| Message ID | 201109241729.26112.ebotcazou@adacore.com |
|---|---|
| State | New |
| Headers | show |
On Sat, Sep 24, 2011 at 5:29 PM, Eric Botcazou <ebotcazou@adacore.com> wrote: >> This is an updated version of the patch. I have 2 new patches and an >> updated testcase which I will sent out individually. >> >> Patch set was bootstrapped and reg-tested on x86_64. >> >> Ok for trunk? >> >> Thanks, >> - Tom >> >> 2011-07-30 Tom de Vries <tom@codesourcery.com> >> >> PR middle-end/43513 >> * Makefile.in (tree-ssa-ccp.o): Add $(PARAMS_H) to rule. >> * tree-ssa-ccp.c (params.h): Include. >> (fold_builtin_alloca_for_var): New function. >> (ccp_fold_stmt): Use fold_builtin_alloca_for_var. > > We have detected another fallout on some Ada code: the transformation replaces > a call to __builtin_alloca with &var, i.e. it introduces an aliased variable, > which invalidates the points-to information of some subsequent call, fooling > DSE into thinking that it can eliminate a live store. Ugh, yeah. I suppose PTA assigned a HEAP var as pointed-to object for the original pointer, even if the transformed stmt orig_ptr_1 = &a; has the points-to information preserved for orig_ptr_1 further propagation of &a will make accesses through orig_ptr_1 have different alias properties. What should work in this special case of a singleton points-to set of orig_ptr_1 (might want to check that) is, do SET_DECL_PT_UID (decl-of-a, DECL_UID (pointed-to orig_ptr_1)); The brute force approach is not acceptable (it'll wreck IPA points-to info). A helper like pt_solution_singleton_p (struct pt_solution *pt, unsigned *uid) whould be nice to have for this. Note that we don't have points-to information computed during the first CCP pass, so the above should be conditional on SSA_NAME_PTR_INFO being present and not ! ->anything (but then assert that we actually do have a singleton, or fail the folding). Richard. > The brute force approach > > Index: tree-ssa-ccp.c > =================================================================== > --- tree-ssa-ccp.c (revision 179038) > +++ tree-ssa-ccp.c (working copy) > @@ -2014,7 +2014,10 @@ do_ssa_ccp (void) > ccp_initialize (); > ssa_propagate (ccp_visit_stmt, ccp_visit_phi_node); > if (ccp_finalize ()) > - return (TODO_cleanup_cfg | TODO_update_ssa | TODO_remove_unused_locals); > + return (TODO_cleanup_cfg > + | TODO_update_ssa > + | TODO_rebuild_alias > + | TODO_remove_unused_locals); > else > return 0; > } > > works, but we might want to be move clever. Thoughts? > > -- > Eric Botcazou >
> Ugh, yeah. I suppose PTA assigned a HEAP var as pointed-to object for the > original pointer, even if the transformed stmt > > orig_ptr_1 = &a; > > has the points-to information preserved for orig_ptr_1 further propagation > of &a will make accesses through orig_ptr_1 have different alias > properties. AFAICS it's an escaping problem: the new variable isn't seen as escaping at some call point. Testcase attached, compile opt22.adb at -O but make sure your tree is up-to-date. > What should work in this special case of a singleton points-to set of > orig_ptr_1 (might want to check that) is, do > > SET_DECL_PT_UID (decl-of-a, DECL_UID (pointed-to orig_ptr_1)); > > The brute force approach is not acceptable (it'll wreck IPA points-to > info). OK, thanks for the tip. 2011-09-26 Eric Botcazou <ebotcazou@adacore.com> * gnat.dg/opt22.adb: New test. * gnat.dg/opt22_pkg.ad[sb]: New helper.
Index: tree-ssa-ccp.c =================================================================== --- tree-ssa-ccp.c (revision 179038) +++ tree-ssa-ccp.c (working copy) @@ -2014,7 +2014,10 @@ do_ssa_ccp (void) ccp_initialize (); ssa_propagate (ccp_visit_stmt, ccp_visit_phi_node); if (ccp_finalize ()) - return (TODO_cleanup_cfg | TODO_update_ssa | TODO_remove_unused_locals); + return (TODO_cleanup_cfg + | TODO_update_ssa + | TODO_rebuild_alias + | TODO_remove_unused_locals); else return 0; }