Merge pull request #573 from larsclausen/short-circuit-logical-operator

Short circuit logical operator

eval_tree.cc

@@ -811,6 +811,36 @@ NetEConst* NetEBLogic::eval_arguments_(const NetExpr*l, const NetExpr*r) const
 
       const NetEConst*lc = dynamic_cast<const NetEConst*>(l);
       const NetEConst*rc = dynamic_cast<const NetEConst*>(r);
+
+      // If the left side is constant and the right side is short circuited
+      // replace the expression with a constant
+      if (rc == 0 && lc != 0) {
+	    verinum v = lc->value();
+	    verinum::V res = verinum::Vx;
+	    switch (op_) {
+		case 'a': // Logical AND (&&)
+		  if (v.is_zero())
+			res = verinum::V0;
+		  break;
+		case 'o': // Logical OR (||)
+		  if (! v.is_zero() && v.is_defined())
+			res = verinum::V1;
+		  break;
+		case 'q': // Logical implication (->)
+		  if (v.is_zero())
+			res = verinum::V1;
+		  break;
+		default:
+		  break;
+	    }
+	    if (res != verinum::Vx) {
+		  NetEConst*tmp = new NetEConst(verinum(res, 1));
+		  ivl_assert(*this, tmp);
+		  eval_debug(this, tmp, false);
+		  return tmp;
+	    }
+      }
+
       if (lc == 0 || rc == 0) return 0;
 
       verinum::V lv = verinum::V0;

tgt-vvp/eval_condit.c

@@ -278,22 +278,39 @@ static int draw_condition_binary_le(ivl_expr_t expr)
 
 static int draw_condition_binary_lor(ivl_expr_t expr)
 {
+      unsigned label_out = local_count++;
+
       ivl_expr_t le = ivl_expr_oper1(expr);
       ivl_expr_t re = ivl_expr_oper2(expr);
 
       int lx = draw_eval_condition(le);
+      int tmp_flag = lx;
+
+      /* Short circuit right hand side if necessary */
+      fprintf(vvp_out, "    %%jmp/1 T_%u.%u, %d;\n", thread_count, label_out, lx);
 
       if (lx < 8) {
-	    int tmp = allocate_flag();
-	    fprintf(vvp_out, "    %%flag_mov %d, %d;\n", tmp, lx);
-	    lx = tmp;
+	  tmp_flag = allocate_flag();
+	  fprintf(vvp_out, "    %%flag_mov %d, %d;\n", tmp_flag, lx);
       }
 
       int rx = draw_eval_condition(re);
 
-      fprintf(vvp_out, "    %%flag_or %d, %d;\n", rx, lx);
-      clr_flag(lx);
-      return rx;
+      /*
+       * The flag needs to be in the same position regardless of whether the
+       * right side is short-cicuited or not.
+       */
+      if (lx == tmp_flag) {
+	  fprintf(vvp_out, "    %%flag_or %d, %d;\n", lx, rx);
+      } else {
+	  fprintf(vvp_out, "    %%flag_or %d, %d;\n", rx, tmp_flag);
+	  if (lx != rx)
+	      fprintf(vvp_out, "    %%flag_mov %d, %d;\n", lx, rx);
+	  clr_flag(tmp_flag);
+      }
+      fprintf(vvp_out, "T_%u.%u;\n", thread_count, label_out);
+      clr_flag(rx);
+      return lx;
 }
 
 static int draw_condition_binary(ivl_expr_t expr)

tgt-vvp/eval_vec4.c

@@ -432,25 +432,45 @@ static void draw_binary_vec4_lequiv(ivl_expr_t expr)
       assert(ivl_expr_width(expr) == 1);
 }
 
-static void draw_binary_vec4_land(ivl_expr_t expr)
+static void draw_binary_vec4_logical(ivl_expr_t expr, char op)
 {
+      const char *opcode;
+      const char *jmp_type;
+
+      switch (op) {
+	  case 'a':
+	    opcode = "and";
+	    jmp_type = "0";
+	    break;
+	  case 'o':
+	    opcode = "or";
+	    jmp_type = "1";
+	    break;
+	  default:
+	    assert(0);
+	    break;
+      }
+
+      unsigned label_out = local_count++;
+
       ivl_expr_t le = ivl_expr_oper1(expr);
       ivl_expr_t re = ivl_expr_oper2(expr);
 
-	/* Push the left expression. Reduce it to a single bit if
-	   necessary. */
-      draw_eval_vec4(le);
-      if (ivl_expr_width(le) > 1)
-	    fprintf(vvp_out, "    %%or/r;\n");
-
-	/* Now push the right expression. Again, reduce to a single
-	   bit if necessary. */
+      /* Evaluate the left expression as a conditon and skip the right expression
+       * if the left is false. */
+      int flag = draw_eval_condition(le);
+      fprintf(vvp_out, "    %%flag_get/vec4 %d;\n", flag);
+      fprintf(vvp_out, "    %%jmp/%s T_%u.%u, %d;\n", jmp_type, thread_count,
+	      label_out, flag);
+      clr_flag(flag);
+      /* Now push the right expression. Reduce to a single bit if necessary. */
       draw_eval_vec4(re);
       if (ivl_expr_width(re) > 1)
 	    fprintf(vvp_out, "    %%or/r;\n");
 
-      fprintf(vvp_out, "    %%and;\n");
+      fprintf(vvp_out, "    %%%s;\n", opcode);
 
+      fprintf(vvp_out, "T_%u.%u;\n", thread_count, label_out);
       if (ivl_expr_width(expr) > 1)
 	    fprintf(vvp_out, "    %%pad/u %u;\n", ivl_expr_width(expr));
 }
@@ -632,29 +652,6 @@ static void draw_binary_vec4_le(ivl_expr_t expr)
       }
 }
 
-static void draw_binary_vec4_lor(ivl_expr_t expr)
-{
-      ivl_expr_t le = ivl_expr_oper1(expr);
-      ivl_expr_t re = ivl_expr_oper2(expr);
-
-	/* Push the left expression. Reduce it to a single bit if
-	   necessary. */
-      draw_eval_vec4(le);
-      if (ivl_expr_width(le) > 1)
-	    fprintf(vvp_out, "    %%or/r;\n");
-
-	/* Now push the right expression. Again, reduce to a single
-	   bit if necessary. */
-      draw_eval_vec4(re);
-      if (ivl_expr_width(re) > 1)
-	    fprintf(vvp_out, "    %%or/r;\n");
-
-      fprintf(vvp_out, "    %%or;\n");
-
-      if (ivl_expr_width(expr) > 1)
-	    fprintf(vvp_out, "    %%pad/u %u;\n", ivl_expr_width(expr));
-}
-
 static void draw_binary_vec4_lrs(ivl_expr_t expr)
 {
       ivl_expr_t le = ivl_expr_oper1(expr);
@@ -695,7 +692,8 @@ static void draw_binary_vec4(ivl_expr_t expr)
 {
       switch (ivl_expr_opcode(expr)) {
 	  case 'a': /* Logical && */
-	    draw_binary_vec4_land(expr);
+	  case 'o': /* || (logical or) */
+	    draw_binary_vec4_logical(expr, ivl_expr_opcode(expr));
 	    break;
 
 	  case '+':
@@ -738,10 +736,6 @@ static void draw_binary_vec4(ivl_expr_t expr)
 	    draw_binary_vec4_lrs(expr);
 	    break;
 
-	  case 'o': /* || (logical or) */
-	    draw_binary_vec4_lor(expr);
-	    break;
-
 	  case 'q': /* -> (logical implication) */
 	    draw_binary_vec4_limpl(expr);
 	    break;