[PATCH][m68k] PR target/36134 Prefer LEA over ADD.L

[PATCH][m68k] PR target/36134 Prefer LEA over ADD.L

[Andrew Stubbs]

[-- Attachment #1: Type: text/plain, Size: 461 bytes --]

The attached patch attempts to resolve pr36134 by preferring the 4-byte 
LEA instruction over the longer 6-byte ADD.L instruction, where appropriate.

This patch differs somewhat from the original patch posted by Gunnar von 
Boehm in that it only prefers LEA when it would not use an extra 
register. It also takes the split condition and scheduling attributes 
into account, and adds a test-case.

I've run a full test with no regressions.

OK?

Andrew Stubbs

[-- Attachment #2: pr36134.patch --]
[-- Type: text/x-diff, Size: 3704 bytes --]

2008-11-13  Maxim Kuvyrkov  <maxim@codesourcery.com>
	    Andrew Stubbs  <ams@codesourcery.com>
	    Gunnar Von Boehn  <gunnar@genesi-usa.com>

	gcc/
	PR target/36134
	* config/m68k/m68k.md (addsi3_5200): Add a new alternative preferring
	the shorter LEA insn over ADD.L where possible.

	gcc/testsuite/
	PR target/36134
	* gcc.target/m68k/pr36134.c: New test.

---
 src/gcc-4.3/gcc/config/m68k/m68k.md                 |   25 ++++++++++----------
 src/gcc-4.3/gcc/testsuite/gcc.target/m68k/pr36134.c |   23 ++++++++++++++++++
 2 files changed, 36 insertions(+), 12 deletions(-)

Index: issue3283/src/gcc-4.3/gcc/config/m68k/m68k.md
===================================================================
--- issue3283.orig/src/gcc-4.3/gcc/config/m68k/m68k.md	2008-11-12 15:07:27.000000000 +0000
+++ issue3283/src/gcc-4.3/gcc/config/m68k/m68k.md	2008-11-13 12:42:12.000000000 +0000
@@ -2368,9 +2368,9 @@
   "* return output_addsi3 (operands);")
 
 (define_insn_and_split "*addsi3_5200"
-  [(set (match_operand:SI 0 "nonimmediate_operand"         "=mr,mr,m,r,  ?a,?a,?a,?a")
-	(plus:SI (match_operand:SI 1 "general_operand"     "%0, 0, 0,0,   a, a, r, a")
-		 (match_operand:SI 2 "general_src_operand" " I, L, d,mrKi,Cj,r, a, J")))]
+  [(set (match_operand:SI 0 "nonimmediate_operand"         "=mr,mr,a,m,r,  ?a, ?a,?a,?a")
+	(plus:SI (match_operand:SI 1 "general_operand"     "%0, 0, 0,0,0,   a,  a, r, a")
+		 (match_operand:SI 2 "general_src_operand" " I, L, J,d,mrKi,Cj, r, a, J")))]
   "TARGET_COLDFIRE"
 {
   switch (which_alternative)
@@ -2382,21 +2382,22 @@
       operands[2] = GEN_INT (- INTVAL (operands[2]));
       return "subq%.l %2,%0";
 
-    case 2:
     case 3:
+    case 4:
       return "add%.l %2,%0";
 
-    case 4:
+    case 5:
       /* move%.l %2,%0\n\tadd%.l %1,%0 */
       return "#";
 
-    case 5:
+    case 6:
       return MOTOROLA ? "lea (%1,%2.l),%0" : "lea %1@(0,%2:l),%0";
 
-    case 6:
+    case 7:
       return MOTOROLA ? "lea (%2,%1.l),%0" : "lea %2@(0,%1:l),%0";
 
-    case 7:
+    case 2:
+    case 8:
       return MOTOROLA ? "lea (%c2,%1),%0" : "lea %1@(%c2),%0";
 
     default:
@@ -2404,16 +2405,16 @@
       return "";
     }
 }
-  "&& reload_completed && (extract_constrain_insn_cached (insn), which_alternative == 4) && !operands_match_p (operands[0], operands[1])"
+  "&& reload_completed && (extract_constrain_insn_cached (insn), which_alternative == 5) && !operands_match_p (operands[0], operands[1])"
   [(set (match_dup 0)
 	(match_dup 2))
    (set (match_dup 0)
 	(plus:SI (match_dup 0)
 		 (match_dup 1)))]
   ""
-  [(set_attr "type" "aluq_l,aluq_l,alu_l,alu_l,*,lea,lea,lea")
-   (set_attr "opy" "2,2,2,2,*,*,*,*")
-   (set_attr "opy_type" "*,*,*,*,*,mem6,mem6,mem5")])
+  [(set_attr "type" "aluq_l,aluq_l,lea,alu_l,alu_l,*,lea,lea,lea")
+   (set_attr "opy" "2,2,*,2,2,*,*,*,*")
+   (set_attr "opy_type" "*,*,mem5,*,*,*,mem6,mem6,mem5")])
 
 (define_insn ""
   [(set (match_operand:SI 0 "nonimmediate_operand" "=a")
Index: issue3283/src/gcc-4.3/gcc/testsuite/gcc.target/m68k/pr36134.c
===================================================================
--- /dev/null	1970-01-01 00:00:00.000000000 +0000
+++ issue3283/src/gcc-4.3/gcc/testsuite/gcc.target/m68k/pr36134.c	2008-11-12 15:10:47.000000000 +0000
@@ -0,0 +1,23 @@
+/* pr36134.c
+
+   This test ensures that the shorter LEA instruction is used in preference
+   to the longer ADD instruction.  */
+
+/* { dg-do compile }  */
+/* { dg-options "-O2" }  */
+/* { dg-final { scan-assembler "lea" } } */
+/* { dg-final { scan-assembler-not "add" } } */
+
+int *a, *b;
+
+void
+f ()
+{
+  while (a > b)
+    {
+      *a++ = *b++;
+      *a++ = *b++;
+      *a++ = *b++;
+      *a++ = *b++;
+    }
+}

Re: [PATCH][m68k] PR target/36134 Prefer LEA over ADD.L

[Jeff Law]

Andrew Stubbs wrote:
> The attached patch attempts to resolve pr36134 by preferring the 
> 4-byte LEA instruction over the longer 6-byte ADD.L instruction, where 
> appropriate.
>
> This patch differs somewhat from the original patch posted by Gunnar 
> von Boehm in that it only prefers LEA when it would not use an extra 
> register. It also takes the split condition and scheduling attributes 
> into account, and adds a test-case.
>
> I've run a full test with no regressions.
>
> OK?
>
> Andrew Stubbs
This is fine.  Please install.

Thanks,
Jeff

Re: [PATCH][m68k] PR target/36134 Prefer LEA over ADD.L

[Andrew Stubbs]

Jeff Law wrote:
> This is fine.  Please install.

Committed, thanks.

Andrew