Re: [PATCH] Make sure SCALAR_INT_MODE_P before invoke try_const_anchors

[Richard Biener <rguenther@suse.de>]

On Mon, 12 Jun 2023, Jiufu Guo wrote:

> Richard Biener <rguenther@suse.de> writes:
> 
> > On Fri, 9 Jun 2023, Jiufu Guo wrote:
> >
> >> 
> >> Hi,
> >> 
> >> Richard Biener <rguenther@suse.de> writes:
> >> 
> >> > On Fri, 9 Jun 2023, Jiufu Guo wrote:
> >> >
> >> >> 
> >> >> Hi,
> >> >> 
> >> >> Richard Biener <rguenther@suse.de> writes:
> >> >> 
> >> >> > On Fri, 9 Jun 2023, Richard Sandiford wrote:
> >> >> >
> >> >> >> guojiufu <guojiufu@linux.ibm.com> writes:
> >> >> >> > Hi,
> >> >> >> >
> >> >> >> > On 2023-06-09 16:00, Richard Biener wrote:
> >> >> >> >> On Fri, 9 Jun 2023, Jiufu Guo wrote:
> >> >> >> >> 
> >> >> >> >>> Hi,
> >> >> >> >>> 
> >> ...
> >> >> >> >>> 
> >> >> >> >>> This patch is raised when drafting below one.
> >> >> >> >>> https://gcc.gnu.org/pipermail/gcc-patches/2022-October/603530.html.
> >> >> >> >>> With that patch, "{[%1:DI]=0;} stack_tie" with BLKmode runs into
> >> >> >> >>> try_const_anchors, and hits the assert/ice.
> >> >> >> >>> 
> >> >> >> >>> Boostrap and regtest pass on ppc64{,le} and x86_64.
> >> >> >> >>> Is this ok for trunk?
> >> >> >> >> 
> >> >> >> >> Iff the correct fix at all (how can a CONST_INT have BLKmode?) then
> >> >> >> >> I suggest to instead fix try_const_anchors to change
> >> >> >> >> 
> >> >> >> >>   /* CONST_INT is used for CC modes, but we should leave those alone.  
> >> >> >> >> */
> >> >> >> >>   if (GET_MODE_CLASS (mode) == MODE_CC)
> >> >> >> >>     return NULL_RTX;
> >> >> >> >> 
> >> >> >> >>   gcc_assert (SCALAR_INT_MODE_P (mode));
> >> >> >> >> 
> >> >> >> >> to
> >> >> >> >> 
> >> >> >> >>   /* CONST_INT is used for CC modes, leave any non-scalar-int mode 
> >> >> >> >> alone.  */
> >> >> >> >>   if (!SCALAR_INT_MODE_P (mode))
> >> >> >> >>     return NULL_RTX;
> >> >> >> >> 
> >> >> >> >
> >> >> >> > This is also able to fix this issue.  there is a "Punt on CC modes" 
> >> >> >> > patch
> >> >> >> > to return NULL_RTX in try_const_anchors.
> >> >> >> >
> >> >> >> >> but as said I wonder how we arrive at a BLKmode CONST_INT and whether
> >> >> >> >> we should have fended this off earlier.  Can you share more complete
> >> >> >> >> RTL of that stack_tie?
> >> >> >> >
> >> >> >> >
> >> >> >> > (insn 15 14 16 3 (parallel [
> >> >> >> >              (set (mem/c:BLK (reg/f:DI 1 1) [1  A8])
> >> >> >> >                  (const_int 0 [0]))
> >> >> >> >          ]) "/home/guojiufu/temp/gdb.c":13:3 922 {stack_tie}
> >> >> >> >       (nil))
> >> >> >> >
> >> >> >> > It is "set (mem/c:BLK (reg/f:DI 1 1) (const_int 0 [0])".
> >> >> >> 
> >> >> >> I'm not convinced this is correct RTL.  (unspec:BLK [(const_int 0)] ...)
> >> >> >> would be though.  It's arguably more accurate too, since the effect
> >> >> >> on the stack locations is unspecified rather than predictable.
> >> >> >
> >> >> > powerpc seems to be the only port with a stack_tie that's not
> >> >> > using an UNSPEC RHS.
> >> >> In rs6000.md, it is
> >> >> 
> >> >> ; This is to explain that changes to the stack pointer should
> >> >> ; not be moved over loads from or stores to stack memory.
> >> >> (define_insn "stack_tie"
> >> >>   [(match_parallel 0 "tie_operand"
> >> >> 		   [(set (mem:BLK (reg 1)) (const_int 0))])]
> >> >>   ""
> >> >>   ""
> >> >>   [(set_attr "length" "0")])
> >> >> 
> >> >> This would be just an placeholder insn, and acts as the comments.
> >> >> UNSPEC_ would works like other targets.  While, I'm wondering
> >> >> the concerns on "set (mem:BLK (reg 1)) (const_int 0)".
> >> >> MODEs between SET_DEST and SET_SRC?
> >> >
> >> > I don't think the issue is the mode but the issue is that
> >> > the patter as-is says some memory is zeroed while that's not
> >> > actually true (not specifying a size means we can't really do
> >> > anything with this MEM, but still).  Using an UNSPEC avoids
> >> > implying anything for the stored value.
> >> >
> >> > Of course I think a MEM SET_DEST without a specified size is bougs
> >> > as well, but there's larger precedent for this...
> >> 
> >> Thanks for your kindly comments!
> >> Using "(set (mem:BLK (reg 1)) (const_int 0))" here, may because this
> >> insn does not generate real thing (not a real store and no asm code),
> >> may like barrier.
> >> 
> >> While I agree that, using UNSPEC may be more clear to avoid mis-reading.
> >
> > Btw, another way to avoid the issue in CSE is to make it not process
> > (aka record anything for optimization) for SET from MEMs with
> > !MEM_SIZE_KNOWN_P
> 
> Thanks! Yes, this would make sense.
> Then, there are two ideas(patches) to handle this issue:
> Which one would be preferable?  This one (from compiling time aspect)?
> 
> And maybe, the changes in rs6000 stack_tie through using unspec
> can be a standalone enhancement besides cse patch.
> 
> Thanks for comments!
> 
> BR,
> Jeff (Jiufu Guo)
> 
> -------------------- patch 1
> diff --git a/gcc/cse.cc b/gcc/cse.cc
> index 2bb63ac4105..06ecdadecbc 100644
> --- a/gcc/cse.cc
> +++ b/gcc/cse.cc
> @@ -4271,6 +4271,8 @@ find_sets_in_insn (rtx_insn *insn, vec<struct set> *psets)
>  	 someplace else, so it isn't worth cse'ing.  */
>        else if (GET_CODE (SET_SRC (x)) == CALL)
>  	;
> +      else if (MEM_P (SET_DEST (x)) && !MEM_SIZE_KNOWN_P (SET_DEST (x)))
> +	;

