Use SImode instead of HImode if aligned.

From-SVN: r13479

3 files changed, 485 insertions, 84 deletions

diff --git a/gcc/config/i386/i386.c b/gcc/config/i386/i386.c
index cfeb5eb..a4591f0 100644
--- a/gcc/config/i386/i386.c
+++ b/gcc/config/i386/i386.c
@@ -388,6 +388,85 @@ optimization_options (level)
 #endif
 }
 
+/* Sign-extend a 16-bit constant */
+
+struct rtx_def *
+i386_sext16_if_const (op)
+     struct rtx_def *op;
+{
+  if (GET_CODE (op) == CONST_INT)
+    {
+      HOST_WIDE_INT val = INTVAL (op);
+      HOST_WIDE_INT sext_val;
+      if (val & 0x8000)
+	sext_val = val | ~0xffff;
+      else
+	sext_val = val & 0xffff;
+      if (sext_val != val)
+	op = GEN_INT (sext_val);
+    }
+  return op;
+}
+
+/* Return nonzero if the rtx is aligned */
+
+static int
+i386_aligned_reg_p (regno)
+     int regno;
+{
+  return (regno == STACK_POINTER_REGNUM
+	  || (!flag_omit_frame_pointer
+	      && regno == FRAME_POINTER_REGNUM));
+}
+
+int
+i386_aligned_p (op)
+     rtx op;
+{
+  /* registers and immediate operands are always "aligned" */
+  if (GET_CODE (op) != MEM)
+    return 1;
+
+  /* Don't even try to do any aligned optimizations with volatiles */
+  if (MEM_VOLATILE_P (op))
+    return 0;
+
+  /* Get address of memory operand */
+  op = XEXP (op, 0);
+
+  switch (GET_CODE (op))
+    {
+    case CONST_INT:
+	if (INTVAL (op) & 3)
+	  break;
+	return 1;
+
+    /* match "reg + offset" */
+    case PLUS:
+	if (GET_CODE (XEXP (op, 1)) != CONST_INT)
+	  break;
+	if (INTVAL (XEXP (op, 1)) & 3)
+	  break;
+	op = XEXP (op, 0);
+	if (GET_CODE (op) != REG)
+	  break;
+	/* fall through */
+    case REG:
+	return i386_aligned_reg_p (REGNO (op));
+    }
+  return 0;
+}
+
+/* Return nonzero if INSN looks like it won't compute useful cc bits
+   as a side effect.  This information is only a hint. */
+
+int
+i386_cc_probably_useless_p (insn)
+     rtx insn;
+{
+  return !next_cc0_user (insn);
+}
+
 /* Return nonzero if IDENTIFIER with arguments ARGS is a valid machine specific
    attribute for DECL.  The attributes in ATTRIBUTES have previously been
    assigned to DECL.  */
@@ -854,7 +933,7 @@ asm_add (n, x)
     output_asm_insn (AS1 (dec%L0,%0), xops);
   else if (n == 1)
     output_asm_insn (AS1 (inc%L0,%0), xops);
-  else if (n < 0)
+  else if (n < 0 || n == 128)
     {
       xops[1] = GEN_INT (-n);
       output_asm_insn (AS2 (sub%L0,%1,%0), xops);
diff --git a/gcc/config/i386/i386.h b/gcc/config/i386/i386.h
index ac08c45..20ec844 100644
--- a/gcc/config/i386/i386.h
+++ b/gcc/config/i386/i386.h
@@ -157,7 +157,8 @@ extern int target_flags;
 #define TARGET_PENTIUMPRO (ix86_cpu == PROCESSOR_PENTIUMPRO)
 #define TARGET_USE_LEAVE (ix86_cpu == PROCESSOR_I386)
 #define TARGET_PUSH_MEMORY (ix86_cpu == PROCESSOR_I386)
-#define TARGET_ZERO_EXTEND_WITH_AND (ix86_cpu != PROCESSOR_I386)
+#define TARGET_ZERO_EXTEND_WITH_AND (ix86_cpu != PROCESSOR_I386 \
+				     && ix86_cpu != PROCESSOR_PENTIUMPRO)
 #define TARGET_DOUBLE_WITH_ADD (ix86_cpu != PROCESSOR_I386)
 #define TARGET_USE_BIT_TEST (ix86_cpu == PROCESSOR_I386)
 #define TARGET_UNROLL_STRLEN (ix86_cpu != PROCESSOR_I386)
@@ -2604,6 +2605,9 @@ do {									\
 extern void override_options ();
 extern void order_regs_for_local_alloc ();
 extern char *output_strlen_unroll ();
+extern struct rtx_def *i386_sext16_if_const ();
+extern int i386_aligned_p ();
+extern int i386_cc_probably_useless_p ();
 extern int i386_valid_decl_attribute_p ();
 extern int i386_valid_type_attribute_p ();
 extern int i386_return_pops_args ();
diff --git a/gcc/config/i386/i386.md b/gcc/config/i386/i386.md
index fa969b5..7d8fb2b 100644
--- a/gcc/config/i386/i386.md
+++ b/gcc/config/i386/i386.md
@@ -770,6 +770,12 @@
 	}
     }
 
