Patchwork [i386] : Rewrite shift splitters to use indirect functions.

login
register
mail settings
Submitter Uros Bizjak
Date Sept. 20, 2010, 7:02 p.m.
Message ID <AANLkTimRDDixAEv9rD9SwrnU_V=CciQSO62oKQ8mTiwJ@mail.gmail.com>
Download mbox | patch
Permalink /patch/65243/
State New
Headers show

Comments

Uros Bizjak - Sept. 20, 2010, 7:02 p.m.
Hello!

Attached patch reduces "if (mode == DImode) ..." tests in shift
splitters to minimum.  No functional change otherwise.

2010-09-20  Uros Bizjak  <ubizjak@gmail.com>

	* config/i386/i386.c (ix86_expand_ashl_const): Rewrite using
	indirect functions.
	(ix86_split_ashl): Ditto.
	(ix86_split_ashr): Ditto.
	(ix86_split_lshr): Ditto.
	(ix86_adjust_counter): Ditto.

Tested on x86_64-pc-linux-gnu {,-m32}, committed to mainline SVN.

Uros.

Patch

Index: i386.c
===================================================================
--- i386.c	(revision 164444)
+++ i386.c	(working copy)
@@ -18057,32 +18057,29 @@  ix86_split_long_move (rtx operands[])
 static void
 ix86_expand_ashl_const (rtx operand, int count, enum machine_mode mode)
 {
-  if (count == 1)
+  rtx (*insn)(rtx, rtx, rtx);
+
+  if (count == 1
+      || (count * ix86_cost->add <= ix86_cost->shift_const
+	  && !optimize_insn_for_size_p ()))
     {
-      emit_insn ((mode == DImode
-		  ? gen_addsi3
-		  : gen_adddi3) (operand, operand, operand));
+      insn = mode == DImode ? gen_addsi3 : gen_adddi3;
+      while (count-- > 0)
+	emit_insn (insn (operand, operand, operand));
     }
-  else if (!optimize_insn_for_size_p ()
-	   && count * ix86_cost->add <= ix86_cost->shift_const)
+  else
     {
-      int i;
-      for (i=0; i<count; i++)
-	{
-	  emit_insn ((mode == DImode
-		      ? gen_addsi3
-		      : gen_adddi3) (operand, operand, operand));
-	}
+      insn = mode == DImode ? gen_ashlsi3 : gen_ashldi3;
+      emit_insn (insn (operand, operand, GEN_INT (count)));
     }
-  else
-    emit_insn ((mode == DImode
-		? gen_ashlsi3
-		: gen_ashldi3) (operand, operand, GEN_INT (count)));
 }
 
 void
 ix86_split_ashl (rtx *operands, rtx scratch, enum machine_mode mode)
 {
+  rtx (*gen_ashl3)(rtx, rtx, rtx);
+  rtx (*gen_shld)(rtx, rtx, rtx);
+
   rtx low[2], high[2];
   int count;
   const int single_width = mode == DImode ? 32 : 64;
@@ -18102,11 +18099,12 @@  ix86_split_ashl (rtx *operands, rtx scra
 	}
       else
 	{
+	  gen_shld = mode == DImode ? gen_x86_shld : gen_x86_64_shld;
+
 	  if (!rtx_equal_p (operands[0], operands[1]))
 	    emit_move_insn (operands[0], operands[1]);
-	  emit_insn ((mode == DImode
-		     ? gen_x86_shld
-		     : gen_x86_64_shld) (high[0], low[0], GEN_INT (count)));
+
+	  emit_insn (gen_shld (high[0], low[0], GEN_INT (count)));
 	  ix86_expand_ashl_const (low[0], count, mode);
 	}
       return;
@@ -18114,6 +18112,8 @@  ix86_split_ashl (rtx *operands, rtx scra
 
   split_double_mode (mode, operands, 1, low, high);
 
+  gen_ashl3 = mode == DImode ? gen_ashlsi3 : gen_ashldi3;
+
   if (operands[1] == const1_rtx)
     {
       /* Assuming we've chosen a QImode capable registers, then 1 << N
@@ -18144,33 +18144,44 @@  ix86_split_ashl (rtx *operands, rtx scra
 	 pentium4 a bit; no one else seems to care much either way.  */
       else
 	{
+	  enum machine_mode half_mode;
+	  rtx (*gen_lshr3)(rtx, rtx, rtx);
+	  rtx (*gen_and3)(rtx, rtx, rtx);
+	  rtx (*gen_xor3)(rtx, rtx, rtx);
+	  HOST_WIDE_INT bits;
 	  rtx x;
 
