| Message ID | alpine.LNX.2.00.1201271453250.4999@zhemvz.fhfr.qr |
|---|---|
| State | New |
| Headers | show |
On Fri, 27 Jan 2012, Richard Guenther wrote: > > Quite obvious (just generates dead code) and quite annoys reading > SRA dumps. > > Bootstrapped on x86_64-unknown-linux-gnu, testing in progress. This regresses gcc.dg/tree-ssa/ipa-split-5.c, ipa-split is confused about clobbers on paths where the variable is unused. The following alternative patch makes SRA remove fully scalarized variables (but it still won't remove such dead clobbers if there are no other uses at the time of SRA). Richard. 2012-01-27 Richard Guenther <rguenther@suse.de> * tree-sra.c (sra_modify_assign): Do not transform clobbers. Index: gcc/tree-sra.c =================================================================== *** gcc/tree-sra.c (revision 183616) --- gcc/tree-sra.c (working copy) *************** sra_modify_constructor_assign (gimple *s *** 2817,2822 **** --- 2817,2836 ---- if (!acc) return SRA_AM_NONE; + if (gimple_clobber_p (*stmt)) + { + /* Remove clobbers of fully scalarized variables, otherwise + do nothing. */ + if (acc->grp_covered) + { + unlink_stmt_vdef (*stmt); + gsi_remove (gsi, true); + return SRA_AM_REMOVED; + } + else + return SRA_AM_NONE; + } + loc = gimple_location (*stmt); if (VEC_length (constructor_elt, CONSTRUCTOR_ELTS (gimple_assign_rhs1 (*stmt))) > 0)
Index: gcc/tree-sra.c =================================================================== --- gcc/tree-sra.c (revision 183616) +++ gcc/tree-sra.c (working copy) @@ -2915,11 +2915,12 @@ sra_modify_assign (gimple *stmt, gimple_ location_t loc; gimple_stmt_iterator orig_gsi = *gsi; - if (!gimple_assign_single_p (*stmt)) + if (!gimple_assign_single_p (*stmt) + || gimple_clobber_p (*stmt)) return SRA_AM_NONE; + lhs = gimple_assign_lhs (*stmt); rhs = gimple_assign_rhs1 (*stmt); - if (TREE_CODE (rhs) == CONSTRUCTOR) return sra_modify_constructor_assign (stmt, gsi);