I don't know CSE enough to decide if this is correct, it depends on
whether pset is only used to record new sets or if it is also used
to invalidate prior recorded sets (in which case these sets have to
be recorded still).  That is, the fix might better be in the caller
who interprets 'pset'

>        else if (GET_CODE (SET_SRC (x)) == CONST_VECTOR
>  	       && GET_MODE_CLASS (GET_MODE (SET_SRC (x))) != MODE_VECTOR_BOOL
>  	       /* Prevent duplicates from being generated if the type is a V1
> @@ -4314,6 +4316,8 @@ find_sets_in_insn (rtx_insn *insn, vec<struct set> *psets)
>  		;
>  	      else if (GET_CODE (SET_SRC (y)) == CALL)
>  		;
> +	      else if (MEM_P (SET_DEST (y)) && !MEM_SIZE_KNOWN_P (SET_DEST (y)))
> +		;
>  	      else
>  		add_to_set (psets, y);
>  	    }
> -----------------------------
> -------------------patch 2
> diff --git a/gcc/cse.cc b/gcc/cse.cc
> index 2bb63ac4105..ddb76fd281d 100644
> --- a/gcc/cse.cc
> +++ b/gcc/cse.cc
> @@ -1312,11 +1312,10 @@ try_const_anchors (rtx src_const, machine_mode mode)
>    rtx lower_exp = NULL_RTX, upper_exp = NULL_RTX;
>    unsigned lower_old, upper_old;
>  
> -  /* CONST_INT is used for CC modes, but we should leave those alone.  */
> -  if (GET_MODE_CLASS (mode) == MODE_CC)
> +  /* CONST_INT is used for CC/BLK modes, leave any non-scalar-int mode alone. */
> +  if (!SCALAR_INT_MODE_P (mode))
>      return NULL_RTX;
>  
> -  gcc_assert (SCALAR_INT_MODE_P (mode));
>    if (!compute_const_anchors (src_const, &lower_base, &lower_offs,
>  			      &upper_base, &upper_offs))
>      return NULL_RTX;
> -------------
> 
> 
> BR,
> Jeff (Jiufu Guo)
> >
> > Richard.
> 

-- 
Richard Biener <rguenther@suse.de>
SUSE Software Solutions Germany GmbH, Frankenstrasse 146, 90461 Nuernberg,
Germany; GF: Ivo Totev, Andrew Myers, Andrew McDonald, Boudien Moerman;
HRB 36809 (AG Nuernberg)