https://gcc.gnu.org/g:71bf3daa8dabe45aa14e7315195a70ad0d883337
commit r15-3957-g71bf3daa8dabe45aa14e7315195a70ad0d883337
Author: Richard Biener <rguent...@suse.de>
Date: Sat Sep 28 14:02:18 2024 +0200
tree-optimization/116842 - vectorizer load hosting breaks UID order
The following fixes the case when vectorizing a load hoists an invariant
load and dependent stmts, thereby breaking UID order of said stmts.
While we duplicate the load we just move the dependences.
PR tree-optimization/116842
* tree-vect-stmts.cc (hoist_defs_of_uses): Sort stmts to hoist
after UID to avoid breaking vect_stmt_dominates_stmt_p.
* g++.dg/torture/pr116842.C: New testcase.
Diff:
---
gcc/testsuite/g++.dg/torture/pr116842.C | 16 +++++++++++++++
gcc/tree-vect-stmts.cc | 36 ++++++++++++++++++++++-----------
2 files changed, 40 insertions(+), 12 deletions(-)
diff --git a/gcc/testsuite/g++.dg/torture/pr116842.C
b/gcc/testsuite/g++.dg/torture/pr116842.C
new file mode 100644
index 000000000000..39821390046c
--- /dev/null
+++ b/gcc/testsuite/g++.dg/torture/pr116842.C
@@ -0,0 +1,16 @@
+// { dg-do compile }
+
+short a, b, c;
+unsigned d(unsigned, int e) { return e; }
+void f(bool g, short e[][3][3][3][3], unsigned h[][3][3], char i[][8],
+ short j[][18][18][18], short k[][18][18][18], short l[][8][8][8][8])
+{
+ for (char m;;)
+ {
+ for (short n = 0; n < 8; n += 5)
+ a = j[m][6][2][m];
+ for (short o(l[m][m][m][m][m] / i[m][m] ?: e[m][m][4][m][2]); o; o = g)
+ for (char p; p < (c && i[g]) + 7; p += 2)
+ b = d(h[6][g][2], k[m][5][g][2] != m);
+ }
+}
diff --git a/gcc/tree-vect-stmts.cc b/gcc/tree-vect-stmts.cc
index 0e75e3b49567..540a9b733081 100644
--- a/gcc/tree-vect-stmts.cc
+++ b/gcc/tree-vect-stmts.cc
@@ -9788,6 +9788,20 @@ permute_vec_elements (vec_info *vinfo,
return data_ref;
}
+/* Comparator for qsort, sorting after GIMPLE UID. */
+
+static int
+sort_after_uid (const void *a_, const void *b_)
+{
+ const gimple *a = *(const gimple * const *)a_;
+ const gimple *b = *(const gimple * const *)b_;
+ if (gimple_uid (a) < gimple_uid (b))
+ return -1;
+ else if (gimple_uid (a) > gimple_uid (b))
+ return 1;
+ return 0;
+}
+
/* Hoist the definitions of all SSA uses on STMT_INFO out of the loop LOOP,
inserting them on the loops preheader edge. Returns true if we
were successful in doing so (and thus STMT_INFO can be moved then),
@@ -9799,7 +9813,7 @@ hoist_defs_of_uses (stmt_vec_info stmt_info, class loop
*loop, bool hoist_p)
{
ssa_op_iter i;
tree op;
- bool any = false;
+ auto_vec<gimple *, 8> to_hoist;
FOR_EACH_SSA_TREE_OPERAND (op, stmt_info->stmt, i, SSA_OP_USE)
{
@@ -9822,26 +9836,24 @@ hoist_defs_of_uses (stmt_vec_info stmt_info, class loop
*loop, bool hoist_p)
&& flow_bb_inside_loop_p (loop, gimple_bb (def_stmt2)))
return false;
}
- any = true;
+ to_hoist.safe_push (def_stmt);
}
}
- if (!any)
+ if (to_hoist.is_empty ())
return true;
if (!hoist_p)
return true;
- FOR_EACH_SSA_TREE_OPERAND (op, stmt_info->stmt, i, SSA_OP_USE)
+ /* Preserve UID order, otherwise we break dominance checks. */
+ to_hoist.qsort (sort_after_uid);
+
+ for (gimple *def_stmt : to_hoist)
{
- gimple *def_stmt = SSA_NAME_DEF_STMT (op);
- if (!gimple_nop_p (def_stmt)
- && flow_bb_inside_loop_p (loop, gimple_bb (def_stmt)))
- {
- gimple_stmt_iterator gsi = gsi_for_stmt (def_stmt);
- gsi_remove (&gsi, false);
- gsi_insert_on_edge_immediate (loop_preheader_edge (loop), def_stmt);
- }
+ gimple_stmt_iterator gsi = gsi_for_stmt (def_stmt);
+ gsi_remove (&gsi, false);
+ gsi_insert_on_edge_immediate (loop_preheader_edge (loop), def_stmt);
}
return true;
