Patchwork [ARM] Turning off 64bits ops in Neon and gfortran/modulo-scheduling problem

login
register
mail settings
Submitter Christophe Lyon
Date Dec. 19, 2012, 3:59 p.m.
Message ID <CAKdteObaABsFAveuh975BXwwYiH=nwpmjwaCUrpG2BZ1sqPfjw@mail.gmail.com>
Download mbox | patch
Permalink /patch/207434/
State New
Headers show

Comments

Christophe Lyon - Dec. 19, 2012, 3:59 p.m.
On 17 December 2012 16:12, Richard Earnshaw <rearnsha@arm.com> wrote:
> On 29/11/12 17:16, Christophe Lyon wrote:
>> On trunk I have noticed a regression in gfortran when using modulo
>> scheduling: sms-1.f90 now fails, but I suspect it's not because of
>> this patch since forcing compilation for armv5t makes the same test
>> fail with and without my patch.
>>
>
> Hmm, that's worrying.  Could you please makesure this is recorded in
> bugzilla.  If this is a regression, please mark it as such.
>
I was about to do so, but after bisecting it turns out that the
problem was introduced by
http://gcc.gnu.org/viewcvs?root=gcc&view=rev&rev=192969 and is very
likely to be another instance of PR55562, which has just been fixed
by http://gcc.gnu.org/ml/gcc-patches/2012-12/msg01137.html.


>
> Now that this optimization is disabled by default, the onlya8 code is
> completely redundant and should be purged, along with the insn alternatives
> that used it.
>
> R.
>
Here is a new version of my patch, with the cleanup you requested.

2012-12-18  Christophe Lyon  <christophe.lyon@linaro.org>

        gcc/
        * config/arm/arm-protos.h (tune_params): Add
        prefer_neon_for_64bits field.
        * config/arm/arm.c (prefer_neon_for_64bits): New variable.
        (arm_slowmul_tune): Default prefer_neon_for_64bits to false.
        (arm_fastmul_tune, arm_strongarm_tune, arm_xscale_tune): Ditto.
        (arm_9e_tune, arm_v6t2_tune, arm_cortex_tune): Ditto.
        (arm_cortex_a5_tune, arm_cortex_a15_tune): Ditto.
        (arm_cortex_a9_tune, arm_fa726te_tune): Ditto.
        (arm_option_override): Handle -mneon-for-64bits new option.
        * config/arm/arm.h (TARGET_PREFER_NEON_64BITS): New macro.
        (prefer_neon_for_64bits): Declare new variable.
        * config/arm/arm.md (arch): Rename neon_onlya8 and neon_nota8 to
        avoid_neon_for_64bits and neon_for_64bits. Remove onlya8 and
        nota8.
        (arch_enabled): Handle new arch types. Remove support for onlya8
        and nota8.
        (one_cmpldi2): Use new arch names.
        * config/arm/arm.opt (mneon-for-64bits): Add option.
        * config/arm/neon.md (adddi3_neon, subdi3_neon, iordi3_neon)
        (anddi3_neon, xordi3_neon, ashldi3_neon, <shift>di3_neon): Use
        neon_for_64bits instead of nota8 and avoid_neon_for_64bits instead
        of onlya8.
        * doc/invoke.texi (-mneon-for-64bits): Document.

        gcc/testsuite/
        * gcc.target/arm/neon-for-64bits-1.c: New tests.
        * gcc.target/arm/neon-for-64bits-2.c: Likewise.
Christophe Lyon - Jan. 8, 2013, 4:24 p.m.
Ping?
http://gcc.gnu.org/ml/gcc-patches/2012-12/msg01197.html

