Hi Sun,
Thank you for the reminder.
Zhou,Feng suggested that we can add the comments into the mu-chi
construction part after this "restricted checkin" period, and he also filed
another bug to keep track of this.
Can we do that later in the future?
Best Regards,
Ann
On Tue, Mar 15, 2011 at 4:10 PM, Sun Chan <sun.c...@gmail.com> wrote:
> and pls put the log message in the code as comments. It will be more
> obvious for people reading code.
> Sun
>
> On Wed, Mar 16, 2011 at 7:06 AM, Suneel Jain <suneel.j...@gmail.com>
> wrote:
> > Please put the reviewer's name in the checkin message.
> > And since we are in a restricted checkin mode (only defect
> > fixes), please include the Bugzilla defect number as well.
> >
> > Also, please add you name to the open64/MAINTAINERS file.
> >
> > - Suneel
> >
> >
> > On Tue, Mar 15, 2011 at 3:44 PM, <s...@open64.net> wrote:
> >> Author: xan
> >> Date: 2011-03-15 18:44:18 -0400 (Tue, 15 Mar 2011)
> >> New Revision: 3519
> >>
> >> Modified:
> >> trunk/osprey/be/opt/opt_dse.cxx
> >> Log:
> >> delete the enhancement of dse which is added from open64-4.2.1, since
> >> the enhancement is not aware of the fact that there is "implied mus"
> >> for all the variables that are aliased with the variable of the mu node.
> >> So the chi that is dead because there is no use of its result is not
> >> really dead.
> >>
> >>
> >>
> >>
> >> Modified: trunk/osprey/be/opt/opt_dse.cxx
> >> ===================================================================
> >> --- trunk/osprey/be/opt/opt_dse.cxx 2011-03-14 20:54:07 UTC (rev
> 3518)
> >> +++ trunk/osprey/be/opt/opt_dse.cxx 2011-03-15 22:44:18 UTC (rev
> 3519)
> >> @@ -87,10 +87,6 @@
> >> #include "opt_ssa.h"
> >> #include "opt_mu_chi.h"
> >> #include "opt_util.h"
> >> -#include <vector>
> >> -using std::vector;
> >> -#include "opt_alias_rule.h"
> >> -#include "idx_32_set.h"
> >>
> >> // ====================================================================
> >> // ====================================================================
> >> @@ -108,7 +104,6 @@
> >> #ifdef TARG_SL
> >> vector <WN *> *_injury_aux_intrnop;
> >> #endif
> >> - vector <WN *> *_last_store_vec;
> >>
> >> //
> >> // access methods
> >> @@ -136,7 +131,7 @@
> >> void Set_Required_VSE( VER_STAB_ENTRY *, BOOL , WN * ) const;
> >> void Set_Required_PHI( VER_STAB_ENTRY *vse, WN *ref_wn ) const;
> >> void Set_Required_MU( MU_NODE *mu, BOOL real_use ) const;
> >> - void Set_Required_CHI( CHI_NODE *chi, BOOL *chi_is_live ) const;
> >> + void Set_Required_CHI( CHI_NODE *chi ) const;
> >> void Set_Required_WN( WN *wn ) const;
> >> void Add_EH_exposed_use(WN *call) const;
> >> #ifdef KEY
> >> @@ -153,18 +148,6 @@
> >> void Repair_Injured_AuxIntrnOP(void) const ;
> >> void Append_Injured_AuxIntrnOp(WN *wn) const {
> _injury_aux_intrnop->insert(_injury_aux_intrnop->begin(), wn);};
> >> #endif
> >> - WN* Last_store(AUX_ID aid) const
> >> - { return (*_last_store_vec)[aid]; }
