PATCH [2/n]: Prepare x32: Convert pointer to TLS symbol if needed

PATCH [2/n]: Prepare x32: Convert pointer to TLS symbol if needed

[H.J. Lu]

Hi,

Backend may promote pointers to Pmode.  Before we force a TLS symbol
to a pseudo, we may need to convert it to proper mode.  OK for trunk?

Thanks.


H.J.
----
2011-06-11  H.J. Lu  <hongjiu.lu@intel.com>

	* calls.c (precompute_register_parameters): Convert pointer to
	TLS symbol if needed.

diff --git a/gcc/calls.c b/gcc/calls.c
index feb98d2..de98267 100644
--- a/gcc/calls.c
+++ b/gcc/calls.c
@@ -706,7 +706,13 @@ precompute_register_parameters (int num_actuals, struct arg_data *args,
 	   pseudo now.  TLS symbols sometimes need a call to resolve.  */
 	if (CONSTANT_P (args[i].value)
 	    && !targetm.legitimate_constant_p (args[i].mode, args[i].value))
-	  args[i].value = force_reg (args[i].mode, args[i].value);
+	  {
+	    if (GET_MODE (args[i].value) != args[i].mode)
+	      args[i].value = convert_to_mode (args[i].mode,
+					       args[i].value,
+					       args[i].unsignedp);
+	    args[i].value = force_reg (args[i].mode, args[i].value);
+	  }
 
 	/* If we are to promote the function arg to a wider mode,
 	   do it now.  */

Re: PATCH [2/n]: Prepare x32: Convert pointer to TLS symbol if needed

[Richard Sandiford]

"H.J. Lu" <hongjiu.lu@intel.com> writes:
> @@ -706,7 +706,13 @@ precompute_register_parameters (int num_actuals, struct arg_data *args,
>  	   pseudo now.  TLS symbols sometimes need a call to resolve.  */
>  	if (CONSTANT_P (args[i].value)
>  	    && !targetm.legitimate_constant_p (args[i].mode, args[i].value))
> -	  args[i].value = force_reg (args[i].mode, args[i].value);
> +	  {
> +	    if (GET_MODE (args[i].value) != args[i].mode)
> +	      args[i].value = convert_to_mode (args[i].mode,
> +					       args[i].value,
> +					       args[i].unsignedp);
> +	    args[i].value = force_reg (args[i].mode, args[i].value);
> +	  }

But if GET_MODE (args[i].value) != args[i].mode, then the call to
targetm.legitimate_constant_p looks wrong.  The mode passed in the
first argument is supposed to the mode of the second argument.

Is there any reason why this and the following:

	/* If we are to promote the function arg to a wider mode,
	   do it now.  */

	if (args[i].mode != TYPE_MODE (TREE_TYPE (args[i].tree_value)))
	  args[i].value
	    = convert_modes (args[i].mode,
			     TYPE_MODE (TREE_TYPE (args[i].tree_value)),
			     args[i].value, args[i].unsignedp);

need to be done in the current order?  I can't think of any off-hand.
If not, would swapping them also fix the bug?

(I can't review this either way, of course.)

Richard

Re: PATCH [2/n]: Prepare x32: Convert pointer to TLS symbol if needed

[H.J. Lu]

