summaryrefslogtreecommitdiff
path: root/src/loop_unroll.cc
diff options
context:
space:
mode:
Diffstat (limited to 'src/loop_unroll.cc')
-rw-r--r--src/loop_unroll.cc118
1 files changed, 97 insertions, 21 deletions
diff --git a/src/loop_unroll.cc b/src/loop_unroll.cc
index 911d900..86ffd84 100644
--- a/src/loop_unroll.cc
+++ b/src/loop_unroll.cc
@@ -20,6 +20,7 @@ std::set<int> Loop::unroll(int stmt_num, int level, int unroll_amount,
std::vector<std::vector<std::string> > idxNames,
int cleanup_split_level) {
// check for sanity of parameters
+ // check for sanity of parameters
if (unroll_amount < 0)
throw std::invalid_argument(
"invalid unroll amount " + to_string(unroll_amount));
@@ -70,6 +71,16 @@ std::set<int> Loop::unroll(int stmt_num, int level, int unroll_amount,
}
dim2 = stmt[*i].loop_level[dim2].payload;
+ /*if (dv.isCarried(dim2)
+ && (dv.hasNegative(dim2) && !dv.quasi))
+ throw loop_error(
+ "loop error: Unrolling is illegal, dependence violation!");
+
+ if (dv.isCarried(dim2)
+ && (dv.hasPositive(dim2) && dv.quasi))
+ throw loop_error(
+ "loop error: Unrolling is illegal, dependence violation!");
+ */
bool safe = false;
if (dv.isCarried(dim2) && dv.hasPositive(dim2)) {
@@ -78,11 +89,20 @@ std::set<int> Loop::unroll(int stmt_num, int level, int unroll_amount,
"loop error: a quasi dependence with a positive carried distance");
if (!dv.quasi) {
if (dv.lbounds[dim2] != posInfinity) {
+ //if (dv.lbounds[dim2] != negInfinity)
if (dv.lbounds[dim2] > unroll_amount)
safe = true;
} else
safe = true;
- }
+ }/* else {
+ if (dv.ubounds[dim2] != negInfinity) {
+ if (dv.ubounds[dim2] != posInfinity)
+ if ((-(dv.ubounds[dim2])) > unroll_amount)
+ safe = true;
+ } else
+ safe = true;
+ }*/
+
if (!safe) {
for (int l = level + 1; l <= (n - 1) / 2; l++) {
int dim3 = l - 1;
@@ -131,7 +151,7 @@ std::set<int> Loop::unroll(int stmt_num, int level, int unroll_amount,
h.update_coef(mapping.output_var(j), -1);
}
hull = Intersection(hull,
- Range(Restrict_Domain(mapping, copy(stmt[*i].IS))));
+ omega::Range(Restrict_Domain(mapping, copy(stmt[*i].IS))));
hull.simplify(2, 4);
}
@@ -288,7 +308,8 @@ std::set<int> Loop::unroll(int stmt_num, int level, int unroll_amount,
overflow_table[i][j][NULL] = int_mod_hat(
overflow_table[i][j][NULL], unroll_amount);
- // Since we don't have MODULO instruction in SUIF yet (only MOD), make all coef positive in the final formula
+ // Since we don't have MODULO instruction in SUIF yet (only MOD),
+ // make all coef positive in the final formula
for (std::map<Variable_ID, int>::iterator k =
overflow_table[i][j].begin();
k != overflow_table[i][j].end(); k++)
@@ -437,14 +458,16 @@ std::set<int> Loop::unroll(int stmt_num, int level, int unroll_amount,
std::vector<CG_outputRepr *> lb_repr_list, ub_repr_list;
for (int i = 0; i < lb_list.size(); i++) {
lb_repr_list.push_back(
- output_lower_bound_repr(ocg, lb_list[i],
+ output_lower_bound_repr(ocg,
+ lb_list[i],
bound.set_var(dim + 1), result.first, result.second,
bound, Relation::True(bound.n_set()),
std::vector<std::pair<CG_outputRepr *, int> >(
bound.n_set(),