> >> - void Set_last_store( AUX_ID aid, WN *store) const
> >> - { (*_last_store_vec)[aid] = store; }
> >> - //OSP_468, remove all Set_Required_Imp_VSE()
> >> - //void Set_Required_Imp_VSE( VER_ID vid, BOOL real_use) const;
> >> - BOOL Mem_WN_equiv_rec(WN *wn1, WN *wn2) const;
> >> - BOOL Mem_WN_equiv(WN *wn1, WN* wn2) const;
> >> - BOOL Same_memloc( WN* store1, WN* store2) const;
> >> - VER_ID Prop_vsym_new_result( VER_ID vid ) const;
> >> - void Propagate_vsym_wn( WN *wn ) const;
> >> - void Propagate_vsym_bb( BB_NODE *bb ) const;
> >>
> >> public:
> >>
> >> @@ -178,11 +161,6 @@
> >> _injury_aux_intrnop = CXX_NEW(vector<WN *>, pool);
> >> #endif
> >>
> >> - //init _last_store_vec
> >> - INT asym_count = Opt_stab()->Lastidx() + 1;
> >> - _last_store_vec = CXX_NEW(vector<WN *>, pool);
> >> - _last_store_vec->insert(_last_store_vec->end(), asym_count,
> (WN*)NULL);
> >> -
> >> }
> >>
> >> ~DSE( void )
> >> @@ -422,8 +400,8 @@
> >> // we only need to propagate this usage if the symbol
> >> // was not already marked as having a use
> >> if ( vse->Any_use() ) return;
> >> -
> >> - BOOL vse_live=TRUE;
> >> + vse->Set_Any_use();
> >> +
> >> // now recursively follow this symbol's use-def chain
> >> switch ( vse->Type() ) {
> >> case WHIRL_STMT:
> >> @@ -437,7 +415,7 @@
> >> Set_Required_PHI( vse, ref_wn );
> >> break;
> >> case CHI_STMT:
> >> - Set_Required_CHI( vse->Chi(), &vse_live );
> >> + Set_Required_CHI( vse->Chi() );
> >> break;
> >> case ENTRY_STMT:
> >> // no definition, value is just live-in to the region
> >> @@ -452,12 +430,8 @@
> >> }
> >>
> >> if ( Tracing() ) {
> >> - if(vse_live)
> >> - fprintf( TFile, "<dse> Required VSE: var:%d version:%d\n",
> >> + fprintf( TFile, "<dse> Required VSE: var:%d version:%d\n",
> >> vse->Aux_id(), vse->Version() );
> >> - else
> >> - fprintf( TFile, "<dse> Not Required VSE: var:%d version:%d\n",
> >> - vse->Aux_id(), vse->Version() );
> >> }
> >> }
> >>
> >> @@ -482,7 +456,6 @@
> >> for ( INT32 opndnum = 0; opndnum < phi->Size(); opndnum++ ) {
> >> VER_ID phi_opnd = phi->Opnd(opndnum);
> >> VER_STAB_ENTRY *sym = Opt_stab()->Ver_stab_entry(phi_opnd);
> >> - Set_last_store( sym->Aux_id(), NULL);
> >> Set_Required_VSE( sym, FALSE, ref_wn );
> >> }
> >> }
> >> @@ -496,7 +469,6 @@
> >> DSE::Set_Required_MU( MU_NODE *mu, BOOL real_use ) const
> >> {
> >> VER_STAB_ENTRY *ver = Opt_stab()->Ver_stab_entry(mu->Opnd());
> >> - Set_last_store(ver->Aux_id(), NULL);
> >> Set_Required_VSE( ver, real_use, NULL );
> >> }
> >>
> >> @@ -515,110 +487,13 @@
> >> opt_stab->Ver_stab_entry(WN_ver(rhs))->Aux_id());
> >> }
> >>
> >> -BOOL
> >> -DSE::Mem_WN_equiv_rec(WN *wn1, WN *wn2) const
> >> -{
> >> - if (!wn1 || !wn2) return FALSE;
> >> - if (!Mem_WN_equiv(wn1,wn2)) {
> >> - return FALSE;
> >> - }
> >> - for (INT i=0; i<WN_kid_count(wn1); i++) {
> >> - if (!Mem_WN_equiv_rec(WN_kid(wn1,i),WN_kid(wn2,i))) {
> >> - return FALSE;
> >> - }
> >> - }
> >> - return TRUE;
> >> -}