On Wed, Jun 29, 2011 at 1:45 AM, Richard Sandiford
<richard.sandiford@linaro.org> wrote:
> "H.J. Lu" <hongjiu.lu@intel.com> writes:
>> @@ -706,7 +706,13 @@ precompute_register_parameters (int num_actuals, struct arg_data *args,
>>          pseudo now.  TLS symbols sometimes need a call to resolve.  */
>>       if (CONSTANT_P (args[i].value)
>>           && !targetm.legitimate_constant_p (args[i].mode, args[i].value))
>> -       args[i].value = force_reg (args[i].mode, args[i].value);
>> +       {
>> +         if (GET_MODE (args[i].value) != args[i].mode)
>> +           args[i].value = convert_to_mode (args[i].mode,
>> +                                            args[i].value,
>> +                                            args[i].unsignedp);
>> +         args[i].value = force_reg (args[i].mode, args[i].value);
>> +       }
>
> But if GET_MODE (args[i].value) != args[i].mode, then the call to
> targetm.legitimate_constant_p looks wrong.  The mode passed in the
> first argument is supposed to the mode of the second argument.
>
> Is there any reason why this and the following:
>
>        /* If we are to promote the function arg to a wider mode,
>           do it now.  */
>
>        if (args[i].mode != TYPE_MODE (TREE_TYPE (args[i].tree_value)))
>          args[i].value
>            = convert_modes (args[i].mode,
>                             TYPE_MODE (TREE_TYPE (args[i].tree_value)),
>                             args[i].value, args[i].unsignedp);
>
> need to be done in the current order?  I can't think of any off-hand.
> If not, would swapping them also fix the bug?
>
> (I can't review this either way, of course.)

It works on the testcase.  I will do a full test.

Thanks.

-- 
H.J.

Re: PATCH [2/n]: Prepare x32: Convert pointer to TLS symbol if needed

[H.J. Lu]

[-- Attachment #1: Type: text/plain, Size: 2136 bytes --]

On Wed, Jun 29, 2011 at 7:06 AM, H.J. Lu <hjl.tools@gmail.com> wrote:
> On Wed, Jun 29, 2011 at 1:45 AM, Richard Sandiford
> <richard.sandiford@linaro.org> wrote:
>> "H.J. Lu" <hongjiu.lu@intel.com> writes:
>>> @@ -706,7 +706,13 @@ precompute_register_parameters (int num_actuals, struct arg_data *args,
>>>          pseudo now.  TLS symbols sometimes need a call to resolve.  */
>>>       if (CONSTANT_P (args[i].value)
>>>           && !targetm.legitimate_constant_p (args[i].mode, args[i].value))
>>> -       args[i].value = force_reg (args[i].mode, args[i].value);
>>> +       {
>>> +         if (GET_MODE (args[i].value) != args[i].mode)
>>> +           args[i].value = convert_to_mode (args[i].mode,
>>> +                                            args[i].value,
>>> +                                            args[i].unsignedp);
>>> +         args[i].value = force_reg (args[i].mode, args[i].value);
>>> +       }
>>
>> But if GET_MODE (args[i].value) != args[i].mode, then the call to
>> targetm.legitimate_constant_p looks wrong.  The mode passed in the
>> first argument is supposed to the mode of the second argument.
>>
>> Is there any reason why this and the following:
>>
>>        /* If we are to promote the function arg to a wider mode,
>>           do it now.  */
>>
>>        if (args[i].mode != TYPE_MODE (TREE_TYPE (args[i].tree_value)))
>>          args[i].value
>>            = convert_modes (args[i].mode,
>>                             TYPE_MODE (TREE_TYPE (args[i].tree_value)),
>>                             args[i].value, args[i].unsignedp);
>>
>> need to be done in the current order?  I can't think of any off-hand.
>> If not, would swapping them also fix the bug?
>>
>> (I can't review this either way, of course.)
>
> It works on the testcase.  I will do a full test.
>

It works.  There are no regressions on Linux/x86-64.
OK for trunk?

Thanks.

-- 
H.J.
----

2011-06-29  H.J. Lu  <hongjiu.lu@intel.com>

	PR middle-end/47715
	* calls.c (precompute_register_parameters): Promote the function
	argument before checking non-legitimate constant.