+	  if (mode == DImode)
+	    {
+	      half_mode = SImode;
+	      gen_lshr3 = gen_lshrsi3;
+	      gen_and3 = gen_andsi3;
+	      gen_xor3 = gen_xorsi3;
+	      bits = 5;
+	    }
+	  else
+	    {
+	      half_mode = DImode;
+	      gen_lshr3 = gen_lshrdi3;
+	      gen_and3 = gen_anddi3;
+	      gen_xor3 = gen_xordi3;
+	      bits = 6;
+	    }
+
 	  if (TARGET_PARTIAL_REG_STALL && !optimize_insn_for_size_p ())
-	    x = gen_rtx_ZERO_EXTEND (mode == DImode ? SImode : DImode, operands[2]);
+	    x = gen_rtx_ZERO_EXTEND (half_mode, operands[2]);
 	  else
-	    x = gen_lowpart (mode == DImode ? SImode : DImode, operands[2]);
+	    x = gen_lowpart (half_mode, operands[2]);
 	  emit_insn (gen_rtx_SET (VOIDmode, high[0], x));
 
-	  emit_insn ((mode == DImode
-		      ? gen_lshrsi3
-		      : gen_lshrdi3) (high[0], high[0],
-				      GEN_INT (mode == DImode ? 5 : 6)));
-	  emit_insn ((mode == DImode
-		      ? gen_andsi3
-		      : gen_anddi3) (high[0], high[0], const1_rtx));
+	  emit_insn (gen_lshr3 (high[0], high[0], GEN_INT (bits)));
+	  emit_insn (gen_and3 (high[0], high[0], const1_rtx));
 	  emit_move_insn (low[0], high[0]);
-	  emit_insn ((mode == DImode
-		      ? gen_xorsi3
-		      : gen_xordi3) (low[0], low[0], const1_rtx));
+	  emit_insn (gen_xor3 (low[0], low[0], const1_rtx));
 	}
 
-      emit_insn ((mode == DImode
-		    ? gen_ashlsi3
-		    : gen_ashldi3) (low[0], low[0], operands[2]));
-      emit_insn ((mode == DImode
-		    ? gen_ashlsi3
-		    : gen_ashldi3) (high[0], high[0], operands[2]));
+      emit_insn (gen_ashl3 (low[0], low[0], operands[2]));
+      emit_insn (gen_ashl3 (high[0], high[0], operands[2]));
       return;
     }
 
@@ -18186,36 +18197,41 @@  ix86_split_ashl (rtx *operands, rtx scra
     }
   else
     {
+      gen_shld = mode == DImode ? gen_x86_shld : gen_x86_64_shld;
+
       if (!rtx_equal_p (operands[0], operands[1]))
 	emit_move_insn (operands[0], operands[1]);
 
       split_double_mode (mode, operands, 1, low, high);
-      emit_insn ((mode == DImode
-		  ? gen_x86_shld
-		  : gen_x86_64_shld) (high[0], low[0], operands[2]));
+      emit_insn (gen_shld (high[0], low[0], operands[2]));
     }
 
-  emit_insn ((mode == DImode
-	      ? gen_ashlsi3
-	      : gen_ashldi3) (low[0], low[0], operands[2]));
+  emit_insn (gen_ashl3 (low[0], low[0], operands[2]));
 
   if (TARGET_CMOVE && scratch)
     {
+      rtx (*gen_x86_shift_adj_1)(rtx, rtx, rtx, rtx)
+	= mode == DImode ? gen_x86_shiftsi_adj_1 : gen_x86_shiftdi_adj_1;
+
       ix86_expand_clear (scratch);
-      emit_insn ((mode == DImode
-		  ? gen_x86_shiftsi_adj_1
-		  : gen_x86_shiftdi_adj_1) (high[0], low[0], operands[2],
-					    scratch));
+      emit_insn (gen_x86_shift_adj_1 (high[0], low[0], operands[2], scratch));
     }
   else
-    emit_insn ((mode == DImode
-		? gen_x86_shiftsi_adj_2
-		: gen_x86_shiftdi_adj_2) (high[0], low[0], operands[2]));
+    {
+      rtx (*gen_x86_shift_adj_2)(rtx, rtx, rtx)
+	= mode == DImode ? gen_x86_shiftsi_adj_2 : gen_x86_shiftdi_adj_2;
+
+      emit_insn (gen_x86_shift_adj_2 (high[0], low[0], operands[2]));
+    }
 }
 