> >> -
> >> -BOOL
> >> -DSE::Mem_WN_equiv(WN *wn1, WN *wn2) const
> >> -{
> >> - if (!WN_Equiv(wn1, wn2))
> >> - return FALSE;
> >> -
> >> - OPERATOR opr = WN_operator(wn1);
> >> - if(OPERATOR_has_field_id(opr)) {
> >> - if(WN_field_id(wn1) != WN_field_id(wn2))
> >> - return FALSE;
> >> - }
> >> -
> >> - if(opr == OPR_STBITS || opr == OPR_ISTBITS || opr == OPR_LDBITS ||
> opr == OPR_ILDBITS) {
> >> - if( WN_bit_offset(wn1) != WN_bit_offset(wn2))
> >> - return FALSE;
> >> - if( WN_bit_size(wn1) != WN_bit_size(wn2))
> >> - return FALSE;
> >> - }
> >> -
> >> - if (WN_has_mu(wn1, Cfg()->Rgn_level())) {
> >> - MU_NODE *mu1 = Opt_stab()->Get_occ(wn1)->Mem_mu_node();
> >> - MU_NODE *mu2 = Opt_stab()->Get_occ(wn2)->Mem_mu_node();
> >> - if(mu1 != NULL && mu2 != NULL && mu1->Opnd() != mu2->Opnd() )
> >> - return FALSE;
> >> - if((mu1==NULL) != (mu2 == NULL))
> >> - return FALSE;
> >> - }
> >> -
> >> - return TRUE;
> >> -}
> >> -
> >> -BOOL
> >> -DSE::Same_memloc( WN* store1, WN* store2) const
> >> -{
> >> - FmtAssert(OPERATOR_is_scalar_istore(WN_operator(store1)) ||
> >> - OPERATOR_is_scalar_store (WN_operator(store1)) ||
> >> - WN_operator(store1) == OPR_MSTORE,
> >> - ("DSE::Same_memloc: store1 is not istore"));
> >> - FmtAssert(OPERATOR_is_scalar_istore(WN_operator(store2)) ||
> >> - OPERATOR_is_scalar_store (WN_operator(store2)) ||
> >> - WN_operator(store2) == OPR_MSTORE,
> >> - ("DSE::Same_memloc: store2 is not istore"));
> >> -
> >> - OCC_TAB_ENTRY *occ1 = Opt_stab()->Get_occ(store1);
> >> - OCC_TAB_ENTRY *occ2 = Opt_stab()->Get_occ(store2);
> >> - FmtAssert(occ1 != NULL && occ2 != NULL, ("DSE::Same_memloc: occ ==
> NULL"));
> >> - POINTS_TO *pt1 = occ1->Points_to();
> >> - POINTS_TO *pt2 = occ2->Points_to();
> >> - FmtAssert(pt1 != NULL && pt2 != NULL, ("DSE::Same_memloc: points_to
> == NULL"));
> >> -
> >> - if (Opt_stab()->Rule()->Same_location(store1, store2, pt1, pt2)) {
> >> -#if defined(TARG_NVISA)
> >> - //for dynamic array, we need to be more conservative
> >> - INT i;
> >> - TY_IDX ty;
> >> - ty = ST_type(pt1->Base());
> >> - if (TY_kind(ty) == KIND_ARRAY) {
> >> - for (i = 0; i < TY_AR_ndims(ty); i++) {
> >> - if ( !TY_AR_const_lbnd(ty, i) ||
> >> - !TY_AR_const_ubnd(ty, i) )
> >> - return FALSE;
> >> - }
> >> - }
> >> -
> >> - ty = ST_type(pt2->Base());
> >> - if (TY_kind(ty) == KIND_ARRAY) {
> >> - for (i = 0; i < TY_AR_ndims(ty); i++) {
> >> - if ( !TY_AR_const_lbnd(ty, i) ||
> >> - !TY_AR_const_ubnd(ty, i) )
> >> - return FALSE;
> >> - }
> >> - }
> >> -#endif
> >> - return TRUE;
> >> - }
> >> - else if (Mem_WN_equiv(store1, store2)) {
> >> - if (WN_kid_count(store1)>1 && Mem_WN_equiv_rec(WN_kid1(store1),
> WN_kid1(store2)))
> >> - return TRUE;
> >> - }
> >> - return FALSE;
> >> -}
> >> -
> >> // ====================================================================
> >> // Recursively set the flags that say this chi statement and the
> >> // variables it references are used.
