Re: Re: [PATCH] Middle-end: Fix bug of induction variable vectorization for RVV

["juzhe.zhong@rivai.ai" <juzhe.zhong@rivai.ai>]

[-- Attachment #1: Type: text/plain, Size: 17734 bytes --]

Thanks a lot. I think I finally understand what you mean now :).

Could you confirm this following codes:?

diff --git a/gcc/tree-vect-loop.cc b/gcc/tree-vect-loop.cc
index 8abc1937d74..5615b16bdcd 100644
--- a/gcc/tree-vect-loop.cc
+++ b/gcc/tree-vect-loop.cc
@@ -10306,10 +10306,39 @@ vectorizable_induction (loop_vec_info loop_vinfo,


   /* Create the vector that holds the step of the induction.  */
+  gimple_stmt_iterator *step_iv_si = NULL;
   if (nested_in_vect_loop)
-    /* iv_loop is nested in the loop to be vectorized. Generate:
-       vec_step = [S, S, S, S]  */
-    new_name = step_expr;
+    {
+      /* iv_loop is nested in the loop to be vectorized. Generate:
+        vec_step = [S, S, S, S]  */
+      new_name = step_expr;
+      /* We expect LOOP_VINFO_USING_SELECT_VL_P to be false in nested loop.  */
+      gcc_assert (!LOOP_VINFO_USING_SELECT_VL_P (loop_vinfo));
+    }
+  else if (LOOP_VINFO_USING_SELECT_VL_P (loop_vinfo))
+    {
+      /* When we're using loop_len produced by SELEC_VL, the non-final
+        iterations are not always processing VF elements.  So vectorize
+        induction variable instead of
+
+          _21 = vect_vec_iv_.6_22 + { VF, ... };
+
+        We should generate:
+
+          _35 = .SELECT_VL (ivtmp_33, VF);
+          vect_cst__22 = [vec_duplicate_expr] _35;
+          _21 = vect_vec_iv_.6_22 + vect_cst__22;  */
+      gimple_seq seq = NULL;
+      vec_loop_lens *lens = &LOOP_VINFO_LENS (loop_vinfo);
+      tree len = vect_get_loop_len (loop_vinfo, NULL, lens, 1, vectype, 0, 0);
+      expr = force_gimple_operand (fold_convert (TREE_TYPE (step_expr),
+                                                unshare_expr (len)),
+                                  &seq, true, NULL_TREE);
+      new_name = gimple_build (&seq, MULT_EXPR, TREE_TYPE (step_expr), expr,
+                              step_expr);
+      gsi_insert_seq_before (&si, seq, GSI_SAME_STMT);
+      step_iv_si = &si;
+    }
   else
     {
       /* iv_loop is the loop to be vectorized. Generate:
@@ -10336,7 +10365,7 @@ vectorizable_induction (loop_vec_info loop_vinfo,
              || TREE_CODE (new_name) == SSA_NAME);
   new_vec = build_vector_from_val (step_vectype, t);
   vec_step = vect_init_vector (loop_vinfo, stmt_info,
-                              new_vec, step_vectype, NULL);
+                              new_vec, step_vectype, step_iv_si);


   /* Create the following def-use cycle:
@@ -10382,6 +10411,8 @@ vectorizable_induction (loop_vec_info loop_vinfo,
       gimple_seq seq = NULL;
       /* FORNOW. This restriction should be relaxed.  */
       gcc_assert (!nested_in_vect_loop);
+      /* We expect LOOP_VINFO_USING_SELECT_VL_P to be false if ncopies > 1.  */
+      gcc_assert (!LOOP_VINFO_USING_SELECT_VL_P (loop_vinfo));


If it is Ok for you. I am gonna testing it on X86 bootstrap + regtest.

Thanks.



juzhe.zhong@rivai.ai
 
From: Richard Biener
Date: 2023-11-10 18:19
To: juzhe.zhong@rivai.ai
CC: Richard Biener; gcc-patches; richard.sandiford; kito.cheng; Kito.cheng
Subject: Re: Re: [PATCH] Middle-end: Fix bug of induction variable vectorization for RVV
On Fri, 10 Nov 2023, juzhe.zhong@rivai.ai wrote:
 
