Re: [PATCH/RFC v2 3/14] Add new optabs for reducing vectors to scalars

[Alan Lawrence <alan dot lawrence at arm dot com>]

Well, even that C source, you'd need to be careful and ensure that the 
vectorized loop never went round more than once, or else the additions within 
the loop would be performed in 8 bits, different from the final reduction...

So: original patch with updated commenting attached...Segher, is there any 
chance you could test this on powerpc too? (in combination with patch 2/14, 
which will need to be applied first; you can skip patch 1, and >=4.)

--Alan

Richard Biener wrote:
> On Thu, Sep 25, 2014 at 4:32 PM, Alan Lawrence <alan.lawrence@arm.com> wrote:
> > Ok, so, I've tried making reduc_plus optab take two modes: that of the
> > vector to reduce, and the result; thus allowing platforms to provide a
> > widening reduction. However, I'm keeping reduc_[us](min|max)_optab with only
> > a single mode, as widening makes no sense there.
> >
> > I've not gone as far as making the vectorizer use any such a widening
> > reduction, however: as previously stated, I'm not really sure what the input
> > source code for that even looks like (maybe in a language other than C?). If
> > we wanted to do a non-widening reduction using such an instruction (by
> > discarding the extra bits), strikes me the platform can/should provide a
> > non-widening optab for that case...
>
> I expect it to apply to sth like
>
> int foo (char *in, int n)
> {
>    int res = 0;
>    for (int i = 0; i < n; ++i)
>      res += *in;
>    return res;
> }
>
> where you'd see
>
>   temc = *in;
>   tem = (int)temc;
>   res += tem;
>
> we probably handle this by widening the chars to ints and unrolling
> the loop enough to make that work (thus for n == 16 it would maybe
> fail to vectorize?).  It should be more efficient to pattern-detect
> this as widening reduction.
>
> > Testing: bootstrapped on x86_64 linux + check-gcc; cross-tested
> > aarch64-none-elf check-gcc; cross-tested aarch64_be-none-elf aarch64.exp +
> > vect.exp.
> >
> > So, my feeling is that the extra complexity here doesn't really buy us
> > anything; and that if we do want to support / use widening reductions in the
> > future, we should do so with a separate, reduc_plus_widen... optab, and
> > stick with the original patch/formulation for now. (In other words: this
> > patch is a guide to how I think a dual-mode reduc_plus_optab looks, but I
> > don't honestly like it!).
> >
> > If you agree, I shall transplant the comments on scalar_reduc_to_vector from
> > this patch into the original, and then post that revised version?
>
> I agree.  We can come back once a target implements such widening
> reduction.
>
> Richard.
>
> > Cheers, Alan
> >
> >
> > Richard Biener wrote:
> > > On Mon, Sep 22, 2014 at 3:26 PM, Alan Lawrence <alan.lawrence@arm.com>
> > > wrote:
> > > > Richard Biener wrote:
> > > > > scalar_reduc_to_vector misses a comment.
> > > >
> > > > Ok to reuse the comment in optabs.h in optabs.c also?
> > >
> > > Sure.
> > >
> > > > > I wonder if at the end we wouldn't transition all backends and then
> > > > > renaming reduc_*_scal_optab back to reduc_*_optab makes sense.
> > > >
> > > > Yes, that sounds like a plan, the _scal is a bit of a mouthful.
> > > >
> > > > > The optabs have only one mode - I wouldn't be surprised if an ISA
> > > > > invents for example v4si -> di reduction?  So do we want to make
> > > > > reduc_plus_scal_optab a little bit more future proof (maybe there
> > > > > is already an ISA that supports this kind of reduction?).
> > > >
> > > > That sounds like a plausible thing for an ISA to do, indeed. However
> > > > given
> > > > these names are only used by the autovectorizer rather than directly, the
> > > > question is what the corresponding source code looks like, and/or what
> > > > changes to the autovectorizer we might have to make to (look for code to)
> > > > exploit such an instruction.
> > >
> > > Ah, indeed.  Would be sth like a REDUC_WIDEN_SUM_EXPR or so.
> > >
> > > > At this point I could go for a
> > > > reduc_{plus,min_max}_scal_<mode><mode> which reduces from the first
> > > > vector
> > > > mode to the second scalar mode, and then make the vectorizer look only
> > > > for
> > > > cases where the second mode was the element type of the first; but I'm
> > > > not
> > > > sure I want to do anything more complicated than that at this stage.
> > > > (However, indeed it would leave the possibility open for the future.)
> > >
> > > Yeah, agreed.  For the min/max case a widen variant isn't useful anyway.
> > >
> > > Thanks,
> > > Richard.
> > >
> > > > --Alan
diff --git a/gcc/doc/md.texi b/gcc/doc/md.texi
index 80e8bd6..84e5261 100644
--- a/gcc/doc/md.texi
+++ b/gcc/doc/md.texi
@@ -4783,29 +4783,48 @@ it is unspecified which of the two operands is returned as the result.
 @cindex @code{reduc_smax_@var{m}} instruction pattern
 @item @samp{reduc_smin_@var{m}}, @samp{reduc_smax_@var{m}}
 Find the signed minimum/maximum of the elements of a vector. The vector is
-operand 1, and the scalar result is stored in the least significant bits of
+operand 1, and the result is stored in the least significant bits of
 operand 0 (also a vector). The output and input vector should have the same
-modes.
+modes. These are legacy optabs, and platforms should prefer to implement
+@samp{reduc_smin_scal_@var{m}} and @samp{reduc_smax_scal_@var{m}}.
 
