[DF] [performance] generate DF_REF_BASE REFs in REGNO order

[Dimitrios Apostolou <jimis@gmx.net>]

[-- Attachment #1: Type: TEXT/PLAIN, Size: 2164 bytes --]

Hello list,

the attached patch achieves a performance improvement by first recording 
DF_REF_BASE DEFs within df_get_call_refs() before the DF_REF_REGULARs are 
recorded in df_defs_record(). BASE DEFs are also recorded in REGNO order. 
Improvement has been measured as follows, for compiling tcp_ipv4.c of 
linux kernel with -O0 optimisation:

trunk  : 1438.4 M instr, 0.627s
patched: 1376.5 M instr, 0.604s

It also includes suggested changes from Paolo discussed on list (subject: 
what can be in a group set). Many thanks to him for the invaluable help 
while writing the patch.

For whoever is interested, you can see the two profiles with fully 
annotated source before and after the change, at the following links. The 
big difference is that qsort() is now called only 33 times instead of 
thousands, from df_sort_and_compress_refs().
Further measurements, comments and ideas for further improvements are 
welcome.

http://gcc.gnu.org/wiki/OptimisingGCC?action=AttachFile&do=view&target=callgrind-tcp_ipv4-trunk-co-109439-prod.txt
http://gcc.gnu.org/wiki/OptimisingGCC?action=AttachFile&do=view&target=callgrind-tcp_ipv4-df2-co-prod.txt


Changelog:

2011-07-29  Dimitrios Apostolou  <jimis@gmx.net>
             Paolo Bonzini  <bonzini@gnu.org>

         (df_def_record_1): Assert a parallel must contain an EXPR_LIST at
 	this point.  Receive the LOC and move its extraction...
         (df_defs_record): ... here. Rewrote logic with a switch statement
 	instead of multiple if-else.
 	(df_find_hard_reg_defs, df_find_hard_reg_defs_1): New functions
 	that duplicate the logic of df_defs_record() and df_def_record_1()
 	but without actually recording any DEFs, only marking them in
 	the defs HARD_REG_SET.
 	(df_get_call_refs): Call df_find_hard_reg_defs() to mark DEFs that
 	are the result of the call. Record DF_REF_BASE DEFs in REGNO
 	order. Use regs_invalidated_by_call HARD_REG_SET instead of
 	regs_invalidated_by_call_regset bitmap.
 	(df_insn_refs_collect): Record DF_REF_REGULAR DEFs after
 	df_get_call_refs().


Thanks,
Dimitris


P.S. maraz: that's 4.3% improvement in instruction count, should you start 
worrying or is it too late? ;-)

[-- Attachment #2: Type: TEXT/PLAIN, Size: 11214 bytes --]

=== modified file 'gcc/df-scan.c'
--- gcc/df-scan.c	2011-02-02 20:08:06 +0000
+++ gcc/df-scan.c	2011-07-29 16:01:50 +0000
@@ -111,7 +111,7 @@ static void df_ref_record (enum df_ref_c
 			   rtx, rtx *,
 			   basic_block, struct df_insn_info *,
 			   enum df_ref_type, int ref_flags);
-static void df_def_record_1 (struct df_collection_rec *, rtx,
+static void df_def_record_1 (struct df_collection_rec *, rtx *,
 			     basic_block, struct df_insn_info *,
 			     int ref_flags);
 static void df_defs_record (struct df_collection_rec *, rtx,
@@ -2916,40 +2916,27 @@ df_read_modify_subreg_p (rtx x)
 }
 
 
-/* Process all the registers defined in the rtx, X.
+/* Process all the registers defined in the rtx pointed by LOC.
    Autoincrement/decrement definitions will be picked up by
    df_uses_record.  */
 
 static void
 df_def_record_1 (struct df_collection_rec *collection_rec,
-                 rtx x, basic_block bb, struct df_insn_info *insn_info,
+                 rtx *loc, basic_block bb, struct df_insn_info *insn_info,
 		 int flags)
 {
-  rtx *loc;
-  rtx dst;
-
- /* We may recursively call ourselves on EXPR_LIST when dealing with PARALLEL
-     construct.  */
-  if (GET_CODE (x) == EXPR_LIST || GET_CODE (x) == CLOBBER)
-    loc = &XEXP (x, 0);
-  else
-    loc = &SET_DEST (x);
-  dst = *loc;
+  rtx dst = *loc;
 