[-- Attachment #2: gcc-x32-pr47715-1.patch --]
[-- Type: text/plain, Size: 1452 bytes --]

2011-06-29  H.J. Lu  <hongjiu.lu@intel.com>

	PR middle-end/47715
	* calls.c (precompute_register_parameters): Promote the function
	argument before checking non-legitimate constant.

diff --git a/gcc/calls.c b/gcc/calls.c
index bba477c..7538e4e 100644
--- a/gcc/calls.c
+++ b/gcc/calls.c
@@ -702,12 +702,6 @@ precompute_register_parameters (int num_actuals, struct arg_data *args,
 	    pop_temp_slots ();
 	  }
 
-	/* If the value is a non-legitimate constant, force it into a
-	   pseudo now.  TLS symbols sometimes need a call to resolve.  */
-	if (CONSTANT_P (args[i].value)
-	    && !targetm.legitimate_constant_p (args[i].mode, args[i].value))
-	  args[i].value = force_reg (args[i].mode, args[i].value);
-
 	/* If we are to promote the function arg to a wider mode,
 	   do it now.  */
 
@@ -717,6 +711,12 @@ precompute_register_parameters (int num_actuals, struct arg_data *args,
 			     TYPE_MODE (TREE_TYPE (args[i].tree_value)),
 			     args[i].value, args[i].unsignedp);
 
+	/* If the value is a non-legitimate constant, force it into a
+	   pseudo now.  TLS symbols sometimes need a call to resolve.  */
+	if (CONSTANT_P (args[i].value)
+	    && !targetm.legitimate_constant_p (args[i].mode, args[i].value))
+	  args[i].value = force_reg (args[i].mode, args[i].value);
+
 	/* If we're going to have to load the value by parts, pull the
 	   parts into pseudos.  The part extraction process can involve
 	   non-trivial computation.  */

Re: PATCH [2/n]: Prepare x32: Convert pointer to TLS symbol if needed

[Richard Sandiford]

"H.J. Lu" <hjl.tools@gmail.com> writes:
> On Wed, Jun 29, 2011 at 7:06 AM, H.J. Lu <hjl.tools@gmail.com> wrote:
>> On Wed, Jun 29, 2011 at 1:45 AM, Richard Sandiford
>> <richard.sandiford@linaro.org> wrote:
>>> "H.J. Lu" <hongjiu.lu@intel.com> writes:
>>>> @@ -706,7 +706,13 @@ precompute_register_parameters (int num_actuals, struct arg_data *args,
>>>>          pseudo now.  TLS symbols sometimes need a call to resolve.  */
>>>>       if (CONSTANT_P (args[i].value)
>>>>           && !targetm.legitimate_constant_p (args[i].mode, args[i].value))
>>>> -       args[i].value = force_reg (args[i].mode, args[i].value);
>>>> +       {
>>>> +         if (GET_MODE (args[i].value) != args[i].mode)
>>>> +           args[i].value = convert_to_mode (args[i].mode,
>>>> +                                            args[i].value,
>>>> +                                            args[i].unsignedp);
>>>> +         args[i].value = force_reg (args[i].mode, args[i].value);
>>>> +       }
>>>
>>> But if GET_MODE (args[i].value) != args[i].mode, then the call to
>>> targetm.legitimate_constant_p looks wrong.  The mode passed in the
>>> first argument is supposed to the mode of the second argument.
>>>
>>> Is there any reason why this and the following:
>>>
>>>        /* If we are to promote the function arg to a wider mode,
>>>           do it now.  */
>>>
>>>        if (args[i].mode != TYPE_MODE (TREE_TYPE (args[i].tree_value)))
>>>          args[i].value
>>>            = convert_modes (args[i].mode,
>>>                             TYPE_MODE (TREE_TYPE (args[i].tree_value)),
>>>                             args[i].value, args[i].unsignedp);
>>>
>>> need to be done in the current order?  I can't think of any off-hand.
>>> If not, would swapping them also fix the bug?
>>>
>>> (I can't review this either way, of course.)
>>
>> It works on the testcase.  I will do a full test.
>>
>
> It works.  There are no regressions on Linux/x86-64.

Great!  I can't approve it, but FWIW, it looks good to me.  The new order
seems to make more conceptual sense: coerce the value into the right mode,
then coerce it into the right type of rtx.

Richard