+  /* use 32-bit test instruction if there are no sign issues */
+  if (GET_CODE (operands[1]) == CONST_INT
+      && !(INTVAL (operands[1]) & ~0x7fff)
+      && i386_aligned_p (operands[0]))
+    return AS2 (test%L0,%1,%k0);
+
   if (CONSTANT_P (operands[1]) || GET_CODE (operands[0]) == MEM)
     return AS2 (test%W0,%1,%0);
 
@@ -931,10 +937,21 @@
 
   if (REG_P (operands[0]))
     {
-      if (REG_P (operands[1]))
-	return AS2 (mov%L0,%k1,%k0);
-      else if (CONSTANT_P (operands[1]))
-	return AS2 (mov%L0,%1,%k0);
+      if (i386_aligned_p (operands[1]))
+	{
+	  operands[1] = i386_sext16_if_const (operands[1]);
+	  return AS2 (mov%L0,%k1,%k0);
+	}
+      if (TARGET_PENTIUMPRO)
+	{
+	  /* movzwl is faster than movw on the Pentium Pro,
+	   * although not as fast as an aligned movl. */
+#ifdef INTEL_SYNTAX
+	  return AS2 (movzx,%1,%k0);
+#else
+	  return AS2 (movz%W0%L0,%1,%k0);
+#endif
+	}
     }
 
   return AS2 (mov%W0,%1,%0);
@@ -1040,7 +1057,7 @@
 {
   rtx link;
   if (operands[1] == const0_rtx && REG_P (operands[0]))
-    return AS2 (xor%B0,%0,%0);
+    return AS2 (xor%L0,%k0,%k0);
 
   if (operands[1] == const1_rtx
       && (link = find_reg_note (insn, REG_WAS_0, 0))
@@ -1747,7 +1764,10 @@
     {
       xops[0] = operands[0];
       xops[1] = gen_rtx (CONST_INT, VOIDmode, 0xffff);
-      output_asm_insn (AS2 (mov%W0,%1,%w0),operands);
+      if (i386_aligned_p (operands[1]))
+	output_asm_insn (AS2 (mov%L0,%k1,%k0),operands);
+      else
+	output_asm_insn (AS2 (mov%W0,%1,%w0),operands);
       output_asm_insn (AS2 (and%L0,%1,%k0), xops);
       RET;
     }
@@ -2752,7 +2772,7 @@
 }")
 