Thanks,

Christophe

On 19 December 2012 16:59, Christophe Lyon <christophe.lyon@linaro.org> wrote:
> On 17 December 2012 16:12, Richard Earnshaw <rearnsha@arm.com> wrote:
>> On 29/11/12 17:16, Christophe Lyon wrote:
>>> On trunk I have noticed a regression in gfortran when using modulo
>>> scheduling: sms-1.f90 now fails, but I suspect it's not because of
>>> this patch since forcing compilation for armv5t makes the same test
>>> fail with and without my patch.
>>>
>>
>> Hmm, that's worrying.  Could you please makesure this is recorded in
>> bugzilla.  If this is a regression, please mark it as such.
>>
> I was about to do so, but after bisecting it turns out that the
> problem was introduced by
> http://gcc.gnu.org/viewcvs?root=gcc&view=rev&rev=192969 and is very
> likely to be another instance of PR55562, which has just been fixed
> by http://gcc.gnu.org/ml/gcc-patches/2012-12/msg01137.html.
>
>
>>
>> Now that this optimization is disabled by default, the onlya8 code is
>> completely redundant and should be purged, along with the insn alternatives
>> that used it.
>>
>> R.
>>
> Here is a new version of my patch, with the cleanup you requested.
>
> 2012-12-18  Christophe Lyon  <christophe.lyon@linaro.org>
>
>         gcc/
>         * config/arm/arm-protos.h (tune_params): Add
>         prefer_neon_for_64bits field.
>         * config/arm/arm.c (prefer_neon_for_64bits): New variable.
>         (arm_slowmul_tune): Default prefer_neon_for_64bits to false.
>         (arm_fastmul_tune, arm_strongarm_tune, arm_xscale_tune): Ditto.
>         (arm_9e_tune, arm_v6t2_tune, arm_cortex_tune): Ditto.
>         (arm_cortex_a5_tune, arm_cortex_a15_tune): Ditto.
>         (arm_cortex_a9_tune, arm_fa726te_tune): Ditto.
>         (arm_option_override): Handle -mneon-for-64bits new option.
>         * config/arm/arm.h (TARGET_PREFER_NEON_64BITS): New macro.
>         (prefer_neon_for_64bits): Declare new variable.
>         * config/arm/arm.md (arch): Rename neon_onlya8 and neon_nota8 to
>         avoid_neon_for_64bits and neon_for_64bits. Remove onlya8 and
>         nota8.
>         (arch_enabled): Handle new arch types. Remove support for onlya8
>         and nota8.
>         (one_cmpldi2): Use new arch names.
>         * config/arm/arm.opt (mneon-for-64bits): Add option.
>         * config/arm/neon.md (adddi3_neon, subdi3_neon, iordi3_neon)
>         (anddi3_neon, xordi3_neon, ashldi3_neon, <shift>di3_neon): Use
>         neon_for_64bits instead of nota8 and avoid_neon_for_64bits instead
>         of onlya8.
>         * doc/invoke.texi (-mneon-for-64bits): Document.
>
>         gcc/testsuite/
>         * gcc.target/arm/neon-for-64bits-1.c: New tests.
>         * gcc.target/arm/neon-for-64bits-2.c: Likewise.
Christophe Lyon - Jan. 16, 2013, 1:46 p.m.
Ping^2 ?


On 8 January 2013 17:24, Christophe Lyon <christophe.lyon@linaro.org> wrote:
> Ping?
> http://gcc.gnu.org/ml/gcc-patches/2012-12/msg01197.html
>
> Thanks,
>
> Christophe
>
> On 19 December 2012 16:59, Christophe Lyon <christophe.lyon@linaro.org> wrote:
>> On 17 December 2012 16:12, Richard Earnshaw <rearnsha@arm.com> wrote:
>>> On 29/11/12 17:16, Christophe Lyon wrote:
>>>> On trunk I have noticed a regression in gfortran when using modulo
>>>> scheduling: sms-1.f90 now fails, but I suspect it's not because of
>>>> this patch since forcing compilation for armv5t makes the same test
>>>> fail with and without my patch.
>>>>
>>>
>>> Hmm, that's worrying.  Could you please makesure this is recorded in
>>> bugzilla.  If this is a regression, please mark it as such.
>>>
>> I was about to do so, but after bisecting it turns out that the
>> problem was introduced by
>> http://gcc.gnu.org/viewcvs?root=gcc&view=rev&rev=192969 and is very
>> likely to be another instance of PR55562, which has just been fixed
>> by http://gcc.gnu.org/ml/gcc-patches/2012-12/msg01137.html.
>>
>>
>>>
>>> Now that this optimization is disabled by default, the onlya8 code is
>>> completely redundant and should be purged, along with the insn alternatives
>>> that used it.
>>>
>>> R.
>>>
>> Here is a new version of my patch, with the cleanup you requested.
>>
>> 2012-12-18  Christophe Lyon  <christophe.lyon@linaro.org>
>>
>>         gcc/
>>         * config/arm/arm-protos.h (tune_params): Add
>>         prefer_neon_for_64bits field.
>>         * config/arm/arm.c (prefer_neon_for_64bits): New variable.
>>         (arm_slowmul_tune): Default prefer_neon_for_64bits to false.
>>         (arm_fastmul_tune, arm_strongarm_tune, arm_xscale_tune): Ditto.
>>         (arm_9e_tune, arm_v6t2_tune, arm_cortex_tune): Ditto.
>>         (arm_cortex_a5_tune, arm_cortex_a15_tune): Ditto.
>>         (arm_cortex_a9_tune, arm_fa726te_tune): Ditto.
>>         (arm_option_override): Handle -mneon-for-64bits new option.
>>         * config/arm/arm.h (TARGET_PREFER_NEON_64BITS): New macro.
>>         (prefer_neon_for_64bits): Declare new variable.
>>         * config/arm/arm.md (arch): Rename neon_onlya8 and neon_nota8 to
>>         avoid_neon_for_64bits and neon_for_64bits. Remove onlya8 and
>>         nota8.
>>         (arch_enabled): Handle new arch types. Remove support for onlya8
>>         and nota8.
>>         (one_cmpldi2): Use new arch names.
>>         * config/arm/arm.opt (mneon-for-64bits): Add option.
>>         * config/arm/neon.md (adddi3_neon, subdi3_neon, iordi3_neon)
>>         (anddi3_neon, xordi3_neon, ashldi3_neon, <shift>di3_neon): Use
>>         neon_for_64bits instead of nota8 and avoid_neon_for_64bits instead
>>         of onlya8.
>>         * doc/invoke.texi (-mneon-for-64bits): Document.
>>
>>         gcc/testsuite/
>>         * gcc.target/arm/neon-for-64bits-1.c: New tests.
>>         * gcc.target/arm/neon-for-64bits-2.c: Likewise.
Christophe Lyon - Jan. 21, 2013, 4:52 p.m.
Ping?
http://gcc.gnu.org/ml/gcc-patches/2012-12/msg01197.html