   /* It is legal to have a set destination be a parallel. */
   if (GET_CODE (dst) == PARALLEL)
     {
       int i;
-
       for (i = XVECLEN (dst, 0) - 1; i >= 0; i--)
 	{
 	  rtx temp = XVECEXP (dst, 0, i);
-	  if (GET_CODE (temp) == EXPR_LIST || GET_CODE (temp) == CLOBBER
-	      || GET_CODE (temp) == SET)
-	    df_def_record_1 (collection_rec,
-                             temp, bb, insn_info,
-			     GET_CODE (temp) == CLOBBER
-			     ? flags | DF_REF_MUST_CLOBBER : flags);
+	  gcc_assert (GET_CODE (temp) == EXPR_LIST);
+	  df_def_record_1 (collection_rec, &XEXP (temp, 0),
+			   bb, insn_info, flags);
 	}
       return;
     }
@@ -3003,26 +2990,98 @@ df_defs_record (struct df_collection_rec
 		int flags)
 {
   RTX_CODE code = GET_CODE (x);
+  int i;
 
-  if (code == SET || code == CLOBBER)
-    {
-      /* Mark the single def within the pattern.  */
-      int clobber_flags = flags;
-      clobber_flags |= (code == CLOBBER) ? DF_REF_MUST_CLOBBER : 0;
-      df_def_record_1 (collection_rec, x, bb, insn_info, clobber_flags);
-    }
-  else if (code == COND_EXEC)
+  switch (code)
     {
+    case SET:
+      df_def_record_1 (collection_rec, &SET_DEST (x), bb, insn_info, flags);
+      break;
+
+    case CLOBBER:
+      flags |= DF_REF_MUST_CLOBBER;
+      df_def_record_1 (collection_rec, &XEXP (x, 0), bb, insn_info, flags);
+      break;
+
+    case COND_EXEC:
       df_defs_record (collection_rec, COND_EXEC_CODE (x),
 		      bb, insn_info, DF_REF_CONDITIONAL);
+      break;
+
+    case PARALLEL:
+      for (i = XVECLEN (x, 0) - 1; i >= 0; i--)
+	df_defs_record (collection_rec, XVECEXP (x, 0, i),
+			bb, insn_info, flags);
+      break;
+    default:
+      /* No DEFs to record in other cases */
+      break;
     }
-  else if (code == PARALLEL)
+}
+
+/* Set the bits in *defs of registers defined in the pattern rtx */
+
+static void
+df_find_hard_reg_defs_1 (rtx *loc, basic_block bb,
+			 int flags, HARD_REG_SET *defs)
+{
+  rtx dst = *loc;
+
+  /* It is legal to have a set destination be a parallel. */
+  if (GET_CODE (dst) == PARALLEL)
     {
       int i;
+      for (i = XVECLEN (dst, 0) - 1; i >= 0; i--)
+	{
+	  rtx temp = XVECEXP (dst, 0, i);
+	  gcc_assert (GET_CODE (temp) == EXPR_LIST);
+	  df_find_hard_reg_defs_1 (&XEXP (temp, 0), bb, flags, defs);
+	}
+      return;
+    }
+
+  if (GET_CODE (dst) == STRICT_LOW_PART)
+      dst = XEXP (dst, 0);
+
+  if (GET_CODE (dst) == ZERO_EXTRACT)
+      dst = XEXP (dst, 0);
 
-      /* Mark the multiple defs within the pattern.  */
+  /* At this point if we do not have a reg or a subreg, just return.  */
+  if (REG_P (dst))
+    SET_HARD_REG_BIT (*defs, REGNO (dst));
+  else if (GET_CODE (dst) == SUBREG && REG_P (SUBREG_REG (dst)))
+    SET_HARD_REG_BIT (*defs, REGNO (SUBREG_REG (dst)));
+}
+
+static void
+df_find_hard_reg_defs (rtx x, basic_block bb, 
+		       int flags, HARD_REG_SET *defs)
+{
+  RTX_CODE code = GET_CODE (x);
+  int i;
+
+  switch (code)
+    {
+    case SET:
+      df_find_hard_reg_defs_1 (&SET_DEST (x), bb, flags, defs);
+      break;
+
+    case CLOBBER:
+      flags |= DF_REF_MUST_CLOBBER;
+      df_find_hard_reg_defs_1 (&XEXP (x, 0), bb, flags, defs);
+      break;
+
+    case COND_EXEC:
+      df_find_hard_reg_defs (COND_EXEC_CODE (x), bb, DF_REF_CONDITIONAL, defs);
+      break;
+
+    case PARALLEL:
       for (i = XVECLEN (x, 0) - 1; i >= 0; i--)
-	df_defs_record (collection_rec, XVECEXP (x, 0, i), bb, insn_info, flags);
+	df_find_hard_reg_defs (XVECEXP (x, 0, i), bb, flags, defs);
+      break;
+    default:
+      /* No DEFs to record in other cases */
+      break;
     }
 }
 