 (define_insn "addsidi3_2"
-  [(set (match_operand:DI 0 "nonimmediate_operand" "=&r,r,o,&r,!&r,r,o,o,!o")
+  [(set (match_operand:DI 0 "nonimmediate_operand" "=&r,r,o,&r,!&r,&r,o,o,!o")
 	(plus:DI (zero_extend:DI (match_operand:SI 2 "general_operand" "o,ri,ri,o,o,ri,ri,i,r"))
 		 (match_operand:DI 1 "general_operand" "0,0,0,iF,ro,roiF,riF,o,o")))
    (clobber (match_scratch:SI 3 "=X,X,X,X,X,X,X,&r,&r"))]
@@ -2952,6 +2972,18 @@
   if (operands[2] == constm1_rtx)
     return AS1 (dec%L0,%0);
 
+  /* subl $-128,%ebx is smaller than addl $128,%ebx. */
+  if (GET_CODE (operands[2]) == CONST_INT
+      && INTVAL (operands[2]) == 128)
+    {
+      /* This doesn't compute the carry bit in the same way
+       * as add%L0, but we use inc and dec above and they
+       * don't set the carry bit at all.  If inc/dec don't need
+       * a CC_STATUS_INIT, this doesn't either... */
+      operands[2] = GEN_INT (-128);
+      return AS2 (sub%L0,%2,%0);
+    }
+
   return AS2 (add%L0,%2,%0);
 }")
 
@@ -3003,9 +3035,11 @@
   "*
 {
   /* ??? what about offsettable memory references? */
-  if (QI_REG_P (operands[0])
+  if (!TARGET_PENTIUMPRO /* partial stalls are just too painful to risk. */
+      && QI_REG_P (operands[0])
       && GET_CODE (operands[2]) == CONST_INT
-      && (INTVAL (operands[2]) & 0xff) == 0)
+      && (INTVAL (operands[2]) & 0xff) == 0
+      && i386_cc_probably_useless_p (insn))
     {
       int byteval = (INTVAL (operands[2]) >> 8) & 0xff;
       CC_STATUS_INIT;
@@ -3019,6 +3053,28 @@
       return AS2 (add%B0,%2,%h0);
     }
 
+  /* Use a 32-bit operation when possible, to avoid the prefix penalty. */
+  if (REG_P (operands[0])
+      && i386_aligned_p (operands[2])
+      && i386_cc_probably_useless_p (insn))
+    {
+      CC_STATUS_INIT;
+
+      if (GET_CODE (operands[2]) == CONST_INT)
+	{
+	  HOST_WIDE_INT intval = 0xffff & INTVAL (operands[2]);
+
+	  if (intval == 1)
+	    return AS1 (inc%L0,%k0);
+
+	  if (intval == 0xffff)
+	    return AS1 (dec%L0,%k0);
+
+	  operands[2] = i386_sext16_if_const (operands[2]);
+	}
+      return AS2 (add%L0,%k2,%k0);
+    }
+
   if (operands[2] == const1_rtx)
     return AS1 (inc%W0,%0);
 
@@ -3246,7 +3302,18 @@
 	(minus:HI (match_operand:HI 1 "nonimmediate_operand" "0,0")
 		  (match_operand:HI 2 "general_operand" "ri,rm")))]
   "ix86_binary_operator_ok (MINUS, HImode, operands)"
-  "* return AS2 (sub%W0,%2,%0);")
+  "*
+{
+  if (REG_P (operands[0])
+      && i386_aligned_p (operands[2])
+      && i386_cc_probably_useless_p (insn))
+    {
+      CC_STATUS_INIT;
+      operands[2] = i386_sext16_if_const (operands[2]);
+      return AS2 (sub%L0,%k2,%k0);
+    }
+ return AS2 (sub%W0,%2,%0);
+}")
 
 (define_expand "subqi3"
   [(set (match_operand:QI 0 "general_operand" "")
@@ -3518,7 +3585,6 @@
 ;; The `r' in `rm' for operand 3 looks redundant, but it causes
 ;; optional reloads to be generated if op 3 is a pseudo in a stack slot.
 