On 16 January 2013 14:46, Christophe Lyon <christophe.lyon@linaro.org> wrote:
> Ping^2 ?
>
>
> On 8 January 2013 17:24, Christophe Lyon <christophe.lyon@linaro.org> wrote:
>> Ping?
>> http://gcc.gnu.org/ml/gcc-patches/2012-12/msg01197.html
>>
>> Thanks,
>>
>> Christophe
>>
>> On 19 December 2012 16:59, Christophe Lyon <christophe.lyon@linaro.org> wrote:
>>> On 17 December 2012 16:12, Richard Earnshaw <rearnsha@arm.com> wrote:
>>>> On 29/11/12 17:16, Christophe Lyon wrote:
>>>>> On trunk I have noticed a regression in gfortran when using modulo
>>>>> scheduling: sms-1.f90 now fails, but I suspect it's not because of
>>>>> this patch since forcing compilation for armv5t makes the same test
>>>>> fail with and without my patch.
>>>>>
>>>>
>>>> Hmm, that's worrying.  Could you please makesure this is recorded in
>>>> bugzilla.  If this is a regression, please mark it as such.
>>>>
>>> I was about to do so, but after bisecting it turns out that the
>>> problem was introduced by
>>> http://gcc.gnu.org/viewcvs?root=gcc&view=rev&rev=192969 and is very
>>> likely to be another instance of PR55562, which has just been fixed
>>> by http://gcc.gnu.org/ml/gcc-patches/2012-12/msg01137.html.
>>>
>>>
>>>>
>>>> Now that this optimization is disabled by default, the onlya8 code is
>>>> completely redundant and should be purged, along with the insn alternatives
>>>> that used it.
>>>>
>>>> R.
>>>>
>>> Here is a new version of my patch, with the cleanup you requested.
>>>
>>> 2012-12-18  Christophe Lyon  <christophe.lyon@linaro.org>
>>>
>>>         gcc/
>>>         * config/arm/arm-protos.h (tune_params): Add
>>>         prefer_neon_for_64bits field.
>>>         * config/arm/arm.c (prefer_neon_for_64bits): New variable.
>>>         (arm_slowmul_tune): Default prefer_neon_for_64bits to false.
>>>         (arm_fastmul_tune, arm_strongarm_tune, arm_xscale_tune): Ditto.
>>>         (arm_9e_tune, arm_v6t2_tune, arm_cortex_tune): Ditto.
>>>         (arm_cortex_a5_tune, arm_cortex_a15_tune): Ditto.
>>>         (arm_cortex_a9_tune, arm_fa726te_tune): Ditto.
>>>         (arm_option_override): Handle -mneon-for-64bits new option.
>>>         * config/arm/arm.h (TARGET_PREFER_NEON_64BITS): New macro.
>>>         (prefer_neon_for_64bits): Declare new variable.
>>>         * config/arm/arm.md (arch): Rename neon_onlya8 and neon_nota8 to
>>>         avoid_neon_for_64bits and neon_for_64bits. Remove onlya8 and
>>>         nota8.
>>>         (arch_enabled): Handle new arch types. Remove support for onlya8
>>>         and nota8.
>>>         (one_cmpldi2): Use new arch names.
>>>         * config/arm/arm.opt (mneon-for-64bits): Add option.
>>>         * config/arm/neon.md (adddi3_neon, subdi3_neon, iordi3_neon)
>>>         (anddi3_neon, xordi3_neon, ashldi3_neon, <shift>di3_neon): Use
>>>         neon_for_64bits instead of nota8 and avoid_neon_for_64bits instead
>>>         of onlya8.
>>>         * doc/invoke.texi (-mneon-for-64bits): Document.
>>>
>>>         gcc/testsuite/
>>>         * gcc.target/arm/neon-for-64bits-1.c: New tests.
>>>         * gcc.target/arm/neon-for-64bits-2.c: Likewise.
Ramana Radhakrishnan - Feb. 1, 2013, 5:43 p.m.
>>
> Here is a new version of my patch, with the cleanup you requested.
>
> 2012-12-18  Christophe Lyon  <christophe.lyon@linaro.org>
>
>          gcc/
>          * config/arm/arm-protos.h (tune_params): Add
>          prefer_neon_for_64bits field.
>          * config/arm/arm.c (prefer_neon_for_64bits): New variable.
>          (arm_slowmul_tune): Default prefer_neon_for_64bits to false.
>          (arm_fastmul_tune, arm_strongarm_tune, arm_xscale_tune): Ditto.
>          (arm_9e_tune, arm_v6t2_tune, arm_cortex_tune): Ditto.
>          (arm_cortex_a5_tune, arm_cortex_a15_tune): Ditto.
>          (arm_cortex_a9_tune, arm_fa726te_tune): Ditto.
>          (arm_option_override): Handle -mneon-for-64bits new option.
>          * config/arm/arm.h (TARGET_PREFER_NEON_64BITS): New macro.
>          (prefer_neon_for_64bits): Declare new variable.
>          * config/arm/arm.md (arch): Rename neon_onlya8 and neon_nota8 to
>          avoid_neon_for_64bits and neon_for_64bits. Remove onlya8 and
>          nota8.
>          (arch_enabled): Handle new arch types. Remove support for onlya8
>          and nota8.
>          (one_cmpldi2): Use new arch names.
>          * config/arm/arm.opt (mneon-for-64bits): Add option.
>          * config/arm/neon.md (adddi3_neon, subdi3_neon, iordi3_neon)
>          (anddi3_neon, xordi3_neon, ashldi3_neon, <shift>di3_neon): Use
>          neon_for_64bits instead of nota8 and avoid_neon_for_64bits instead
>          of onlya8.
>          * doc/invoke.texi (-mneon-for-64bits): Document.
>
>          gcc/testsuite/
>          * gcc.target/arm/neon-for-64bits-1.c: New tests.
>          * gcc.target/arm/neon-for-64bits-2.c: Likewise.
>