@@ -3308,7 +3367,7 @@ df_get_conditional_uses (struct df_colle
 }
 
 
-/* Get call's extra defs and uses. */
+/* Get call's extra defs and uses (track caller-saved registers). */
 
 static void
 df_get_call_refs (struct df_collection_rec * collection_rec,
@@ -3317,20 +3376,50 @@ df_get_call_refs (struct df_collection_r
                   int flags)
 {
   rtx note;
-  bitmap_iterator bi;
-  unsigned int ui;
   bool is_sibling_call;
   unsigned int i;
-  df_ref def;
-  bitmap_head defs_generated;
+  HARD_REG_SET defs_generated;
 
-  bitmap_initialize (&defs_generated, &df_bitmap_obstack);
+  CLEAR_HARD_REG_SET (defs_generated);
+  df_find_hard_reg_defs (PATTERN (insn_info->insn), bb, 
+			 0, &defs_generated);
 
-  /* Do not generate clobbers for registers that are the result of the
-     call.  This causes ordering problems in the chain building code
-     depending on which def is seen first.  */
-  FOR_EACH_VEC_ELT (df_ref, collection_rec->def_vec, i, def)
-    bitmap_set_bit (&defs_generated, DF_REF_REGNO (def));
+  is_sibling_call = SIBLING_CALL_P (insn_info->insn);
+
+  for (i = 0; i < FIRST_PSEUDO_REGISTER; i++)
+    {
+      if (i == STACK_POINTER_REGNUM)
+	/* The stack ptr is used (honorarily) by a CALL insn.  */
+	df_ref_record (DF_REF_BASE, collection_rec, regno_reg_rtx[i],
+		       NULL, bb, insn_info, DF_REF_REG_USE,
+		       DF_REF_CALL_STACK_USAGE | flags);
+      else if (global_regs[i])
+	{
+	  /* Calls to const functions cannot access any global registers and
+	     calls to pure functions cannot set them.  All other calls may
+	     reference any of the global registers, so they are recorded as
+	     used. */
+	  if (!RTL_CONST_CALL_P (insn_info->insn))
+	    {
+	      df_ref_record (DF_REF_BASE, collection_rec, regno_reg_rtx[i],
+			     NULL, bb, insn_info, DF_REF_REG_USE, flags);
+	      if (!RTL_PURE_CALL_P (insn_info->insn))
+		df_ref_record (DF_REF_BASE, collection_rec, regno_reg_rtx[i],
+			       NULL, bb, insn_info, DF_REF_REG_DEF, flags);
+	    }
+	}
+      else
+	if (TEST_HARD_REG_BIT(regs_invalidated_by_call, i)
+	    /* no clobbers for regs that are the result of the call */
+	    && !TEST_HARD_REG_BIT (defs_generated, i)
+	    && (!is_sibling_call
+		|| !bitmap_bit_p (df->exit_block_uses, i)
+		|| refers_to_regno_p (i, i+1,
+				      crtl->return_rtx, NULL)))
+	  df_ref_record (DF_REF_BASE, collection_rec, regno_reg_rtx[i],
+			 NULL, bb, insn_info, DF_REF_REG_DEF,
+			 DF_REF_MAY_CLOBBER | flags);
+    }
 
