[RFC] [autovect patch] Implement vectorization hints

Devang Patel dpatel@apple.com
Fri Mar 25 16:34:00 GMT 2005


Now, I have a patch (it bootstraps on powerpc-darwin, and needs make  
check test & other usual testing) to initialize loop structure early.


     368   /* FIXME: SRA may generate arbitrary gimple code, exposing  
new
     369      aliased and call-clobbered variables.  As mentioned below,
     370      pass_may_alias should be a TODO item.  */
     371   NEXT_PASS (pass_may_alias);
     372   NEXT_PASS (pass_rename_ssa_copies);
     373   NEXT_PASS (pass_loop_init);         /* <--- Initialize loops
     374   NEXT_PASS (pass_dominator);         /* <--- do jump  
threading and maintain loop structure
     375   NEXT_PASS (pass_redundant_phi);
     376   NEXT_PASS (pass_dce);
     377   NEXT_PASS (pass_dse);
     378   NEXT_PASS (pass_may_alias);
     379   NEXT_PASS (pass_forwprop);
     380   NEXT_PASS (pass_phiopt);
     381   NEXT_PASS (pass_ccp);
     382   NEXT_PASS (pass_redundant_phi);
     383   NEXT_PASS (pass_fold_builtins);
     384   /* FIXME: May alias should a TODO but for 4.0.0,
     385      we add may_alias right after fold builtins
     386      which can create arbitrary GIMPLE.  */
     387   NEXT_PASS (pass_may_alias);
     388   NEXT_PASS (pass_split_crit_edges);
     389   NEXT_PASS (pass_pre);
     390   NEXT_PASS (pass_sink_code);
     391   NEXT_PASS (pass_loop);
     392   NEXT_PASS (pass_dominator);
     393   NEXT_PASS (pass_redundant_phi);

There are couple of passes that will need some tweakings before I can  
initialize loop at the very beginning of tree ssa optimization passes  
list.

more later,
-
Devang



More information about the Gcc-patches mailing list