 @cindex @code{reduc_umin_@var{m}} instruction pattern
 @cindex @code{reduc_umax_@var{m}} instruction pattern
 @item @samp{reduc_umin_@var{m}}, @samp{reduc_umax_@var{m}}
 Find the unsigned minimum/maximum of the elements of a vector. The vector is
-operand 1, and the scalar result is stored in the least significant bits of
+operand 1, and the result is stored in the least significant bits of
 operand 0 (also a vector). The output and input vector should have the same
-modes.
+modes. These are legacy optabs, and platforms should prefer to implement
+@samp{reduc_umin_scal_@var{m}} and @samp{reduc_umax_scal_@var{m}}.
 
 @cindex @code{reduc_splus_@var{m}} instruction pattern
-@item @samp{reduc_splus_@var{m}}
-Compute the sum of the signed elements of a vector. The vector is operand 1,
-and the scalar result is stored in the least significant bits of operand 0
-(also a vector). The output and input vector should have the same modes.
-
 @cindex @code{reduc_uplus_@var{m}} instruction pattern
-@item @samp{reduc_uplus_@var{m}}
-Compute the sum of the unsigned elements of a vector. The vector is operand 1,
-and the scalar result is stored in the least significant bits of operand 0
+@item @samp{reduc_splus_@var{m}}, @samp{reduc_uplus_@var{m}}
+Compute the sum of the signed/unsigned elements of a vector. The vector is
+operand 1, and the result is stored in the least significant bits of operand 0
 (also a vector). The output and input vector should have the same modes.
+These are legacy optabs, and platforms should prefer to implement
+@samp{reduc_plus_scal_@var{m}}.
+
+@cindex @code{reduc_smin_scal_@var{m}} instruction pattern
+@cindex @code{reduc_smax_scal_@var{m}} instruction pattern
+@item @samp{reduc_smin_scal_@var{m}}, @samp{reduc_smax_scal_@var{m}}
+Find the signed minimum/maximum of the elements of a vector. The vector is
+operand 1, and operand 0 is the scalar result, with mode equal to the mode of
+the elements of the input vector.
+
+@cindex @code{reduc_umin_scal_@var{m}} instruction pattern
+@cindex @code{reduc_umax_scal_@var{m}} instruction pattern
+@item @samp{reduc_umin_scal_@var{m}}, @samp{reduc_umax_scal_@var{m}}
+Find the unsigned minimum/maximum of the elements of a vector. The vector is
+operand 1, and operand 0 is the scalar result, with mode equal to the mode of
+the elements of the input vector.
+
+@cindex @code{reduc_plus_scal_@var{m}} instruction pattern
+@item @samp{reduc_plus_scal_@var{m}}
+Compute the sum of the elements of a vector. The vector is operand 1, and
+operand 0 is the scalar result, with mode equal to the mode of the elements of
+the input vector.
 