   /* Record the registers used to pass arguments, and explicitly
      noted as clobbered.  */
@@ -3345,7 +3434,7 @@ df_get_call_refs (struct df_collection_r
 	  if (REG_P (XEXP (XEXP (note, 0), 0)))
 	    {
 	      unsigned int regno = REGNO (XEXP (XEXP (note, 0), 0));
-	      if (!bitmap_bit_p (&defs_generated, regno))
+	      if (!TEST_HARD_REG_BIT (defs_generated, regno))
 		df_defs_record (collection_rec, XEXP (note, 0), bb,
 				insn_info, flags);
 	    }
@@ -3355,40 +3444,6 @@ df_get_call_refs (struct df_collection_r
 	}
     }
 
-  /* The stack ptr is used (honorarily) by a CALL insn.  */
-  df_ref_record (DF_REF_BASE, collection_rec, regno_reg_rtx[STACK_POINTER_REGNUM],
-		 NULL, bb, insn_info, DF_REF_REG_USE,
-		 DF_REF_CALL_STACK_USAGE | flags);
-
-  /* Calls to const functions cannot access any global registers and calls to
-     pure functions cannot set them.  All other calls may reference any of the
-     global registers, so they are recorded as used.  */
-  if (!RTL_CONST_CALL_P (insn_info->insn))
-    for (i = 0; i < FIRST_PSEUDO_REGISTER; i++)
-      if (global_regs[i])
-	{
-	  df_ref_record (DF_REF_BASE, collection_rec, regno_reg_rtx[i],
-			 NULL, bb, insn_info, DF_REF_REG_USE, flags);
-	  if (!RTL_PURE_CALL_P (insn_info->insn))
-	    df_ref_record (DF_REF_BASE, collection_rec, regno_reg_rtx[i],
-			   NULL, bb, insn_info, DF_REF_REG_DEF, flags);
-	}
-
-  is_sibling_call = SIBLING_CALL_P (insn_info->insn);
-  EXECUTE_IF_SET_IN_BITMAP (regs_invalidated_by_call_regset, 0, ui, bi)
-    {
-      if (!global_regs[ui]
-	  && (!bitmap_bit_p (&defs_generated, ui))
-	  && (!is_sibling_call
-	      || !bitmap_bit_p (df->exit_block_uses, ui)
-	      || refers_to_regno_p (ui, ui+1,
-				    crtl->return_rtx, NULL)))
-        df_ref_record (DF_REF_BASE, collection_rec, regno_reg_rtx[ui],
-		       NULL, bb, insn_info, DF_REF_REG_DEF,
-		       DF_REF_MAY_CLOBBER | flags);
-    }
-
-  bitmap_clear (&defs_generated);
   return;
 }
 
@@ -3398,7 +3453,7 @@ df_get_call_refs (struct df_collection_r
    and reg chains. */
 
 static void
-df_insn_refs_collect (struct df_collection_rec* collection_rec,
+df_insn_refs_collect (struct df_collection_rec *collection_rec,
 		      basic_block bb, struct df_insn_info *insn_info)
 {
   rtx note;
@@ -3410,9 +3465,6 @@ df_insn_refs_collect (struct df_collecti
   VEC_truncate (df_ref, collection_rec->eq_use_vec, 0);
   VEC_truncate (df_mw_hardreg_ptr, collection_rec->mw_vec, 0);
 
-  /* Record register defs.  */
-  df_defs_record (collection_rec, PATTERN (insn_info->insn), bb, insn_info, 0);
-
   /* Process REG_EQUIV/REG_EQUAL notes.  */
   for (note = REG_NOTES (insn_info->insn); note;
        note = XEXP (note, 1))
@@ -3444,12 +3496,17 @@ df_insn_refs_collect (struct df_collecti
     }
 
   if (CALL_P (insn_info->insn))
+    /* Record DF_REF_BASE register defs for CALL_INSNs. */
     df_get_call_refs (collection_rec, bb, insn_info,
 		      (is_cond_exec) ? DF_REF_CONDITIONAL : 0);
 
+  /* Record DF_REF_REGULAR defs and uses.  */
+  df_defs_record (collection_rec, PATTERN (insn_info->insn),
+		  bb, insn_info, 0);
+
   /* Record the register uses.  */
-  df_uses_record (collection_rec,
-		  &PATTERN (insn_info->insn), DF_REF_REG_USE, bb, insn_info, 0);
+  df_uses_record (collection_rec, &PATTERN (insn_info->insn),
+		  DF_REF_REG_USE, bb, insn_info, 0);
 
   /* DF_REF_CONDITIONAL needs corresponding USES. */
   if (is_cond_exec)