#! /bin/sh -e

# DP: <your description>

dir=
if [ $# -eq 3 -a "$2" = '-d' ]; then
    pdir="-d $3"
    dir="$3/"
elif [ $# -ne 1 ]; then
    echo >&2 "`basename $0`: script expects -patch|-unpatch as argument"
    exit 1
fi
case "$1" in
    -patch)
        patch $pdir -f --no-backup-if-mismatch -p0 < $0
        ;;
    -unpatch)
        patch $pdir -f --no-backup-if-mismatch -R -p0 < $0
        ;;
    *)
        echo >&2 "`basename $0`: script expects -patch|-unpatch as argument"
        exit 1
esac
exit 0

From: "H.J. Lu" <hjl@lucon.org>
Sender: gcc-patches-owner@gcc.gnu.org
To: gcc-patches@gcc.gnu.org
Subject: PATCH: PR target/30961: [4.1/4.2/4.3 regression] redundant reg/mem 	stores/moves
Date: Mon, 27 Aug 2007 11:34:12 -0700

We start with

(note:HI 3 4 22 2 NOTE_INSN_FUNCTION_BEG)

(insn:HI 6 3 10 2 c.c:3 (set (reg:DF 58 [ <result> ])
        (subreg:DF (reg/v:DI 59 [ in ]) 0)) 102 {*movdf_integer_rex64} (expr_list:REG_DEAD (reg/v:DI 59 [ in ])
        (nil)))

(insn:HI 10 6 16 2 c.c:7 (set (reg/i:DF 21 xmm0 [ <result> ])
        (reg:DF 58 [ <result> ])) 102 {*movdf_integer_rex64} (expr_list:REG_DEAD (reg:DF 58 [ <result> ])
        (nil)))

(insn:HI 16 10 0 2 c.c:7 (use (reg/i:DF 21 xmm0 [ <result> ])) -1 (nil))

we are trying to allocate registers for insn 6 and we allocate
xmm0 for the return value. Reload doesn't check if xmm0 can be used for
DF 59, it allocates xmm1 for DF 59 and generates:

Reloads for insn # 6
Reload 0: reload_in (DF) = (reg:DF 5 di)
        SSE_REGS, RELOAD_FOR_INPUT (opnum = 1), can't combine
        reload_in_reg: (subreg:DF (reg/v:DI 5 di [orig:59 in ] [59]) 0)
        reload_reg_rtx: (reg:DF 22 xmm1)
...

(note:HI 4 1 3 2 [bb 2] NOTE_INSN_BASIC_BLOCK)

(note:HI 3 4 22 2 NOTE_INSN_FUNCTION_BEG)

(insn 22 3 23 2 c.c:3 (set (mem/c:DF (plus:DI (reg/f:DI 7 sp)
                (const_int -8 [0xfffffffffffffff8])) [0 S8 A8])
        (reg:DF 5 di)) 102 {*movdf_integer_rex64} (nil))

(insn 23 22 6 2 c.c:3 (set (reg:DF 22 xmm1)
        (mem/c:DF (plus:DI (reg/f:DI 7 sp)
                (const_int -8 [0xfffffffffffffff8])) [0 S8 A8])) 102 {*movdf_integer_rex64} (nil))

(insn:HI 6 23 16 2 c.c:3 (set (reg:DF 21 xmm0 [orig:58 <result> ] [58])
        (reg:DF 22 xmm1)) 102 {*movdf_integer_rex64} (nil))

(insn 16 6 21 2 c.c:7 (use (reg/i:DF 21 xmm0 [ <result> ])) -1 (nil))

This patch tries to use the destination register when reloading for input. It
generates

Reloads for insn # 6
Reload 0: reload_in (DF) = (reg:DF 5 di)
        SSE_REGS, RELOAD_FOR_INPUT (opnum = 1), can't combine
        reload_in_reg: (subreg:DF (reg/v:DI 5 di [orig:59 in ] [59]) 0)
        reload_reg_rtx: (reg:DF 21 xmm0)
...
(note:HI 4 1 3 2 [bb 2] NOTE_INSN_BASIC_BLOCK)

(note:HI 3 4 22 2 NOTE_INSN_FUNCTION_BEG)

