diff --git a/compiler/semdata.nim b/compiler/semdata.nim
index b25f72f2d9..91a1ebf86a 100644
--- a/compiler/semdata.nim
+++ b/compiler/semdata.nim
@@ -47,7 +47,7 @@ type
     efLValue, efWantIterator, efInTypeof,
     efWantStmt, efAllowStmt, efDetermineType,
     efAllowDestructor, efWantValue, efOperand, efNoSemCheck,
-    efNoProcvarCheck, efFromHlo
+    efNoProcvarCheck, efNoEvaluateGeneric, efInCall, efFromHlo
   TExprFlags* = set[TExprFlag]
 
   TTypeAttachedOp* = enum
diff --git a/compiler/semexprs.nim b/compiler/semexprs.nim
index 7135dcf346..57fb3ceed1 100644
--- a/compiler/semexprs.nim
+++ b/compiler/semexprs.nim
@@ -717,6 +717,34 @@ proc resolveIndirectCall(c: PContext; n, nOrig: PNode;
       initCandidate(c, result, t)
       matches(c, n, nOrig, result)
 
+proc bracketedMacro(n: PNode): PSym =
+  if n.len >= 1 and n[0].kind == nkSym:
+    result = n[0].sym
+    if result.kind notin {skMacro, skTemplate}:
+      result = nil
+
+proc semBracketedMacro(c: PContext; outer, inner: PNode; s: PSym;
+                       flags: TExprFlags): PNode =
+  # We received untransformed bracket expression coming from macroOrTmpl[].
+  # Transform it to macro or template call, where first come normal
+  # arguments, next come generic template arguments.
+  var sons = newSeq[PNode]()
+  sons.add inner.sons[0]
+  # Normal arguments:
+  for i in 1..<outer.len:
+    sons.add outer.sons[i]
+  # Generic template arguments from bracket expression:
+  for i in 1..<inner.len:
+    sons.add inner.sons[i]
+  shallowCopy(outer.sons, sons)
+  # FIXME: Shouldn't we check sfImmediate and call semDirectOp?
+  # However passing to semDirectOp doesn't work here.
+  case s.kind
+  of skMacro: result = semMacroExpr(c, outer, outer, s, flags)
+  of skTemplate: result = semTemplateExpr(c, outer, s, flags)
+  else: assert(false)
+  return
+
 proc semIndirectOp(c: PContext, n: PNode, flags: TExprFlags): PNode =
   result = nil
   checkMinSonsLen(n, 1)
@@ -732,10 +760,15 @@ proc semIndirectOp(c: PContext, n: PNode, flags: TExprFlags): PNode =
       for i in countup(1, sonsLen(n) - 1): addSon(result, n.sons[i])
       return semExpr(c, result, flags)
   else:
-    n.sons[0] = semExpr(c, n.sons[0])
+    n.sons[0] = semExpr(c, n.sons[0], {efInCall})
     let t = n.sons[0].typ
     if t != nil and t.kind == tyVar:
       n.sons[0] = newDeref(n.sons[0])
+    elif n.sons[0].kind == nkBracketExpr:
+      let s = bracketedMacro(n.sons[0])
+      if s != nil:
+        return semBracketedMacro(c, n, n.sons[0], s, flags)
+
   let nOrig = n.copyTree
   semOpAux(c, n)
   var t: PType = nil
@@ -965,8 +998,20 @@ proc semSym(c: PContext, n: PNode, s: PSym, flags: TExprFlags): PNode =
       else: result = newSymNode(s, n.info)
     else:
       result = newSymNode(s, n.info)
-  of skMacro: result = semMacroExpr(c, n, n, s, flags)
-  of skTemplate: result = semTemplateExpr(c, n, s, flags)
+  of skMacro:
+    if efNoEvaluateGeneric in flags and s.ast[genericParamsPos].len > 0:
+      markUsed(n.info, s)
+      styleCheckUse(n.info, s)
+      result = newSymNode(s, n.info)
+    else:
+      result = semMacroExpr(c, n, n, s, flags)
+  of skTemplate:
+    if efNoEvaluateGeneric in flags and s.ast[genericParamsPos].len > 0:
