diff mbox

[i386] : Remaining FP moves cleanups

Message ID BANLkTinDbrWcYi6x1zDsaC34M1tA1iC1oQ@mail.gmail.com
State New
Headers show

Commit Message

Uros Bizjak June 5, 2011, 8:46 p.m. UTC
Hello!

2011-06-05  Uros Bizjak  <ubizjak@gmail.com>

	* config/i386/i386.md (*movdf_internal_rex64) <case 8,9,10>:
	Remove MODE_TI handling.  Remove SSE1 handling in attribute "mode"
	calculation.
	(*movdf_internal_rex64) <case 6,7,8>: Remove MODE_TI handling.
	Simplify MODE_V1DF and MODE_V2SF handling.
	(*movsf_internal): Remove x constraint from operand 1 alternative 7.
	Simplify MODE_SF handling.

Patch was bootstrapped and regression tested on x86_64-pc-linux-gnuj
{,-m32} AVX target. Committed to mainline SVN.

Uros.

Comments

H.J. Lu June 7, 2011, 1:42 p.m. UTC | #1
On Sun, Jun 5, 2011 at 1:46 PM, Uros Bizjak <ubizjak@gmail.com> wrote:
> Hello!
>
> 2011-06-05  Uros Bizjak  <ubizjak@gmail.com>
>
>        * config/i386/i386.md (*movdf_internal_rex64) <case 8,9,10>:
>        Remove MODE_TI handling.  Remove SSE1 handling in attribute "mode"
>        calculation.
>        (*movdf_internal_rex64) <case 6,7,8>: Remove MODE_TI handling.
>        Simplify MODE_V1DF and MODE_V2SF handling.
>        (*movsf_internal): Remove x constraint from operand 1 alternative 7.
>        Simplify MODE_SF handling.
>
> Patch was bootstrapped and regression tested on x86_64-pc-linux-gnuj
> {,-m32} AVX target. Committed to mainline SVN.
>

Hi Uros,

The new *movsf_internal patter has

 [(set (match_operand:SF 0 "nonimmediate_operand"
          "=f,m,f,?r ,?m,x,x,x,m,!*y,!m,!*y,?Yi,?r,!*Ym,!r")
        (match_operand:SF 1 "general_operand"
          "fm,f,G,rmF,Fr,C,x,m,x,m  ,*y,*y ,r  ,Yi,r   ,*Ym"))]
...
    case 7:
    case 8:
      if (TARGET_AVX && REG_P (operands[0]) && REG_P (operands[1]))
        return "vmovss\t{%1, %0, %0|%0, %0, %1}";
      return "%vmovss\t{%1, %0|%0, %1}";

Since 7 is store and 8 is load,

if (TARGET_AVX && REG_P (operands[0]) && REG_P (operands[1]))

will always be false.  Should we remove it?

Thanks.
Uros Bizjak June 7, 2011, 2:02 p.m. UTC | #2
On Tue, Jun 7, 2011 at 3:42 PM, H.J. Lu <hjl.tools@gmail.com> wrote:

>> 2011-06-05  Uros Bizjak  <ubizjak@gmail.com>
>>
>>        * config/i386/i386.md (*movdf_internal_rex64) <case 8,9,10>:
>>        Remove MODE_TI handling.  Remove SSE1 handling in attribute "mode"
>>        calculation.
>>        (*movdf_internal_rex64) <case 6,7,8>: Remove MODE_TI handling.
>>        Simplify MODE_V1DF and MODE_V2SF handling.
>>        (*movsf_internal): Remove x constraint from operand 1 alternative 7.
>>        Simplify MODE_SF handling.
>>
>> Patch was bootstrapped and regression tested on x86_64-pc-linux-gnuj
>> {,-m32} AVX target. Committed to mainline SVN.
>>
>
> Hi Uros,
>
> The new *movsf_internal patter has
>
>  [(set (match_operand:SF 0 "nonimmediate_operand"
>          "=f,m,f,?r ,?m,x,x,x,m,!*y,!m,!*y,?Yi,?r,!*Ym,!r")
>        (match_operand:SF 1 "general_operand"
>          "fm,f,G,rmF,Fr,C,x,m,x,m  ,*y,*y ,r  ,Yi,r   ,*Ym"))]
> ...
>    case 7:
>    case 8:
>      if (TARGET_AVX && REG_P (operands[0]) && REG_P (operands[1]))
>        return "vmovss\t{%1, %0, %0|%0, %0, %1}";
>      return "%vmovss\t{%1, %0|%0, %1}";
>
> Since 7 is store and 8 is load,
>
> if (TARGET_AVX && REG_P (operands[0]) && REG_P (operands[1]))
>
> will always be false.  Should we remove it?