(insn 22 3 23 2 c.c:3 (set (mem/c:DF (plus:DI (reg/f:DI 7 sp)
                (const_int -8 [0xfffffffffffffff8])) [0 S8 A8])
        (reg:DF 5 di)) 102 {*movdf_integer_rex64} (nil))

(insn 23 22 6 2 c.c:3 (set (reg:DF 21 xmm0)
        (mem/c:DF (plus:DI (reg/f:DI 7 sp)
                (const_int -8 [0xfffffffffffffff8])) [0 S8 A8])) 102 {*movdf_integer_rex64} (nil))

(insn:HI 6 23 10 2 c.c:3 (set (reg:DF 22 xmm1 [orig:58 <result> ] [58])
        (reg:DF 21 xmm0)) 102 {*movdf_integer_rex64} (nil))

(insn:HI 10 6 16 2 c.c:7 (set (reg/i:DF 21 xmm0 [ <result> ])
        (reg:DF 22 xmm1 [orig:58 <result> ] [58])) 102 {*movdf_integer_rex64} (nil))

(insn 16 10 21 2 c.c:7 (use (reg/i:DF 21 xmm0 [ <result> ])) -1 (nil))


H.J.
----
gcc/

2007-08-27  H.J. Lu  <hongjiu.lu@intel.com>

	PR target/30961
	* reload1.c (find_reg): Favor the hard register in destination
	if it is usable and a memory location is needed for reload
	input.

gcc/testsuite/

2007-08-27  H.J. Lu  <hongjiu.lu@intel.com>

	PR target/30961
	* gcc.target/i386/pr30961-1.c: New.

--- gcc/reload1.c.second	2007-08-27 09:35:08.000000000 -0700
+++ gcc/reload1.c	2007-08-27 09:36:33.000000000 -0700
@@ -1781,6 +1781,20 @@ find_reg (struct insn_chain *chain, int 
   HARD_REG_SET not_usable;
   HARD_REG_SET used_by_other_reload;
   reg_set_iterator rsi;
+#ifdef SECONDARY_MEMORY_NEEDED
+  rtx body = PATTERN (chain->insn);
+  unsigned int dest_reg = FIRST_PSEUDO_REGISTER;
+
+  if (GET_CODE (body) == SET)
+    {
+      rtx dest = SET_DEST (body);
+
+      if ((REG_P (dest)
+	   || (GET_CODE (dest) == SUBREG
+	       && REG_P (SUBREG_REG (dest)))))
+	dest_reg = reg_or_subregno (dest);
+    }
+#endif
 
   COPY_HARD_REG_SET (not_usable, bad_spill_regs);
   IOR_HARD_REG_SET (not_usable, bad_spill_regs_global);
@@ -1821,6 +1835,18 @@ find_reg (struct insn_chain *chain, int 
 	    this_cost--;
 	  if (rl->out && REG_P (rl->out) && REGNO (rl->out) == regno)
 	    this_cost--;
+#ifdef SECONDARY_MEMORY_NEEDED
+	  /* If a memory location is needed for rl->in and dest_reg
+	     is usable, we will favor it.  */
+	  else if (dest_reg == regno
+		   && rl->in
+		   && REG_P (rl->in)
+		   && REGNO (rl->in) < FIRST_PSEUDO_REGISTER
+		   && SECONDARY_MEMORY_NEEDED (REGNO_REG_CLASS (REGNO (rl->in)),
+					       rl->class,
+					       rl->mode))
+	    this_cost = 0;
+#endif
 	  if (this_cost < best_cost
 	      /* Among registers with equal cost, prefer caller-saved ones, or
 		 use REG_ALLOC_ORDER if it is defined.  */
--- gcc/testsuite/gcc.target/i386/pr30961-1.c.second	2007-08-27 11:01:59.000000000 -0700
+++ gcc/testsuite/gcc.target/i386/pr30961-1.c	2007-08-27 11:02:51.000000000 -0700
@@ -0,0 +1,13 @@
+/* { dg-do compile } */
+/* { dg-require-effective-target lp64 } */
+/* { dg-options "-O2" } */
+
+double
+convert (long long in)
+{
+  double f;
+  __builtin_memcpy( &f, &in, sizeof( in ) );
+  return f;
+}
+
+/* { dg-final { scan-assembler-not "movapd" } } */