[V1] HIGH part of symbol ref is invalid for constant pool

  Hi,

In patch https://gcc.gnu.org/pipermail/gcc-patches/2022-July/597712.html,
test case was not added.  After more check, a testcase is added for it.

The high part of the symbol address is invalid for the constant pool.
In function rs6000_cannot_force_const_mem, we already return true for
"HIGH with UNSPEC" rtx.  Below are some examples also indicate the high
part of a symbol_ref:
(high:DI (const:DI (plus:DI (symbol_ref:DI ("xx") (const_int 12 [0xc])))))
(high:DI (symbol_ref:DI ("var_1")..)))

This patch updates rs6000_cannot_force_const_mem to return true for
rtx with HIGH code.

Bootstrapped and regtested on ppc64le and ppc64.
Is it ok for trunk?

BR,
Jeff(Jiufu)


gcc/ChangeLog:

	* config/rs6000/rs6000.cc (rs6000_cannot_force_const_mem):
	Return true for HIGH code rtx.

gcc/testsuite/ChangeLog:

	* gcc.target/powerpc/constpoolcheck.c: New test.

---
 gcc/config/rs6000/rs6000.cc                       |  7 +++++--
 gcc/testsuite/gcc.target/powerpc/constpoolcheck.c | 11 +++++++++++
 2 files changed, 16 insertions(+), 2 deletions(-)
 create mode 100644 gcc/testsuite/gcc.target/powerpc/constpoolcheck.c
  

Hi Jeff,

on 2022/7/19 22:30, Jiufu Guo wrote:
> Hi,
> 
> In patch https://gcc.gnu.org/pipermail/gcc-patches/2022-July/597712.html,
> test case was not added.  After more check, a testcase is added for it.
> 

Good to see that you constructed one actual test case, nice!  :)

> The high part of the symbol address is invalid for the constant pool.
> In function rs6000_cannot_force_const_mem, we already return true for
> "HIGH with UNSPEC" rtx.  Below are some examples also indicate the high
> part of a symbol_ref:
> (high:DI (const:DI (plus:DI (symbol_ref:DI ("xx") (const_int 12 [0xc])))))
> (high:DI (symbol_ref:DI ("var_1")..)))
> 
> This patch updates rs6000_cannot_force_const_mem to return true for
> rtx with HIGH code.
> 
> Bootstrapped and regtested on ppc64le and ppc64.
> Is it ok for trunk?

I think this patch is OK with some nits below tweaked.

> 
> BR,
> Jeff(Jiufu)
> 
> 
> gcc/ChangeLog:
> 
> 	* config/rs6000/rs6000.cc (rs6000_cannot_force_const_mem):
> 	Return true for HIGH code rtx.
> 
> gcc/testsuite/ChangeLog:
> 
> 	* gcc.target/powerpc/constpoolcheck.c: New test.
> 
> ---
>  gcc/config/rs6000/rs6000.cc                       |  7 +++++--
>  gcc/testsuite/gcc.target/powerpc/constpoolcheck.c | 11 +++++++++++
>  2 files changed, 16 insertions(+), 2 deletions(-)
>  create mode 100644 gcc/testsuite/gcc.target/powerpc/constpoolcheck.c
> 
> diff --git a/gcc/config/rs6000/rs6000.cc b/gcc/config/rs6000/rs6000.cc
> index 0af2085adc0..d56832ebbfc 100644
> --- a/gcc/config/rs6000/rs6000.cc
> +++ b/gcc/config/rs6000/rs6000.cc
> @@ -9704,8 +9704,11 @@ rs6000_init_stack_protect_guard (void)
>  static bool
>  rs6000_cannot_force_const_mem (machine_mode mode ATTRIBUTE_UNUSED, rtx x)
>  {
> -  if (GET_CODE (x) == HIGH
> -      && GET_CODE (XEXP (x, 0)) == UNSPEC)
> +  /* High part of a symbol ref/address can not be put into constant pool. e.g.