> Hi, Richard.
> 
> I am sorry for bothering you. I am trying to understand what you mean.
> 
> Is this following codes that you want ?
> 
>   /* Create the vector that holds the step of the induction.  */
>   if (nested_in_vect_loop)
>     {
>       /* iv_loop is nested in the loop to be vectorized. Generate:
>          vec_step = [S, S, S, S]  */
>       new_name = step_expr;
>       /* We expect LOOP_VINFO_USING_SELECT_VL_P to be false in nested loop.  */
>       gcc_assert (!LOOP_VINFO_USING_SELECT_VL_P (loop_vinfo));
>       t = unshare_expr (new_name);
>       gcc_assert (CONSTANT_CLASS_P (new_name)
>                   || TREE_CODE (new_name) == SSA_NAME);
>       new_vec = build_vector_from_val (step_vectype, t);
>       vec_step
>         = vect_init_vector (loop_vinfo, stmt_info, new_vec, step_vectype, NULL);
>     }
>   else if (LOOP_VINFO_USING_SELECT_VL_P (loop_vinfo))
>     {
>       /* When we're using loop_len produced by SELEC_VL, the non-final
>          iterations are not always processing VF elements.  So vectorize
>          induction variable instead of
> 
>            _21 = vect_vec_iv_.6_22 + { VF, ... };
> 
>          We should generate:
> 
>            _35 = .SELECT_VL (ivtmp_33, VF);
>            vect_cst__22 = [vec_duplicate_expr] _35;
>            _21 = vect_vec_iv_.6_22 + vect_cst__22;  */
>       vec_loop_lens *lens = &LOOP_VINFO_LENS (loop_vinfo);
>       tree len = vect_get_loop_len (loop_vinfo, NULL, lens, 1, vectype, 0, 0);
>       expr = force_gimple_operand (fold_convert (TREE_TYPE (step_expr),
>                                                  unshare_expr (len)),
>                                    &seq, true, NULL_TREE);
>       gsi_insert_seq_before (&si, seq, GSI_SAME_STMT);
>       t = unshare_expr (new_name);
>       gcc_assert (CONSTANT_CLASS_P (new_name)
>                   || TREE_CODE (new_name) == SSA_NAME);
>       new_vec = build_vector_from_val (step_vectype, t);
>       vec_step
>         = vect_init_vector (loop_vinfo, stmt_info, new_vec, step_vectype, &si);
>     }
>   else
>     {
>       /* iv_loop is the loop to be vectorized. Generate:
>           vec_step = [VF*S, VF*S, VF*S, VF*S]  */
>       gimple_seq seq = NULL;
>       if (SCALAR_FLOAT_TYPE_P (TREE_TYPE (step_expr)))
>         {
>           expr = build_int_cst (integer_type_node, vf);
>           expr = gimple_build (&seq, FLOAT_EXPR, TREE_TYPE (step_expr), expr);
>         }
>       else
>         expr = build_int_cst (TREE_TYPE (step_expr), vf);
>       new_name = gimple_build (&seq, MULT_EXPR, TREE_TYPE (step_expr),
>                                expr, step_expr);
>       if (seq)
>         {
>           new_bb = gsi_insert_seq_on_edge_immediate (pe, seq);
>           gcc_assert (!new_bb);
>         }
>       t = unshare_expr (new_name);
>       gcc_assert (CONSTANT_CLASS_P (new_name)
>                   || TREE_CODE (new_name) == SSA_NAME);
>       new_vec = build_vector_from_val (step_vectype, t);
>       vec_step
>         = vect_init_vector (loop_vinfo, stmt_info, new_vec, step_vectype, NULL);
>     }
> 
> It seems that this following codes:
> 
>       t = unshare_expr (new_name);
>       gcc_assert (CONSTANT_CLASS_P (new_name)
>                   || TREE_CODE (new_name) == SSA_NAME);
>       new_vec = build_vector_from_val (step_vectype, t);
>       vec_step
>         = vect_init_vector 
> 
> appears 3 times.  I am not sure whether it is the way you want?
 
I'd avoid that particular bit by having
 
   gimple_stmt_iterator *si = NULL;
 
before the if () and set that accordingly only in the
LOOP_VINFO_USING_SELECT_VL_P path.  But otherwise yes.
 
Richard.
 
