Hi, Could a gatekeeper please review the attached patch.
This patch fixes a problem in load folding optimization where the memory location used in the load operation is overwritten by a store operation. A check is added to prevent folding of such load operations. The test-case and the patch is attached with the mail. Regards, Shivaram AMD Open Source Compiler Engineering
testcase.f
Description: testcase.f
Index: be/cg/x8664/ebo_special.cxx
===================================================================
--- be/cg/x8664/ebo_special.cxx (revision 3725)
+++ be/cg/x8664/ebo_special.cxx (working copy)
@@ -10891,6 +10891,8 @@
#if Is_True_On
if (!(EBO_Opt_Mask & EBO_FOLD_LOAD_DUPLICATE)) return FALSE;
#endif
+ ARC_LIST *true_arcs, *arcs;
+ OPS ops = OPS_EMPTY;
BOOL do_unpckhpd = FALSE;
if (OP_code(op) != TOP_fmovddup && OP_code(op) != TOP_vmovddup)
@@ -10983,7 +10985,41 @@
return FALSE;
if (index_loc >= 0 && !Pred_Opnd_Avail(op, loaded_tn_info, index_loc))
return FALSE;
+ // for this bb, obtain the dependence graph so that we can
+ // walk this op's expression tree to check dependencies.
+ // Make sure that load operand is not overwritten
+ // by a store operation before op using this dependency graph.
+ CG_DEP_Compute_Graph ( op->bb,
+ NO_ASSIGNED_REG_DEPS,
+ NON_CYCLIC,
+ NO_MEMREAD_ARCS,
+ INCLUDE_MEMIN_ARCS,
+ NO_CONTROL_ARCS,
+ NULL);
+
+
+ const INT hash_value = EBO_hash_op( load, NULL );
+ EBO_OP_INFO* opinfo = EBO_opinfo_table[hash_value];
+
+ while( opinfo != NULL ) {
+ OP* next_op = opinfo->in_op;
+ if( next_op == load ) break;
+ if( next_op && OP_store( next_op ) ) {
+ for (arcs = OP_succs(load); arcs != NULL; arcs = ARC_LIST_rest(arcs)) {
+ ARC *arc = ARC_LIST_first(arcs);
+ if (ARC_kind(arc) != CG_DEP_MEMANTI) continue;
+ OP *succ_op = ARC_succ(arc);
+ if ((succ_op == next_op) && OP_Precedes(next_op, op)) {
+ CG_DEP_Delete_Graph (op->bb);
+ return FALSE;
+ }
+ }
+ }
+ opinfo = opinfo->same;
+ }
+ CG_DEP_Delete_Graph (op->bb);
+
TOP topcode;
if (base && offset && index && scale) {
new_op = Mk_OP (TOP_fmovddupxx,
------------------------------------------------------------------------------ EMC VNX: the world's simplest storage, starting under $10K The only unified storage solution that offers unified management Up to 160% more powerful than alternatives and 25% more efficient. Guaranteed. http://p.sf.net/sfu/emc-vnx-dev2dev
_______________________________________________ Open64-devel mailing list [email protected] https://lists.sourceforge.net/lists/listinfo/open64-devel
