diff mbox

[ARM] Keep constants in register when expanding

Message ID CACgzC7B8QeqxKurQTTUgLnL=24asBDjFtFw+ZoXUG01Wk7=ctg@mail.gmail.com
State New
Headers show

Commit Message

Zhenqiang Chen Aug. 13, 2014, 7:46 a.m. UTC 
On 12 August 2014 17:24, Richard Earnshaw <rearnsha@arm.com> wrote:
> On 05/08/14 10:31, Zhenqiang Chen wrote:
>> Hi,
>>
>> For some large constants, ARM will split them during expanding, which
>> makes impossible to hoist them out the loop or shared by different
>> references (refer the test case in the patch).
>>
>> The patch keeps some constants in registers. If the constant can not
>> be optimized, the cprop and combine passes can optimize them as what
>> we do in current expand pass with
>>
>> define_insn_and_split "*arm_subsi3_insn"
>> define_insn_and_split "*arm_andsi3_insn"
>> define_insn_and_split "*iorsi3_insn"
>> define_insn_and_split "*arm_xorsi3"
>>
>> The patch does not modify addsi3 since the define_insn_and_split
>> "*arm_addsi3" is only valid when (reload_completed ||
>> !arm_eliminable_register (operands[1])). The cprop and combine passes
>> can not optimize the large constant if we put it in register, which
>> will lead to regression.
>>
>> For logic operators, the patch skips changes for constants:
>>
>> INTVAL (operands[2]) < 0 && const_ok_for_arm (-INTVAL (operands[2])
>>
>> since expand pass always uses "sign-extend" to get the value
>> (trunc_int_for_mode called from immed_wide_int_const) for rtl, and
>> logs show most negative values are UNSIGNED when they are TREE node.
>> And combine pass is smart enough to recover the negative value to
>> positive value.
>>
>> Bootstrap and no make check regression on Chrome book.
>> For coremark, dhrystone and eembcv1, no any code size and performance
>> change on Cortex-M4.
>> No any file in CSiBE has code size change for Cortex-A15 and Cortex-M4.
>> No Spec2000 performance regression on Chrome book and dumped assemble
>> codes only show very few difference.
>>
>> OK for trunk?
>>
>
> Not yet.
>
> I think the principle is sound, but there are some issues that need
> sorting out.
>
> 1) We shouldn't do this when not optimizing; there's no advantage to
> hoisting out the constants in that case.  Indeed, it may not be
> worth-while at -O1 either.

Updated. Only apply it when optimize >= 2

> 2) I think you should be using const_ok_for_op rather than working
> through all the permitted values.  If that function isn't doing the
> right thing, then it probably needs extending to handle those cases as well.

There are some difference. I want to filter cases for which
arm_split_constant will get better result. Patch is updated to add a
function const_ok_for_split, in which

(i < 0) && const_ok_for_arm (-i) and 0xffff can not fit into const_ok_for_op.

> 3) Why haven't you handled addsi3?

As I had mentioned in the original mail, since the
define_insn_and_split "*arm_addsi3" is only valid when
(reload_completed || !arm_eliminable_register (operands[1])). The
cprop and combine passes can not optimize the large constant if we put
it in register (and it is not hoisted out of the loop), which will
lead to regression.

Any reason for *arm_addsi3 with additional condition (reload_completed
|| !arm_eliminable_register (operands[1]))? Without it, addsi3 can be
handled as subsi3.

> See below for some specific comments.

All are updated.