> 
> Thanks.
> 
> 
> 
> juzhe.zhong@rivai.ai
>  
> From: Richard Biener
> Date: 2023-11-10 17:46
> To: ???
> CC: richard.guenther; gcc-patches; richard.sandiford; kito.cheng; kito.cheng
> Subject: Re: Re: [PATCH] Middle-end: Fix bug of induction variable vectorization for RVV
> On Thu, 9 Nov 2023, ??? wrote:
>  
> > Hi, Richard.
> > 
> > >> I think it would be better to split out building a tree from VF from both
> > >> arms and avoid using 'vf' when LOOP_VINFO_USING_SELECT_VL_P.
> > 
> > I am trying to split out building tree from both arms as you suggested..
> > Could you take a look the following codes ?
> > 
> > diff --git a/gcc/tree-vect-loop.cc b/gcc/tree-vect-loop.cc
> > index 8abc1937d74..24a86187d11 100644
> > --- a/gcc/tree-vect-loop.cc
> > +++ b/gcc/tree-vect-loop.cc
> > @@ -10315,19 +10315,47 @@ vectorizable_induction (loop_vec_info loop_vinfo,
> >        /* iv_loop is the loop to be vectorized. Generate:
> >     vec_step = [VF*S, VF*S, VF*S, VF*S]  */
> >        gimple_seq seq = NULL;
> > -      if (SCALAR_FLOAT_TYPE_P (TREE_TYPE (step_expr)))
> > +      if (LOOP_VINFO_USING_SELECT_VL_P (loop_vinfo))
> >   {
> > -   expr = build_int_cst (integer_type_node, vf);
> > -   expr = gimple_build (&seq, FLOAT_EXPR, TREE_TYPE (step_expr), expr);
> > +   /* When we're using loop_len produced by SELEC_VL, the non-final
> > +      iterations are not always processing VF elements.  So vectorize
> > +      induction variable instead of
> > +
> > +        _21 = vect_vec_iv_.6_22 + { VF, ... };
> > +
> > +      We should generate:
> > +
> > +       _35 = .SELECT_VL (ivtmp_33, VF);
> > +       vect_cst__22 = [vec_duplicate_expr] _35;
> > +       _21 = vect_vec_iv_.6_22 + vect_cst__22;  */
> > +   vec_loop_lens *lens = &LOOP_VINFO_LENS (loop_vinfo);
> > +   tree len
> > +     = vect_get_loop_len (loop_vinfo, NULL, lens, 1, vectype, 0, 0);
> > +   expr = force_gimple_operand (fold_convert (TREE_TYPE (step_expr),
> > +      unshare_expr (len)),
> > +        &seq, true, NULL_TREE);
> >   }
> >        else
> > - expr = build_int_cst (TREE_TYPE (step_expr), vf);
> > + {
> > +   bool float_p = SCALAR_FLOAT_TYPE_P (TREE_TYPE (step_expr));
> > +   expr = build_int_cst (float_p ? integer_type_node
> > + : TREE_TYPE (step_expr),
> > + vf);
> > +   if (float_p)
> > +     expr = gimple_build (&seq, FLOAT_EXPR, TREE_TYPE (step_expr), expr);
> > + }
> > +
>  
> I meant you keep the existing flow in the function, specifically
> I think you should handle SCALAR_FLOAT_TYPE_P like it was previously
> handled, just build 'vf' in the dynamic way.
>  
> >        new_name = gimple_build (&seq, MULT_EXPR, TREE_TYPE (step_expr),
> >          expr, step_expr);
> >        if (seq)
> >   {
> > -   new_bb = gsi_insert_seq_on_edge_immediate (pe, seq);
> > -   gcc_assert (!new_bb);
> > +   if (LOOP_VINFO_USING_SELECT_VL_P (loop_vinfo))
> > +     gsi_insert_seq_before (&si, seq, GSI_SAME_STMT);
> > +   else
> > +     {
> > +       new_bb = gsi_insert_seq_on_edge_immediate (pe, seq);
> > +       gcc_assert (!new_bb);
> > +     }
> >   }
> >      }
> >  
> > @@ -10335,9 +10363,9 @@ vectorizable_induction (loop_vec_info loop_vinfo,
> >    gcc_assert (CONSTANT_CLASS_P (new_name)
> >         || TREE_CODE (new_name) == SSA_NAME);
> >    new_vec = build_vector_from_val (step_vectype, t);
> > -  vec_step = vect_init_vector (loop_vinfo, stmt_info,
> > -        new_vec, step_vectype, NULL);
> > -
> > +  vec_step
> > +    = vect_init_vector (loop_vinfo, stmt_info, new_vec, step_vectype,