Ok for 4.9 stage1 now.

regards
Ramana

Patch

diff --git a/gcc/config/arm/arm-protos.h b/gcc/config/arm/arm-protos.h
index d942c5b..c92f055 100644
--- a/gcc/config/arm/arm-protos.h
+++ b/gcc/config/arm/arm-protos.h
@@ -247,6 +247,8 @@  struct tune_params
      performance. The first element covers Thumb state and the second one
      is for ARM state.  */
   bool logical_op_non_short_circuit[2];
+  /* Prefer Neon for 64-bit bitops.  */
+  bool prefer_neon_for_64bits;
 };
 
 extern const struct tune_params *current_tune;
diff --git a/gcc/config/arm/arm.c b/gcc/config/arm/arm.c
index 84ce56f..5e99436 100644
--- a/gcc/config/arm/arm.c
+++ b/gcc/config/arm/arm.c
@@ -816,6 +816,10 @@  int arm_arch_thumb2;
 int arm_arch_arm_hwdiv;
 int arm_arch_thumb_hwdiv;
 
+/* Nonzero if we should use Neon to handle 64-bits operations rather
+   than core registers.  */
+int prefer_neon_for_64bits = 0;
+
 /* In case of a PRE_INC, POST_INC, PRE_DEC, POST_DEC memory reference,
    we must report the mode of the memory reference from
    TARGET_PRINT_OPERAND to TARGET_PRINT_OPERAND_ADDRESS.  */
@@ -895,6 +899,7 @@  const struct tune_params arm_slowmul_tune =
   arm_default_branch_cost,
   false,					/* Prefer LDRD/STRD.  */
   {true, true},					/* Prefer non short circuit.  */
+  false                                         /* Prefer Neon for 64-bits bitops.  */
 };
 
 const struct tune_params arm_fastmul_tune =
@@ -908,6 +913,7 @@  const struct tune_params arm_fastmul_tune =
   arm_default_branch_cost,
   false,					/* Prefer LDRD/STRD.  */
   {true, true},					/* Prefer non short circuit.  */
+  false                                         /* Prefer Neon for 64-bits bitops.  */
 };
 