Thanks!
-Zhenqiang

>>
>> ChangeLog:
>> 2014-08-05  Zhenqiang Chen  <zhenqiang.chen@linaro.org>
>>
>>         * config/arm/arm.md (subsi3, andsi3, iorsi3, xorsi3): Keep some
>>         large constants in register other than split them.
>>
>> testsuite/ChangeLog:
>> 2014-08-05  Zhenqiang Chen  <zhenqiang.chen@linaro.org>
>>
>>         * gcc.target/arm/maskdata.c: New test.
>>
>>
>> skip-split.patch
>>
>>
>> diff --git a/gcc/config/arm/arm.md b/gcc/config/arm/arm.md
>> index bd8ea8f..c8b3001 100644
>> --- a/gcc/config/arm/arm.md
>> +++ b/gcc/config/arm/arm.md
>> @@ -1162,9 +1162,16 @@
>>      {
>>        if (TARGET_32BIT)
>>          {
>> -          arm_split_constant (MINUS, SImode, NULL_RTX,
>> -                           INTVAL (operands[1]), operands[0],
>> -                           operands[2], optimize && can_create_pseudo_p ());
>> +       if (!const_ok_for_arm (INTVAL (operands[1]))
>> +           && can_create_pseudo_p ())
>> +         {
>> +           operands[1] = force_reg (SImode, operands[1]);
>> +           emit_insn (gen_subsi3 (operands[0], operands[1], operands[2]));
>
> The final emit_insn shouldn't be be needed.  Once you've forced
> operands[1] into a register, you can just fall out the bottom of the
> pattern and let the default expansion take over.  But...
>> +         }
>> +       else
>> +         arm_split_constant (MINUS, SImode, NULL_RTX,
>> +                             INTVAL (operands[1]), operands[0], operands[2],
>> +                             optimize && can_create_pseudo_p ());
>>            DONE;
>
> ... You'll need to move the DONE inside the else clause.
>
>>       }
>>        else /* TARGET_THUMB1 */
>> @@ -2077,6 +2084,17 @@
>>             emit_insn (gen_thumb2_zero_extendqisi2_v6 (operands[0],
>>                                                        operands[1]));
>>           }
>> +       else if (!(const_ok_for_arm (INTVAL (operands[2]))
>> +                  || const_ok_for_arm (~INTVAL (operands[2]))
>> +                     /* zero_extendhi instruction is efficient enough.  */
>> +                  || INTVAL (operands[2]) == 0xffff
>> +                  || (INTVAL (operands[2]) < 0
>> +                      && const_ok_for_arm (-INTVAL (operands[2]))))
>> +                && can_create_pseudo_p ())
>> +         {
>> +           operands[2] = force_reg (SImode, operands[2]);
>> +           emit_insn (gen_andsi3 (operands[0], operands[1], operands[2]));
>
> Same here.
>
>> +         }
>>         else
>>           arm_split_constant (AND, SImode, NULL_RTX,
>>                               INTVAL (operands[2]), operands[0],
>> @@ -2882,9 +2900,20 @@
>>      {
>>        if (TARGET_32BIT)
>>          {
>> -          arm_split_constant (IOR, SImode, NULL_RTX,
>> -                           INTVAL (operands[2]), operands[0], operands[1],
>> -                           optimize && can_create_pseudo_p ());
>> +       if (!(const_ok_for_arm (INTVAL (operands[2]))
>> +             || (TARGET_THUMB2
>> +                 && const_ok_for_arm (~INTVAL (operands[2])))
>> +             || (INTVAL (operands[2]) < 0
>> +                 && const_ok_for_arm (-INTVAL (operands[2]))))
>> +           && can_create_pseudo_p ())
>> +         {
>> +           operands[2] = force_reg (SImode, operands[2]);
>> +           emit_insn (gen_iorsi3 (operands[0], operands[1], operands[2]));
>
> And here.
>
>> +         }
>> +       else
>> +         arm_split_constant (IOR, SImode, NULL_RTX,
>> +                             INTVAL (operands[2]), operands[0], operands[1],
>> +                             optimize && can_create_pseudo_p ());
>>            DONE;
>>       }
>>        else /* TARGET_THUMB1 */
>> @@ -3052,9 +3081,18 @@
>>      {
>>        if (TARGET_32BIT)
>>          {
>> -          arm_split_constant (XOR, SImode, NULL_RTX,
>> -                           INTVAL (operands[2]), operands[0], operands[1],
>> -                           optimize && can_create_pseudo_p ());
>> +       if (!(const_ok_for_arm (INTVAL (operands[2]))
>> +             || (INTVAL (operands[2]) < 0
>> +                  && const_ok_for_arm (-INTVAL (operands[2]))))
>> +           && can_create_pseudo_p ())
>> +         {
>> +           operands[2] = force_reg (SImode, operands[2]);
>> +           emit_insn (gen_xorsi3 (operands[0], operands[1], operands[2]));
>
> And here.
>
>> +         }
>> +       else
>> +         arm_split_constant (XOR, SImode, NULL_RTX,
>> +                             INTVAL (operands[2]), operands[0], operands[1],
>> +                             optimize && can_create_pseudo_p ());
>>            DONE;
>>       }
>>        else /* TARGET_THUMB1 */
>>
>> diff --git a/gcc/testsuite/gcc.target/arm/maskdata.c b/gcc/testsuite/gcc.target/arm/maskdata.c
>> new file mode 100644
>> index 0000000..b4231a4
>> --- /dev/null
>> +++ b/gcc/testsuite/gcc.target/arm/maskdata.c
>> @@ -0,0 +1,16 @@
>> +/* { dg-do compile } */
>> +/* { dg-options " -O2 -fno-gcse " }  */
>> +/* { dg-require-effective-target arm_thumb2_ok } */
>> +
>> +#define MASK 0xfe00ff
>> +void maskdata (int * data, int len)
>> +{
>> +   int i = len;
>> +   for (; i > 0; i -= 2)
>> +    {
>> +      data[i] &= MASK;
>> +      data[i + 1] &= MASK;
>> +    }
>> +}
>> +/* { dg-final { scan-assembler "254" } } */
>> +/* { dg-final { scan-assembler "255" } } */
>>
>
>
diff mbox

