http://gcc.gnu.org/bugzilla/show_bug.cgi?id=52091
Ira Rosen <irar at il dot ibm.com> changed: What |Removed |Added ---------------------------------------------------------------------------- CC| |irar at il dot ibm.com --- Comment #5 from Ira Rosen <irar at il dot ibm.com> 2012-02-02 13:22:53 UTC --- (In reply to comment #4) I think that the problem here is that D.2030_19 = prephitmp.17_21 & 65535 is falsely initialized as reduction, while it isn't a reduction. The mistake occurs because its def_stmt, prephitmp.17_21 = PHI <d.3_44(7), pretmp.16_28(2)> , is a phi node marked as double reduction: <bb 3>: # f.6_36 = PHI <f.7_13(7), 0(2)> # prephitmp.17_21 = PHI <d.3_44(7), pretmp.16_28(2)> - double reduction phi ... D.2030_19 = prephitmp.17_21 & 65535; - not a reduction stmt ... <bb 4>: # b.4_41 = PHI <b.5_11(5), 0(3)> ... d.3_9 = d_lsm.23_31 & 1; ... <bb 6>: # d.3_44 = PHI <d.3_9(4)> - double reduction stmt I think we should fail to vectorize D.2030_19 = prephitmp.17_21 & 65535, or any other non-phi/not vect_double_reduction_def stmt with a double reduction phi as a def_stmt. We can either check this in every vectorizable_* for every operand, like this: Index: tree-vect-stmts.c =================================================================== --- tree-vect-stmts.c (revision 183125) +++ tree-vect-stmts.c (working copy) @@ -3326,7 +3326,8 @@ vectorizable_operation (gimple stmt, gimple_stmt_i op0 = gimple_assign_rhs1 (stmt); if (!vect_is_simple_use_1 (op0, loop_vinfo, bb_vinfo, - &def_stmt, &def, &dt[0], &vectype)) + &def_stmt, &def, &dt[0], &vectype) + || dt[0] == vect_double_reduction_def) { if (vect_print_dump_info (REPORT_DETAILS)) fprintf (vect_dump, "use not simple."); or pass stmt or stmt_info to vect_is_simple_use and check it there. > OT, it is strange that we are creating a reduction for a loop which loops > exactly as many times as there are units in the vector, that doesn't seem to > be > profitable. > Right, but doesn't cost model catch this?