RE: [Patch,rtl Optimization]: Better register pressure estimate for Loop Invariant Code Motion.

[Ajit Kumar Agarwal <ajit.kumar.agarwal@xilinx.com>]

[-- Attachment #1: Type: text/plain, Size: 5755 bytes --]

The patch is modified based on the review comments from Richard, Bernd and David. 

The following changes are done to incorporate the comments received on the previous mail.

1. With this patch the liveness of the Loop is not stored on the LOOPDATA structures. The liveness is calculated based on the current loops
 for which the invariant is checked and the regs_used is calculated. 
2. Memory leaks are fixed.
3. Reworked on the comments section based on Bernd's comments.

Bootstrapped and regtested for i386 and Microblaze target.

SPEC CPU 2000 benchmarks are run on i386 target and following is the summary of the results.

SPEC CPU 2000 INT benchmarks.

( Gemoean Score without the change vs Geomean score with reg pressure change = 3745.193 vs 3745.328)

SPEC CPU 2000 FP benchmarks.

( Gemoean Score without the change vs Geomean score with reg pressure change = 4741.825 vs 4748.364).


    [Patch,rtl Optimization]: Better register pressure estimate for loop invariant code motion

    Calculate the loop liveness used for regs for calculating the register pressure
    in the cost estimation.  Loop liveness is based on the following properties.
    We only need to find the set of objects that are live at the birth or the header
    of the loop. We don't need to calculate the live through the loop by considering
    live in and live out of all the basic blocks of the loop. This is based on the
    point that the set of objects that are live-in at the birth or header of the loop
    will be live-in at every node in the loop.

    If a v live is out at the header of the loop then the variable is live-in at every node
    in the loop. To prove this, consider a loop L with header h such that the variable v
    defined at d is live-in at h. Since v is live at h, d is not part of L. This follows i
    from the dominance property, i.e. h is strictly dominated by d. Furthermore, there
    exists a path from h to a use of v which does not go through d. For every node p in
    the loop, since the loop is strongly connected and node is a component of the CFG,
    there exists a path, consisting only of nodes of L from p to h. Concatenating these
    two paths proves that v is live-in and live-out of p.

    Calculate the live-out and live-in for the exit edge of the loop. This patch considers
    liveness for not only the loop latch but also the liveness outside the loops.

    ChangeLog:
    2015-12-15  Ajit Agarwal  <ajitkum@xilinx.com>

        * loop-invariant.c
        (find_invariants_to_move): Add the logic of regs_used based
        on liveness.
        * cfgloopanal.c
        (estimate_reg_pressure_cost): Update the heuristics in presence
        of call_p.

    Signed-off-by:Ajit Agarwal ajitkum@xilinx.com.

Thanks & Regards
Ajit

-----Original Message-----
From: Bernd Schmidt [mailto:bschmidt@redhat.com] 
Sent: Wednesday, December 09, 2015 7:34 PM
To: Ajit Kumar Agarwal; Richard Biener
Cc: Jeff Law; GCC Patches; Vinod Kathail; Shail Aditya Gupta; Vidhumouli Hunsigida; Nagaraju Mekala
Subject: Re: [Patch,rtl Optimization]: Better register pressure estimate for Loop Invariant Code Motion.

On 12/09/2015 12:22 PM, Ajit Kumar Agarwal wrote:
>
> This is because the available_regs = 6 and the regs_needed = 1 and 
> new_regs = 0 and the regs_used = 10.  As the reg_used that are based 
> on the Liveness given above is greater than the available_regs, then
 > it's candidate of spill and the estimate_register_pressure calculates  > the spill cost. This spill cost is greater than inv_cost and gain  > comes to be negative. The disables the loop invariant for the above  > testcase.

As far as I can tell this loop does not lead to a spill. Hence, failure of this testcase would suggest there is something wrong with your idea.

>> +  FOR_EACH_LOOP (loop, LI_FROM_INNERMOST)  {

Formatting.

>> +    /* Loop Liveness is based on the following proprties.