 void
 ix86_split_ashr (rtx *operands, rtx scratch, enum machine_mode mode)
 {
+  rtx (*gen_ashr3)(rtx, rtx, rtx)
+    = mode == DImode ? gen_ashrsi3 : gen_ashrdi3;
+  rtx (*gen_shrd)(rtx, rtx, rtx);
+
   rtx low[2], high[2];
   int count;
   const int single_width = mode == DImode ? 32 : 64;
@@ -18228,10 +18244,8 @@  ix86_split_ashr (rtx *operands, rtx scra
       if (count == single_width * 2 - 1)
 	{
 	  emit_move_insn (high[0], high[1]);
-	  emit_insn ((mode == DImode
-		      ? gen_ashrsi3
-		      : gen_ashrdi3) (high[0], high[0],
-				      GEN_INT (single_width - 1)));
+	  emit_insn (gen_ashr3 (high[0], high[0],
+				GEN_INT (single_width - 1)));
 	  emit_move_insn (low[0], high[0]);
 
 	}
@@ -18239,64 +18253,64 @@  ix86_split_ashr (rtx *operands, rtx scra
 	{
 	  emit_move_insn (low[0], high[1]);
 	  emit_move_insn (high[0], low[0]);
-	  emit_insn ((mode == DImode
-		      ? gen_ashrsi3
-		      : gen_ashrdi3) (high[0], high[0],
-				      GEN_INT (single_width - 1)));
+	  emit_insn (gen_ashr3 (high[0], high[0],
+				GEN_INT (single_width - 1)));
+
 	  if (count > single_width)
-	    emit_insn ((mode == DImode
-			? gen_ashrsi3
-			: gen_ashrdi3) (low[0], low[0],
-					GEN_INT (count - single_width)));
+	    emit_insn (gen_ashr3 (low[0], low[0],
+				  GEN_INT (count - single_width)));
 	}
       else
 	{
+	  gen_shrd = mode == DImode ? gen_x86_shrd : gen_x86_64_shrd;
+
 	  if (!rtx_equal_p (operands[0], operands[1]))
 	    emit_move_insn (operands[0], operands[1]);
-	  emit_insn ((mode == DImode
-		      ? gen_x86_shrd
-		      : gen_x86_64_shrd) (low[0], high[0], GEN_INT (count)));
-	  emit_insn ((mode == DImode
-		      ? gen_ashrsi3
-		      : gen_ashrdi3) (high[0], high[0], GEN_INT (count)));
+
+	  emit_insn (gen_shrd (low[0], high[0], GEN_INT (count)));
+	  emit_insn (gen_ashr3 (high[0], high[0], GEN_INT (count)));
 	}
     }
   else
     {
-      if (!rtx_equal_p (operands[0], operands[1]))
+      gen_shrd = mode == DImode ? gen_x86_shrd : gen_x86_64_shrd;
+
+     if (!rtx_equal_p (operands[0], operands[1]))
 	emit_move_insn (operands[0], operands[1]);
 
       split_double_mode (mode, operands, 1, low, high);
 
-      emit_insn ((mode == DImode
-		  ? gen_x86_shrd
-		  : gen_x86_64_shrd) (low[0], high[0], operands[2]));
-      emit_insn ((mode == DImode
-		  ? gen_ashrsi3
-		  : gen_ashrdi3)  (high[0], high[0], operands[2]));
+      emit_insn (gen_shrd (low[0], high[0], operands[2]));
+      emit_insn (gen_ashr3 (high[0], high[0], operands[2]));
 
       if (TARGET_CMOVE && scratch)
 	{
+	  rtx (*gen_x86_shift_adj_1)(rtx, rtx, rtx, rtx)
+	    = mode == DImode ? gen_x86_shiftsi_adj_1 : gen_x86_shiftdi_adj_1;
+
 	  emit_move_insn (scratch, high[0]);
-	  emit_insn ((mode == DImode
-		      ? gen_ashrsi3
-		      : gen_ashrdi3) (scratch, scratch,
-				      GEN_INT (single_width - 1)));
-	  emit_insn ((mode == DImode
-		      ? gen_x86_shiftsi_adj_1
-		      : gen_x86_shiftdi_adj_1) (low[0], high[0], operands[2],
-						scratch));
+	  emit_insn (gen_ashr3 (scratch, scratch,
+				GEN_INT (single_width - 1)));
+	  emit_insn (gen_x86_shift_adj_1 (low[0], high[0], operands[2],
+					  scratch));
 	}
       else
-	emit_insn ((mode == DImode
-		    ? gen_x86_shiftsi_adj_3
-		    : gen_x86_shiftdi_adj_3) (low[0], high[0], operands[2]));
+	{
+	  rtx (*gen_x86_shift_adj_3)(rtx, rtx, rtx)
+	    = mode == DImode ? gen_x86_shiftsi_adj_3 : gen_x86_shiftdi_adj_3;
+
+	  emit_insn (gen_x86_shift_adj_3 (low[0], high[0], operands[2]));
+	}
     }
 }
 