std::make_pair(
static_cast<CG_outputRepr *>(NULL),
- 0))));
+ 0)),
+ uninterpreted_symbols[stmt_num]));
GEQ_Handle h = cond_lower.and_with_GEQ(lb_list[i]);
}
for (int i = 0; i < ub_list.size(); i++) {
@@ -455,21 +478,28 @@ std::set<int> Loop::unroll(int stmt_num, int level, int unroll_amount,
bound.n_set(),
std::make_pair(
static_cast<CG_outputRepr *>(NULL),
- 0))));
+ 0)),
+ uninterpreted_symbols[stmt_num]));
GEQ_Handle h = cond_upper.and_with_GEQ(ub_list[i]);
h.update_coef(cond_upper.get_local(over_free_var), -stride);
}
- CG_outputRepr *lbRepr, *ubRepr;
- if (lb_repr_list.size() > 1)
+ CG_outputRepr *lbRepr, *ubRepr;
+ if (lb_repr_list.size() > 1) {
+ //fprintf(stderr, "loop_unroll.cc createInvoke( max )\n");
lbRepr = ocg->CreateInvoke("max", lb_repr_list);
- else if (lb_repr_list.size() == 1)
+ }
+ else if (lb_repr_list.size() == 1) {
lbRepr = lb_repr_list[0];
+ }
- if (ub_repr_list.size() > 1)
+ if (ub_repr_list.size() > 1) {
+ //fprintf(stderr, "loop_unroll.cc createInvoke( min )\n");
ubRepr = ocg->CreateInvoke("min", ub_repr_list);
- else if (ub_repr_list.size() == 1)
+ }
+ else if (ub_repr_list.size() == 1) {
ubRepr = ub_repr_list[0];
+ }
// create overflow assignment
CG_outputRepr *rhs = ocg->CreatePlus(ocg->CreateMinus(ubRepr, lbRepr),
@@ -503,7 +533,7 @@ std::set<int> Loop::unroll(int stmt_num, int level, int unroll_amount,
h.update_coef(mapping.output_var(i), 1);
h.update_coef(mapping.input_var(i), -1);
}
- Relation overflow_IS = Range(Restrict_Domain(mapping, copy(hull)));
+ Relation overflow_IS = omega::Range(Restrict_Domain(mapping, copy(hull)));
for (int i = 1; i < cleanup_split_level; i++)
overflow_IS.name_set_var(i, hull.set_var(i)->name());
overflow_IS.setup_names();
@@ -548,7 +578,11 @@ std::set<int> Loop::unroll(int stmt_num, int level, int unroll_amount,
stmt[stmt_num].loop_level[i].parallel_level;
}
+ fprintf(stderr, "loop_unroll.cc L581 adding stmt %d\n", stmt.size());
stmt.push_back(overflow_stmt);
+
+ uninterpreted_symbols.push_back(uninterpreted_symbols[stmt_num]);
+ uninterpreted_symbols_stringrepr.push_back(uninterpreted_symbols_stringrepr[stmt_num]);
dep.insert();
overflow_stmt_num = stmt.size() - 1;
overflow[overflow_stmt_num] = over_var_list;
@@ -567,8 +601,7 @@ std::set<int> Loop::unroll(int stmt_num, int level, int unroll_amount,
dv.type = DEP_W2W;
{
IR_ScalarSymbol *overflow_sym = NULL;
- std::vector<IR_ScalarRef *> scalars = ir->FindScalarRef(
- overflow_code);
+ std::vector<IR_ScalarRef *> scalars = ir->FindScalarRef(overflow_code);
for (int i = scalars.size() - 1; i >= 0; i--)
if (scalars[i]->is_write()) {
overflow_sym = scalars[i]->symbol();
@@ -703,7 +736,12 @@ std::set<int> Loop::unroll(int stmt_num, int level, int unroll_amount,
new_stmt.loop_level = stmt[*i].loop_level;
new_stmt.ir_stmt_node = NULL;
+
+ fprintf(stderr, "loop_unroll.cc L740 adding stmt %d\n", stmt.size());
stmt.push_back(new_stmt);
+
+ uninterpreted_symbols.push_back(uninterpreted_symbols[stmt_num]);
+ uninterpreted_symbols_stringrepr.push_back(uninterpreted_symbols_stringrepr[stmt_num]);
dep.insert();