Nit: two spaces after the period in "... pool.".

> +     (high:DI (symbol_ref:DI ("var")..)) or

Nit: You have one "or" at the end of the above line, I think it's better to
keep the below line consistent by either removing the above " or" or adding
one "or" at the end of the below line.

> +     (high:DI (unspec:DI [(symbol_ref/u:DI ("*.LC0")..)


> +     (high:DI (const:DI (plus:DI (symbol_ref:DI ("xx")) (const_int 12)))).  */
> +  if (GET_CODE (x) == HIGH)
>      return true;
> 
>    /* A TLS symbol in the TOC cannot contain a sum.  */
> diff --git a/gcc/testsuite/gcc.target/powerpc/constpoolcheck.c b/gcc/testsuite/gcc.target/powerpc/constpoolcheck.c
> new file mode 100644
> index 00000000000..ed7a994827b
> --- /dev/null
> +++ b/gcc/testsuite/gcc.target/powerpc/constpoolcheck.c

Maybe it's good to name it to "const-pool-check.c" or "not-force-const-mem.c".

> @@ -0,0 +1,11 @@
> +/* { dg-do compile { target powerpc*-*-* } } */

Nit: this "dg-do" line isn't needed since all here are default.

BR,
Kewen

> +/* { dg-options "-O1 -mdejagnu-cpu=power10" } */
> +/* (high:DI (symbol_ref:DI ("var_48")..))) should not cause ICE. */
> +extern short var_48;
> +void
> +foo (double *r)
> +{
> +  if (var_48)
> +    *r = 1234.5678;
> +}
> +
  

"Kewen.Lin" <linkw@linux.ibm.com> writes:

> Hi Jeff,
>
> on 2022/7/19 22:30, Jiufu Guo wrote:
>> Hi,
>> 
>> In patch https://gcc.gnu.org/pipermail/gcc-patches/2022-July/597712.html,
>> test case was not added.  After more check, a testcase is added for it.
>> 
>
> Good to see that you constructed one actual test case, nice!  :)
>
>> The high part of the symbol address is invalid for the constant pool.
>> In function rs6000_cannot_force_const_mem, we already return true for
>> "HIGH with UNSPEC" rtx.  Below are some examples also indicate the high
>> part of a symbol_ref:
>> (high:DI (const:DI (plus:DI (symbol_ref:DI ("xx") (const_int 12 [0xc])))))
>> (high:DI (symbol_ref:DI ("var_1")..)))
>> 
>> This patch updates rs6000_cannot_force_const_mem to return true for
>> rtx with HIGH code.
>> 
>> Bootstrapped and regtested on ppc64le and ppc64.
>> Is it ok for trunk?
>
> I think this patch is OK with some nits below tweaked.

Thanks so much for your time to review and helpful comments!
I will update accordingly before commit.

BR,
Jeff(Jiufu)