No, because it also handles fallthru from case 6.

OTOH, you are right, this statement can be moved to case 6, without
the check for registers.

Thanks,
Uros.
Uros Bizjak June 7, 2011, 2:12 p.m. UTC | #3
On Tue, Jun 7, 2011 at 4:07 PM, H.J. Lu <hjl.tools@gmail.com> wrote:

>> OTOH, you are right, this statement can be moved to case 6, without
>> the check for registers.
>>
>
> Like this?  OK for trunk with a ChangeLog entry?
>
> Thanks.
>
> --
> H.J.
> ----
> diff --git a/gcc/config/i386/i386.md b/gcc/config/i386/i386.md
> index 58de87b..a61bffb 100644
> --- a/gcc/config/i386/i386.md
> +++ b/gcc/config/i386/i386.md
> @@ -3207,11 +3207,11 @@
>     case 6:
>       if (get_attr_mode (insn) == MODE_V4SF)
>        return "%vmovaps\t{%1, %0|%0, %1}";
> +      else if (TARGET_AVX && REG_P (operands[0]) && REG_P (operands[1]))

if (TARGET_AVX)

> +       return "vmovss\t{%1, %0, %0|%0, %0, %1}";
>
>     case 7:
>     case 8:
> -      if (TARGET_AVX && REG_P (operands[0]) && REG_P (operands[1]))
> -       return "vmovss\t{%1, %0, %0|%0, %0, %1}";
>       return "%vmovss\t{%1, %0|%0, %1}";
>
>     case 9:
>

OK with this change and suitable ChangeLog entry.

Thanks,
Uros.
diff mbox

Patch

Index: i386.md
===================================================================
--- i386.md	(revision 174655)
+++ i386.md	(working copy)
@@ -2956,9 +2956,6 @@ 
     case 10:
       switch (get_attr_mode (insn))
 	{
-	case MODE_TI:
-	  if (!TARGET_SSE_PACKED_SINGLE_INSN_OPTIMAL)
-	    return "%vmovdqa\t{%1, %0|%0, %1}";
 	case MODE_V2DF:
 	  if (!TARGET_SSE_PACKED_SINGLE_INSN_OPTIMAL)
 	    return "%vmovapd\t{%1, %0|%0, %1}";
@@ -2970,8 +2967,7 @@ 
 	case MODE_DF:
 	  if (TARGET_AVX && REG_P (operands[0]) && REG_P (operands[1]))
 	    return "vmovsd\t{%1, %0, %0|%0, %0, %1}";
-	  else
-	    return "%vmovsd\t{%1, %0|%0, %1}";
+	  return "%vmovsd\t{%1, %0|%0, %1}";
 	case MODE_V1DF:
 	  return "%vmovlpd\t{%1, %d0|%d0, %1}";
 	case MODE_V2SF:
@@ -3014,13 +3010,6 @@ 
 	       (eq_attr "alternative" "3,4,5,6,11,12")
 		 (const_string "DI")
 
