Barrier instruction support.

src/enc.cpp

@@ -164,7 +164,11 @@ Instruction *ByteDecoder::decode(const vector<Byte> &v, Size &n) {
       inst.setDestPReg(readByte(v, n));
       inst.setSrcPReg(readByte(v, n));
       break;
-    default:
+    case Instruction::AC_2REGSRC:
+      inst.setSrcReg(readByte(v, n));
+      inst.setSrcReg(readByte(v, n));
+      break;
+   default:
       decodeError("Unknown argument class.");
   }
 
@@ -331,6 +335,10 @@ Instruction *WordDecoder::decode(const std::vector<Byte> &v, Size &idx) {
       inst.setSrcPReg((code>>i3)&pMask);
       inst.setSrcPReg((code>>(i3-r))&pMask);
       break;
+    case Instruction::AC_2REGSRC:
+      inst.setSrcReg((code>>i2)&rMask);
+      inst.setSrcReg((code>>i3)&rMask);
+      break;
     defualt:
       cout << "Unrecognized argument class in word decoder.\n";
       exit(1);

src/include/core.h

@@ -7,6 +7,8 @@
 #include <string>
 #include <vector>
 #include <stack>
+#include <map>
+#include <set>
 
 #include "types.h"
 #include "archdef.h"
@@ -84,6 +86,7 @@ namespace Harp {
     Word interruptEntry;
 
     std::vector<Warp> w;
+    std::map<Word, std::set<Warp *> > b; // Barriers
   };
 
   class Warp {

src/include/instruction.h

@@ -31,7 +31,7 @@ namespace Harp {
                   JMPRT, LD, ST, LDI, RTOP, ANDP, ORP, XORP, NOTP, ISNEG, 
                   ISZERO, HALT, TRAP, JMPRU, SKEP, RETI, TLBRM,
                   ITOF, FTOI, FADD, FSUB, FMUL, FDIV, FNEG, WSPAWN,
-                  SPLIT, JOIN };
+                  SPLIT, JOIN, BAR };
     enum ArgClass {
       AC_NONE, AC_2REG, AC_2IMM, AC_3REG, AC_3PREG, AC_3IMM, AC_3REGSRC, 
       AC_1IMM, AC_1REG, AC_3IMMSRC, AC_PREG_REG, AC_2PREG, AC_2REGSRC

src/instruction.cpp

@@ -360,6 +360,29 @@ void Instruction::executeOn(Warp &c) {
                      }
                    }
                    break;
+      case BAR: if (sjOnce) {
+                  sjOnce = false;
+                  Word id(reg[rsrc[0]]), n(reg[rsrc[1]]);
+                  set<Warp*> &b(c.core->b[id]);
+
+                  // Add current warp to the barrier and halt.
+                  b.insert(&c);
+                  c.shadowActiveThreads = c.activeThreads;
+                  nextActiveThreads = 0;
+
+                  D(2, "Barrier " << id << ' ' << b.size() << " of " << n);
+
+                  // If the barrier's full, reactivate warps waiting at it
+                  if (b.size() == n) {
+                    set<Warp*>::iterator it;
+                    for (it = b.begin(); it != b.end(); ++it)
+                      (*it)->activeThreads = (*it)->shadowActiveThreads;
+                    c.core->b.erase(id);
+                    nextActiveThreads = c.shadowActiveThreads;
+                  }
+		}
+                break;
+
       default:
         cout << "ERROR: Unsupported instruction: " << *this << "\n";
         exit(1);