>
>> 
>> BR,
>> Jeff(Jiufu)
>> 
>> 
>> gcc/ChangeLog:
>> 
>> 	* config/rs6000/rs6000.cc (rs6000_cannot_force_const_mem):
>> 	Return true for HIGH code rtx.
>> 
>> gcc/testsuite/ChangeLog:
>> 
>> 	* gcc.target/powerpc/constpoolcheck.c: New test.
>> 
>> ---
>>  gcc/config/rs6000/rs6000.cc                       |  7 +++++--
>>  gcc/testsuite/gcc.target/powerpc/constpoolcheck.c | 11 +++++++++++
>>  2 files changed, 16 insertions(+), 2 deletions(-)
>>  create mode 100644 gcc/testsuite/gcc.target/powerpc/constpoolcheck.c
>> 
>> diff --git a/gcc/config/rs6000/rs6000.cc b/gcc/config/rs6000/rs6000.cc
>> index 0af2085adc0..d56832ebbfc 100644
>> --- a/gcc/config/rs6000/rs6000.cc
>> +++ b/gcc/config/rs6000/rs6000.cc
>> @@ -9704,8 +9704,11 @@ rs6000_init_stack_protect_guard (void)
>>  static bool
>>  rs6000_cannot_force_const_mem (machine_mode mode ATTRIBUTE_UNUSED, rtx x)
>>  {
>> -  if (GET_CODE (x) == HIGH
>> -      && GET_CODE (XEXP (x, 0)) == UNSPEC)
>> +  /* High part of a symbol ref/address can not be put into constant pool. e.g.
>
> Nit: two spaces after the period in "... pool.".
Thanks!
>
>> +     (high:DI (symbol_ref:DI ("var")..)) or
>
> Nit: You have one "or" at the end of the above line, I think it's better to
> keep the below line consistent by either removing the above " or" or adding
> one "or" at the end of the below line.
Thanks!
>
>> +     (high:DI (unspec:DI [(symbol_ref/u:DI ("*.LC0")..)
>
>
>> +     (high:DI (const:DI (plus:DI (symbol_ref:DI ("xx")) (const_int 12)))).  */
>> +  if (GET_CODE (x) == HIGH)
>>      return true;
>> 
>>    /* A TLS symbol in the TOC cannot contain a sum.  */
>> diff --git a/gcc/testsuite/gcc.target/powerpc/constpoolcheck.c b/gcc/testsuite/gcc.target/powerpc/constpoolcheck.c
>> new file mode 100644
>> index 00000000000..ed7a994827b
>> --- /dev/null
>> +++ b/gcc/testsuite/gcc.target/powerpc/constpoolcheck.c
>
> Maybe it's good to name it to "const-pool-check.c" or "not-force-const-mem.c".
Great sugguestion! Thanks.
>
>> @@ -0,0 +1,11 @@
>> +/* { dg-do compile { target powerpc*-*-* } } */
>
> Nit: this "dg-do" line isn't needed since all here are default.
Thanks for your comments!

>
> BR,
> Kewen
>
>> +/* { dg-options "-O1 -mdejagnu-cpu=power10" } */
>> +/* (high:DI (symbol_ref:DI ("var_48")..))) should not cause ICE. */
>> +extern short var_48;
>> +void
>> +foo (double *r)
>> +{
>> +  if (var_48)
>> +    *r = 1234.5678;
>> +}
>> +
  

Hi!

On Tue, Jul 19, 2022 at 10:30:54PM +0800, Jiufu Guo wrote:
> In patch https://gcc.gnu.org/pipermail/gcc-patches/2022-July/597712.html,
> test case was not added.  After more check, a testcase is added for it.
> 
> The high part of the symbol address is invalid for the constant pool.

Invalid, how so?  Is there a PR related here?

But it is not particularly useful ever, either: we do not know two
different addresses will have the same HIGH unless we know the exact
address, and then we don't need HIGH anyway.

> 	* config/rs6000/rs6000.cc (rs6000_cannot_force_const_mem):
> 	Return true for HIGH code rtx.

	* config/rs6000/rs6000.cc (rs6000_cannot_force_const_mem): Return true
	for HIGH code rtx.

Please don't wrap lines early: changelog lines are 80 positions long,
including the leading tab (which counts as eight positions).

>  static bool
>  rs6000_cannot_force_const_mem (machine_mode mode ATTRIBUTE_UNUSED, rtx x)
>  {
> -  if (GET_CODE (x) == HIGH
> -      && GET_CODE (XEXP (x, 0)) == UNSPEC)
> +  /* High part of a symbol ref/address can not be put into constant pool. e.g.
> +     (high:DI (symbol_ref:DI ("var")..)) or
> +     (high:DI (unspec:DI [(symbol_ref/u:DI ("*.LC0")..)
> +     (high:DI (const:DI (plus:DI (symbol_ref:DI ("xx")) (const_int 12)))).  */
> +  if (GET_CODE (x) == HIGH)
>      return true;

I'm not sure the new comment is helpful at all?  Are these examples of
where the compiler (or assembler perhaps) will choke?

