[gcc r11-9584] tree-optimization/103544 - SLP reduction chain as SLP reduction issue

[Richard Biener]

https://gcc.gnu.org/g:d1dc0f6222ecb6eda674e708235c7bc0d90fa987

commit r11-9584-gd1dc0f6222ecb6eda674e708235c7bc0d90fa987
Author: Richard Biener <rguenther@suse.de>
Date:   Mon Dec 6 11:43:28 2021 +0100

    tree-optimization/103544 - SLP reduction chain as SLP reduction issue
    
    When SLP reduction chain vectorization support added handling of
    an outer conversion in the chain picking a failed reduction up
    as SLP reduction that broke the invariant that the whole reduction
    was forward reachable.  The following plugs that hole noting
    a future enhancement possibility.
    
    2021-12-06  Richard Biener  <rguenther@suse.de>
    
            PR tree-optimization/103544
            * tree-vect-slp.c (vect_analyze_slp): Only add a SLP reduction
            opportunity if the stmt in question is the reduction root.
    
            * gcc.dg/vect/pr103544.c: New testcase.
    
    (cherry picked from commit ee01694151edc7e8aef84dc3c484469e2ae443a0)

Diff:
---
 gcc/testsuite/gcc.dg/vect/pr103544.c | 24 ++++++++++++++++++++++++
 gcc/tree-vect-slp.c                  |  9 +++++++--
 2 files changed, 31 insertions(+), 2 deletions(-)

diff --git a/gcc/testsuite/gcc.dg/vect/pr103544.c b/gcc/testsuite/gcc.dg/vect/pr103544.c
new file mode 100644
index 00000000000..c8bdee86e77
--- /dev/null
+++ b/gcc/testsuite/gcc.dg/vect/pr103544.c
@@ -0,0 +1,24 @@
+/* { dg-do compile } */
+/* { dg-additional-options "-O3" } */
+/* { dg-additional-options "-march=haswell" { target x86_64-*-* i?86-*-* } } */
+
+int crash_me(char* ptr, unsigned long size)
+{
+  short result[16] = {0};
+
+  unsigned long no_iters = 0;
+  for(unsigned long i = 0; i < size - 12; i+= 13){
+      for(unsigned long j = 0; j < 12; j++){
+	  result[j] += ptr[i + j] - '0';
+      }
+      no_iters++;
+  }
+
+  int result_int = 0;
+  for(int j = 0; j < 12; j++){
+      int bit_value = result[j] > no_iters/2 ? 1 : 0;
+      result_int |= bit_value;
+  }
+
+  return result_int;
+}
diff --git a/gcc/tree-vect-slp.c b/gcc/tree-vect-slp.c
index 230ff4081a5..ee39527eb69 100644
--- a/gcc/tree-vect-slp.c
+++ b/gcc/tree-vect-slp.c
@@ -2915,8 +2915,13 @@ vect_analyze_slp (vec_info *vinfo, unsigned max_tree_size)
 		vinfo = next;
 	      }
 	    STMT_VINFO_DEF_TYPE (first_element) = vect_internal_def;
-	    /* It can be still vectorized as part of an SLP reduction.  */
-	    loop_vinfo->reductions.safe_push (last);
+	    /* It can be still vectorized as part of an SLP reduction.
+	       ???  But only if we didn't skip a conversion around the group.
+	       In that case we'd have to reverse engineer that conversion
+	       stmt following the chain using reduc_idx and from the PHI
+	       using reduc_def.  */
+	    if (STMT_VINFO_DEF_TYPE (last) == vect_reduction_def)
+	      loop_vinfo->reductions.safe_push (last);
 	  }
 
       /* Find SLP sequences starting from groups of reductions.  */