what_stmt_num[*i].push_back(stmt.size() - 1);
}
@@ -990,33 +1028,70 @@ std::set<int> Loop::unroll(int stmt_num, int level, int unroll_amount,
Statement new_stmt;
new_stmt.code = NULL;
- for (int j = 1; j < unroll_amount; j++)
+ for (int j = 1; j < unroll_amount; j++) {
for (int i = 0; i < stmt_order.size(); i++) {
std::vector<std::string> loop_vars;
std::vector<CG_outputRepr *> subs;
+
+ //fprintf(stderr, "loop_unroll.cc, will replace '%s with '%s+%d' ??\n",
+ // stmt[stmt_order[i].second].IS.set_var(level)->name().c_str(),
+ // stmt[stmt_order[i].second].IS.set_var(level)->name().c_str(), j * stride);
+
loop_vars.push_back(
stmt[stmt_order[i].second].IS.set_var(level)->name());
subs.push_back(
- ocg->CreatePlus(
- ocg->CreateIdent(
- stmt[stmt_order[i].second].IS.set_var(
- level)->name()),
- ocg->CreateInt(j * stride)));
+ ocg->CreatePlus(ocg->CreateIdent(stmt[stmt_order[i].second].IS.set_var(level)->name()),
+ ocg->CreateInt(j * stride))); // BUG HERE
+ //fprintf(stderr, "loop_unroll.cc subs now has %d parts\n", subs.size());
+ //for (int k=0; k< subs.size(); k++) //fprintf(stderr, "subs[%d] = 0x%x\n", k, subs[k]);
+
+ //fprintf(stderr, "ij %d %d ", i, j);
+ //fprintf(stderr, "old src was =\n");
+ //stmt[stmt_order[i].second].code->dump(); fflush(stdout); //fprintf(stderr, "\n");
+
+
+
CG_outputRepr *code = ocg->CreateSubstitutedStmt(0,
- stmt[stmt_order[i].second].code->clone(), loop_vars,
+ stmt[stmt_order[i].second].code->clone(),
+ loop_vars,
subs);
+
+ //fprintf(stderr, "old src is =\n");
+ //stmt[stmt_order[i].second].code->dump(); fflush(stdout); //fprintf(stderr, "\n");
+
+ //fprintf(stderr, "substituted copy is =\n");
+ //code->dump(); //fprintf(stderr, "\n\n");
+
+
new_stmt.code = ocg->StmtListAppend(new_stmt.code, code);
+ //fprintf(stderr, "appended code =\n");
+ //new_stmt.code->dump();
+
}
+ }
+
+
+ //fprintf(stderr, "new_stmt.IS = \n");
new_stmt.IS = copy(stmt[stmt_num].IS);
new_stmt.xform = copy(stmt[stmt_num].xform);
assign_const(new_stmt.xform, 2 * max_level,
stmt_order[stmt_order.size() - 1].first + 1);
new_stmt.loop_level = stmt[stmt_num].loop_level;
new_stmt.ir_stmt_node = NULL;
+
+ new_stmt.has_inspector = false; // ?? or from copied stmt?
+ if (stmt[stmt_num].has_inspector) fprintf(stderr, "OLD STMT HAS INSPECTOR\n");
+ else fprintf(stderr, "OLD STMT DOES NOT HAVE INSPECTOR\n");
+
+ fprintf(stderr, "loop_unroll.cc L1083 adding stmt %d\n", stmt.size());
stmt.push_back(new_stmt);
+
+ uninterpreted_symbols.push_back(uninterpreted_symbols[stmt_num]);
+ uninterpreted_symbols_stringrepr.push_back(uninterpreted_symbols_stringrepr[stmt_num]);
dep.insert();
+ //fprintf(stderr, "update dependence graph\n");
// update dependence graph
if (stmt[stmt_num].loop_level[level - 1].type == LoopLevelOriginal) {
int dep_dim = stmt[stmt_num].loop_level[level - 1].payload;
@@ -1140,6 +1215,7 @@ std::set<int> Loop::unroll(int stmt_num, int level, int unroll_amount,
}
}
+ //fprintf(stderr, " loop_unroll.cc returning new_stmts\n");
return new_stmts;
}
generated by cgit v1.2.3-70-g09d2 (git 2.47.1) at 2025-02-02 02:45:01 +0000