Patch

diff --git a/gcc/config/arm/arm-protos.h b/gcc/config/arm/arm-protos.h
index be5e72a..f913035 100644
--- a/gcc/config/arm/arm-protos.h
+++ b/gcc/config/arm/arm-protos.h
@@ -51,6 +51,7 @@  extern bool arm_modes_tieable_p (enum machine_mode, enum machine_mode);
 extern int const_ok_for_arm (HOST_WIDE_INT);
 extern int const_ok_for_op (HOST_WIDE_INT, enum rtx_code);
 extern int const_ok_for_dimode_op (HOST_WIDE_INT, enum rtx_code);
+extern int const_ok_for_split (HOST_WIDE_INT, enum rtx_code);
 extern int arm_split_constant (RTX_CODE, enum machine_mode, rtx,
 			       HOST_WIDE_INT, rtx, rtx, int);
 extern int legitimate_pic_operand_p (rtx);
diff --git a/gcc/config/arm/arm.c b/gcc/config/arm/arm.c
index 7f62ca4..78a624d 100644
--- a/gcc/config/arm/arm.c
+++ b/gcc/config/arm/arm.c
@@ -3584,6 +3584,38 @@  const_ok_for_dimode_op (HOST_WIDE_INT i, enum rtx_code code)
     }
 }
 