> > + LOOP_VINFO_USING_SELECT_VL_P (loop_vinfo) ? &si : NULL);
>  
> again this makes the flow hard to follow.  I suppose refactoring this
> overall to
>  
>     if (nested_in_vect_loop)
> ...
>     else if (LOOP_VINFO_USING_SELECT_VL_P (..))
> ...
>     else
> ...
>  
> and duplicate this tail into the cases makes it easier to follow.
>  
> For nested_in_vect_loop we never have LOOP_VINFO_USING_SELECT_VL_P?
>  
> Richard.
>  
>  
> > Thanks.
> > 
> > 
> > juzhe.zhong@rivai.ai
> >  
> > From: Richard Biener
> > Date: 2023-11-09 20:16
> > To: Juzhe-Zhong
> > CC: gcc-patches; richard.sandiford; rguenther; kito.cheng; kito.cheng
> > Subject: Re: [PATCH] Middle-end: Fix bug of induction variable vectorization for RVV
> > On Wed, Nov 8, 2023 at 11:53?AM Juzhe-Zhong <juzhe.zhong@rivai.ai> wrote:
> > >
> > > PR: https://gcc.gnu.org/bugzilla/show_bug.cgi?id=112438
> > >
> > > SELECT_VL result is not necessary always VF in non-final iteration.
> > >
> > > Current GIMPLE IR is wrong:
> > >
> > > # vect_vec_iv_.21_25 = PHI <_24(4), { 0, 1, 2, ... }(3)>
> > > ...
> > > _24 = vect_vec_iv_.21_25 + { POLY_INT_CST [4, 4], ... };
> > >
> > > After this patch which is correct for SELECT_VL:
> > >
> > > # vect_vec_iv_.8_22 = PHI <_21(4), { 0, 1, 2, ... }(3)>
> > > ...
> > > _35 = .SELECT_VL (ivtmp_33, POLY_INT_CST [4, 4]);
> > > _21 = vect_vec_iv_.8_22 + { POLY_INT_CST [4, 4], ... };
> > >
> > > kito, could you give more explanation ?
> > >
> > >         PR middle/112438
> > >
> > > gcc/ChangeLog:
> > >
> > >         * tree-vect-loop.cc (vectorizable_induction): Fix bug.
> > >
> > > gcc/testsuite/ChangeLog:
> > >
> > >         * gcc.target/riscv/rvv/autovec/pr112438.c: New test.
> > >
> > > ---
> > >  .../gcc.target/riscv/rvv/autovec/pr112438.c   | 35 +++++++++++++++++
> > >  gcc/tree-vect-loop.cc                         | 39 +++++++++++++++----
> > >  2 files changed, 67 insertions(+), 7 deletions(-)
> > >  create mode 100644 gcc/testsuite/gcc.target/riscv/rvv/autovec/pr112438.c
> > >
> > > diff --git a/gcc/testsuite/gcc.target/riscv/rvv/autovec/pr112438.c b/gcc/testsuite/gcc.target/riscv/rvv/autovec/pr112438.c
> > > new file mode 100644
> > > index 00000000000..b326d56a52c
> > > --- /dev/null
> > > +++ b/gcc/testsuite/gcc.target/riscv/rvv/autovec/pr112438.c
> > > @@ -0,0 +1,35 @@
> > > +/* { dg-do compile } */
> > > +/* { dg-options "-march=rv64gcv -mabi=lp64d -O3 -fno-vect-cost-model -ffast-math -fdump-tree-optimized-details" } */
> > > +
> > > +void
> > > +foo (int n, int *__restrict in, int *__restrict out)
> > > +{
> > > +  for (int i = 0; i < n; i += 1)
> > > +    {
> > > +      out[i] = in[i] + i;
> > > +    }
> > > +}
> > > +
> > > +void
> > > +foo2 (int n, float * __restrict in,
> > > +float * __restrict out)
> > > +{
> > > +  for (int i = 0; i < n; i += 1)
> > > +    {
> > > +      out[i] = in[i] + i;
> > > +    }
> > > +}
> > > +
> > > +void
> > > +foo3 (int n, float * __restrict in,
> > > +float * __restrict out, float x)
> > > +{
> > > +  for (int i = 0; i < n; i += 1)
> > > +    {
> > > +      out[i] = in[i] + i* i;
> > > +    }
> > > +}
> > > +
> > > +/* We don't want to see vect_vec_iv_.21_25 + { POLY_INT_CST [4, 4], ... }.  */
> > > +/* { dg-final { scan-tree-dump-not "\\+ \{ POLY_INT_CST" "optimized" } } */
> > > +
> > > diff --git a/gcc/tree-vect-loop.cc b/gcc/tree-vect-loop.cc
> > > index a544bc9b059..3e103946168 100644
> > > --- a/gcc/tree-vect-loop.cc
> > > +++ b/gcc/tree-vect-loop.cc
