[gcc r14-2156] tree-optimization/110443 - prevent SLP splat of gathers

[gcc r14-2156] tree-optimization/110443 - prevent SLP splat of gathers

[Richard Biener]

https://gcc.gnu.org/g:51c8cbc6bba387f953d9be48c4a4c8b657dd54a5

commit r14-2156-g51c8cbc6bba387f953d9be48c4a4c8b657dd54a5
Author: Richard Biener <rguenther@suse.de>
Date:   Wed Jun 28 10:16:57 2023 +0200

    tree-optimization/110443 - prevent SLP splat of gathers
    
    The following prevents non-grouped load SLP in case the element
    to splat is from a gather operation.  While it should be possible
    to support this it is not similar to the single element interleaving
    case I was trying to mimic here.
    
            PR tree-optimization/110443
            * tree-vect-slp.cc (vect_build_slp_tree_1): Reject non-grouped
            gather loads.
    
            * gcc.dg/torture/pr110443.c: New testcase.

Diff:
---
 gcc/testsuite/gcc.dg/torture/pr110443.c | 21 +++++++++++++++++++++
 gcc/tree-vect-slp.cc                    |  3 ++-
 2 files changed, 23 insertions(+), 1 deletion(-)

diff --git a/gcc/testsuite/gcc.dg/torture/pr110443.c b/gcc/testsuite/gcc.dg/torture/pr110443.c
new file mode 100644
index 00000000000..61cf705869c
--- /dev/null
+++ b/gcc/testsuite/gcc.dg/torture/pr110443.c
@@ -0,0 +1,21 @@
+/* { dg-do compile } */
+
+typedef struct {
+  float real;
+  float imag;
+} complex_t;
+extern unsigned char fftorder[];
+float *a52_imdct_256_data;
+int a52_imdct_256_i, a52_imdct_256_k;
+float a52_imdct_256_b_r;
+void a52_imdct_256()
+{
+  complex_t buf1[64];
+  a52_imdct_256_i = 0;
+  for (; a52_imdct_256_i < 64; a52_imdct_256_i++) {
+    a52_imdct_256_k = fftorder[a52_imdct_256_i];
+    buf1[a52_imdct_256_i].real = buf1[a52_imdct_256_i].imag =
+        a52_imdct_256_data[a52_imdct_256_k];
+  }
+  a52_imdct_256_b_r = buf1[0].real * buf1[0].imag;
+}
diff --git a/gcc/tree-vect-slp.cc b/gcc/tree-vect-slp.cc
index 8cb1ac1f319..355d078d66e 100644
--- a/gcc/tree-vect-slp.cc
+++ b/gcc/tree-vect-slp.cc
@@ -1291,7 +1291,8 @@ vect_build_slp_tree_1 (vec_info *vinfo, unsigned char *swap,
 		 vectorization.  For loop vectorization we can handle
 		 splats the same we handle single element interleaving.  */
 	      && (is_a <bb_vec_info> (vinfo)
-		  || stmt_info != first_stmt_info))
+		  || stmt_info != first_stmt_info
+		  || STMT_VINFO_GATHER_SCATTER_P (stmt_info)))
 	    {
 	      /* Not grouped load.  */
 	      if (dump_enabled_p ())