 /* StrongARM has early execution of branches, so a sequence that is worth
@@ -924,6 +930,7 @@  const struct tune_params arm_strongarm_tune =
   arm_default_branch_cost,
   false,					/* Prefer LDRD/STRD.  */
   {true, true},					/* Prefer non short circuit.  */
+  false                                         /* Prefer Neon for 64-bits bitops.  */
 };
 
 const struct tune_params arm_xscale_tune =
@@ -937,6 +944,7 @@  const struct tune_params arm_xscale_tune =
   arm_default_branch_cost,
   false,					/* Prefer LDRD/STRD.  */
   {true, true},					/* Prefer non short circuit.  */
+  false                                         /* Prefer Neon for 64-bits bitops.  */
 };
 
 const struct tune_params arm_9e_tune =
@@ -950,6 +958,7 @@  const struct tune_params arm_9e_tune =
   arm_default_branch_cost,
   false,					/* Prefer LDRD/STRD.  */
   {true, true},					/* Prefer non short circuit.  */
+  false                                         /* Prefer Neon for 64-bits bitops.  */
 };
 
 const struct tune_params arm_v6t2_tune =
@@ -963,6 +972,7 @@  const struct tune_params arm_v6t2_tune =
   arm_default_branch_cost,
   false,					/* Prefer LDRD/STRD.  */
   {true, true},					/* Prefer non short circuit.  */
+  false                                         /* Prefer Neon for 64-bits bitops.  */
 };
 
 /* Generic Cortex tuning.  Use more specific tunings if appropriate.  */
@@ -977,6 +987,7 @@  const struct tune_params arm_cortex_tune =
   arm_default_branch_cost,
   false,					/* Prefer LDRD/STRD.  */
   {true, true},					/* Prefer non short circuit.  */
+  false                                         /* Prefer Neon for 64-bits bitops.  */
 };
 
 const struct tune_params arm_cortex_a15_tune =
@@ -990,6 +1001,7 @@  const struct tune_params arm_cortex_a15_tune =
   arm_default_branch_cost,
   true,						/* Prefer LDRD/STRD.  */
   {true, true},					/* Prefer non short circuit.  */
+  false                                         /* Prefer Neon for 64-bits bitops.  */
 };
 
 /* Branches can be dual-issued on Cortex-A5, so conditional execution is
@@ -1006,6 +1018,7 @@  const struct tune_params arm_cortex_a5_tune =
   arm_cortex_a5_branch_cost,
   false,					/* Prefer LDRD/STRD.  */
   {false, false},				/* Prefer non short circuit.  */
+  false                                         /* Prefer Neon for 64-bits bitops.  */
 };
 
 const struct tune_params arm_cortex_a9_tune =
@@ -1019,6 +1032,7 @@  const struct tune_params arm_cortex_a9_tune =
   arm_default_branch_cost,
   false,					/* Prefer LDRD/STRD.  */
   {true, true},					/* Prefer non short circuit.  */
+  false                                         /* Prefer Neon for 64-bits bitops.  */
 };
 
 /* The arm_v6m_tune is duplicated from arm_cortex_tune, rather than
@@ -1034,6 +1048,7 @@  const struct tune_params arm_v6m_tune =
   arm_default_branch_cost,
   false,					/* Prefer LDRD/STRD.  */
   {false, false},				/* Prefer non short circuit.  */
+  false                                         /* Prefer Neon for 64-bits bitops.  */
 };
 
 const struct tune_params arm_fa726te_tune =
@@ -1047,6 +1062,7 @@  const struct tune_params arm_fa726te_tune =
   arm_default_branch_cost,
   false,					/* Prefer LDRD/STRD.  */
   {true, true},					/* Prefer non short circuit.  */
+  false                                         /* Prefer Neon for 64-bits bitops.  */
 };
 
 
@@ -2077,6 +2093,12 @@  arm_option_override (void)
                            global_options.x_param_values,
                            global_options_set.x_param_values);
 
+  /* Use Neon to perform 64-bits operations rather than core
+     registers.  */
+  prefer_neon_for_64bits = current_tune->prefer_neon_for_64bits;
+  if (use_neon_for_64bits == 1)
+     prefer_neon_for_64bits = true;
+
   /* Use the alternative scheduling-pressure algorithm by default.  */
   maybe_set_param_value (PARAM_SCHED_PRESSURE_ALGORITHM, 2,
                          global_options.x_param_values,
diff --git a/gcc/config/arm/arm.h b/gcc/config/arm/arm.h
index d0f351d..bcaa890 100644
--- a/gcc/config/arm/arm.h
+++ b/gcc/config/arm/arm.h
@@ -356,6 +356,9 @@  extern void (*arm_lang_output_object_attributes_hook)(void);
 #define TARGET_IDIV		((TARGET_ARM && arm_arch_arm_hwdiv) \
 				 || (TARGET_THUMB2 && arm_arch_thumb_hwdiv))
 