-;; ??? What if we only change one byte of an offsettable memory reference?
 (define_insn "andsi3"
   [(set (match_operand:SI 0 "nonimmediate_operand" "=rm,r")
 	(and:SI (match_operand:SI 1 "nonimmediate_operand" "%0,0")
@@ -3526,6 +3592,7 @@
   ""
   "*
 {
+  HOST_WIDE_INT intval;
   if (!rtx_equal_p (operands[0], operands[1])
       && rtx_equal_p (operands[0], operands[2]))
     {
@@ -3534,10 +3601,14 @@
       operands[1] = operands[2];
       operands[2] = tmp;
     }
-  if (GET_CODE (operands[2]) == CONST_INT
-      && ! (GET_CODE (operands[0]) == MEM && MEM_VOLATILE_P (operands[0])))
+  switch (GET_CODE (operands[2]))
     {
-      if (INTVAL (operands[2]) == 0xffff && REG_P (operands[0])
+    case CONST_INT:
+      if (GET_CODE (operands[0]) == MEM && MEM_VOLATILE_P (operands[0]))
+	break;
+      intval = INTVAL (operands[2]);
+      /* zero-extend 16->32? */
+      if (intval == 0xffff && REG_P (operands[0])
 	  && (! REG_P (operands[1])
 	      || REGNO (operands[0]) != 0 || REGNO (operands[1]) != 0)
 	  && (!TARGET_ZERO_EXTEND_WITH_AND || ! rtx_equal_p (operands[0], operands[1])))
@@ -3552,7 +3623,8 @@
 #endif
 	}
 
-      if (INTVAL (operands[2]) == 0xff && REG_P (operands[0])
+      /* zero extend 8->32? */
+      if (intval == 0xff && REG_P (operands[0])
 	  && !(REG_P (operands[1]) && NON_QI_REG_P (operands[1]))
 	  && (! REG_P (operands[1])
 	      || REGNO (operands[0]) != 0 || REGNO (operands[1]) != 0)
@@ -3568,39 +3640,99 @@
 #endif
 	}
 
-      if (QI_REG_P (operands[0]) && ~(INTVAL (operands[2]) | 0xff) == 0)
-	{
-	  CC_STATUS_INIT;
+      /* Check partial bytes.. non-QI-regs are not available */
+      if (REG_P (operands[0]) && ! QI_REG_P (operands[0]))
+	break;
 
-	  if (INTVAL (operands[2]) == 0xffffff00)
+      /* only low byte has zero bits? */
+      if (~(intval | 0xff) == 0)
+	{
+	  intval &= 0xff;
+	  if (REG_P (operands[0]))
 	    {
-	      operands[2] = const0_rtx;
-	      return AS2 (mov%B0,%2,%b0);
+	      if (intval == 0)
+		{
+		  CC_STATUS_INIT;
+		  return AS2 (xor%B0,%b0,%b0);
+		}
+
+	      /* we're better off with the 32-bit version if reg != EAX */
+	      /* the value is sign-extended in 8 bits */
+	      if (REGNO (operands[0]) != 0 && (intval & 0x80))
+		break;
 	    }
 
-	  operands[2] = GEN_INT (INTVAL (operands[2]) & 0xff);
+	  CC_STATUS_INIT;
+
+	  operands[2] = GEN_INT (intval);
+
+	  if (intval == 0)
+	    return AS2 (mov%B0,%2,%b0);
+
 	  return AS2 (and%B0,%2,%b0);
 	}
 
-      if (QI_REG_P (operands[0]) && ~(INTVAL (operands[2]) | 0xff00) == 0)
+      /* only second byte has zero? */
+      if (~(intval | 0xff00) == 0)
 	{
 	  CC_STATUS_INIT;
 
-	  if (INTVAL (operands[2]) == 0xffff00ff)
+	  intval = (intval >> 8) & 0xff;
+	  operands[2] = GEN_INT (intval);
+	  if (intval == 0)
 	    {
-	      operands[2] = const0_rtx;
-	      return AS2 (mov%B0,%2,%h0);
+	      if (REG_P (operands[0]))
+		return AS2 (xor%B0,%h0,%h0);
+	      operands[0] = adj_offsettable_operand (operands[0], 1);
+	      return AS2 (mov%B0,%2,%b0);
 	    }
 
-	  operands[2] = GEN_INT ((INTVAL (operands[2]) >> 8) & 0xff);
-	  return AS2 (and%B0,%2,%h0);
+	  if (REG_P (operands[0]))
+	    return AS2 (and%B0,%2,%h0);
+
+	  operands[0] = adj_offsettable_operand (operands[0], 1);
+	  return AS2 (and%B0,%2,%b0);
 	}
 
-      if (GET_CODE (operands[0]) == MEM && INTVAL (operands[2]) == 0xffff0000)
+      if (REG_P (operands[0]))
+	break;
+
+      /* third byte has zero bits? */
+      if (~(intval | 0xff0000) == 0)
+	{
+	  intval = (intval >> 16) & 0xff;
+	  operands[0] = adj_offsettable_operand (operands[0], 2);
+byte_and_operation:
+	  CC_STATUS_INIT;
+	  operands[2] = GEN_INT (intval);
+	  if (intval == 0)
+	    return AS2 (mov%B0,%2,%b0);
+	  return AS2 (and%B0,%2,%b0);
+	}
+
+      /* fourth byte has zero bits? */
+      if (~(intval | 0xff000000) == 0)
+	{
+	  intval = (intval >> 24) & 0xff;
+	  operands[0] = adj_offsettable_operand (operands[0], 3);
+	  goto byte_and_operation;
+	}
+
+      /* Low word is zero? */
+      if (intval == 0xffff0000)
         {
+word_zero_and_operation:
+	  CC_STATUS_INIT;
 	  operands[2] = const0_rtx;
 	  return AS2 (mov%W0,%2,%w0);
 	}
+
+      /* High word is zero? */
+      if (intval == 0x0000ffff)
+        {
+	  operands[0] = adj_offsettable_operand (operands[0], 2);
+	  goto word_zero_and_operation;
+	}
     }
 
   return AS2 (and%L0,%2,%0);
@@ -3647,6 +3779,38 @@
 	  operands[2] = GEN_INT ((INTVAL (operands[2]) >> 8) & 0xff);
 	  return AS2 (and%B0,%2,%h0);
 	}
+
+      /* use 32-bit ops on registers when there are no sign issues.. */
+      if (REG_P (operands[0]))
+	{
+	  if (!(INTVAL (operands[2]) & ~0x7fff))
+	    return AS2 (and%L0,%2,%k0);
+	}
+    }
+
+  if (REG_P (operands[0])
+      && i386_aligned_p (operands[2]))
+    {
+      CC_STATUS_INIT;
+      /* If op[2] is constant, we should zero-extend it and */
+      /* make a note that op[0] has been zero-extended, so  */
+      /* that we could use 32-bit ops on it forthwith, but  */
+      /* there is no such reg-note available. Instead we do */
+      /* a sign extension as that can result in shorter asm */
+      operands[2] = i386_sext16_if_const (operands[2]);
+      return AS2 (and%L0,%k2,%k0);
+    }
+
+  /* Use a 32-bit word with the upper bits set, invalidate CC */
+  if (GET_CODE (operands[2]) == CONST_INT
+      && i386_aligned_p (operands[0]))
+    {
+      HOST_WIDE_INT val = INTVAL (operands[2]);
+      CC_STATUS_INIT;
+      val |= ~0xffff;
+      if (val != INTVAL (operands[2]))
+	operands[2] = GEN_INT (val);
+      return AS2 (and%L0,%k2,%k0);
     }
 
   return AS2 (and%W0,%2,%0);
@@ -3685,7 +3849,10 @@
 
 ;;- Bit set (inclusive or) instructions
 
-;; ??? What if we only change one byte of an offsettable memory reference?
+;; This optimizes known byte-wide operations to memory, and in some cases
+;; to QI registers.. Note that we don't want to use the QI registers too
+;; aggressively, because often the 32-bit register instruction is the same
+;; size, and likely to be faster on PentiumPro.
 (define_insn "iorsi3"
   [(set (match_operand:SI 0 "nonimmediate_operand" "=rm,r")
 	(ior:SI (match_operand:SI 1 "nonimmediate_operand" "%0,0")
@@ -3693,29 +3860,76 @@
   ""
   "*
 {
-  if (GET_CODE (operands[2]) == CONST_INT
-      && ! (GET_CODE (operands[0]) == MEM && MEM_VOLATILE_P (operands[0])))
+  HOST_WIDE_INT intval;
+  switch (GET_CODE (operands[2]))
     {
-      if ((! REG_P (operands[0]) || QI_REG_P (operands[0]))
-	  && (INTVAL (operands[2]) & ~0xff) == 0)
-	{
-	  CC_STATUS_INIT;
+    case CONST_INT:
 
-	  if (INTVAL (operands[2]) == 0xff)
-	    return AS2 (mov%B0,%2,%b0);
+      if (REG_P (operands[0]) && ! QI_REG_P (operands[0]))
+	break;
+
+      /* don't try to optimize volatile accesses */
+      if (GET_CODE (operands[0]) == MEM && MEM_VOLATILE_P (operands[0]))
+	break;
+
+      intval = INTVAL (operands[2]);
+      if ((intval & ~0xff) == 0)
+        {
+	  if (REG_P (operands[0]))
+	    {
+	      /* Do low byte access only for %eax or when high bit is set */
+	      if (REGNO (operands[0]) != 0 && !(intval & 0x80))
+	        break;
+	    }
 
-	  return AS2 (or%B0,%2,%b0);
+byte_or_operation:
+	    CC_STATUS_INIT;
+
+	    if (intval != INTVAL (operands[2]))
+	      operands[2] = GEN_INT (intval);
+
+	    if (intval == 0xff)
+	      return AS2 (mov%B0,%2,%b0);
+
+	    return AS2 (or%B0,%2,%b0);
 	}
 
-      if (QI_REG_P (operands[0]) && (INTVAL (operands[2]) & ~0xff00) == 0)
+      /* second byte? */
+      if ((intval & ~0xff00) == 0)
 	{
-	  CC_STATUS_INIT;
-	  operands[2] = GEN_INT (INTVAL (operands[2]) >> 8);
+	  intval >>= 8;
 
-	  if (INTVAL (operands[2]) == 0xff)
-	    return AS2 (mov%B0,%2,%h0);
+	  if (REG_P (operands[0]))
+	    {
+	      CC_STATUS_INIT;
+	      operands[2] = GEN_INT (intval);
+	      if (intval == 0xff)
+		return AS2 (mov%B0,%2,%h0);
 
-	  return AS2 (or%B0,%2,%h0);
+	      return AS2 (or%B0,%2,%h0);
+	    }
+
+	  operands[0] = adj_offsettable_operand (operands[0], 1);
+	  goto byte_or_operation;
+	}
+
+      if (REG_P (operands[0]))
+	break;
+
+      /* third byte? */
+      if ((intval & ~0xff0000) == 0)
+	{
+	  intval >>= 16;
+	  operands[0] = adj_offsettable_operand (operands[0], 2);
+	  goto byte_or_operation;
+	}
+
+      /* fourth byte? */
+      if ((intval & ~0xff000000) == 0)
+	{
+	  intval = (intval >> 24) & 0xff;
+	  operands[0] = adj_offsettable_operand (operands[0], 3);
+	  goto byte_or_operation;
 	}
     }
 
@@ -3729,38 +3943,77 @@
   ""
   "*
 {
-  if (GET_CODE (operands[2]) == CONST_INT
-      && ! (GET_CODE (operands[0]) == MEM && MEM_VOLATILE_P (operands[0])))
+  HOST_WIDE_INT intval;
+  switch (GET_CODE (operands[2]))
     {
-      /* Can we ignore the upper byte? */
-      if ((! REG_P (operands[0]) || QI_REG_P (operands[0]))
-	  && (INTVAL (operands[2]) & 0xff00) == 0)
-	{
-	  CC_STATUS_INIT;
-	  if (INTVAL (operands[2]) & 0xffff0000)
-	    operands[2] = GEN_INT (INTVAL (operands[2]) & 0xffff);
+    case CONST_INT:
 
-	  if (INTVAL (operands[2]) == 0xff)
-	    return AS2 (mov%B0,%2,%b0);
+      if (REG_P (operands[0]) && ! QI_REG_P (operands[0]))
+	break;
+
+      /* don't try to optimize volatile accesses */
+      if (GET_CODE (operands[0]) == MEM && MEM_VOLATILE_P (operands[0]))
+	break;
+
+      intval = 0xffff & INTVAL (operands[2]);
+
+      if ((intval & 0xff00) == 0)
+        {
+	  if (REG_P (operands[0]))
+	    {
+	      /* Do low byte access only for %eax or when high bit is set */
+	      if (REGNO (operands[0]) != 0 && !(intval & 0x80))
+	        break;
+	    }
+
+byte_or_operation:
+	    CC_STATUS_INIT;
+
+	    if (intval == 0xff)
+	      return AS2 (mov%B0,%2,%b0);
 
-	  return AS2 (or%B0,%2,%b0);
+	    return AS2 (or%B0,%2,%b0);
 	}
 
-      /* Can we ignore the lower byte? */
-      /* ??? what about offsettable memory references? */
-      if (QI_REG_P (operands[0])
-	  && (INTVAL (operands[2]) & 0xff) == 0)
+      /* high byte? */
+      if ((intval & 0xff) == 0)
 	{
-	  CC_STATUS_INIT;
-	  operands[2] = GEN_INT ((INTVAL (operands[2]) >> 8) & 0xff);
+	  intval >>= 8;
+	  operands[2] = GEN_INT (intval);
 
-	  if (INTVAL (operands[2]) == 0xff)
-	    return AS2 (mov%B0,%2,%h0);
+	  if (REG_P (operands[0]))
+	    {
+	      CC_STATUS_INIT;
+	      if (intval == 0xff)
+		return AS2 (mov%B0,%2,%h0);
+
+	      return AS2 (or%B0,%2,%h0);
+	    }
+
+	  operands[0] = adj_offsettable_operand (operands[0], 1);
 
-	  return AS2 (or%B0,%2,%h0);
+	  goto byte_or_operation;
 	}
     }
 
+  if (REG_P (operands[0])
+      && i386_aligned_p (operands[2]))
+    {
+      CC_STATUS_INIT;
+      operands[2] = i386_sext16_if_const (operands[2]);
+      return AS2 (or%L0,%k2,%k0);
+    }
+
+  if (GET_CODE (operands[2]) == CONST_INT
+      && i386_aligned_p (operands[0]))
+    {
+      CC_STATUS_INIT;
+      intval = 0xffff & INTVAL (operands[2]);
+      if (intval != INTVAL (operands[2]))
+	operands[2] = GEN_INT (intval);
+      return AS2 (or%L0,%2,%k0);
+    }
+
   return AS2 (or%W0,%2,%0);
 }")
 
@@ -3773,7 +4026,6 @@
 
 ;;- xor instructions
 
-;; ??? What if we only change one byte of an offsettable memory reference?
 (define_insn "xorsi3"
   [(set (match_operand:SI 0 "nonimmediate_operand" "=rm,r")
 	(xor:SI (match_operand:SI 1 "nonimmediate_operand" "%0,0")
@@ -3781,29 +4033,76 @@
   ""
   "*
 {
-  if (GET_CODE (operands[2]) == CONST_INT
-      && ! (GET_CODE (operands[0]) == MEM && MEM_VOLATILE_P (operands[0])))
+  HOST_WIDE_INT intval;
+  switch (GET_CODE (operands[2]))
     {
-      if ((! REG_P (operands[0]) || QI_REG_P (operands[0]))
-	  && (INTVAL (operands[2]) & ~0xff) == 0)
-	{
-	  CC_STATUS_INIT;
+    case CONST_INT:
 
-	  if (INTVAL (operands[2]) == 0xff)
-	    return AS1 (not%B0,%b0);
+      if (REG_P (operands[0]) && ! QI_REG_P (operands[0]))
+	break;
 
-	  return AS2 (xor%B0,%2,%b0);
+      /* don't try to optimize volatile accesses */
+      if (GET_CODE (operands[0]) == MEM && MEM_VOLATILE_P (operands[0]))
+	break;
+
+      intval = INTVAL (operands[2]);
+      if ((intval & ~0xff) == 0)
+        {
+	  if (REG_P (operands[0]))
+	    {
+	      /* Do low byte access only for %eax or when high bit is set */
+	      if (REGNO (operands[0]) != 0 && !(intval & 0x80))
+	        break;
+	    }
+
+byte_xor_operation:
+	    CC_STATUS_INIT;
+	      
+	    if (intval == 0xff)
+	      return AS1 (not%B0,%b0);
+
+	    if (intval != INTVAL (operands[2]))
+	      operands[2] = GEN_INT (intval);
+	    return AS2 (xor%B0,%2,%b0);
 	}
 
-      if (QI_REG_P (operands[0]) && (INTVAL (operands[2]) & ~0xff00) == 0)
+      /* second byte? */
+      if ((intval & ~0xff00) == 0)
 	{
-	  CC_STATUS_INIT;
-	  operands[2] = GEN_INT (INTVAL (operands[2]) >> 8);
+	  intval >>= 8;
 
-	  if (INTVAL (operands[2]) == 0xff)
-	    return AS1 (not%B0,%h0);
+	  if (REG_P (operands[0]))
+	    {
+	      CC_STATUS_INIT;
+	      if (intval == 0xff)
+		return AS1 (not%B0,%h0);
 
-	  return AS2 (xor%B0,%2,%h0);
+	      operands[2] = GEN_INT (intval);
+	      return AS2 (xor%B0,%2,%h0);
+	    }
+
+	  operands[0] = adj_offsettable_operand (operands[0], 1);
+
+	  goto byte_xor_operation;
+	}
+
+      if (REG_P (operands[0]))
+	break;
+
+      /* third byte? */
+      if ((intval & ~0xff0000) == 0)
+	{
+	  intval >>= 16;
+	  operands[0] = adj_offsettable_operand (operands[0], 2);
+	  goto byte_xor_operation;
+	}
+
+      /* fourth byte? */
+      if ((intval & ~0xff000000) == 0)
+	{
+	  intval = (intval >> 24) & 0xff;
+	  operands[0] = adj_offsettable_operand (operands[0], 3);
+	  goto byte_xor_operation;
 	}
     }
 
@@ -3849,6 +4148,25 @@
 	}
     }
 
+  if (REG_P (operands[0])
+      && i386_aligned_p (operands[2]))
+    {
+      CC_STATUS_INIT;
+      operands[2] = i386_sext16_if_const (operands[2]);
+      return AS2 (xor%L0,%k2,%k0);
+    }
+
+  if (GET_CODE (operands[2]) == CONST_INT
+      && i386_aligned_p (operands[0]))
+    {
+      HOST_WIDE_INT intval;
+      CC_STATUS_INIT;
+      intval = 0xffff & INTVAL (operands[2]);
+      if (intval != INTVAL (operands[2]))
+	operands[2] = GEN_INT (intval);
+      return AS2 (xor%L0,%2,%k0);
+    }
+
   return AS2 (xor%W0,%2,%0);
 }")