"properties"

>> +       we only require to calculate the set of objects that are live at
>> +       the birth or the header of the loop.
>> +       We don't need to calculate the live through the Loop considering
>> +       Live in and Live out of all the basic blocks of the Loop. This is
>> +       because the set of objects. That are live-in at the birth or header
>> +       of the loop will be live-in at every node in the Loop.
>> +       If a v live out at the header of the loop then the variable is live-in
>> +       at every node in the Loop. To prove this, Consider a Loop L with header
>> +       h such that The variable v defined at d is live-in at h. Since v is live
>> +       at h, d is not part of L. This follows from the dominance property, i.e.
>> +       h is strictly dominated by d. Furthermore, there exists a path from h to
>> +       a use of v which does not go through d. For every node of the loop, p,
>> +       since the loop is strongly connected Component of the CFG, there exists
>> +       a path, consisting only of nodes of L from p to h. Concatenating those
>> +       two paths prove that v is live-in and live-out Of p.  */

Please Refrain From Randomly Capitalizing Words, and please also fix the grammar ("set of objects. That are live-in"). These problems make this comment (and also your emails) extremely hard to read.

Partly for that reason, I can't quite make up my mind whether this patch makes things better or just different. The testcase failure makes me think it's probably not an improvement.

>> +    bitmap_ior_into (&LOOP_DATA (loop)->regs_live, DF_LR_IN (loop->header));
>> +    bitmap_ior_into (&LOOP_DATA (loop)->regs_live, DF_LR_OUT 
>> + (loop->header));

Formatting.


Bernd

[-- Attachment #2: pressure-licm.patch --]
[-- Type: application/octet-stream, Size: 7222 bytes --]

From b5107db65d487765d0bf0ba184d01f5211e20006 Mon Sep 17 00:00:00 2001
From: Ajit Kumar Agarwal <ajitkum@xilix.com>
Date: Tue, 15 Dec 2015 13:25:04 +0530
Subject: [PATCH] [Patch,rtl Optimization]: Better register pressure estimate for loop invariant code motion

Calculate the loop liveness used for regs for calculating the register pressure
in the cost estimation.  Loop liveness is based on the following properties.
We only need to find the set of objects that are live at the birth or the header
of the loop. We don't need to calculate the live through the loop by considering
live in and live out of all the basic blocks of the loop. This is based on the
point that the set of objects that are live-in at the birth or header of the loop
will be live-in at every node in the loop.

If a v live is out at the header of the loop then the variable is live-in at every node
in the loop. To prove this, consider a loop L with header h such that the variable v
defined at d is live-in at h. Since v is live at h, d is not part of L. This follows i
from the dominance property, i.e. h is strictly dominated by d. Furthermore, there
exists a path from h to a use of v which does not go through d. For every node p in
the loop, since the loop is strongly connected and node is a component of the CFG,
there exists a path, consisting only of nodes of L from p to h. Concatenating these
two paths proves that v is live-in and live-out of p.

Calculate the live-out and live-in for the exit edge of the loop. This patch considers
liveness for not only the loop latch but also the liveness outside the loops.

ChangeLog:
2015-12-15  Ajit Agarwal  <ajitkum@xilinx.com>

	* loop-invariant.c
	(find_invariants_to_move): Add the logic of regs_used based
	on liveness.
	* cfgloopanal.c
	(estimate_reg_pressure_cost): Update the heuristics in presence
	of call_p.

Signed-off-by:Ajit Agarwal ajitkum@xilinx.com.
---
 gcc/cfgloopanal.c    |   12 +++++++-
 gcc/loop-invariant.c |   70 +++++++++++++++++++++++++++++++++++--------------
 2 files changed, 60 insertions(+), 22 deletions(-)

diff --git a/gcc/cfgloopanal.c b/gcc/cfgloopanal.c
index 2220e86..c2b2947 100644
--- a/gcc/cfgloopanal.c
+++ b/gcc/cfgloopanal.c
@@ -373,15 +373,23 @@ estimate_reg_pressure_cost (unsigned n_new, unsigned n_old, bool speed,
 
   /* If there is a call in the loop body, the call-clobbered registers
      are not available for loop invariants.  */
+
   if (call_p)
     available_regs = available_regs - target_clobbered_regs;
-
+  
   /* If we have enough registers, we should use them and not restrict
      the transformations unnecessarily.  */
   if (regs_needed + target_res_regs <= available_regs)
     return 0;
 