+/* Should NEON be used for 64-bits bitops.  */
+#define TARGET_PREFER_NEON_64BITS (prefer_neon_for_64bits)
+
 /* True iff the full BPABI is being used.  If TARGET_BPABI is true,
    then TARGET_AAPCS_BASED must be true -- but the converse does not
    hold.  TARGET_BPABI implies the use of the BPABI runtime library,
@@ -541,6 +544,10 @@  extern int arm_arch_arm_hwdiv;
 /* Nonzero if chip supports integer division instruction in Thumb mode.  */
 extern int arm_arch_thumb_hwdiv;
 
+/* Nonzero if we should use Neon to handle 64-bits operations rather
+   than core registers.  */
+extern int prefer_neon_for_64bits;
+
 #ifndef TARGET_DEFAULT
 #define TARGET_DEFAULT  (MASK_APCS_FRAME)
 #endif
diff --git a/gcc/config/arm/arm.md b/gcc/config/arm/arm.md
index 649e901..e24c2b4 100644
--- a/gcc/config/arm/arm.md
+++ b/gcc/config/arm/arm.md
@@ -96,7 +96,7 @@ 
 ; for ARM or Thumb-2 with arm_arch6, and nov6 for ARM without
 ; arm_arch6.  This attribute is used to compute attribute "enabled",
 ; use type "any" to enable an alternative in all cases.
-(define_attr "arch" "any,a,t,32,t1,t2,v6,nov6,onlya8,neon_onlya8,nota8,neon_nota8,iwmmxt,iwmmxt2"
+(define_attr "arch" "any,a,t,32,t1,t2,v6,nov6,neon_for_64bits,avoid_neon_for_64bits,iwmmxt,iwmmxt2"
   (const_string "any"))
 
 (define_attr "arch_enabled" "no,yes"
@@ -131,22 +131,14 @@ 
 	      (match_test "TARGET_32BIT && !arm_arch6"))
 	 (const_string "yes")
 
-	 (and (eq_attr "arch" "onlya8")
-	      (eq_attr "tune" "cortexa8"))
+	 (and (eq_attr "arch" "avoid_neon_for_64bits")
+	      (match_test "TARGET_NEON")
+	      (not (match_test "TARGET_PREFER_NEON_64BITS")))
 	 (const_string "yes")
 
-	 (and (eq_attr "arch" "neon_onlya8")
-	      (eq_attr "tune" "cortexa8")
-	      (match_test "TARGET_NEON"))
-	 (const_string "yes")
-
-	 (and (eq_attr "arch" "nota8")
-	      (not (eq_attr "tune" "cortexa8")))
-	 (const_string "yes")
-
-	 (and (eq_attr "arch" "neon_nota8")
-	      (not (eq_attr "tune" "cortexa8"))
-	      (match_test "TARGET_NEON"))
+	 (and (eq_attr "arch" "neon_for_64bits")
+	      (match_test "TARGET_NEON")
+	      (match_test "TARGET_PREFER_NEON_64BITS"))
 	 (const_string "yes")
 
 	 (and (eq_attr "arch" "iwmmxt2")
@@ -4326,7 +4318,7 @@ 
   [(set_attr "length" "*,8,8,*")
    (set_attr "predicable" "no,yes,yes,no")
    (set_attr "neon_type" "neon_int_1,*,*,neon_int_1")
-   (set_attr "arch" "neon_nota8,*,*,neon_onlya8")]
+   (set_attr "arch" "neon_for_64bits,*,*,avoid_neon_for_64bits")]
 )
 
 (define_expand "one_cmplsi2"
diff --git a/gcc/config/arm/arm.opt b/gcc/config/arm/arm.opt
index 61d2d2f..c4ede22 100644
--- a/gcc/config/arm/arm.opt
+++ b/gcc/config/arm/arm.opt
@@ -247,3 +247,7 @@  that may trigger Cortex-M3 errata.
 munaligned-access
 Target Report Var(unaligned_access) Init(2)
 Enable unaligned word and halfword accesses to packed data.
+
+mneon-for-64bits
+Target Report RejectNegative Var(use_neon_for_64bits) Init(0)
+Use Neon to perform 64-bits operations rather than core registers.
diff --git a/gcc/config/arm/neon.md b/gcc/config/arm/neon.md
index fc38269..6e13d1a 100644
--- a/gcc/config/arm/neon.md
+++ b/gcc/config/arm/neon.md
@@ -488,7 +488,7 @@ 
   [(set_attr "neon_type" "neon_int_1,*,*,neon_int_1,*,*,*")
    (set_attr "conds" "*,clob,clob,*,clob,clob,clob")
    (set_attr "length" "*,8,8,*,8,8,8")
-   (set_attr "arch" "nota8,*,*,onlya8,*,*,*")]
+   (set_attr "arch" "neon_for_64bits,*,*,avoid_neon_for_64bits,*,*,*")]
 )
 