+/* Return true if I is a valid constant for split with the operation CODE.
+   The condition should align with the constrain of the corresponding
+   define_insn_and_split pattern to make sure later pass can optimize
+   the constants.  */
+int
+const_ok_for_split (HOST_WIDE_INT i, enum rtx_code code)
+{
+  if (optimize < 2
+      || !can_create_pseudo_p ()
+      || const_ok_for_arm (i)
+	 /* Since expand pass always uses "sign-extend" to get the value
+	    (trunc_int_for_mode called from immed_wide_int_const) for rtl,
+	    and logs show most negative values are UNSIGNED when they are
+	    TREE node. And combine pass is smart enough to recover the
+	    negative value to positive value.  */
+      || ((i < 0) && const_ok_for_arm (-i)))
+    return 1;
+
+  switch (code)
+    {
+    case AND:
+      /* zero_extendhi instruction is efficient.  */
+      return const_ok_for_arm (~i) || (i == 0xffff);
+
+    case IOR:
+      return TARGET_THUMB2 && const_ok_for_arm (~i);
+
+    default:
+      return 1;
+    }
+}
+
 /* Emit a sequence of insns to handle a large constant.
    CODE is the code of the operation required, it can be any of SET, PLUS,
    IOR, AND, XOR, MINUS;
diff --git a/gcc/config/arm/arm.md b/gcc/config/arm/arm.md
index cd9ab6c..f3ca849 100644
--- a/gcc/config/arm/arm.md
+++ b/gcc/config/arm/arm.md
@@ -1162,10 +1162,16 @@ 
     {
       if (TARGET_32BIT)
         {
-          arm_split_constant (MINUS, SImode, NULL_RTX,
-	                      INTVAL (operands[1]), operands[0],
-	  		      operands[2], optimize && can_create_pseudo_p ());
-          DONE;
+	  if (!const_ok_for_split (INTVAL (operands[1]), MINUS))
+	    operands[1] = force_reg (SImode, operands[1]);
+	  else
+	    {
+	      arm_split_constant (MINUS, SImode, NULL_RTX,
+				  INTVAL (operands[1]), operands[0],
+				  operands[2],
+				  optimize && can_create_pseudo_p ());
+	      DONE;
+	    }
 	}
       else /* TARGET_THUMB1 */
         operands[1] = force_reg (SImode, operands[1]);
@@ -2076,14 +2082,19 @@ 
 	      operands[1] = convert_to_mode (QImode, operands[1], 1);
 	      emit_insn (gen_thumb2_zero_extendqisi2_v6 (operands[0],
 							 operands[1]));
+	      DONE;
 	    }
+	  else if (!const_ok_for_split (INTVAL (operands[2]), AND))
+	    operands[2] = force_reg (SImode, operands[2]);
 	  else
-	    arm_split_constant (AND, SImode, NULL_RTX,
-				INTVAL (operands[2]), operands[0],
-				operands[1],
-				optimize && can_create_pseudo_p ());
+	    {
+	      arm_split_constant (AND, SImode, NULL_RTX,
+				  INTVAL (operands[2]), operands[0],
+				  operands[1],
+				  optimize && can_create_pseudo_p ());
 
-          DONE;
+	      DONE;
+	    }
         }
     }
   else /* TARGET_THUMB1 */
@@ -2882,10 +2893,16 @@ 
     {
       if (TARGET_32BIT)
         {
-          arm_split_constant (IOR, SImode, NULL_RTX,
-	                      INTVAL (operands[2]), operands[0], operands[1],
-			      optimize && can_create_pseudo_p ());
-          DONE;
+	  if (!const_ok_for_split (INTVAL (operands[2]), IOR))
+	    operands[2] = force_reg (SImode, operands[2]);
+	  else
+	    {
+	      arm_split_constant (IOR, SImode, NULL_RTX,
+				  INTVAL (operands[2]), operands[0],
+				  operands[1],
+				  optimize && can_create_pseudo_p ());
+	      DONE;
+	    }
 	}
       else /* TARGET_THUMB1 */
         {
@@ -3052,10 +3069,16 @@ 
     {
       if (TARGET_32BIT)
         {
-          arm_split_constant (XOR, SImode, NULL_RTX,
-	                      INTVAL (operands[2]), operands[0], operands[1],
-			      optimize && can_create_pseudo_p ());
-          DONE;
+	  if (!const_ok_for_split (INTVAL (operands[2]), XOR))
+	    operands[2] = force_reg (SImode, operands[2]);
+	  else
+	    {
+	      arm_split_constant (XOR, SImode, NULL_RTX,
+				  INTVAL (operands[2]), operands[0],
+				  operands[1],
+				  optimize && can_create_pseudo_p ());
+	      DONE;
+	    }
 	}
       else /* TARGET_THUMB1 */
         {
diff --git a/gcc/testsuite/gcc.target/arm/maskdata.c b/gcc/testsuite/gcc.target/arm/maskdata.c
new file mode 100644
index 0000000..6960c75
--- /dev/null
+++ b/gcc/testsuite/gcc.target/arm/maskdata.c
@@ -0,0 +1,15 @@ 
+/* { dg-do compile } */
+/* { dg-options " -O2 -fno-gcse " }  */
+/* { dg-require-effective-target arm_thumb2_ok } */
+
+#define MASK 0xfe00ff
+void maskdata (int * data, int len)
+{
+   int i = len;
+   for (; i > 0; i -= 2)
+    {
+      data[i] &= MASK;
+      data[i + 1] &= MASK;
+    }
+}
+/* { dg-final { scan-assembler-not "65536" } } */