-  if (regs_needed <= available_regs)
+  /* Estimation of target_clobbered register is based on the call_used
+     arrays which is not the right estimate for the clobbered register
+     used in called function. Instead of considering the spill cost we
+     consider only the reg_cost aggressively.  */
+
+  if ((regs_needed <= available_regs) 
+      || (call_p && (regs_needed <= 
+          (available_regs + target_clobbered_regs))))
     /* If we are close to running out of registers, try to preserve
        them.  */
     cost = target_reg_cost [speed] * n_new;
diff --git a/gcc/loop-invariant.c b/gcc/loop-invariant.c
index 53d1377..62896dd 100644
--- a/gcc/loop-invariant.c
+++ b/gcc/loop-invariant.c
@@ -1365,8 +1365,7 @@ gain_for_invariant (struct invariant *inv, unsigned *regs_needed,
       else
 	size_cost = 0;
     }
-
-  return comp_cost - size_cost;
+  return comp_cost - size_cost + 1;
 }
 
 /* Finds invariant with best gain for moving.  Returns the gain, stores
@@ -1460,23 +1459,54 @@ find_invariants_to_move (bool speed, bool call_p)
     /* REGS_USED is actually never used when the flag is on.  */
     regs_used = 0;
   else
-    /* We do not really do a good job in estimating number of
-       registers used; we put some initial bound here to stand for
-       induction variables etc.  that we do not detect.  */
+    /* The logic used in estimating the number of regs_used is changed.
+       Now it will be based on liveness of the loop. */
     {
-      unsigned int n_regs = DF_REG_SIZE (df);
-
-      regs_used = 2;
-
-      for (i = 0; i < n_regs; i++)
-	{
-	  if (!DF_REGNO_FIRST_DEF (i) && DF_REGNO_LAST_USE (i))
-	    {
-	      /* This is a value that is used but not changed inside loop.  */
-	      regs_used++;
-	    }
-	}
-    }
+      int  i;
+      edge e;
+      vec<edge> edges;
+      bitmap_head regs_live;
+
+      bitmap_initialize (&regs_live, &reg_obstack);
+      edges = get_loop_exit_edges (curr_loop);
+
+      /* Loop liveness is based on the following properties.
+         We only need to find the set of objects that are live at the
+         birth or the header of the loop.
+         We don't need to calculate the live through the loop considering
+         live-in and live-out of all the basic blocks of the loop. This is
+         based on the point that the set of objects that are live-in at the
+         birth or header of the loop will be live-in at every block in the
+         loop.
+
+         If a v live out at the header of the loop then the variable is
+         live-in at every node in the Loop. To prove this, consider a loop
+         L with header h such that the variable v defined at d is live-in
+         at h. Since v is live at h, d is not part of L. This follows from
+         the dominance property, i.e. h is strictly dominated by d. Furthermore,
+         there exists a path from h to a use of v which does not go through d.
+         For every node of the loop, p, since the loop is strongly connected
+         component of the CFG, there exists a path, consisting only of nodes
+         of L from p to h. Concatenating these two paths prove that v is
+         live-in and live-out of p.  */
+
+       bitmap_ior_into (&regs_live, DF_LR_IN (curr_loop->header));
+       bitmap_ior_into (&regs_live, DF_LR_OUT (curr_loop->header));
+
+       /* Calculate the live-out and live-in for the exit edge of the loop.
+          This considers liveness for not only the loop latch but also the
+          liveness outside the loops.  */
+
+       FOR_EACH_VEC_ELT (edges, i, e)
+         {
+           bitmap_ior_into (&regs_live, DF_LR_OUT (e->src));
+           bitmap_ior_into (&regs_live, DF_LR_IN (e->dest));
+         }
+
+       regs_used = bitmap_count_bits (&regs_live) + 2;
+       bitmap_clear (&regs_live);
+       edges.release ();
+     }
 
   if (! flag_ira_loop_pressure)
     new_regs[0] = regs_needed[0] = 0;
@@ -1967,6 +1997,7 @@ mark_ref_regs (rtx x)
 }
 
 /* Calculate register pressure in the loops.  */
+
 static void
 calculate_loop_reg_pressure (void)
 {
@@ -2085,9 +2116,7 @@ calculate_loop_reg_pressure (void)
       fprintf (dump_file, "\n");
     }
 }
-
 \f
-
 /* Move the invariants out of the loops.  */
 
 void
@@ -2103,6 +2132,7 @@ move_loop_invariants (void)
       calculate_loop_reg_pressure ();
       regstat_free_n_sets_and_refs ();
     }
+
   df_set_flags (DF_EQ_NOTES + DF_DEFER_INSN_RESCAN);
   /* Process the loops, innermost first.  */
   FOR_EACH_LOOP (loop, LI_FROM_INNERMOST)
-- 
1.7.1