[arm] Early split simple DImode equality comparisons

This is the first step of early splitting all the DImode comparison
operations.  We start by factoring the DImode handling out of
arm_gen_compare_reg into its own function.

Simple DImode equality comparisions (such as equality with zero, or
equality with a constant that is zero in one of the two word values
that it comprises) can be done using a single subtract followed by an
ORRS instruction.  This avoids the need for conditional execution.

For example, (r0 != 5) can be written as

	SUB	Rt, R0, gcc-mirror#5
	ORRS	Rt, Rt, R1

The ORRS is now expanded using an SImode pattern that already exists
in the MD file and this gives the register allocator more freedom to
select registers (consecutive pairs are no-longer required).
Furthermore, we can then delete the arm_cmpdi_zero pattern as it is
no-longer required.  We use SUB for the value adjustment as this has a
generally more flexible range of immediates than XOR and what's more
has the opportunity to be relaxed in thumb2 to a 16-bit SUBS
instruction.

	* config/arm/arm.c (arm_select_cc_mode): For DImode equality tests
	return CC_Zmode if comparing against a constant where one word is
	zero.
	(arm_gen_compare_reg): Split DImode handling to ...
	(arm_gen_dicompare_reg): ... here.  Handle equality comparisons
	against simple constants.
	* config/arm/arm.md (arm_cmpdi_zero): Delete pattern.

git-svn-id: svn+ssh://gcc.gnu.org/svn/gcc/trunk@277177 138bc75d-0d04-0410-961f-82ee72b054a4

gcc/ChangeLog

@@ -1,3 +1,13 @@
+2019-10-18  Richard Earnshaw  <rearnsha@arm.com>
+
+	* config/arm/arm.c (arm_select_cc_mode): For DImode equality tests
+	return CC_Zmode if comparing against a constant where one word is
+	zero.
+	(arm_gen_compare_reg): Split DImode handling to ...
+	(arm_gen_dicompare_reg): ... here.  Handle equality comparisons
+	against simple constants.
+	* config/arm/arm.md (arm_cmpdi_zero): Delete pattern.
+
 2019-10-18  Richard Earnshaw  <rearnsha@arm.com>
 
 	* config/arm/arm.md (subsi3_carryin_shift_alt): New pattern.

gcc/config/arm/arm.c

@@ -15350,8 +15350,14 @@ arm_select_cc_mode (enum rtx_code op, rtx x, rtx y)
 	case EQ:
 	case NE:
 	  /* A DImode comparison against zero can be implemented by
-	     or'ing the two halves together.  */
-	  if (y == const0_rtx)
+	     or'ing the two halves together.  We can also handle
+	     immediates where one word of that value is zero by
+	     subtracting the non-zero word from the corresponding word
+	     in the other register and then ORRing it with the other
+	     word.  */
+	  if (CONST_INT_P (y)
+	      && ((UINTVAL (y) & 0xffffffff) == 0
+		  || (UINTVAL (y) >> 32) == 0))
 	    return CC_Zmode;
 
 	  /* We can do an equality test in three Thumb instructions.  */
@@ -15393,37 +15399,64 @@ arm_select_cc_mode (enum rtx_code op, rtx x, rtx y)
   return CCmode;
 }
 
-/* X and Y are two things to compare using CODE.  Emit the compare insn and
-   return the rtx for register 0 in the proper mode.  FP means this is a
-   floating point compare: I don't think that it is needed on the arm.  */
-rtx
-arm_gen_compare_reg (enum rtx_code code, rtx x, rtx y, rtx scratch)
+/* X and Y are two (DImode) things to compare for the condition CODE.  Emit
+   the sequence of instructions needed to generate a suitable condition
+   code register.  Return the CC register result.  */
+static rtx
+arm_gen_dicompare_reg (rtx_code code, rtx x, rtx y, rtx scratch)
 {
-  machine_mode mode;
-  rtx cc_reg;
-  int dimode_comparison = GET_MODE (x) == DImode || GET_MODE (y) == DImode;
+  /* We don't currently handle DImode in thumb1, but rely on libgcc.  */
+  gcc_assert (TARGET_32BIT);
 
   /* We might have X as a constant, Y as a register because of the predicates
      used for cmpdi.  If so, force X to a register here.  */
-  if (dimode_comparison && !REG_P (x))
+  if (!REG_P (x))
     x = force_reg (DImode, x);
 