> >> // ====================================================================
> >>
> >> void
> >> -DSE::Set_Required_CHI( CHI_NODE *chi, BOOL *chi_is_live ) const
> >> +DSE::Set_Required_CHI( CHI_NODE *chi ) const
> >> {
> >> AUX_ID vaux = chi->Aux_id();
> >> BOOL real_use = FALSE;
> >> @@ -642,37 +517,8 @@
> >> Opt_stab()->Ver_stab_entry(chi->Result())->Set_Real_use();
> >> }
> >>
> >> - *chi_is_live = TRUE;
> >> - if (OPERATOR_is_scalar_istore(WN_operator(chiwn)) ||
> >> - OPERATOR_is_scalar_store(WN_operator(chiwn)) ||
> >> - WN_operator(chiwn) == OPR_MSTORE) {
> >> - WN *last_st = Last_store(vaux);
> >> - if( last_st != NULL && chiwn != last_st ) {
> >> - if( Same_memloc(chiwn, last_st) ) {
> >> - *chi_is_live = FALSE;
> >> - if ( Tracing() ) {
> >> - fprintf ( TFile, "DSE::Set_Required_CHI, current chiwn is not
> set live:\n" );
> >> - fdump_tree_no_st(TFile, chiwn);
> >> - }
> >> - }
> >> - }
> >> - if(*chi_is_live)
> >> - Set_last_store(vaux, chiwn);
> >> - }
> >> + Set_Required_WN(chiwn);
> >>
> >> - if(*chi_is_live) {
> >> - VER_STAB_ENTRY *vsym = Opt_stab()->Ver_stab_entry(chi->Result());
> >> - vsym->Set_Any_use();
> >> - Set_Required_WN(chiwn);
> >> - }
> >> -
> >> - if (OPERATOR_is_scalar_istore(WN_operator(chiwn)) ||
> >> - OPERATOR_is_scalar_store(WN_operator(chiwn)) ||
> >> - WN_operator(chiwn) == OPR_MSTORE) {
> >> - if(*chi_is_live)
> >> - Set_last_store(vaux, chiwn);
> >> - }
> >> -
> >> // The following breaks the use-def chain. The definition
> >> // of the chi operand can become dse-dead. It violates assertions
> >> // in opt_verify.cxx. We can change the verifier to check this
> >> @@ -732,11 +578,6 @@
> >>
> >> if ( WN_has_ver(wn) ) {
> >> VER_STAB_ENTRY *sym = Opt_stab()->Ver_stab_entry(WN_ver(wn));
> >> - if(OPERATOR_is_scalar_load( WN_operator(wn) )) {
> >> - //OSP_468
> >> - //Set_Required_Imp_VSE(WN_ver(wn), TRUE);
> >> - Set_last_store(sym->Aux_id(), NULL);
> >> - }
> >> Set_Required_VSE( sym, TRUE, wn );
> >> }
> >>
> >> @@ -877,130 +718,6 @@
> >> return FALSE;
> >> }
> >>
> >> -
> >> -VER_ID
> >> -DSE::Prop_vsym_new_result( VER_ID vid ) const
> >> -{
> >> - VER_STAB_ENTRY *vse = Opt_stab()->Ver_stab_entry(vid);
> >> - if ( vse->Type() == PHI_STMT ) {
> >> - // we assume that the result is correct
> >> - return vse->Phi()->Result();
> >> - }
> >> - else if ( vse->Type() == CHI_STMT ) {
> >> - // is this statement live?
> >> - if (vse->Chi()->Live()) {
> >> - return vse->Chi()->Result();
> >> - }
> >> - else {
> >> - return Prop_vsym_new_result(vse->Chi()->Opnd());
> >> - }
> >> - }
> >> - else {
> >> - return vid;
> >> - }
> >> -}
> >> -
> >> -void
> >> -DSE::Propagate_vsym_wn( WN *wn ) const
> >> -{
> >> - if ( WN_has_ver(wn) ) {
> >> - WN_set_ver(wn, Prop_vsym_new_result(WN_ver(wn)));
> >> - }
> >> -
> >> - if ( WN_has_mu(wn, Cfg()->Rgn_level()) ) {
> >> - // process the mu operand as use
> >> - OCC_TAB_ENTRY *occ = Opt_stab()->Get_occ(wn);
> >> - MU_NODE *mu = occ->Mem_mu_node();
> >> - mu->Set_opnd( Prop_vsym_new_result(mu->Opnd()) );
> >> - }
> >> - return;
> >> -}
> >> -
> >> -// ====================================================================
> >> -// After we've deleted some globals that had chis with a virtual
> >> -// variable in them, we need to go back and fix up the use-def chain
> >> -// to be correct.
> >> -//
> >> -// This involves "skipping" over dead statements and updating the
> >> -// references. The algorithm does update the dead statements as well
> >> -// however, so each reference only has to look at its definition to see
> >> -// what the propagated value is.
> >> -//
> >> -// Do this in dominator-tree order, so it must be called for the first
> >> -// time with the entry-bb that dominates all blocks.