 @cindex @code{sdot_prod@var{m}} instruction pattern
 @item @samp{sdot_prod@var{m}}
diff --git a/gcc/expr.c b/gcc/expr.c
index c792028..3763614 100644
--- a/gcc/expr.c
+++ b/gcc/expr.c
@@ -9045,6 +9045,24 @@ expand_expr_real_2 (sepops ops, rtx target, enum machine_mode tmode,
         op0 = expand_normal (treeop0);
         this_optab = optab_for_tree_code (code, type, optab_default);
         enum machine_mode vec_mode = TYPE_MODE (TREE_TYPE (treeop0));
+
+	if (optab_handler (this_optab, vec_mode) != CODE_FOR_nothing)
+	  {
+	    struct expand_operand ops[2];
+	    enum insn_code icode = optab_handler (this_optab, vec_mode);
+
+	    create_output_operand (&ops[0], target, mode);
+	    create_input_operand (&ops[1], op0, vec_mode);
+	    if (maybe_expand_insn (icode, 2, ops))
+	      {
+		target = ops[0].value;
+		if (GET_MODE (target) != mode)
+		  return gen_lowpart (tmode, target);
+		return target;
+	      }
+	  }
+	/* Fall back to optab with vector result, and then extract scalar.  */
+	this_optab = scalar_reduc_to_vector (this_optab, type);
         temp = expand_unop (vec_mode, this_optab, op0, NULL_RTX, unsignedp);
         gcc_assert (temp);
         /* The tree code produces a scalar result, but (somewhat by convention)
diff --git a/gcc/optabs.c b/gcc/optabs.c
index 605615d..cfb408c 100644
--- a/gcc/optabs.c
+++ b/gcc/optabs.c
@@ -506,13 +506,15 @@ optab_for_tree_code (enum tree_code code, const_tree type,
       return fma_optab;
 
     case REDUC_MAX_EXPR:
-      return TYPE_UNSIGNED (type) ? reduc_umax_optab : reduc_smax_optab;
+      return TYPE_UNSIGNED (type)
+	     ? reduc_umax_scal_optab : reduc_smax_scal_optab;
 
     case REDUC_MIN_EXPR:
-      return TYPE_UNSIGNED (type) ? reduc_umin_optab : reduc_smin_optab;
+      return TYPE_UNSIGNED (type)
+	     ? reduc_umin_scal_optab : reduc_smin_scal_optab;
 
     case REDUC_PLUS_EXPR:
-      return TYPE_UNSIGNED (type) ? reduc_uplus_optab : reduc_splus_optab;
+      return reduc_plus_scal_optab;
 
     case VEC_LSHIFT_EXPR:
       return vec_shl_optab;
@@ -608,7 +610,26 @@ optab_for_tree_code (enum tree_code code, const_tree type,
       return unknown_optab;
     }
 }
-
+
+/* Given optab UNOPTAB that reduces a vector to a scalar, find instead the old
+   optab that produces a vector with the reduction result in one element,
+   for a tree with type TYPE.  */
+
+optab
+scalar_reduc_to_vector (optab unoptab, const_tree type)
+{
+  switch (unoptab)
+    {
+    case reduc_plus_scal_optab:
+      return TYPE_UNSIGNED (type) ? reduc_uplus_optab : reduc_splus_optab;
+
+    case reduc_smin_scal_optab: return reduc_smin_optab;
+    case reduc_umin_scal_optab: return reduc_umin_optab;
+    case reduc_smax_scal_optab: return reduc_smax_optab;
+    case reduc_umax_scal_optab: return reduc_umax_optab;
+    default: return unknown_optab;
+    }
+}
 