-  mode = SELECT_CC_MODE (code, x, y);
-  cc_reg = gen_rtx_REG (mode, CC_REGNUM);
+  machine_mode mode = SELECT_CC_MODE (code, x, y);
+  rtx cc_reg = gen_rtx_REG (mode, CC_REGNUM);
 
-  if (dimode_comparison
-      && mode != CC_CZmode)
+  if (mode != CC_CZmode)
     {
       rtx clobber, set;
 
       /* To compare two non-zero values for equality, XOR them and
 	 then compare against zero.  Not used for ARM mode; there
 	 CC_CZmode is cheaper.  */
-      if (mode == CC_Zmode && y != const0_rtx)
+      if (mode == CC_Zmode)
 	{
-	  gcc_assert (!reload_completed);
-	  x = expand_binop (DImode, xor_optab, x, y, NULL_RTX, 0, OPTAB_WIDEN);
-	  y = const0_rtx;
+	  mode = CC_NOOVmode;
+	  PUT_MODE (cc_reg, mode);
+	  if (y != const0_rtx)
+	    {
+	      gcc_assert (CONST_INT_P (y));
+	      rtx xlo, xhi, ylo, yhi;
+	      arm_decompose_di_binop (x, y, &xlo, &xhi, &ylo, &yhi);
+	      if (!scratch)
+		scratch = gen_reg_rtx (SImode);
+	      if (ylo == const0_rtx)
+		{
+		  yhi = GEN_INT (-INTVAL(yhi));
+		  if (!arm_add_operand (yhi, SImode))
+		    yhi = force_reg (SImode, yhi);
+		  emit_insn (gen_addsi3 (scratch, xhi, yhi));
+		  y = xlo;
+		}
+	      else
+		{
+		  gcc_assert (yhi == const0_rtx);
+		  ylo = GEN_INT (-INTVAL(ylo));
+		  if (!arm_add_operand (ylo, SImode))
+		    ylo = force_reg (SImode, ylo);
+		  emit_insn (gen_addsi3 (scratch, xlo, ylo));
+		  y = xhi;
+		}
+	      x = gen_rtx_IOR (SImode, scratch, y);
+	      y = const0_rtx;
+	    }
+	  else
+	    x = gen_rtx_IOR (SImode, gen_lowpart (SImode, x),
+			     gen_highpart (SImode, x));
 	}
 
       /* A scratch register is required.  */
@@ -15442,6 +15475,22 @@ arm_gen_compare_reg (enum rtx_code code, rtx x, rtx y, rtx scratch)
   return cc_reg;
 }
 
+/* X and Y are two things to compare using CODE.  Emit the compare insn and
+   return the rtx for register 0 in the proper mode.  */
+rtx
+arm_gen_compare_reg (rtx_code code, rtx x, rtx y, rtx scratch)
+{
+  if (GET_MODE (x) == DImode || GET_MODE (y) == DImode)
+    return arm_gen_dicompare_reg (code, x, y, scratch);
+
+  machine_mode mode = SELECT_CC_MODE (code, x, y);
+  rtx cc_reg = gen_rtx_REG (mode, CC_REGNUM);
+
+  emit_set_insn (cc_reg, gen_rtx_COMPARE (mode, x, y));
+
+  return cc_reg;
+}
+
 /* Generate a sequence of insns that will generate the correct return
    address mask depending on the physical architecture that the program
    is running on.  */

gcc/config/arm/arm.md

@@ -6518,17 +6518,6 @@
    (set_attr "type" "multiple")]
 )
 
-(define_insn "*arm_cmpdi_zero"
-  [(set (reg:CC_Z CC_REGNUM)
-	(compare:CC_Z (match_operand:DI 0 "s_register_operand" "r")
-		      (const_int 0)))
-   (clobber (match_scratch:SI 1 "=r"))]
-  "TARGET_32BIT"
-  "orrs%?\\t%1, %Q0, %R0"
-  [(set_attr "conds" "set")
-   (set_attr "type" "logics_reg")]
-)
-
 ; This insn allows redundant compares to be removed by cse, nothing should
 ; ever appear in the output file since (set (reg x) (reg x)) is a no-op that
 ; is deleted later on. The match_dup will match the mode here, so that