tilelink2: allow 0-stage backpressure in combinational regmap

src/main/scala/uncore/tilelink2/RegField.scala

@@ -10,14 +10,20 @@ object RegReadFn
   // (ivalid: Bool, oready: Bool) => (iready: Bool, ovalid: Bool, data: UInt)
   // iready may combinationally depend on oready
   // all other combinational dependencies forbidden (e.g. ovalid <= ivalid)
-  // effects must become visible only on the cycle after ovalid && oready
+  // effects must become visible on the cycle after ovalid && oready
   implicit def apply(x: (Bool, Bool) => (Bool, Bool, UInt)) =
     new RegReadFn(false, x)
-  // (ofire: Bool) => (data: UInt)
-  // effects must become visible on the cycle after ofire
-  implicit def apply(x: Bool => UInt) =
+  // (ready: Bool) => (valid: Bool, data: UInt)
+  // valid must not combinationally depend on ready
+  // effects must become visible on the cycle after valid && ready
+  // ready is only guaranteed to stay high if fed by an Irrevocable (eg: TL2 or concurrency > 0)
+  // ... which is irrelevant if you can service the read immediately
+  // ... if irrevocable, you may start reading on rising ready and raise valid when done
+  // ... otherwise, use the more general in&out ready-valid method above
+  implicit def apply(x: Bool => (Bool, UInt)) =
     new RegReadFn(true, { case (_, oready) =>
-      (Bool(true), Bool(true), x(oready))
+      val (ovalid, data) = x(oready)
+      (Bool(true), ovalid, data)
     })
   // read from a register
   implicit def apply(x: UInt) =
@@ -37,15 +43,20 @@ object RegWriteFn
   // (ivalid: Bool, oready: Bool, data: UInt) => (iready: Bool, ovalid: Bool)
   // iready may combinationally depend on both oready and data
   // all other combinational dependencies forbidden (e.g. ovalid <= ivalid)
-  // effects must become visible only on the cycle after ovalid && oready
+  // effects must become visible on the cycle after ovalid && oready
   implicit def apply(x: (Bool, Bool, UInt) => (Bool, Bool)) =
     new RegWriteFn(false, x)
-  // (ofire: Bool, data: UInt) => ()
-  // effects must become visible on the cycle after ofire
-  implicit def apply(x: (Bool, UInt) => Unit) =
+  // (valid: Bool, data: UInt) => (ready: Bool)
+  // ready may combinationally depend on data (but not valid)
+  // effects must become visible on the cycle after valid && ready
+  // valid is only guaranteed to stay high if fed by an Irrevocable (eg: TL2 or concurrency > 0)
+  // ... which is irrelevant if you can service the write immediately
+  // ... if irrevocable, you may start writing on rising valid and raise ready when done
+  // ... otherwise, use the more general in&out ready-valid method above
+  implicit def apply(x: (Bool, UInt) => Bool) =
+    // combinational => data valid on oready
     new RegWriteFn(true, { case (_, oready, data) =>
-      x(oready, data)
-      (Bool(true), Bool(true))
+      (Bool(true), x(oready, data))
     })
   // updates a register
   implicit def apply(x: UInt) =