-	       /* For SSE1, we have many fewer alternatives.  */
-	       (eq (symbol_ref "TARGET_SSE2") (const_int 0))
-		 (cond [(eq_attr "alternative" "7,8")
-			  (const_string "V4SF")
-		       ]
-		   (const_string "V2SF"))
-
 	       /* xorps is one byte shorter.  */
 	       (eq_attr "alternative" "7")
 		 (cond [(ne (symbol_ref "optimize_function_for_size_p (cfun)")
@@ -3099,9 +3088,6 @@ 
     case 8:
       switch (get_attr_mode (insn))
 	{
-	case MODE_TI:
-	  if (!TARGET_SSE_PACKED_SINGLE_INSN_OPTIMAL)
-	    return "%vmovdqa\t{%1, %0|%0, %1}";
 	case MODE_V2DF:
 	  if (!TARGET_SSE_PACKED_SINGLE_INSN_OPTIMAL)
 	    return "%vmovapd\t{%1, %0|%0, %1}";
@@ -3113,18 +3099,11 @@ 
 	case MODE_DF:
 	  if (TARGET_AVX && REG_P (operands[0]) && REG_P (operands[1]))
 	    return "vmovsd\t{%1, %0, %0|%0, %0, %1}";
-	  else
-	    return "%vmovsd\t{%1, %0|%0, %1}";
+	  return "%vmovsd\t{%1, %0|%0, %1}";
 	case MODE_V1DF:
-	  if (TARGET_AVX && REG_P (operands[0]))
-	    return "vmovlpd\t{%1, %0, %0|%0, %0, %1}";
-	  else
-	    return "%vmovlpd\t{%1, %0|%0, %1}";
+	  return "%vmovlpd\t{%1, %d0|%d0, %1}";
 	case MODE_V2SF:
-	  if (TARGET_AVX && REG_P (operands[0]))
-	    return "vmovlps\t{%1, %0, %0|%0, %0, %1}";
-	  else
-	    return "%vmovlps\t{%1, %0|%0, %1}";
+	  return "%vmovlps\t{%1, %d0|%d0, %1}";
 	default:
 	  gcc_unreachable ();
 	}
@@ -3150,9 +3129,9 @@ 
 
 	       /* For SSE1, we have many fewer alternatives.  */
 	       (eq (symbol_ref "TARGET_SSE2") (const_int 0))
-		 (cond [(eq_attr "alternative" "5,6")
-			  (const_string "V4SF")
-		       ]
+		 (if_then_else
+		   (eq_attr "alternative" "5,6")
+		   (const_string "V4SF")
 		   (const_string "V2SF"))
 
 	       /* xorps is one byte shorter.  */
@@ -3195,9 +3174,9 @@ 
 
 (define_insn "*movsf_internal"
   [(set (match_operand:SF 0 "nonimmediate_operand"
-	  "=f,m,f,?r ,?m,x,x,x ,m,!*y,!m,!*y,?Yi,?r,!*Ym,!r")
+	  "=f,m,f,?r ,?m,x,x,x,m,!*y,!m,!*y,?Yi,?r,!*Ym,!r")
 	(match_operand:SF 1 "general_operand"
-	  "fm,f,G,rmF,Fr,C,x,xm,x,m  ,*y,*y ,r  ,Yi,r   ,*Ym"))]
+	  "fm,f,G,rmF,Fr,C,x,m,x,m  ,*y,*y ,r  ,Yi,r   ,*Ym"))]
   "!(MEM_P (operands[0]) && MEM_P (operands[1]))
    && (!can_create_pseudo_p ()
        || (ix86_cmodel == CM_MEDIUM || ix86_cmodel == CM_LARGE)
@@ -3228,23 +3207,24 @@ 
     case 6:
       if (get_attr_mode (insn) == MODE_V4SF)
 	return "%vmovaps\t{%1, %0|%0, %1}";
-      else
-	return "%vmovss\t{%1, %d0|%d0, %1}";
+
     case 7:
-      if (TARGET_AVX && REG_P (operands[1]))
-	return "vmovss\t{%1, %0, %0|%0, %0, %1}";
-      else
-	return "%vmovss\t{%1, %0|%0, %1}";
     case 8:
+      if (TARGET_AVX && REG_P (operands[0]) && REG_P (operands[1]))
+	return "vmovss\t{%1, %0, %0|%0, %0, %1}";
       return "%vmovss\t{%1, %0|%0, %1}";
 
-    case 9: case 10: case 14: case 15:
+    case 9:
+    case 10:
+    case 14:
+    case 15:
       return "movd\t{%1, %0|%0, %1}";
 
     case 11:
       return "movq\t{%1, %0|%0, %1}";
 
-    case 12: case 13:
+    case 12:
+    case 13:
       return "%vmovd\t{%1, %0|%0, %1}";
 
     default: