diff mbox

[ARM] Thumb2 constant loading optimization

Message ID 4CFF8170.6020403@codesourcery.com
State New
Headers show

Commit Message

Andrew Stubbs Dec. 8, 2010, 1 p.m. UTC 
Here is a patch I'd like reviewed for mainline GCC post 4.6. I don't 
think it's suitable for stage 3.

At present, the support for constant loading via immediate operands (as 
opposed to constant pools) is not well tuned for Thumb2. There are a few 
separate issues:

    * 8-bit immediates can have arbitrary shifts applied, but are 
currently limited to even shift offsets. (This appears to be a bug 
rather than a deliberate state since half the support was there, but the 
other half missing.)

    * Addw / subw support is completely missing.

    * Replicated constants are recognised, but constant splitting never 
takes advantage of them.

    * Constants that can be inverted/negated are only identified by very 
crude heuristics, sometimes harmful even in the existing code, and not 
at all suited to replicated constants.

My patch addresses all of these issues. Here are some before and after 
examples of generated code:

Example 1: subw

    a - 0xfff

    Before:
          sub     r0, r0, #4064 ; 0xfe0
          subs    r0, r0, #31   ; 0x01f
    After:
          subw    r0, r0, #4095 ; 0xfff

Example 2: addw

    a + 0xfffff

    Before:
          movw    r3, #65535       ; 0x0ffff
          movt    r3, 15           ; 0xf0000
          adds    r3, r0, r3
    After:
          add     r0, r0, #1044480 ; 0xff000
          addw    r0, r0, #4095    ; 0x00fff

Example 3: arbitrary shifts bug fix

    a - 0xfff1

    Before:
          sub     r0, r0, #65024   ; 0xfe00
          sub     r0, r0, #496     ; 0x01f0
          sub     r0, r0, #1       ; 0x0001
    After:
          sub     r0, r0, #65280   ; 0xff00
          sub     r0, r0, #241     ; 0x00f1

Example 4: 16-bit replicated patterns

    a + 0x44004401

    Before:
          movw    r3, #17409          ; 0x00004401
          movt    r3, 17408           ; 0x44000000
          adds    r3, r0, r3
    After:
          add     r0, r0, #1140868096 ; 0x44004400
          adds    r0, r0, #1          ; 0x00000001

Example 5: 32-bit replicated patterns

    a & 0xaaaaaa00

    Before:
          mov     r3, #43520           ; 0x0000aa00
          movt    r3, 43690            ; 0xaaaa0000
          and     r3, r0, r3
    After:
          and     r0, r0, #-1431655766 ; 0xaaaaaaaa
          bic     r0, r0, #170         ; 0x000000aa

The constant splitting code was duplicated in two places, and I needed 
to find a new way to tackle the negated/inverted constants cases 
(replicated constants render the old rather dumb heuristics completely 
useless), so I have rearranged the code somewhat. Not as much has 
changed as it looks like, but hopefully it's a bit easier to maintain 
now, and I think it now reliably always chooses the most efficient sense 
to encode the constant.

There is one point in this patch that I am uncertain about: I've renamed 
the 'j' constraint to 'ja'. I did this because it allowed me to add jb 
and jB which are similar, but different. The 'j' constraint was not 
documented anywhere, but I think it is possible it may have been used by 
third party code? Is this a problem?

Is the patch OK to commit, once stage 1 returns?

There are still a few extra optimizations that might be worth looking at 
in future:
    * For SET and PLUS operations only, there are some advantages in 
splitting constants arithmetically, rather than purely bitwise (this is 
not true of non-replicated constants so not relevant to ARM/Thumb1).
      - e.g. 0x01010201 = 0x01010101 + 0x00000100
    * 16-bit replicated constants when there isn't an exact match 
(similar to the 32-bit constants used in this patch).
    * something else?

Andrew

Comments

Andrew Stubbs April 12, 2011, 9:22 a.m. UTC | #1 
Ping.

On 08/12/10 13:00, Andrew Stubbs wrote:
> Here is a patch I'd like reviewed for mainline GCC post 4.6. I don't
> think it's suitable for stage 3.
>
> At present, the support for constant loading via immediate operands (as
> opposed to constant pools) is not well tuned for Thumb2. There are a few
> separate issues:
>
> * 8-bit immediates can have arbitrary shifts applied, but are currently
> limited to even shift offsets. (This appears to be a bug rather than a
> deliberate state since half the support was there, but the other half
> missing.)
>
> * Addw / subw support is completely missing.
>
> * Replicated constants are recognised, but constant splitting never
> takes advantage of them.
>
> * Constants that can be inverted/negated are only identified by very
> crude heuristics, sometimes harmful even in the existing code, and not
> at all suited to replicated constants.
>
> My patch addresses all of these issues. Here are some before and after
> examples of generated code:
>
> Example 1: subw
>
> a - 0xfff
>
> Before:
> sub r0, r0, #4064 ; 0xfe0
> subs r0, r0, #31 ; 0x01f
> After:
> subw r0, r0, #4095 ; 0xfff
>
> Example 2: addw
>
> a + 0xfffff
>
> Before:
> movw r3, #65535 ; 0x0ffff
> movt r3, 15 ; 0xf0000
> adds r3, r0, r3
> After:
> add r0, r0, #1044480 ; 0xff000
> addw r0, r0, #4095 ; 0x00fff
>
> Example 3: arbitrary shifts bug fix
>
> a - 0xfff1
>
> Before:
> sub r0, r0, #65024 ; 0xfe00
> sub r0, r0, #496 ; 0x01f0
> sub r0, r0, #1 ; 0x0001
> After:
> sub r0, r0, #65280 ; 0xff00
> sub r0, r0, #241 ; 0x00f1
>
> Example 4: 16-bit replicated patterns
>
> a + 0x44004401
>
> Before:
> movw r3, #17409 ; 0x00004401
> movt r3, 17408 ; 0x44000000
> adds r3, r0, r3
> After:
> add r0, r0, #1140868096 ; 0x44004400
> adds r0, r0, #1 ; 0x00000001
>
> Example 5: 32-bit replicated patterns
>
> a & 0xaaaaaa00
>
> Before:
> mov r3, #43520 ; 0x0000aa00
> movt r3, 43690 ; 0xaaaa0000
> and r3, r0, r3
> After:
> and r0, r0, #-1431655766 ; 0xaaaaaaaa
> bic r0, r0, #170 ; 0x000000aa
>
> The constant splitting code was duplicated in two places, and I needed
> to find a new way to tackle the negated/inverted constants cases
> (replicated constants render the old rather dumb heuristics completely
> useless), so I have rearranged the code somewhat. Not as much has
> changed as it looks like, but hopefully it's a bit easier to maintain
> now, and I think it now reliably always chooses the most efficient sense
> to encode the constant.
>
> There is one point in this patch that I am uncertain about: I've renamed
> the 'j' constraint to 'ja'. I did this because it allowed me to add jb
> and jB which are similar, but different. The 'j' constraint was not
> documented anywhere, but I think it is possible it may have been used by
> third party code? Is this a problem?
>
> Is the patch OK to commit, once stage 1 returns?
>
> There are still a few extra optimizations that might be worth looking at
> in future:
> * For SET and PLUS operations only, there are some advantages in
> splitting constants arithmetically, rather than purely bitwise (this is
> not true of non-replicated constants so not relevant to ARM/Thumb1).
> - e.g. 0x01010201 = 0x01010101 + 0x00000100
> * 16-bit replicated constants when there isn't an exact match (similar
> to the 32-bit constants used in this patch).
> * something else?
>
> Andrew
>
Richard Earnshaw April 12, 2011, 10:02 a.m. UTC | #2 
On Wed, 2010-12-08 at 13:00 +0000, Andrew Stubbs wrote:
> Here is a patch I'd like reviewed for mainline GCC post 4.6. I don't 
> think it's suitable for stage 3.
> 
> At present, the support for constant loading via immediate operands (as 
> opposed to constant pools) is not well tuned for Thumb2. There are a few 
> separate issues:
> 
>     * 8-bit immediates can have arbitrary shifts applied, but are 
> currently limited to even shift offsets. (This appears to be a bug 
> rather than a deliberate state since half the support was there, but the 
> other half missing.)
> 
>     * Addw / subw support is completely missing.
> 
>     * Replicated constants are recognised, but constant splitting never 
> takes advantage of them.
> 
>     * Constants that can be inverted/negated are only identified by very 
> crude heuristics, sometimes harmful even in the existing code, and not 
> at all suited to replicated constants.
> 
> My patch addresses all of these issues. Here are some before and after 
> examples of generated code:
> 
> Example 1: subw
> 
>     a - 0xfff
> 
>     Before:
>           sub     r0, r0, #4064 ; 0xfe0
>           subs    r0, r0, #31   ; 0x01f
>     After:
>           subw    r0, r0, #4095 ; 0xfff
> 
> Example 2: addw
> 
>     a + 0xfffff
> 
>     Before:
>           movw    r3, #65535       ; 0x0ffff
>           movt    r3, 15           ; 0xf0000
>           adds    r3, r0, r3
>     After:
>           add     r0, r0, #1044480 ; 0xff000
>           addw    r0, r0, #4095    ; 0x00fff
> 
> Example 3: arbitrary shifts bug fix
> 
>     a - 0xfff1
> 
>     Before:
>           sub     r0, r0, #65024   ; 0xfe00
>           sub     r0, r0, #496     ; 0x01f0
>           sub     r0, r0, #1       ; 0x0001
>     After:
>           sub     r0, r0, #65280   ; 0xff00
>           sub     r0, r0, #241     ; 0x00f1
> 
> Example 4: 16-bit replicated patterns
> 
>     a + 0x44004401
> 
>     Before:
>           movw    r3, #17409          ; 0x00004401
>           movt    r3, 17408           ; 0x44000000
>           adds    r3, r0, r3
>     After:
>           add     r0, r0, #1140868096 ; 0x44004400
>           adds    r0, r0, #1          ; 0x00000001
> 
> Example 5: 32-bit replicated patterns
> 
>     a & 0xaaaaaa00
> 
>     Before:
>           mov     r3, #43520           ; 0x0000aa00
>           movt    r3, 43690            ; 0xaaaa0000
>           and     r3, r0, r3
>     After:
>           and     r0, r0, #-1431655766 ; 0xaaaaaaaa
>           bic     r0, r0, #170         ; 0x000000aa
> 
> The constant splitting code was duplicated in two places, and I needed 
> to find a new way to tackle the negated/inverted constants cases 
> (replicated constants render the old rather dumb heuristics completely 
> useless), so I have rearranged the code somewhat. Not as much has 
> changed as it looks like, but hopefully it's a bit easier to maintain 
> now, and I think it now reliably always chooses the most efficient sense 
> to encode the constant.
> 
> There is one point in this patch that I am uncertain about: I've renamed 
> the 'j' constraint to 'ja'. I did this because it allowed me to add jb 
> and jB which are similar, but different. The 'j' constraint was not 
> documented anywhere, but I think it is possible it may have been used by 
> third party code? Is this a problem?
> 
> Is the patch OK to commit, once stage 1 returns?
> 
> There are still a few extra optimizations that might be worth looking at 
> in future:
>     * For SET and PLUS operations only, there are some advantages in 
> splitting constants arithmetically, rather than purely bitwise (this is 
> not true of non-replicated constants so not relevant to ARM/Thumb1).
>       - e.g. 0x01010201 = 0x01010101 + 0x00000100
>     * 16-bit replicated constants when there isn't an exact match 
> (similar to the 32-bit constants used in this patch).
>     * something else?
> 
> Andrew


I'd be happier trying to review this patch if it could be broken down
into smaller hunks.  One immediate problem that stands out is:

 (define_insn_and_split "*arm_subsi3_insn"
-  [(set (match_operand:SI           0 "s_register_operand" "=r,r,rk,r,r")
-       (minus:SI (match_operand:SI 1 "reg_or_int_operand" "rI,r,k,?n,r")
-                 (match_operand:SI 2 "reg_or_int_operand" "r,rI,r, r,?n")))]
+  [(set (match_operand:SI           0 "s_register_operand" "=r,r,rk,r, k, r,r")
+       (minus:SI (match_operand:SI 1 "reg_or_int_operand" "rI,r,k, rk,k, ?n,r")
+                 (match_operand:SI 2 "reg_or_int_operand" "r,rI,r, jb,jb,r,?n")))]
   "TARGET_32BIT"
   "@
    rsb%?\\t%0, %2, %1
    sub%?\\t%0, %1, %2
    sub%?\\t%0, %1, %2
+   subw%?\\t%0, %1, %2
+   subw%?\\t%0, %1, %2
    #
    #"
   "&& ((GET_CODE (operands[1]) == CONST_INT
-               && !const_ok_for_arm (INTVAL (operands[1])))
+               && !(const_ok_for_arm (INTVAL (operands[1]))
+            || satisfies_constraint_jb (operands[2])))
        || (GET_CODE (operands[2]) == CONST_INT
-          && !const_ok_for_arm (INTVAL (operands[2]))))"
+          && !(const_ok_for_arm (INTVAL (operands[2]))
+               || satisfies_constraint_jb (operands[2]))))"
   [(clobber (const_int 0))]
   "
   arm_split_constant (MINUS, SImode, curr_insn,
                       INTVAL (operands[1]), operands[0], operands[2], 0);
   DONE;
   "
-  [(set_attr "length" "4,4,4,16,16")
+  [(set_attr "length" "4,4,4,4,4,16,16")
    (set_attr "predicable" "yes")]
 )
 

Firstly, MINUS (REG, CONST) is not canonical (the compiler should
generate PLUS (REG, -CONST)), so your additional patterns should, in
theory, never match.  If they are matching, then really we should work
out why and then fix those rather than adding more code here...

Secondly, these new alternatives seem to be unconditionally enabled.  It
looks as though you are relying on the constraints never matching when
not compiling for Thumb-2 and thus the alternative not being selected.
I'd prefer that the alternative was properly disabled when it wasn't
available by using the 'enabled' attribute.  The same applies to the
addition operation.

Also, your change to use a double-letter sequence beginning with 'j'
means any hand-written inline assembly code using a single 'j' will
break (that's a backwards compatibility issue for users); is there
really no other letter that can be used to prefix the operation?  Also,
think very carefully about whether your new constraints should be
internal only or really public.  I think it would be better to leave 'j'
as is, and add some more P-prefixed constraints that are internal only
for the additional cases.

A final note is that you may have missed some cases.  Now that we have
movw, 
	reg & ~(16-bit const)
can now be done in at most 2 insns:
	movw t1, #16-bit const
	bic  Rd, reg, t1

On thumb-2 you can also use ORN that way as well.

R.

R.
Andrew Stubbs April 12, 2011, 10:14 a.m. UTC | #3 
[Ignoring the other issues for now ...]

On 12/04/11 11:02, Richard Earnshaw wrote:
> Also, your change to use a double-letter sequence beginning with 'j'
> means any hand-written inline assembly code using a single 'j' will
> break (that's a backwards compatibility issue for users); is there
> really no other letter that can be used to prefix the operation?  Also,
> think very carefully about whether your new constraints should be
> internal only or really public.  I think it would be better to leave 'j'
> as is, and add some more P-prefixed constraints that are internal only
> for the additional cases.

The constraints were part of the patch I expected to have to fix up 
following review. :)

I chose to modify the 'j' constraint because it was not documented, last 
I looked, and the new use was related to the meaning of 'j'. I thought 
this might make the patterns a little easier to read. Do we have to 
assume that all constraints might be used by third parties, documented 
or otherwise?

I can switch to using P-prefixed constraints, of course.

Andrew
Richard Earnshaw April 12, 2011, 12:20 p.m. UTC | #4 
On Tue, 2011-04-12 at 11:14 +0100, Andrew Stubbs wrote:
> [Ignoring the other issues for now ...]
> 
> On 12/04/11 11:02, Richard Earnshaw wrote:
> > Also, your change to use a double-letter sequence beginning with 'j'
> > means any hand-written inline assembly code using a single 'j' will
> > break (that's a backwards compatibility issue for users); is there
> > really no other letter that can be used to prefix the operation?  Also,
> > think very carefully about whether your new constraints should be
> > internal only or really public.  I think it would be better to leave 'j'
> > as is, and add some more P-prefixed constraints that are internal only
> > for the additional cases.
> 
> The constraints were part of the patch I expected to have to fix up 
> following review. :)
> 
> I chose to modify the 'j' constraint because it was not documented, last 
> I looked, and the new use was related to the meaning of 'j'. I thought 
> this might make the patterns a little easier to read. Do we have to 
> assume that all constraints might be used by third parties, documented 
> or otherwise?
> 

It may not be in invoke.texi, but it isn't marked @internal in
constraints.md.

> I can switch to using P-prefixed constraints, of course.
> 

I think that would be better.  I think there's a reasonable argument for
keeping 'j' /and/ having it 'public'.

R.

> Andrew
>
diff mbox

Patch

2010-12-08  Andrew Stubbs  <ams@codesourcery.com>

	gcc/
	* config/arm/arm-protos.h (const_ok_for_op): Add prototype.
	* config/arm/arm.c (const_ok_for_op): Add support for movw/addw/subw.
	(count_insns_for_constant, find_best_start): Delete functions.
	(optimal_immediate_sequence): New function.
	(optimal_immediate_sequence_1): New function.
	(arm_gen_constant): Remove old movw support.
	Move constant splitting code to optimal_immediate_sequence.
	Rewrite constant negation/invertion code.
	* config/arm/arm.md (*arm_addsi3): Add addw/subw support.
	(*arm_subsi3_insn): Add subw support.
	(*arm_movsi_insn): Change 'j' constraint to 'ja'.
	* config/arm/constraints.md (j): Rename ...
	(ja): ... the this.
	(jb, jB): New constraints.
	* config/arm/thumb2.md (*thumb2_movsi_insn): Change 'j' constraint
	to 'ja'.
	* config/arm/vfp.md (*arm_movsi_vfp): Likewise.

	gcc/testsuite/
	* gcc.target/arm/thumb2-replicated-constant1.c: New file.
	* gcc.target/arm/thumb2-replicated-constant2.c: New file.
	* gcc.target/arm/thumb2-replicated-constant3.c: New file.
	* gcc.target/arm/thumb2-replicated-constant4.c: New file.

---
 src/gcc-mainline/gcc/config/arm/arm-protos.h       |    1 
 src/gcc-mainline/gcc/config/arm/arm.c              |  456 ++++++++++++--------
 src/gcc-mainline/gcc/config/arm/arm.md             |   35 +-
 src/gcc-mainline/gcc/config/arm/constraints.md     |   18 +
 src/gcc-mainline/gcc/config/arm/thumb2.md          |    4 
 src/gcc-mainline/gcc/config/arm/vfp.md             |    8 
 .../gcc.target/arm/thumb2-replicated-constant1.c   |   27 +
 .../gcc.target/arm/thumb2-replicated-constant2.c   |   75 +++
 .../gcc.target/arm/thumb2-replicated-constant3.c   |   28 +
 .../gcc.target/arm/thumb2-replicated-constant4.c   |   22 +
 10 files changed, 476 insertions(+), 198 deletions(-)
 create mode 100644 src/gcc-mainline/gcc/testsuite/gcc.target/arm/thumb2-replicated-constant1.c
 create mode 100644 src/gcc-mainline/gcc/testsuite/gcc.target/arm/thumb2-replicated-constant2.c
 create mode 100644 src/gcc-mainline/gcc/testsuite/gcc.target/arm/thumb2-replicated-constant3.c
 create mode 100644 src/gcc-mainline/gcc/testsuite/gcc.target/arm/thumb2-replicated-constant4.c

diff --git a/src/gcc-mainline/gcc/config/arm/arm-protos.h b/src/gcc-mainline/gcc/config/arm/arm-protos.h
index 53923bd..db405ac 100644
--- a/src/gcc-mainline/gcc/config/arm/arm-protos.h
+++ b/src/gcc-mainline/gcc/config/arm/arm-protos.h
@@ -46,6 +46,7 @@  extern bool arm_vector_mode_supported_p (enum machine_mode);
 extern bool arm_small_register_classes_for_mode_p (enum machine_mode);
 extern int arm_hard_regno_mode_ok (unsigned int, 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 arm_split_constant (RTX_CODE, enum machine_mode, rtx,
 			       HOST_WIDE_INT, rtx, rtx, int);
 extern RTX_CODE arm_canonicalize_comparison (RTX_CODE, rtx *, rtx *);
diff --git a/src/gcc-mainline/gcc/config/arm/arm.c b/src/gcc-mainline/gcc/config/arm/arm.c
index 88c43e3..cb0130e 100644
--- a/src/gcc-mainline/gcc/config/arm/arm.c
+++ b/src/gcc-mainline/gcc/config/arm/arm.c
@@ -81,7 +81,6 @@  inline static int thumb1_index_register_rtx_p (rtx, int);
 static bool arm_legitimate_address_p (enum machine_mode, rtx, bool);
 static int thumb_far_jump_used_p (void);
 static bool thumb_force_lr_save (void);
-static int const_ok_for_op (HOST_WIDE_INT, enum rtx_code);
 static rtx emit_sfm (int, int);
 static unsigned arm_size_return_regs (void);
 static bool arm_assemble_integer (rtx, unsigned int, int);
@@ -129,7 +128,12 @@  static void thumb1_output_function_prologue (FILE *, HOST_WIDE_INT);
 static int arm_comp_type_attributes (const_tree, const_tree);
 static void arm_set_default_type_attributes (tree);
 static int arm_adjust_cost (rtx, rtx, rtx, int);
-static int count_insns_for_constant (HOST_WIDE_INT, int);
+static int optimal_immediate_sequence (enum rtx_code code,
+				       unsigned HOST_WIDE_INT val,
+				       int return_sequence[]);
+static int optimal_immediate_sequence_1 (enum rtx_code code,
+					 unsigned HOST_WIDE_INT val,
+					 int return_sequence[], int i);
 static int arm_get_strip_length (int);
 static bool arm_function_ok_for_sibcall (tree, tree);
 static enum machine_mode arm_promote_function_mode (const_tree,
@@ -2426,7 +2430,7 @@  const_ok_for_arm (HOST_WIDE_INT i)
 }
 
 /* Return true if I is a valid constant for the operation CODE.  */
-static int
+int
 const_ok_for_op (HOST_WIDE_INT i, enum rtx_code code)
 {
   if (const_ok_for_arm (i))
@@ -2434,7 +2438,21 @@  const_ok_for_op (HOST_WIDE_INT i, enum rtx_code code)
 
   switch (code)
     {
+    case SET:
+      /* See if we can use movw.  */
+      if (arm_arch_thumb2 && (i & 0xffff0000) == 0)
+	  return 1;
+      else
+	  return 0;
+
     case PLUS:
+      /* See if we can use addw or subw.  */
+      if (TARGET_THUMB2
+	  && ((i & 0xfffff000) == 0
+	      || ((-i) & 0xfffff000) == 0))
+	return 1;
+      /* else fall through.  */
+
     case COMPARE:
     case EQ:
     case NE:
@@ -2550,68 +2568,42 @@  arm_split_constant (enum rtx_code code, enum machine_mode mode, rtx insn,
 			   1);
 }
 
-/* Return the number of instructions required to synthesize the given
-   constant, if we start emitting them from bit-position I.  */
+/* Return a sequence of integers, in RETURN_SEQUENCE that fit into
+   ARM/THUMB2 immediates, and add up to VAL.
+   RETURN_SEQUENCE must be an int[4].
+   Thr function return value gives the number of insns required.  */
 static int
-count_insns_for_constant (HOST_WIDE_INT remainder, int i)
-{
-  HOST_WIDE_INT temp1;
-  int step_size = TARGET_ARM ? 2 : 1;
-  int num_insns = 0;
-
-  gcc_assert (TARGET_ARM || i == 0);
-
-  do
-    {
-      int end;
-
-      if (i <= 0)
-	i += 32;
-      if (remainder & (((1 << step_size) - 1) << (i - step_size)))
-	{
-	  end = i - 8;
-	  if (end < 0)
-	    end += 32;
-	  temp1 = remainder & ((0x0ff << end)
-				    | ((i < end) ? (0xff >> (32 - end)) : 0));
-	  remainder &= ~temp1;
-	  num_insns++;
-	  i -= 8 - step_size;
-	}
-      i -= step_size;
-    } while (remainder);
-  return num_insns;
-}
-
-static int
-find_best_start (unsigned HOST_WIDE_INT remainder)
+optimal_immediate_sequence (enum rtx_code code, unsigned HOST_WIDE_INT val,
+			    int return_sequence[])
 {
   int best_consecutive_zeros = 0;
   int i;
   int best_start = 0;
+  int insns1, insns2;
+  int tmp_sequence[4];
 
   /* If we aren't targetting ARM, the best place to start is always at
-     the bottom.  */
-  if (! TARGET_ARM)
-    return 0;
-
-  for (i = 0; i < 32; i += 2)
+     the bottom, otherwise look more closely.  */
+  if (TARGET_ARM)
     {
-      int consecutive_zeros = 0;
-
-      if (!(remainder & (3 << i)))
+      for (i = 0; i < 32; i += 2)
 	{
-	  while ((i < 32) && !(remainder & (3 << i)))
-	    {
-	      consecutive_zeros += 2;
-	      i += 2;
-	    }
-	  if (consecutive_zeros > best_consecutive_zeros)
+	  int consecutive_zeros = 0;
+
+	  if (!(val & (3 << i)))
 	    {
-	      best_consecutive_zeros = consecutive_zeros;
-	      best_start = i - consecutive_zeros;
+	      while ((i < 32) && !(val & (3 << i)))
+		{
+		  consecutive_zeros += 2;
+		  i += 2;
+		}
+	      if (consecutive_zeros > best_consecutive_zeros)
+		{
+		  best_consecutive_zeros = consecutive_zeros;
+		  best_start = i - consecutive_zeros;
+		}
+	      i -= 2;
 	    }
-	  i -= 2;
 	}
     }
 
@@ -2638,13 +2630,161 @@  find_best_start (unsigned HOST_WIDE_INT remainder)
      the constant starting from `best_start', and also starting from
      zero (i.e. with bit 31 first to be output).  If `best_start' doesn't
      yield a shorter sequence, we may as well use zero.  */
+  insns1 = optimal_immediate_sequence_1 (code, val, return_sequence, best_start);
   if (best_start != 0
-      && ((((unsigned HOST_WIDE_INT) 1) << best_start) < remainder)
-      && (count_insns_for_constant (remainder, 0) <=
-	  count_insns_for_constant (remainder, best_start)))
-    best_start = 0;
+      && ((((unsigned HOST_WIDE_INT) 1) << best_start) < val))
+    {
+      insns2 = optimal_immediate_sequence_1 (code, val, tmp_sequence, 0);
+      if (insns2 <= insns1)
+	{
+	  memcpy (return_sequence, tmp_sequence, sizeof(tmp_sequence));
+	  insns1 = insns2;
+	}
+    }
+
+  return insns1;
+}
+
+/* As for optimal_immediate_sequence, but starting at bit-position I.  */
+static int
+optimal_immediate_sequence_1 (enum rtx_code code, unsigned HOST_WIDE_INT val,
+			     int return_sequence[], int i)
+{
+  int remainder = val & 0xffffffff;
+  int insns = 0;
+
+  /* Try and find a way of doing the job in either two or three
+     instructions.
+     
+     In ARM mode we can use 8-bit constants, rotated to any 2-bit aligned
+     location.  We start at position I.  This may be the MSB, or
+     optimial_immediate_sequence may have positioned it at the largest block 
+     of zeros that are aligned on a 2-bit boundary. We then fill up the temps,
+     wrapping around to the top of the word when we drop off the bottom.
+     In the worst case this code should produce no more than four insns.
+
+     In Thumb2 mode, we can use 32/16-bit replicated constants, and 8-bit
+     constants, shifted to any arbitrary location.  We should always start
+     at the MSB.  */
+  do
+    {
+      int end;
+      int b1, b2, b3, b4;
+      unsigned HOST_WIDE_INT result;
+      int loc;
+
+      gcc_assert (insns < 4);
+
+      if (i <= 0)
+	i += 32;
+
+      /* First, find the next normal 12/8-bit shifted/rotated immediate.  */
+      if (remainder & ((TARGET_ARM ? (3 << (i - 2)) : (1 << (i - 1)))))
+	{
+	  loc = i;
+	  if (i <= 12 && TARGET_THUMB2 && code == PLUS)
+	    /* We can use addw/subw for the last 12 bits.  */
+	    result = remainder;
+	  else
+	    {
+	      /* Use an 8-bit shifted/rotated immediate.  */
+	      end = i - 8;
+	      if (end < 0)
+		end += 32;
+	      result = remainder & ((0x0ff << end)
+				   | ((i < end) ? (0xff >> (32 - end))
+						: 0));
+	      i -= 8;
+	    }
+	}
+      else
+	{
+	  /* Arm allows rotates by a multiple of two. Thumb-2 allows
+	     arbitrary shifts.  */
+	  i -= TARGET_ARM ? 2 : 1;
+	  continue;
+	}
+
+      /* Next, see if we can do a better job with a thumb2 replicated
+	 constant.
+       
+         We do it this way around to catch the cases like 0x01F001E0 where
+	 two 8-bit immediates would work, but a replicated constant would
+	 make it worse.
+       
+         TODO: 16-bit constants that don't clear all the bits, but still win.
+         TODO: Arithmetic splitting for set/add/sub, rather than bitwise.  */
+      if (TARGET_THUMB2)
+	{
+	  b1 = (remainder & 0xff000000) >> 24;
+	  b2 = (remainder & 0x00ff0000) >> 16;
+	  b3 = (remainder & 0x0000ff00) >> 8;
+	  b4 = remainder & 0xff;
 
-  return best_start;
+	  if (loc > 24)
+	    {
+	      /* The 8-bit immediate already found clears b1 (and maybe b2),
+		 but must leave b3 and b4 alone.  */
+
+	      /* First try to find a 32-bit replicated constant that clears
+		 almost everything.  We can assume that we can't do it in one,
+		 or else we wouldn't be here.  */
+	      unsigned int tmp = b1 & b2 & b3 & b4;
+	      unsigned int tmp2 = tmp + (tmp << 8) + (tmp << 16)
+				  + (tmp << 24);
+	      unsigned int matching_bytes = (tmp == b1) + (tmp == b2)
+					    + (tmp == b3) + (tmp == b4);
+	      if (tmp
+		  && (matching_bytes >= 3
+		      || (matching_bytes == 2
+			  && const_ok_for_op (remainder & ~tmp2, code))))
+		{
+		  /* At least 3 of the bytes match, and the fourth has at 
+		     least as many bits set, or two of the bytes match
+		     and it will only require one more insn to finish.  */
+		  result = tmp2;
+		  i = tmp != b1 ? 32
+		      : tmp != b2 ? 24
+		      : tmp != b3 ? 16
+		      : 8;
+		}
+
+	      /* Second, try to find a 16-bit replicated constant that can
+		 leave three of the bytes clear.  If b2 or b4 is already
+		 zero, then we can.  If the 8-bit from above would not
+		 clear b2 anyway, then we still win.  */
+	      else if (b1 == b3 && (!b2 || !b4
+			       || (remainder & 0x00ff0000 & ~result)))
+		{
+		  result = remainder & 0xff00ff00;
+		  i = 24;
+		}
+	    }
+	  else if (loc > 16)
+	    {
+	      /* The 8-bit immediate already found clears b2 (and maybe b3)
+		 and we don't get here unless b1 is alredy clear, but it will
+		 leave b4 unchanged.  */
+
+	      /* If we can clear b2 and b4 at once, then we win, since the
+		 8-bits couldn't possibly reach that far.  */
+	      if (b2 == b4)
+		{
+		  result = remainder & 0x00ff00ff;
+		  i = 16;
+		}
+	    }
+	}
+
+      return_sequence[insns++] = result;
+      remainder &= ~result;
+
+      if (code == SET || code == MINUS)
+	code = PLUS;
+    }
+  while (remainder);
+
+  return insns;
 }
 
 /* Emit an instruction with the indicated PATTERN.  If COND is
@@ -2661,7 +2801,6 @@  emit_constant_insn (rtx cond, rtx pattern)
 
 /* As above, but extra parameter GENERATE which, if clear, suppresses
    RTL generation.  */
-/* ??? This needs more work for thumb2.  */
 
 static int
 arm_gen_constant (enum rtx_code code, enum machine_mode mode, rtx cond,
@@ -2673,15 +2812,14 @@  arm_gen_constant (enum rtx_code code, enum machine_mode mode, rtx cond,
   int final_invert = 0;
   int can_negate_initial = 0;
   int i;
-  int num_bits_set = 0;
   int set_sign_bit_copies = 0;
   int clear_sign_bit_copies = 0;
   int clear_zero_bit_copies = 0;
   int set_zero_bit_copies = 0;
-  int insns = 0;
+  int insns = 0, neg_insns, inv_insns;
   unsigned HOST_WIDE_INT temp1, temp2;
   unsigned HOST_WIDE_INT remainder = val & 0xffffffff;
-  int step_size = TARGET_ARM ? 2 : 1;
+  int immediates[4], neg_immediates[4], inv_immediates[4];
 
   /* Find out which operations are safe for a given CODE.  Also do a quick
      check for degenerate cases; these can occur when DImode operations
@@ -2793,9 +2931,7 @@  arm_gen_constant (enum rtx_code code, enum machine_mode mode, rtx cond,
     }
 
   /* If we can do it in one insn get out quickly.  */
-  if (const_ok_for_arm (val)
-      || (can_negate_initial && const_ok_for_arm (-val))
-      || (can_invert && const_ok_for_arm (~val)))
+  if (const_ok_for_op (val, code))
     {
       if (generate)
 	emit_constant_insn (cond,
@@ -2848,15 +2984,6 @@  arm_gen_constant (enum rtx_code code, enum machine_mode mode, rtx cond,
   switch (code)
     {
     case SET:
-      /* See if we can use movw.  */
-      if (arm_arch_thumb2 && (remainder & 0xffff0000) == 0)
-	{
-	  if (generate)
-	    emit_constant_insn (cond, gen_rtx_SET (VOIDmode, target,
-						   GEN_INT (val)));
-	  return 1;
-	}
-
       /* See if we can do this by sign_extending a constant that is known
 	 to be negative.  This is a good, way of doing it, since the shift
 	 may well merge into a subsequent insn.  */
@@ -3207,121 +3334,100 @@  arm_gen_constant (enum rtx_code code, enum machine_mode mode, rtx cond,
       break;
     }
 
-  for (i = 0; i < 32; i++)
-    if (remainder & (1 << i))
-      num_bits_set++;
+  /* Calculate what the instruction sequences would be if we generated it
+     normally, negated, or inverted.  */
+  if (code == AND)
+    /* AND cannot be split into multiple insns, so invert and use BIC.  */
+    insns = 99;
+  else
+    insns = optimal_immediate_sequence (code, remainder, immediates);
 
-  if ((code == AND)
-      || (code != IOR && can_invert && num_bits_set > 16))
-    remainder ^= 0xffffffff;
-  else if (code == PLUS && num_bits_set > 16)
-    remainder = (-remainder) & 0xffffffff;
+  if (can_negate)
+    neg_insns = optimal_immediate_sequence (code, (-remainder) & 0xffffffff,
+					    neg_immediates);
+  else
+    neg_insns = 99;
+
+  if (can_invert)
+    inv_insns = optimal_immediate_sequence (code, remainder ^ 0xffffffff,
+					    inv_immediates);
+  else
+    inv_insns = 99;
 
-  /* For XOR, if more than half the bits are set and there's a sequence
-     of more than 8 consecutive ones in the pattern then we can XOR by the
-     inverted constant and then invert the final result; this may save an
-     instruction and might also lead to the final mvn being merged with
-     some other operation.  */
-  else if (code == XOR && num_bits_set > 16
-	   && (count_insns_for_constant (remainder ^ 0xffffffff,
-					 find_best_start
-					 (remainder ^ 0xffffffff))
-	       < count_insns_for_constant (remainder,
-					   find_best_start (remainder))))
+  /* Is the negated immediate sequence more efficient?  */
+  if (neg_insns < insns && neg_insns <= inv_insns)
     {
-      remainder ^= 0xffffffff;
-      final_invert = 1;
+      insns = neg_insns;
+      memcpy (immediates, neg_immediates, sizeof (immediates));
     }
   else
+    can_negate = 0;
+
+  /* Is the inverted immediate sequence more efficient?
+     We must allow for an extra NOT instruction for XOR operations, although
+     there is some chance that the final 'mvn' will get optimized later.  */
+  if (inv_insns < insns && (code != XOR || (inv_insns + 1) < insns))
     {
-      can_invert = 0;
-      can_negate = 0;
-    }
+      insns = inv_insns;
+      memcpy (immediates, inv_immediates, sizeof (immediates));
 
-  /* Now try and find a way of doing the job in either two or three
-     instructions.
-     We start by looking for the largest block of zeros that are aligned on
-     a 2-bit boundary, we then fill up the temps, wrapping around to the
-     top of the word when we drop off the bottom.
-     In the worst case this code should produce no more than four insns.
-     Thumb-2 constants are shifted, not rotated, so the MSB is always the
-     best place to start.  */
+      if (code == XOR)
+	final_invert = 1;
+    }
+  else
+    can_invert = 0;
 
-  /* ??? Use thumb2 replicated constants when the high and low halfwords are
-     the same.  */
-  {
-    /* Now start emitting the insns.  */
-    i = find_best_start (remainder);
-    do
-      {
-	int end;
+  /* Now output the chosen sequence as instructions.  */
+  if (generate)
+    {
+      for (i = 0; i < insns; i++)
+	{
+	  rtx new_src, temp1_rtx;
 
-	if (i <= 0)
-	  i += 32;
-	if (remainder & (3 << (i - 2)))
-	  {
-	    end = i - 8;
-	    if (end < 0)
-	      end += 32;
-	    temp1 = remainder & ((0x0ff << end)
-				 | ((i < end) ? (0xff >> (32 - end)) : 0));
-	    remainder &= ~temp1;
-
-	    if (generate)
-	      {
-		rtx new_src, temp1_rtx;
+	  temp1 = immediates[i];
 
-		if (code == SET || code == MINUS)
-		  {
-		    new_src = (subtargets ? gen_reg_rtx (mode) : target);
-		    if (can_invert && code != MINUS)
-		      temp1 = ~temp1;
-		  }
-		else
-		  {
-		    if ((final_invert || remainder) && subtargets)
-		      new_src = gen_reg_rtx (mode);
-		    else
-		      new_src = target;
-		    if (can_invert)
-		      temp1 = ~temp1;
-		    else if (can_negate)
-		      temp1 = -temp1;
-		  }
+	  if (code == SET || code == MINUS)
+	    {
+	      new_src = (subtargets ? gen_reg_rtx (mode) : target);
+	      if (can_invert && code != MINUS)
+		temp1 = ~temp1;
+	    }
+	  else
+	    {
+	      if ((final_invert || i < (insns - 1)) && subtargets)
+		new_src = gen_reg_rtx (mode);
+	      else
+		new_src = target;
+	      if (can_invert)
+		temp1 = ~temp1;
+	      else if (can_negate)
+		temp1 = -temp1;
+	    }
 
-		temp1 = trunc_int_for_mode (temp1, mode);
-		temp1_rtx = GEN_INT (temp1);
+	  temp1 = trunc_int_for_mode (temp1, mode);
+	  temp1_rtx = GEN_INT (temp1);
 
-		if (code == SET)
-		  ;
-		else if (code == MINUS)
-		  temp1_rtx = gen_rtx_MINUS (mode, temp1_rtx, source);
-		else
-		  temp1_rtx = gen_rtx_fmt_ee (code, mode, source, temp1_rtx);
+	  if (code == SET)
+	    ;
+	  else if (code == MINUS)
+	    temp1_rtx = gen_rtx_MINUS (mode, temp1_rtx, source);
+	  else
+	    temp1_rtx = gen_rtx_fmt_ee (code, mode, source, temp1_rtx);
 
-		emit_constant_insn (cond,
-				    gen_rtx_SET (VOIDmode, new_src,
-						 temp1_rtx));
-		source = new_src;
-	      }
+	  emit_constant_insn (cond,
+			      gen_rtx_SET (VOIDmode, new_src,
+					   temp1_rtx));
+	  source = new_src;
 
-	    if (code == SET)
-	      {
-		can_invert = 0;
-		code = PLUS;
-	      }
-	    else if (code == MINUS)
+	  if (code == SET)
+	    {
+	      can_invert = 0;
 	      code = PLUS;
-
-	    insns++;
-	    i -= 8 - step_size;
-	  }
-	/* Arm allows rotates by a multiple of two. Thumb-2 allows arbitrary
-	   shifts.  */
-	i -= step_size;
-      }
-    while (remainder);
-  }
+	    }
+	  else if (code == MINUS)
+	    code = PLUS;
+	}
+    }
 
   if (final_invert)
     {
diff --git a/src/gcc-mainline/gcc/config/arm/arm.md b/src/gcc-mainline/gcc/config/arm/arm.md
index 889b86f..8bc9926 100644
--- a/src/gcc-mainline/gcc/config/arm/arm.md
+++ b/src/gcc-mainline/gcc/config/arm/arm.md
@@ -696,21 +696,24 @@ 
 ;;  (plus (reg rN) (reg sp)) into (reg rN).  In this case reload will
 ;; put the duplicated register first, and not try the commutative version.
 (define_insn_and_split "*arm_addsi3"
-  [(set (match_operand:SI          0 "s_register_operand" "=r, k,r,r, k,r")
-	(plus:SI (match_operand:SI 1 "s_register_operand" "%rk,k,r,rk,k,rk")
-		 (match_operand:SI 2 "reg_or_int_operand" "rI,rI,k,L, L,?n")))]
+  [(set (match_operand:SI          0 "s_register_operand" "=r, k,r,r, k, r, k,r, k, r")
+	(plus:SI (match_operand:SI 1 "s_register_operand" "%rk,k,r,rk,k, rk,k,rk,k, rk")
+		 (match_operand:SI 2 "reg_or_int_operand" "rI,rI,k,jb,jb,L, L,jB,jB,?n")))]
   "TARGET_32BIT"
   "@
    add%?\\t%0, %1, %2
    add%?\\t%0, %1, %2
    add%?\\t%0, %2, %1
+   addw%?\\t%0, %1, %2
+   addw%?\\t%0, %1, %2
    sub%?\\t%0, %1, #%n2
    sub%?\\t%0, %1, #%n2
+   subw%?\\t%0, %1, #%n2
+   subw%?\\t%0, %1, #%n2
    #"
   "TARGET_32BIT
    && GET_CODE (operands[2]) == CONST_INT
-   && !(const_ok_for_arm (INTVAL (operands[2]))
-        || const_ok_for_arm (-INTVAL (operands[2])))
+   && !const_ok_for_op (INTVAL (operands[2]), PLUS)
    && (reload_completed || !arm_eliminable_register (operands[1]))"
   [(clobber (const_int 0))]
   "
@@ -719,7 +722,7 @@ 
 		      operands[1], 0);
   DONE;
   "
-  [(set_attr "length" "4,4,4,4,4,16")
+  [(set_attr "length" "4,4,4,4,4,4,4,4,4,16")
    (set_attr "predicable" "yes")]
 )
 
@@ -1173,27 +1176,31 @@ 
 
 ; ??? Check Thumb-2 split length
 (define_insn_and_split "*arm_subsi3_insn"
-  [(set (match_operand:SI           0 "s_register_operand" "=r,r,rk,r,r")
-	(minus:SI (match_operand:SI 1 "reg_or_int_operand" "rI,r,k,?n,r")
-		  (match_operand:SI 2 "reg_or_int_operand" "r,rI,r, r,?n")))]
+  [(set (match_operand:SI           0 "s_register_operand" "=r,r,rk,r, k, r,r")
+	(minus:SI (match_operand:SI 1 "reg_or_int_operand" "rI,r,k, rk,k, ?n,r")
+		  (match_operand:SI 2 "reg_or_int_operand" "r,rI,r, jb,jb,r,?n")))]
   "TARGET_32BIT"
   "@
    rsb%?\\t%0, %2, %1
    sub%?\\t%0, %1, %2
    sub%?\\t%0, %1, %2
+   subw%?\\t%0, %1, %2
+   subw%?\\t%0, %1, %2
    #
    #"
   "&& ((GET_CODE (operands[1]) == CONST_INT
-       	&& !const_ok_for_arm (INTVAL (operands[1])))
+       	&& !(const_ok_for_arm (INTVAL (operands[1]))
+	     || satisfies_constraint_jb (operands[2])))
        || (GET_CODE (operands[2]) == CONST_INT
-	   && !const_ok_for_arm (INTVAL (operands[2]))))"
+	   && !(const_ok_for_arm (INTVAL (operands[2]))
+	        || satisfies_constraint_jb (operands[2]))))"
   [(clobber (const_int 0))]
   "
   arm_split_constant (MINUS, SImode, curr_insn,
                       INTVAL (operands[1]), operands[0], operands[2], 0);
   DONE;
   "
-  [(set_attr "length" "4,4,4,16,16")
+  [(set_attr "length" "4,4,4,4,4,16,16")
    (set_attr "predicable" "yes")]
 )
 
@@ -5128,8 +5135,8 @@ 
 )
 
 (define_insn "*arm_movsi_insn"
-  [(set (match_operand:SI 0 "nonimmediate_operand" "=rk,r,r,r,rk,m")
-	(match_operand:SI 1 "general_operand"      "rk, I,K,j,mi,rk"))]
+  [(set (match_operand:SI 0 "nonimmediate_operand" "=rk,r,r,r, rk,m")
+	(match_operand:SI 1 "general_operand"      "rk, I,K,ja,mi,rk"))]
   "TARGET_ARM && ! TARGET_IWMMXT
    && !(TARGET_HARD_FLOAT && TARGET_VFP)
    && (   register_operand (operands[0], SImode)
diff --git a/src/gcc-mainline/gcc/config/arm/constraints.md b/src/gcc-mainline/gcc/config/arm/constraints.md
index 41a0663..0c1dce2 100644
--- a/src/gcc-mainline/gcc/config/arm/constraints.md
+++ b/src/gcc-mainline/gcc/config/arm/constraints.md
@@ -25,13 +25,13 @@ 
 ;; In ARM state, 'l' is an alias for 'r'
 
 ;; The following normal constraints have been used:
-;; in ARM/Thumb-2 state: G, H, I, j, J, K, L, M
+;; in ARM/Thumb-2 state: G, H, I, J, K, L, M
 ;; in Thumb-1 state: I, J, K, L, M, N, O
 
 ;; The following multi-letter normal constraints have been used:
 ;; in ARM/Thumb-2 state: Da, Db, Dc, Dn, Dl, DL, Dv, Dy, Di, Dz
 ;; in Thumb-1 state: Pa, Pb, Pc, Pd
-;; in Thumb-2 state: Ps, Pt, Pu, Pv, Pw, Px
+;; in Thumb-2 state: Ps, Pt, Pu, Pv, Pw, Px, ja, jb, jB
 
 ;; The following memory constraints have been used:
 ;; in ARM/Thumb-2 state: Q, Ut, Uv, Uy, Un, Um, Us
@@ -67,13 +67,25 @@ 
 (define_register_constraint "h" "TARGET_THUMB ? HI_REGS : NO_REGS"
  "In Thumb state the core registers @code{r8}-@code{r15}.")
 
-(define_constraint "j"
+(define_constraint "ja"
  "A constant suitable for a MOVW instruction. (ARM/Thumb-2)"
  (and (match_test "TARGET_32BIT && arm_arch_thumb2")
       (ior (match_code "high")
 	   (and (match_code "const_int")
                 (match_test "(ival & 0xffff0000) == 0")))))
 
+(define_constraint "jb"
+ "A 12-bit constant suitable for an ADDW or SUBW instruction. (Thumb-2)"
+ (and (match_test "TARGET_THUMB2")
+      (and (match_code "const_int")
+	   (match_test "(ival & 0xfffff000) == 0"))))
+
+(define_constraint "jB"
+ "A constant that satisfies the jb constrant if negated."
+ (and (match_test "TARGET_THUMB2")
+      (and (match_code "const_int")
+	   (match_test "((-ival) & 0xfffff000) == 0"))))
+
 (define_register_constraint "k" "STACK_REG"
  "@internal The stack register.")
 
diff --git a/src/gcc-mainline/gcc/config/arm/thumb2.md b/src/gcc-mainline/gcc/config/arm/thumb2.md
index 9ccb4d8..e6014f7 100644
--- a/src/gcc-mainline/gcc/config/arm/thumb2.md
+++ b/src/gcc-mainline/gcc/config/arm/thumb2.md
@@ -165,8 +165,8 @@ 
 ;; regs.  The high register alternatives are not taken into account when
 ;; choosing register preferences in order to reflect their expense.
 (define_insn "*thumb2_movsi_insn"
-  [(set (match_operand:SI 0 "nonimmediate_operand" "=rk,r,r,r,l ,*hk,m,*m")
-	(match_operand:SI 1 "general_operand"	   "rk ,I,K,j,mi,*mi,l,*hk"))]
+  [(set (match_operand:SI 0 "nonimmediate_operand" "=rk,r,r,r, l ,*hk,m,*m")
+	(match_operand:SI 1 "general_operand"	   "rk ,I,K,ja,mi,*mi,l,*hk"))]
   "TARGET_THUMB2 && ! TARGET_IWMMXT
    && !(TARGET_HARD_FLOAT && TARGET_VFP)
    && (   register_operand (operands[0], SImode)
diff --git a/src/gcc-mainline/gcc/config/arm/vfp.md b/src/gcc-mainline/gcc/config/arm/vfp.md
index 4e7b801..b1b5850 100644
--- a/src/gcc-mainline/gcc/config/arm/vfp.md
+++ b/src/gcc-mainline/gcc/config/arm/vfp.md
@@ -50,8 +50,8 @@ 
 ;; ??? For now do not allow loading constants into vfp regs.  This causes
 ;; problems because small constants get converted into adds.
 (define_insn "*arm_movsi_vfp"
-  [(set (match_operand:SI 0 "nonimmediate_operand" "=rk,r,r,r,rk,m ,*t,r,*t,*t, *Uv")
-      (match_operand:SI 1 "general_operand"	   "rk, I,K,j,mi,rk,r,*t,*t,*Uvi,*t"))]
+  [(set (match_operand:SI 0 "nonimmediate_operand" "=rk,r,r,r, rk,m ,*t,r,*t,*t, *Uv")
+      (match_operand:SI 1 "general_operand"	   "rk, I,K,ja,mi,rk,r,*t,*t,*Uvi,*t"))]
   "TARGET_ARM && TARGET_VFP && TARGET_HARD_FLOAT
    && (   s_register_operand (operands[0], SImode)
        || s_register_operand (operands[1], SImode))"
@@ -90,8 +90,8 @@ 
 ;; See thumb2.md:thumb2_movsi_insn for an explanation of the split
 ;; high/low register alternatives for loads and stores here.
 (define_insn "*thumb2_movsi_vfp"
-  [(set (match_operand:SI 0 "nonimmediate_operand" "=rk,r,r,r, l,*hk,m, *m,*t, r,*t,*t,  *Uv")
-	(match_operand:SI 1 "general_operand"	   "rk, I,K,j,mi,*mi,l,*hk, r,*t,*t,*Uvi,*t"))]
+  [(set (match_operand:SI 0 "nonimmediate_operand" "=rk,r,r,r,  l,*hk,m, *m,*t, r,*t,*t,  *Uv")
+	(match_operand:SI 1 "general_operand"	   "rk, I,K,ja,mi,*mi,l,*hk, r,*t,*t,*Uvi,*t"))]
   "TARGET_THUMB2 && TARGET_VFP && TARGET_HARD_FLOAT
    && (   s_register_operand (operands[0], SImode)
        || s_register_operand (operands[1], SImode))"
diff --git a/src/gcc-mainline/gcc/testsuite/gcc.target/arm/thumb2-replicated-constant1.c b/src/gcc-mainline/gcc/testsuite/gcc.target/arm/thumb2-replicated-constant1.c
new file mode 100644
index 0000000..e10ea03
--- /dev/null
+++ b/src/gcc-mainline/gcc/testsuite/gcc.target/arm/thumb2-replicated-constant1.c
@@ -0,0 +1,27 @@ 
+/* Ensure simple replicated constant immediates work.  */
+/* { dg-options "-mthumb -O2" } */
+/* { dg-require-effective-target arm_thumb2_ok } */
+
+int
+foo1 (int a)
+{
+  return a + 0xfefefefe;
+}
+
+/* { dg-final { scan-assembler "add.*#-16843010" } } */
+
+int
+foo2 (int a)
+{
+  return a - 0xab00ab00;
+}
+
+/* { dg-final { scan-assembler "sub.*#-1426019584" } } */
+
+int
+foo3 (int a)
+{
+  return a & 0x00cd00cd;
+}
+
+/* { dg-final { scan-assembler "and.*#13435085" } } */
diff --git a/src/gcc-mainline/gcc/testsuite/gcc.target/arm/thumb2-replicated-constant2.c b/src/gcc-mainline/gcc/testsuite/gcc.target/arm/thumb2-replicated-constant2.c
new file mode 100644
index 0000000..3739adb
--- /dev/null
+++ b/src/gcc-mainline/gcc/testsuite/gcc.target/arm/thumb2-replicated-constant2.c
@@ -0,0 +1,75 @@ 
+/* Ensure split constants can use replicated patterns.  */
+/* { dg-options "-mthumb -O2" } */
+/* { dg-require-effective-target arm_thumb2_ok } */
+
+int
+foo1 (int a)
+{
+  return a + 0xfe00fe01;
+}
+
+/* { dg-final { scan-assembler "add.*#-33489408" } } */
+/* { dg-final { scan-assembler "add.*#1" } } */
+
+int
+foo2 (int a)
+{
+  return a + 0xdd01dd00;
+}
+
+/* { dg-final { scan-assembler "add.*#-587145984" } } */
+/* { dg-final { scan-assembler "add.*#65536" } } */
+
+int
+foo3 (int a)
+{
+  return a + 0x00443344;
+}
+
+/* { dg-final { scan-assembler "add.*#4456516" } } */
+/* { dg-final { scan-assembler "add.*#13056" } } */
+
+int
+foo4 (int a)
+{
+  return a + 0x77330033;
+}
+
+/* { dg-final { scan-assembler "add.*#1996488704" } } */
+/* { dg-final { scan-assembler "add.*#3342387" } } */
+
+int
+foo5 (int a)
+{
+  return a + 0x11221122;
+}
+
+/* { dg-final { scan-assembler "add.*#285217024" } } */
+/* { dg-final { scan-assembler "add.*#2228258" } } */
+
+int
+foo6 (int a)
+{
+  return a + 0x66666677;
+}
+
+/* { dg-final { scan-assembler "add.*#1717986918" } } */
+/* { dg-final { scan-assembler "add.*#17" } } */
+
+int
+foo7 (int a)
+{
+  return a + 0x99888888;
+}
+
+/* { dg-final { scan-assembler "add.*#-2004318072" } } */
+/* { dg-final { scan-assembler "add.*#285212672" } } */
+
+int
+foo8 (int a)
+{
+  return a + 0xdddddfff;
+}
+
+/* { dg-final { scan-assembler "add.*#-572662307" } } */
+/* { dg-final { scan-assembler "addw.*#546" } } */
diff --git a/src/gcc-mainline/gcc/testsuite/gcc.target/arm/thumb2-replicated-constant3.c b/src/gcc-mainline/gcc/testsuite/gcc.target/arm/thumb2-replicated-constant3.c
new file mode 100644
index 0000000..eb6ad44
--- /dev/null
+++ b/src/gcc-mainline/gcc/testsuite/gcc.target/arm/thumb2-replicated-constant3.c
@@ -0,0 +1,28 @@ 
+/* Ensure negated/inverted replicated constant immediates work.  */
+/* { dg-options "-mthumb -O2" } */
+/* { dg-require-effective-target arm_thumb2_ok } */
+
+int
+foo1 (int a)
+{
+  return a | 0xffffff00;
+}
+
+/* { dg-final { scan-assembler "orn.*#255" } } */
+
+int
+foo2 (int a)
+{
+  return a & 0xffeeffee;
+}
+
+/* { dg-final { scan-assembler "bic.*#1114129" } } */
+
+int
+foo3 (int a)
+{
+  return a & 0xaaaaaa00;
+}
+
+/* { dg-final { scan-assembler "and.*#-1431655766" } } */
+/* { dg-final { scan-assembler "bic.*#170" } } */
diff --git a/src/gcc-mainline/gcc/testsuite/gcc.target/arm/thumb2-replicated-constant4.c b/src/gcc-mainline/gcc/testsuite/gcc.target/arm/thumb2-replicated-constant4.c
new file mode 100644
index 0000000..24efdcf
--- /dev/null
+++ b/src/gcc-mainline/gcc/testsuite/gcc.target/arm/thumb2-replicated-constant4.c
@@ -0,0 +1,22 @@ 
+/* Ensure replicated constants don't make things worse.  */
+/* { dg-options "-mthumb -O2" } */
+/* { dg-require-effective-target arm_thumb2_ok } */
+
+int
+foo1 (int a)
+{
+  /* It might be tempting to use 0x01000100, but it wouldn't help. */
+  return a + 0x01f001e0;
+}
+
+/* { dg-final { scan-assembler "add.*#32505856" } } */
+/* { dg-final { scan-assembler "add.*#480" } } */
+
+int
+foo2 (int a)
+{
+  return a + 0x0f100e10;
+}
+
+/* { dg-final { scan-assembler "add.*#252706816" } } */
+/* { dg-final { scan-assembler "add.*#3600" } } */