rs6000: Make CSE'ing __tls_get_addr calls possible

CSE does not consider calls, not even const calls.  This patch puts a
REG_EQUAL note on the pseudo we assign the __tls_get_addr result to,
so that those pseudos can be CSE'd and the extra calls deleted as dead
code.

CSE should really handle const calls directly, but it is stage 4.

Tested on powerpc64-linux {-m32,-m64}.  I'll also test on powerpc64le-linux,
and will commit it to trunk if that works.


Segher


2019-03-23  Segher Boessenkool  <segher@kernel.crashing.org>

	* config/rs6000/rs6000.c (rs6000_legitimize_tls_address): Add REG_EQUAL
	notes for the result of the __tls_get_addr calls.
	* config/rs6000/rs6000.md (unspec UNSPEC_TLS_GET_ADDR): New.

---
 gcc/config/rs6000/rs6000.c  | 10 ++++++++++
 gcc/config/rs6000/rs6000.md |  1 +
 2 files changed, 11 insertions(+)

On Sun, Mar 24, 2019 at 12:46 AM Segher Boessenkool wrote:
>
> CSE does not consider calls, not even const calls.  This patch puts a
> REG_EQUAL note on the pseudo we assign the __tls_get_addr result to,
> so that those pseudos can be CSE'd and the extra calls deleted as dead
> code.

Hi Segher,

There were REG_EQUAL notes on these tls calls in the past, but I
recall removing them for one reason or another. So watch out for
fall-out from this patch! ;-)

Ciao!
Steven

Hi Steven,

On Sun, Mar 24, 2019 at 11:02:20AM +0100, Steven Bosscher wrote:
> On Sun, Mar 24, 2019 at 12:46 AM Segher Boessenkool wrote:
> >
> > CSE does not consider calls, not even const calls.  This patch puts a
> > REG_EQUAL note on the pseudo we assign the __tls_get_addr result to,
> > so that those pseudos can be CSE'd and the extra calls deleted as dead
> > code.
> 
> There were REG_EQUAL notes on these tls calls in the past, but I
> recall removing them for one reason or another.

I cannot find this in history.  History is pretty big of course, so perhaps
I'm not looking at the right corner :-)

> So watch out for
> fall-out from this patch! ;-)

Thanks for the warning!

The patch doesn't change anything, in the sense that these calls were const
calls before already, but yeah :-)


Segher

> There were REG_EQUAL notes on these tls calls in the past, but I
> recall removing them for one reason or another. So watch out for
> fall-out from this patch! ;-)

Others ports have them though, for example i386, ARM and SPARC.

Hi all,

On Sat, Mar 23, 2019 at 11:46:12PM +0000, Segher Boessenkool wrote:
> CSE does not consider calls, not even const calls.  This patch puts a
> REG_EQUAL note on the pseudo we assign the __tls_get_addr result to,
> so that those pseudos can be CSE'd and the extra calls deleted as dead
> code.
> 
> CSE should really handle const calls directly, but it is stage 4.

I am backporting this to GCC 8.


Segher


> 2019-03-23  Segher Boessenkool  <segher@kernel.crashing.org>
> 
> 	* config/rs6000/rs6000.c (rs6000_legitimize_tls_address): Add REG_EQUAL
> 	notes for the result of the __tls_get_addr calls.
> 	* config/rs6000/rs6000.md (unspec UNSPEC_TLS_GET_ADDR): New.
> 
> ---
>  gcc/config/rs6000/rs6000.c  | 10 ++++++++++
>  gcc/config/rs6000/rs6000.md |  1 +
>  2 files changed, 11 insertions(+)
> 
> diff --git a/gcc/config/rs6000/rs6000.c b/gcc/config/rs6000/rs6000.c
> index 15811bd..35aedf4 100644
> --- a/gcc/config/rs6000/rs6000.c
> +++ b/gcc/config/rs6000/rs6000.c
> @@ -8661,6 +8661,11 @@ rs6000_legitimize_tls_address (rtx addr, enum tls_model model)
>  	  else
>  	    emit_library_call_value (tga, dest, LCT_CONST, Pmode);
>  	  global_tlsarg = NULL_RTX;
> +
> +	  /* Make a note so that the result of this call can be CSEd.  */
> +	  rtvec vec = gen_rtvec (1, copy_rtx (arg));
> +	  rtx uns = gen_rtx_UNSPEC (Pmode, vec, UNSPEC_TLS_GET_ADDR);
> +	  set_unique_reg_note (get_last_insn (), REG_EQUAL, uns);
>  	}
>        else if (model == TLS_MODEL_LOCAL_DYNAMIC)
>  	{
> @@ -8679,6 +8684,11 @@ rs6000_legitimize_tls_address (rtx addr, enum tls_model model)
>  	    emit_library_call_value (tga, tmp1, LCT_CONST, Pmode);
>  	  global_tlsarg = NULL_RTX;
>  
> +	  /* Make a note so that the result of this call can be CSEd.  */
> +	  rtvec vec = gen_rtvec (1, copy_rtx (arg));
> +	  rtx uns = gen_rtx_UNSPEC (Pmode, vec, UNSPEC_TLS_GET_ADDR);
> +	  set_unique_reg_note (get_last_insn (), REG_EQUAL, uns);
> +
>  	  if (rs6000_tls_size == 16)
>  	    {
>  	      if (TARGET_64BIT)
> diff --git a/gcc/config/rs6000/rs6000.md b/gcc/config/rs6000/rs6000.md
> index f2faef8..815077c 100644
> --- a/gcc/config/rs6000/rs6000.md
> +++ b/gcc/config/rs6000/rs6000.md
> @@ -79,6 +79,7 @@ (define_c_enum "unspec"
>     UNSPEC_MPIC_CORRECT		; macho_correct_pic
>     UNSPEC_TLSGD
>     UNSPEC_TLSLD
> +   UNSPEC_TLS_GET_ADDR
>     UNSPEC_MOVESI_FROM_CR
>     UNSPEC_MOVESI_TO_CR
>     UNSPEC_TLSDTPREL
> -- 
> 1.8.3.1

On Sat, Aug 31, 2019 at 02:17:30PM -0500, Segher Boessenkool wrote:
> Hi all,
> 
> On Sat, Mar 23, 2019 at 11:46:12PM +0000, Segher Boessenkool wrote:
> > CSE does not consider calls, not even const calls.  This patch puts a
> > REG_EQUAL note on the pseudo we assign the __tls_get_addr result to,
> > so that those pseudos can be CSE'd and the extra calls deleted as dead
> > code.
> > 
> > CSE should really handle const calls directly, but it is stage 4.
> 
> I am backporting this to GCC 8.

Actually, that is not trivial at all, so I'm not doing that.


Segher