 (define_insn "*sub<mode>3_neon"
@@ -525,7 +525,7 @@ 
   [(set_attr "neon_type" "neon_int_2,*,*,*,neon_int_2")
    (set_attr "conds" "*,clob,clob,clob,*")
    (set_attr "length" "*,8,8,8,*")
-   (set_attr "arch" "nota8,*,*,*,onlya8")]
+   (set_attr "arch" "neon_for_64bits,*,*,*,avoid_neon_for_64bits")]
 )
 
 (define_insn "*mul<mode>3_neon"
@@ -700,7 +700,7 @@ 
 }
   [(set_attr "neon_type" "neon_int_1,neon_int_1,*,*,neon_int_1,neon_int_1")
    (set_attr "length" "*,*,8,8,*,*")
-   (set_attr "arch" "nota8,nota8,*,*,onlya8,onlya8")]
+   (set_attr "arch" "neon_for_64bits,neon_for_64bits,*,*,avoid_neon_for_64bits,avoid_neon_for_64bits")]
 )
 
 ;; The concrete forms of the Neon immediate-logic instructions are vbic and
@@ -745,7 +745,7 @@ 
 }
   [(set_attr "neon_type" "neon_int_1,neon_int_1,*,*,neon_int_1,neon_int_1")
    (set_attr "length" "*,*,8,8,*,*")
-   (set_attr "arch" "nota8,nota8,*,*,onlya8,onlya8")]
+   (set_attr "arch" "neon_for_64bits,neon_for_64bits,*,*,avoid_neon_for_64bits,avoid_neon_for_64bits")]
 )
 
 (define_insn "orn<mode>3_neon"
@@ -841,7 +841,7 @@ 
    veor\t%P0, %P1, %P2"
   [(set_attr "neon_type" "neon_int_1,*,*,neon_int_1")
    (set_attr "length" "*,8,8,*")
-   (set_attr "arch" "nota8,*,*,onlya8")]
+   (set_attr "arch" "neon_for_64bits,*,*,avoid_neon_for_64bits")]
 )
 
 (define_insn "one_cmpl<mode>2"
@@ -1163,7 +1163,7 @@ 
       }
     DONE;
   }"
-  [(set_attr "arch" "nota8,nota8,*,*,onlya8,onlya8")
+  [(set_attr "arch" "neon_for_64bits,neon_for_64bits,*,*,avoid_neon_for_64bits,avoid_neon_for_64bits")
    (set_attr "opt" "*,*,speed,speed,*,*")]
 )
 
@@ -1264,7 +1264,7 @@ 
 
     DONE;
   }"
-  [(set_attr "arch" "nota8,nota8,*,*,onlya8,onlya8")
+  [(set_attr "arch" "neon_for_64bits,neon_for_64bits,*,*,avoid_neon_for_64bits,avoid_neon_for_64bits")
    (set_attr "opt" "*,*,speed,speed,*,*")]
 )
 
diff --git a/gcc/doc/invoke.texi b/gcc/doc/invoke.texi
index 06ba770..01faecd 100644
--- a/gcc/doc/invoke.texi
+++ b/gcc/doc/invoke.texi
@@ -513,7 +513,8 @@  Objective-C and Objective-C++ Dialects}.
 -mtp=@var{name} -mtls-dialect=@var{dialect} @gol
 -mword-relocations @gol
 -mfix-cortex-m3-ldrd @gol
--munaligned-access}
+-munaligned-access @gol
+-mneon-for-64bits}
 
 @emph{AVR Options}
 @gccoptlist{-mmcu=@var{mcu} -maccumulate-args -mbranch-cost=@var{cost} @gol
@@ -11499,6 +11500,11 @@  setting of this option.  If unaligned access is enabled then the
 preprocessor symbol @code{__ARM_FEATURE_UNALIGNED} will also be
 defined.
 
+@item -mneon-for-64bits
+@opindex mneon-for-64bits
+Enables using Neon to handle scalar 64-bits operations. This is
+disabled by default since the cost of moving data from core registers
+to Neon is high.
 @end table
 