 void
 ix86_split_lshr (rtx *operands, rtx scratch, enum machine_mode mode)
 {
+  rtx (*gen_lshr3)(rtx, rtx, rtx)
+    = mode == DImode ? gen_lshrsi3 : gen_lshrdi3;
+  rtx (*gen_shrd)(rtx, rtx, rtx);
+
   rtx low[2], high[2];
   int count;
   const int single_width = mode == DImode ? 32 : 64;
@@ -18312,50 +18326,48 @@  ix86_split_lshr (rtx *operands, rtx scra
 	  ix86_expand_clear (high[0]);
 
 	  if (count > single_width)
-	    emit_insn ((mode == DImode
-			? gen_lshrsi3
-			: gen_lshrdi3) (low[0], low[0],
-					GEN_INT (count - single_width)));
+	    emit_insn (gen_lshr3 (low[0], low[0],
+				  GEN_INT (count - single_width)));
 	}
       else
 	{
+	  gen_shrd = mode == DImode ? gen_x86_shrd : gen_x86_64_shrd;
+
 	  if (!rtx_equal_p (operands[0], operands[1]))
 	    emit_move_insn (operands[0], operands[1]);
-	  emit_insn ((mode == DImode
-		      ? gen_x86_shrd
-		      : gen_x86_64_shrd) (low[0], high[0], GEN_INT (count)));
-	  emit_insn ((mode == DImode
-		      ? gen_lshrsi3
-		      : gen_lshrdi3) (high[0], high[0], GEN_INT (count)));
+
+	  emit_insn (gen_shrd (low[0], high[0], GEN_INT (count)));
+	  emit_insn (gen_lshr3 (high[0], high[0], GEN_INT (count)));
 	}
     }
   else
     {
+      gen_shrd = mode == DImode ? gen_x86_shrd : gen_x86_64_shrd;
+
       if (!rtx_equal_p (operands[0], operands[1]))
 	emit_move_insn (operands[0], operands[1]);
 
       split_double_mode (mode, operands, 1, low, high);
 
-      emit_insn ((mode == DImode
-		  ? gen_x86_shrd
-		  : gen_x86_64_shrd) (low[0], high[0], operands[2]));
-      emit_insn ((mode == DImode
-		  ? gen_lshrsi3
-		  : gen_lshrdi3) (high[0], high[0], operands[2]));
+      emit_insn (gen_shrd (low[0], high[0], operands[2]));
+      emit_insn (gen_lshr3 (high[0], high[0], operands[2]));
 
-      /* Heh.  By reversing the arguments, we can reuse this pattern.  */
       if (TARGET_CMOVE && scratch)
 	{
+	  rtx (*gen_x86_shift_adj_1)(rtx, rtx, rtx, rtx)
+	    = mode == DImode ? gen_x86_shiftsi_adj_1 : gen_x86_shiftdi_adj_1;
+
 	  ix86_expand_clear (scratch);
-	  emit_insn ((mode == DImode
-		      ? gen_x86_shiftsi_adj_1
-		      : gen_x86_shiftdi_adj_1) (low[0], high[0], operands[2],
-						scratch));
+	  emit_insn (gen_x86_shift_adj_1 (low[0], high[0], operands[2],
+					  scratch));
 	}
       else
-	emit_insn ((mode == DImode
-		    ? gen_x86_shiftsi_adj_2
-		    : gen_x86_shiftdi_adj_2) (low[0], high[0], operands[2]));
+	{
+	  rtx (*gen_x86_shift_adj_2)(rtx, rtx, rtx)
+	    = mode == DImode ? gen_x86_shiftsi_adj_2 : gen_x86_shiftdi_adj_2;
+
+	  emit_insn (gen_x86_shift_adj_2 (low[0], high[0], operands[2]));
+	}
     }
 }
 
@@ -18392,10 +18404,10 @@  ix86_expand_aligntest (rtx variable, int
 static void
 ix86_adjust_counter (rtx countreg, HOST_WIDE_INT value)
 {
-  if (GET_MODE (countreg) == DImode)
-    emit_insn (gen_adddi3 (countreg, countreg, GEN_INT (-value)));
-  else
-    emit_insn (gen_addsi3 (countreg, countreg, GEN_INT (-value)));
+  rtx (*gen_add)(rtx, rtx, rtx)
+    = GET_MODE (countreg) == DImode ? gen_adddi3 : gen_addsi3;
+
+  emit_insn (gen_add (countreg, countreg, GEN_INT (-value)));
 }
 
 /* Zero extend possibly SImode EXP to Pmode register.  */