 /* Expand vector widening operations.
 
diff --git a/gcc/optabs.def b/gcc/optabs.def
index b755470..131ea04 100644
--- a/gcc/optabs.def
+++ b/gcc/optabs.def
@@ -243,12 +243,20 @@ OPTAB_D (sin_optab, "sin$a2")
 OPTAB_D (sincos_optab, "sincos$a3")
 OPTAB_D (tan_optab, "tan$a2")
 
+/* Vector reduction to a scalar.  */
+OPTAB_D (reduc_smax_scal_optab, "reduc_smax_scal_$a")
+OPTAB_D (reduc_smin_scal_optab, "reduc_smin_scal_$a")
+OPTAB_D (reduc_plus_scal_optab, "reduc_plus_scal_$a")
+OPTAB_D (reduc_umax_scal_optab, "reduc_umax_scal_$a")
+OPTAB_D (reduc_umin_scal_optab, "reduc_umin_scal_$a")
+/* (Old) Vector reduction, returning a vector with the result in one lane.  */
 OPTAB_D (reduc_smax_optab, "reduc_smax_$a")
 OPTAB_D (reduc_smin_optab, "reduc_smin_$a")
 OPTAB_D (reduc_splus_optab, "reduc_splus_$a")
 OPTAB_D (reduc_umax_optab, "reduc_umax_$a")
 OPTAB_D (reduc_umin_optab, "reduc_umin_$a")
 OPTAB_D (reduc_uplus_optab, "reduc_uplus_$a")
+
 OPTAB_D (sdot_prod_optab, "sdot_prod$I$a")
 OPTAB_D (ssum_widen_optab, "widen_ssum$I$a3")
 OPTAB_D (udot_prod_optab, "udot_prod$I$a")
diff --git a/gcc/optabs.h b/gcc/optabs.h
index 089b15a..1f8e44d 100644
--- a/gcc/optabs.h
+++ b/gcc/optabs.h
@@ -162,6 +162,11 @@ enum optab_subtype
    vector shifts and rotates */
 extern optab optab_for_tree_code (enum tree_code, const_tree, enum optab_subtype);
 
+/* Given an optab that reduces a vector to a scalar, find instead the old
+   optab that produces a vector with the reduction result in one element,
+   for a tree with the specified type.  */
+extern optab scalar_reduc_to_vector (optab, const_tree type);
+
 /* The various uses that a comparison can have; used by can_compare_p:
    jumps, conditional moves, store flag operations.  */
 enum can_compare_purpose
diff --git a/gcc/tree-vect-loop.c b/gcc/tree-vect-loop.c
index 8d97e17..e8b108d 100644
--- a/gcc/tree-vect-loop.c
+++ b/gcc/tree-vect-loop.c
@@ -5102,15 +5102,17 @@ vectorizable_reduction (gimple stmt, gimple_stmt_iterator *gsi,
 
           epilog_reduc_code = ERROR_MARK;
         }
-
-      if (reduc_optab
-          && optab_handler (reduc_optab, vec_mode) == CODE_FOR_nothing)
+      else if (optab_handler (reduc_optab, vec_mode) == CODE_FOR_nothing)
         {
-          if (dump_enabled_p ())
-	    dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location,
-			     "reduc op not supported by target.\n");
+          optab = scalar_reduc_to_vector (reduc_optab, vectype_out);
+          if (optab_handler (optab, vec_mode) == CODE_FOR_nothing)
+            {
+              if (dump_enabled_p ())
+	        dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location,
+				 "reduc op not supported by target.\n");
 
-          epilog_reduc_code = ERROR_MARK;
+	      epilog_reduc_code = ERROR_MARK;
+	    }
         }
     }
   else