+      markUsed(n.info, s)
+      styleCheckUse(n.info, s)
+      result = newSymNode(s, n.info)
+    else:
+      result = semTemplateExpr(c, n, s, flags)
   of skParam:
     markUsed(n.info, s)
     styleCheckUse(n.info, s)
@@ -1193,7 +1238,9 @@ proc semSubscript(c: PContext, n: PNode, flags: TExprFlags): PNode =
     result.add(x[0])
     return
   checkMinSonsLen(n, 2)
-  n.sons[0] = semExprWithType(c, n.sons[0], {efNoProcvarCheck})
+  # make sure we don't evaluate generic macros/templates
+  n.sons[0] = semExprWithType(c, n.sons[0],
+                              {efNoProcvarCheck, efNoEvaluateGeneric})
   let arr = skipTypes(n.sons[0].typ, {tyGenericInst, tyVar, tyPtr, tyRef})
   case arr.kind
   of tyArray, tyOpenArray, tyVarargs, tyArrayConstr, tySequence, tyString,
@@ -1236,12 +1283,30 @@ proc semSubscript(c: PContext, n: PNode, flags: TExprFlags): PNode =
     let s = if n.sons[0].kind == nkSym: n.sons[0].sym
             elif n[0].kind in nkSymChoices: n.sons[0][0].sym
             else: nil
-    if s != nil and s.kind in {skProc, skMethod, skConverter, skIterator}:
-      # type parameters: partial generic specialization
-      n.sons[0] = semSymGenericInstantiation(c, n.sons[0], s)
-      result = explicitGenericInstantiation(c, n, s)
-    elif s != nil and s.kind == skType:
-      result = symNodeFromType(c, semTypeNode(c, n, nil), n.info)
+    if s != nil:
+      case s.kind
+      of skProc, skMethod, skConverter, skIterator:
+        # type parameters: partial generic specialization
+        n.sons[0] = semSymGenericInstantiation(c, n.sons[0], s)
+        result = explicitGenericInstantiation(c, n, s)
+      of skMacro, skTemplate:
+        if efInCall in flags:
+          # We are processing macroOrTmpl[] in macroOrTmpl[](...) call.
+          # Return as is, so it can be transformed into complete macro or
+          # template call in semIndirectOp caller.
+          result = n
+        else:
+          # We are processing macroOrTmpl[] not in call. Transform it to the
+          # macro or template call with generic arguments here.
+          n.kind = nkCall
+          case s.kind
+          of skMacro: result = semMacroExpr(c, n, n, s, flags)
+          of skTemplate: result = semTemplateExpr(c, n, s, flags)
+          else: discard
+      of skType:
+        result = symNodeFromType(c, semTypeNode(c, n, nil), n.info)
+      else:
+        c.p.bracketExpr = n.sons[0]
     else:
       c.p.bracketExpr = n.sons[0]
 
diff --git a/tests/generics/tgenerictmpl2.nim b/tests/generics/tgenerictmpl2.nim
new file mode 100644
index 0000000000..0ecaf9ded0
--- /dev/null
+++ b/tests/generics/tgenerictmpl2.nim
@@ -0,0 +1,31 @@
+discard """
+  output: '''1
+1
+1
+1
+999
+999
+999
+2'''
+"""
+
+# test if we can pass explicit generic arguments to generic templates
+# based on bug report #3496
+
+proc     tproc[T](t: T = 999) = echo t
+template ttmpl[T](t: T = 999) = echo t
+
+tproc(1)
+tproc[int](1)
+ttmpl(1)
+ttmpl[int](1) #<- crash case #1
+
+tproc[int]()
+discard tproc[int]
+ttmpl[int]()  #<- crash case #2
+ttmpl[int]    #<- crash case #3
+
+# but still allow normal use of [] on non-generic templates
+
+template tarr: expr = [1, 2, 3, 4]
+echo tarr[1]