more cleaning

include/triton/codegen/analysis/liveness.h

@@ -42,21 +42,20 @@ struct segment {
 
 class liveness {
 private:
-  typedef std::map<ir::value*, slot_index> indices_map_t;
   typedef std::map<layout_t*, segment>    intervals_map_t;
 
 public:
+  // constructor
   liveness(layout *l): layouts_(l){ }
   // accessors
-  const intervals_map_t& intervals()  const { return intervals_; }
-  segment get_interval(layout_t* v)  const { return intervals_.at(v); }
+  const intervals_map_t& get()  const { return intervals_; }
+  segment get(layout_t* v)  const { return intervals_.at(v); }
   // run
   void run(ir::module &mod);
 
 private:
   // analysis
   layout *layouts_;
-  indices_map_t indices;
   intervals_map_t intervals_;
 };
 

lib/codegen/analysis/allocation.cc

@@ -22,7 +22,7 @@ void allocation::run(ir::module &mod) {
   typedef std::multimap<unsigned, segment> triples_map_type;
 
   std::vector<layout_t*> I;
-  for(auto x: liveness_->intervals())
+  for(auto x: liveness_->get())
     I.push_back(x.first);
   std::vector<layout_t*> J = I;
 
@@ -37,7 +37,7 @@ void allocation::run(ir::module &mod) {
     segment xh = h_it->second;
     H.erase(h_it);
     auto j_it = std::find_if(J.begin(), J.end(), [&](layout_t* JJ){
-      segment xj = liveness_->get_interval(JJ);
+      segment xj = liveness_->get(JJ);
       bool res = xj.intersect(xh);
       for(auto val: H)
         res = res && !val.second.intersect(xj);
@@ -45,7 +45,7 @@ void allocation::run(ir::module &mod) {
     });
     if(j_it != J.end()){
       unsigned size = (*j_it)->size;
-      segment xj = liveness_->get_interval(*j_it);
+      segment xj = liveness_->get(*j_it);
       starts[*j_it] = w;
       H.insert({w + size, segment{max(xh.start, xj.start), min(xh.end, xj.end)}});
       if(xh.start < xj.start)
@@ -68,7 +68,7 @@ void allocation::run(ir::module &mod) {
     unsigned NY = y->size;
     segment XS = {X0, X0 + NX};
     segment YS = {Y0, Y0 + NY};
-    if(liveness_->get_interval(x).intersect(liveness_->get_interval(y))
+    if(liveness_->get(x).intersect(liveness_->get(y))
         && XS.intersect(YS))
       interferences[x].insert(y);
   }

lib/codegen/analysis/layout.cc

@@ -188,17 +188,6 @@ layout_hmma_884_t::layout_hmma_884_t(size_t num_warps,
     throw std::runtime_error("cannot create a kernel with this amount of warps");
 }
 
-inline bool is_loop_latch(ir::phi_node *phi, ir::instruction *terminator){
-  if(phi->get_parent() != terminator->get_parent())
-    return false;
-  if(auto *br = dynamic_cast<ir::cond_branch_inst*>(terminator))
-    return br->get_true_dest() == phi->get_parent()
-           || br->get_false_dest() == phi->get_parent();
-  else if(dynamic_cast<ir::uncond_branch_inst*>(terminator))
-    return false;
-  else
-    throw std::runtime_error("unreachable");
-}
 
 
 
@@ -230,6 +219,19 @@ layout_scanline_t::layout_scanline_t(size_t num_warps,
     throw std::runtime_error("cannot create a kernel with this amount of warps");
 }
 
+inline bool is_loop_latch(ir::phi_node *phi, ir::instruction *terminator){
+  if(phi->get_parent() != terminator->get_parent())
+    return false;
+  if(auto *br = dynamic_cast<ir::cond_branch_inst*>(terminator))
+    return br->get_true_dest() == phi->get_parent()
+           || br->get_false_dest() == phi->get_parent();
+  else if(dynamic_cast<ir::uncond_branch_inst*>(terminator))
+    return false;
+  else
+    throw std::runtime_error("unreachable");
+}
+
+
 void extract_double_bufferable(ir::value *v, std::shared_ptr<double_buffer_info_t>& res) {
   auto* phi = dynamic_cast<ir::phi_node*>(v);
   if(!phi || phi->get_num_incoming() != 2)
@@ -303,7 +305,7 @@ layout_shared_t::layout_shared_t(const layout_t *arg,
     pad = 24 - shapes[row ? 1 : 0] % 32;
   }
   else if(order != arg->order) {
-    pad = 16;
+    pad = 4;
   }
 
   // size
@@ -316,6 +318,7 @@ layout_shared_t::layout_shared_t(const layout_t *arg,
     size *= 2;
 }
 
+// layout factory method
 void layout::create(size_t id, const std::vector<ir::value*>& values) {
   auto it_hmma_c = std::find_if(values.begin(), values.end(), &is_hmma_c);
   auto cmp = [](ir::value* x, ir::value *y) {

lib/codegen/analysis/liveness.cc

@@ -17,13 +17,11 @@ namespace codegen{
 namespace analysis{
 
 
-// Entry point
 void liveness::run(ir::module &mod) {
-  indices.clear();
   intervals_.clear();
 
-
   // Assigns index to each instruction
+  std::map<ir::value*, slot_index> indices;
   for(ir::function *fn: mod.get_function_list()){
     slot_index index = 0;
     for(ir::basic_block *block: fn->blocks())
@@ -33,6 +31,7 @@ void liveness::run(ir::module &mod) {
     }
   }
 
+  // create live intervals
   for(auto &x: layouts_->get_all()) {
     layout_t* layout = x.second;
     if(layout->type != SHARED)