 @node AVR Options
diff --git a/gcc/testsuite/gcc.target/arm/neon-for-64bits-1.c b/gcc/testsuite/gcc.target/arm/neon-for-64bits-1.c
new file mode 100644
index 0000000..a2a4103
--- /dev/null
+++ b/gcc/testsuite/gcc.target/arm/neon-for-64bits-1.c
@@ -0,0 +1,54 @@ 
+/* Check that Neon is *not* used by default to handle 64-bits scalar
+   operations.  */
+
+/* { dg-do compile } */
+/* { dg-require-effective-target arm_neon_ok } */
+/* { dg-options "-O2" } */
+/* { dg-add-options arm_neon } */
+
+typedef long long i64;
+typedef unsigned long long u64;
+typedef unsigned int u32;
+typedef int i32;
+
+/* Unary operators */
+#define UNARY_OP(name, op) \
+  void unary_##name(u64 *a, u64 *b) { *a = op (*b + 0x1234567812345678ULL) ; }
+
+/* Binary operators */
+#define BINARY_OP(name, op) \
+  void binary_##name(u64 *a, u64 *b, u64 *c) { *a = *b op *c ; }
+
+/* Unsigned shift */
+#define SHIFT_U(name, op, amount) \
+  void ushift_##name(u64 *a, u64 *b, int c) { *a = *b op amount; }
+
+/* Signed shift */
+#define SHIFT_S(name, op, amount) \
+  void sshift_##name(i64 *a, i64 *b, int c) { *a = *b op amount; }
+
+UNARY_OP(not, ~)
+
+BINARY_OP(add, +)
+BINARY_OP(sub, -)
+BINARY_OP(and, &)
+BINARY_OP(or, |)
+BINARY_OP(xor, ^)
+
+SHIFT_U(right1, >>, 1)
+SHIFT_U(right2, >>, 2)
+SHIFT_U(right5, >>, 5)
+SHIFT_U(rightn, >>, c)
+
+SHIFT_S(right1, >>, 1)
+SHIFT_S(right2, >>, 2)
+SHIFT_S(right5, >>, 5)
+SHIFT_S(rightn, >>, c)
+
+/* { dg-final {scan-assembler-times "vmvn" 0} }  */
+/* { dg-final {scan-assembler-times "vadd" 0} }  */
+/* { dg-final {scan-assembler-times "vsub" 0} }  */
+/* { dg-final {scan-assembler-times "vand" 0} }  */
+/* { dg-final {scan-assembler-times "vorr" 0} }  */
+/* { dg-final {scan-assembler-times "veor" 0} }  */
+/* { dg-final {scan-assembler-times "vshr" 0} }  */
diff --git a/gcc/testsuite/gcc.target/arm/neon-for-64bits-2.c b/gcc/testsuite/gcc.target/arm/neon-for-64bits-2.c
new file mode 100644
index 0000000..035bfb7
--- /dev/null
+++ b/gcc/testsuite/gcc.target/arm/neon-for-64bits-2.c
@@ -0,0 +1,57 @@ 
+/* Check that Neon is used to handle 64-bits scalar operations.  */
+
+/* { dg-do compile } */
+/* { dg-require-effective-target arm_neon_ok } */
+/* { dg-options "-O2 -mneon-for-64bits" } */
+/* { dg-add-options arm_neon } */
+
+typedef long long i64;
+typedef unsigned long long u64;
+typedef unsigned int u32;
+typedef int i32;
+
+/* Unary operators */
+#define UNARY_OP(name, op) \
+  void unary_##name(u64 *a, u64 *b) { *a = op (*b + 0x1234567812345678ULL) ; }
+
+/* Binary operators */
+#define BINARY_OP(name, op) \
+  void binary_##name(u64 *a, u64 *b, u64 *c) { *a = *b op *c ; }
+
+/* Unsigned shift */
+#define SHIFT_U(name, op, amount) \
+  void ushift_##name(u64 *a, u64 *b, int c) { *a = *b op amount; }
+
+/* Signed shift */
+#define SHIFT_S(name, op, amount) \
+  void sshift_##name(i64 *a, i64 *b, int c) { *a = *b op amount; }
+
+UNARY_OP(not, ~)
+
+BINARY_OP(add, +)
+BINARY_OP(sub, -)
+BINARY_OP(and, &)
+BINARY_OP(or, |)
+BINARY_OP(xor, ^)
+
+SHIFT_U(right1, >>, 1)
+SHIFT_U(right2, >>, 2)
+SHIFT_U(right5, >>, 5)
+SHIFT_U(rightn, >>, c)
+
+SHIFT_S(right1, >>, 1)
+SHIFT_S(right2, >>, 2)
+SHIFT_S(right5, >>, 5)
+SHIFT_S(rightn, >>, c)
+
+/* { dg-final {scan-assembler-times "vmvn" 1} }  */
+/* Two vadd: 1 in unary_not, 1 in binary_add */
+/* { dg-final {scan-assembler-times "vadd" 2} }  */
+/* { dg-final {scan-assembler-times "vsub" 1} }  */
+/* { dg-final {scan-assembler-times "vand" 1} }  */
+/* { dg-final {scan-assembler-times "vorr" 1} }  */
+/* { dg-final {scan-assembler-times "veor" 1} }  */
+/* 6 vshr for right shifts by constant, and variable right shift uses
+   vshl with a negative amount in register.  */
+/* { dg-final {scan-assembler-times "vshr" 6} }  */
+/* { dg-final {scan-assembler-times "vshl" 2} }  */