> > > @@ -10309,10 +10309,30 @@ vectorizable_induction (loop_vec_info loop_vinfo,
> > >      new_name = step_expr;
> > >    else
> > >      {
> > > +      gimple_seq seq = NULL;
> > > +      if (LOOP_VINFO_USING_SELECT_VL_P (loop_vinfo))
> > > +       {
> > > +         /* When we're using loop_len produced by SELEC_VL, the non-final
> > > +            iterations are not always processing VF elements.  So vectorize
> > > +            induction variable instead of
> > > +
> > > +              _21 = vect_vec_iv_.6_22 + { VF, ... };
> > > +
> > > +            We should generate:
> > > +
> > > +              _35 = .SELECT_VL (ivtmp_33, VF);
> > > +              vect_cst__22 = [vec_duplicate_expr] _35;
> > > +              _21 = vect_vec_iv_.6_22 + vect_cst__22;  */
> > > +         vec_loop_lens *lens = &LOOP_VINFO_LENS (loop_vinfo);
> > > +         tree len
> > > +           = vect_get_loop_len (loop_vinfo, NULL, lens, 1, vectype, 0, 0);
> > > +         expr = force_gimple_operand (fold_convert (TREE_TYPE (step_expr),
> > > +                                                    unshare_expr (len)),
> > > +                                      &seq, true, NULL_TREE);
> > > +       }
> >  
> > I think it would be better to split out building a tree from VF from both
> > arms and avoid using 'vf' when LOOP_VINFO_USING_SELECT_VL_P.
> >  
> > Btw, you are not patching the SLP path here which I believe has the same
> > problem but is currently exempt from non-constant VF at least.
> >  
> > Richard.
> >  
> > >        /* iv_loop is the loop to be vectorized. Generate:
> > >           vec_step = [VF*S, VF*S, VF*S, VF*S]  */
> > > -      gimple_seq seq = NULL;
> > > -      if (SCALAR_FLOAT_TYPE_P (TREE_TYPE (step_expr)))
> > > +      else if (SCALAR_FLOAT_TYPE_P (TREE_TYPE (step_expr)))
> > >         {
> > >           expr = build_int_cst (integer_type_node, vf);
> > >           expr = gimple_build (&seq, FLOAT_EXPR, TREE_TYPE (step_expr), expr);
> > > @@ -10323,8 +10343,13 @@ vectorizable_induction (loop_vec_info loop_vinfo,
> > >                                expr, step_expr);
> > >        if (seq)
> > >         {
> > > -         new_bb = gsi_insert_seq_on_edge_immediate (pe, seq);
> > > -         gcc_assert (!new_bb);
> > > +         if (LOOP_VINFO_USING_SELECT_VL_P (loop_vinfo))
> > > +           gsi_insert_seq_before (&si, seq, GSI_SAME_STMT);
> > > +         else
> > > +           {
> > > +             new_bb = gsi_insert_seq_on_edge_immediate (pe, seq);
> > > +             gcc_assert (!new_bb);
> > > +           }
> > >         }
> > >      }
> > >
> > > @@ -10332,9 +10357,9 @@ vectorizable_induction (loop_vec_info loop_vinfo,
> > >    gcc_assert (CONSTANT_CLASS_P (new_name)
> > >               || TREE_CODE (new_name) == SSA_NAME);
> > >    new_vec = build_vector_from_val (step_vectype, t);
> > > -  vec_step = vect_init_vector (loop_vinfo, stmt_info,
> > > -                              new_vec, step_vectype, NULL);
> > > -
> > > +  vec_step
> > > +    = vect_init_vector (loop_vinfo, stmt_info, new_vec, step_vectype,
> > > +                       LOOP_VINFO_USING_SELECT_VL_P (loop_vinfo) ? &si : NULL);
> > >
> > >    /* Create the following def-use cycle:
> > >       loop prolog:
> > > --
> > > 2.36.3
> > >
> >  
> > 
>  
> 
 
-- 
Richard Biener <rguenther@suse.de>
SUSE Software Solutions Germany GmbH,
Frankenstrasse 146, 90461 Nuernberg, Germany;
GF: Ivo Totev, Andrew McDonald, Werner Knoblich; (HRB 36809, AG Nuernberg)