> >> -// ====================================================================
> >> -
> >> -void
> >> -DSE::Propagate_vsym_bb( BB_NODE *bb ) const
> >> -{
> >> - // propagate into the phi-nodes, if any
> >> - PHI_LIST_ITER phi_iter;
> >> - PHI_NODE *phi;
> >> - FOR_ALL_ELEM ( phi, phi_iter, Init(bb->Phi_list()) ) {
> >> - if (phi->Live() ) {
> >> - for ( INT pkid = 0; pkid < phi->Size(); pkid++ ) {
> >> - phi->Set_opnd(pkid, Prop_vsym_new_result(phi->Opnd(pkid)));
> >> - }
> >> - }
> >> - }
> >> -
> >> - // handle the block's statements, INCLUDING non-live ones
> >> - STMT_ITER stmt_iter;
> >> - WN *stmt;
> >> - FOR_ALL_ELEM(stmt, stmt_iter, Init(bb->Firststmt(),bb->Laststmt())) {
> >> - // handle mu references only on live statements
> >> - if ( !Is_deleted_statement(stmt) ) {
> >> - // handle the statement's mu list
> >> - if ( WN_has_mu(stmt, Cfg()->Rgn_level()) ) {
> >> - MU_LIST *mu_list = _opt_stab->Get_stmt_mu_list(stmt);
> >> - if (mu_list) {
> >> - MU_LIST_ITER mu_iter;
> >> - MU_NODE *mu;
> >> - FOR_ALL_NODE( mu, mu_iter, Init(mu_list)) {
> >> - mu->Set_opnd( Prop_vsym_new_result(mu->Opnd()) );
> >> - }
> >> - }
> >> - }
> >> -
> >> - for ( INT32 kidnum = 0; kidnum < WN_kid_count(stmt); kidnum++ ) {
> >> - Propagate_vsym_wn( WN_kid(stmt, kidnum) );
> >> - }
> >> -
> >> - // need to handle all chi statements, dead or not
> >> - // handle the statement's chi list
> >> - if ( WN_has_chi(stmt, Cfg()->Rgn_level())) {
> >> - CHI_LIST *chi_list = _opt_stab->Get_generic_chi_list(stmt);
> >> - if (chi_list) {
> >> - CHI_LIST_ITER chi_iter;
> >> - CHI_NODE *chi;
> >> - FOR_ALL_NODE( chi, chi_iter, Init(chi_list)) {
> >> - // propagate into the chi node's operand
> >> - //
> >> - // only do this for a non-dead chi; otherwise, we can have
> >> - // live-range overlap when we resurrenct this dead chi:
> >> - // resurrecting only updates along the def chain, not the
> >> - // use chain.
> >> - // not updating the dead chi here is ok since all the uses
> >> - // below this should be dead (or we have an overlapped
> >> - // live range already).
> >> - //
> >> - if (chi->Live())
> >> - chi->Set_opnd( Prop_vsym_new_result( chi->Opnd() ));
> >> - }
> >> - }
> >> - }
> >> - }
> >> - }
> >> -
> >> - // do copy propagation for this block's dominated nodes
> >> - BB_NODE *dom_bb;
> >> - BB_LIST_ITER dom_bb_iter;
> >> - FOR_ALL_ELEM(dom_bb, dom_bb_iter, Init(bb->Dom_bbs())) {
> >> - Propagate_vsym_bb(dom_bb);
> >> - }
> >> -}
> >> -
> >> -
> >> #if defined(TARG_SL)
> >> /* result = intrnsic_slave_op( master, ar2,ar3...)
> >> * the master is the result of master intrnsic op
> >> @@ -1192,8 +909,6 @@
> >> }
> >> }
> >>
> >> - Propagate_vsym_bb( Cfg()->Entry_bb() );
> >> -
> >> if ( Tracing() ) {
> >> fprintf ( TFile, "SSA::Dead_store_elim (after dse)\n" );
> >> FOR_ALL_NODE( ssa_id, ver_stab_iter, Init() ) {
> >>
> >>
> >>
> ------------------------------------------------------------------------------
> >> Colocation vs. Managed Hosting
> >> A question and answer guide to determining the best fit
> >> for your organization - today and in the future.
> >> http://p.sf.net/sfu/internap-sfd2d
> >> _______________________________________________
> >> Open64-devel mailing list
> >> Open64-devel@lists.sourceforge.net
> >> https://lists.sourceforge.net/lists/listinfo/open64-devel
> >>
> >
> >
> ------------------------------------------------------------------------------
> > Colocation vs. Managed Hosting
> > A question and answer guide to determining the best fit
> > for your organization - today and in the future.
> > http://p.sf.net/sfu/internap-sfd2d
> > _______________________________________________
> > Open64-devel mailing list
> > Open64-devel@lists.sourceforge.net
> > https://lists.sourceforge.net/lists/listinfo/open64-devel
> >
>
------------------------------------------------------------------------------
Colocation vs. Managed Hosting
A question and answer guide to determining the best fit
for your organization - today and in the future.
http://p.sf.net/sfu/internap-sfd2d
_______________________________________________
Open64-devel mailing list
Open64-devel@lists.sourceforge.net
https://lists.sourceforge.net/lists/listinfo/open64-devel