> --- /dev/null
> +++ b/gcc/testsuite/gcc.target/powerpc/constpoolcheck.c
> @@ -0,0 +1,11 @@
> +/* { dg-do compile { target powerpc*-*-* } } */

Everything in gcc.target/powerpc is target powerpc* always.

> +/* { dg-options "-O1 -mdejagnu-cpu=power10" } */
> +/* (high:DI (symbol_ref:DI ("var_48")..))) should not cause ICE. */

Ah, so there is an ICE, I see.  Please open a PR, and mention that in
the testcase as well as in the commit message and changelog.

I agree with what the patch does, it just needs a little more work :-)


Segher
  

Segher Boessenkool <segher@kernel.crashing.org> writes:

Thanks a lot for your review!

> Hi!
>
> On Tue, Jul 19, 2022 at 10:30:54PM +0800, Jiufu Guo wrote:
>> In patch https://gcc.gnu.org/pipermail/gcc-patches/2022-July/597712.html,
>> test case was not added.  After more check, a testcase is added for it.
>> 
>> The high part of the symbol address is invalid for the constant pool.
>
> Invalid, how so?  Is there a PR related here?
Thanks, I just opened PR106460 for this issue.
>
> But it is not particularly useful ever, either: we do not know two
> different addresses will have the same HIGH unless we know the exact
> address, and then we don't need HIGH anyway.
>
>> 	* config/rs6000/rs6000.cc (rs6000_cannot_force_const_mem):
>> 	Return true for HIGH code rtx.
>
> 	* config/rs6000/rs6000.cc (rs6000_cannot_force_const_mem): Return true
> 	for HIGH code rtx.
>
> Please don't wrap lines early: changelog lines are 80 positions long,
> including the leading tab (which counts as eight positions).
Thanks for your suggestion!
>
>>  static bool
>>  rs6000_cannot_force_const_mem (machine_mode mode ATTRIBUTE_UNUSED, rtx x)
>>  {
>> -  if (GET_CODE (x) == HIGH
>> -      && GET_CODE (XEXP (x, 0)) == UNSPEC)
>> +  /* High part of a symbol ref/address can not be put into constant pool. e.g.
>> +     (high:DI (symbol_ref:DI ("var")..)) or
>> +     (high:DI (unspec:DI [(symbol_ref/u:DI ("*.LC0")..)
>> +     (high:DI (const:DI (plus:DI (symbol_ref:DI ("xx")) (const_int 12)))).  */
>> +  if (GET_CODE (x) == HIGH)
>>      return true;
>
> I'm not sure the new comment is helpful at all?  Are these examples of
> where the compiler (or assembler perhaps) will choke?
I debugged this function with the source code from GCC bootstrap and
regtest, and then figured out these examples.
In the next version patch, I updated the comments a little, hope that
is more meaningful. :-)
>
>> --- /dev/null
>> +++ b/gcc/testsuite/gcc.target/powerpc/constpoolcheck.c
>> @@ -0,0 +1,11 @@
>> +/* { dg-do compile { target powerpc*-*-* } } */
>
> Everything in gcc.target/powerpc is target powerpc* always.
Thanks! I would remove this line.
>
>> +/* { dg-options "-O1 -mdejagnu-cpu=power10" } */
>> +/* (high:DI (symbol_ref:DI ("var_48")..))) should not cause ICE. */
>
> Ah, so there is an ICE, I see.  Please open a PR, and mention that in
> the testcase as well as in the commit message and changelog.
Thanks! I should open PR ealry :)
In the updated patch,  a testcase is named as pr106460.c, and memtioned
in commit message and changelog.
>
> I agree with what the patch does, it just needs a little more work :-)
I submitted a new version patch:
https://gcc.gnu.org/pipermail/gcc-patches/2022-July/598980.html

Thanks in advance for any comments!

BR,
Jeff(Jiufu)

>
>
> Segher