http://gcc.gnu.org/bugzilla/show_bug.cgi?id=52198
Bug #: 52198
Summary: SLP vectorization does not consider swapped operands
consistently
Classification: Unclassified
Product: gcc
Version: 4.7.0
Status: UNCONFIRMED
Keywords: missed-optimization
Severity: normal
Priority: P3
Component: tree-optimization
AssignedTo: [email protected]
ReportedBy: [email protected]
While vect_get_and_check_slp_defs swaps operands of binary operations in some
cases it does not do so when vect_build_slp_tree would reject the tree because
the operands have different operations in def stmts, for example with
D.1751_23 = s_3(D) * D.1737_13;
D.1753_25 = s_3(D) * D.1734_16;
D.1755_27 = s_3(D) * D.1731_19;
D.1757_29 = s_3(D) * D.1728_22;
D.1777_7 = MEM[(const struct LorentzVector &)res_4(D)].theT;
D.1775_31 = D.1777_7 + D.1751_23;
D.1774_32 = MEM[(const struct LorentzVector &)res_4(D)].theZ;
D.1772_34 = D.1753_25 + D.1774_32;
D.1771_35 = MEM[(const struct LorentzVector &)res_4(D)].theY;
D.1769_37 = D.1755_27 + D.1771_35;
D.1768_38 = MEM[(const struct LorentzVector &)res_4(D)].theX;
D.1766_40 = D.1757_29 + D.1768_38;
res_4(D)->theX = D.1766_40;
res_4(D)->theY = D.1769_37;
res_4(D)->theZ = D.1772_34;
res_4(D)->theT = D.1775_31;
where you can see that the first addition stmt has its operands swapped
(because of SSA name numbering contributing to canonicalization here).
This issue triggers in g++.dg/vect/slp-pr50819.cc if you install a patch
to properly release virtual operands of the call stmt during inlining
which frees up that SSA name for re-use:
Index: gcc/tree-inline.c
===================================================================
--- gcc/tree-inline.c (revision 184088)
+++ gcc/tree-inline.c (working copy)
@@ -4002,6 +4002,9 @@ expand_call_inline (basic_block bb, gimp
/* Unlink the calls virtual operands before replacing it. */
unlink_stmt_vdef (stmt);
+ if (gimple_vdef (stmt)
+ && TREE_CODE (gimple_vdef (stmt)) == SSA_NAME)
+ release_ssa_name (gimple_vdef (stmt));
/* If the inlined function returns a result that we care about,
substitute the GIMPLE_CALL with an assignment of the return
