full_costs[FIRST_PSEUDO_REGISTER];
 #ifdef STACK_REGS
   bool no_stack_reg_p;
 #endif
@@ -1713,9 +1716,7 @@  assign_hard_reg (ira_allocno_t a, bool r
   aclass = ALLOCNO_CLASS (a);
   class_size = ira_class_hard_regs_num[aclass];
   best_hard_regno = -1;
-  memset (full_costs, 0, sizeof (int) * class_size);
   mem_cost = 0;
-  memset (costs, 0, sizeof (int) * class_size);
   memset (full_costs, 0, sizeof (int) * class_size);
 #ifdef STACK_REGS
   no_stack_reg_p = false;
@@ -1733,15 +1734,9 @@  assign_hard_reg (ira_allocno_t a, bool r
   cost = ALLOCNO_UPDATED_CLASS_COST (a);
   for (i = 0; i < class_size; i++)
     if (a_costs != NULL)
-      {
-	costs[i] += a_costs[i];
-	full_costs[i] += a_costs[i];
-      }
+      full_costs[i] += a_costs[i];
     else
-      {
-	costs[i] += cost;
-	full_costs[i] += cost;
-      }
+      full_costs[i] += cost;
   nwords = ALLOCNO_NUM_OBJECTS (a);
   curr_allocno_process++;
   for (word = 0; word < nwords; word++)
@@ -1861,7 +1856,7 @@  assign_hard_reg (ira_allocno_t a, bool r
       queue_update_cost (a, NULL,  COST_HOP_DIVISOR);
       update_conflict_hard_regno_costs (full_costs, aclass, false);
     }
-  min_cost = min_full_cost = INT_MAX;
+  min_full_cost = INT_MAX;
   /* We don't care about giving callee saved registers to allocnos no
      living through calls because call clobbered registers are
      allocated first (it is usual practice to put them first in
@@ -1878,25 +1873,15 @@  assign_hard_reg (ira_allocno_t a, bool r
       if (! check_hard_reg_p (a, hard_regno,
 			      conflicting_regs, profitable_hard_regs))
 	continue;
-      cost = costs[i];
+
       full_cost = full_costs[i];
       if (!HONOR_REG_ALLOC_ORDER)
 	{
-	  if ((saved_nregs = calculate_saved_nregs (hard_regno, mode)) != 0)
-	  /* We need to save/restore the hard register in
-	     epilogue/prologue.  Therefore we increase the cost.  */
-	  {
-	    rclass = REGNO_REG_CLASS (hard_regno);
-	    add_cost = ((ira_memory_move_cost[mode][rclass][0]
-		         + ira_memory_move_cost[mode][rclass][1])
-		        * saved_nregs / hard_regno_nregs (hard_regno,
-							  mode) - 1);
-	    cost += add_cost;
-	    full_cost += add_cost;
-	  }
+	  /* If we need to save/restore the hard register in epilogue/prologue,
+	     then increase its cost.  */
+	  full_cost += calculate_saved_nregs_cost (hard_regno, mode);
 	}
-      if (min_cost > cost)
-	min_cost = cost;
+
       if (min_full_cost > full_cost)
 	{
 	  min_full_cost = full_cost;
@@ -2971,12 +2956,18 @@  improve_allocation (void)
       for (j = 0; j < class_size; j++)
 	{
 	  hregno = ira_class_hard_regs[aclass][j];
+	  int hregno_cost = costs[hregno];
+
+	  /* Include the cost of saving and restoring HREGNO in the function
+	     prologue/epilogue.  */
+	  hregno_cost += calculate_saved_nregs_cost (hregno, mode);
+
 	  if (check_hard_reg_p (a, hregno,
 				conflicting_regs, profitable_hard_regs)
-	      && min_cost > costs[hregno])
+	      && min_cost > hregno_cost)
 	    {
 	      best = hregno;
-	      min_cost = costs[hregno];
+	      min_cost = hregno_cost;
 	    }
 	}
       if (min_cost >= 0)