diff --git a/litex/build/openocd.py b/litex/build/openocd.py
index 87354a4ee..f0de91b90 100644
--- a/litex/build/openocd.py
+++ b/litex/build/openocd.py
@@ -40,19 +40,42 @@ class OpenOCD(GenericProgrammer):
         self.call(["openocd", "-f", config, "-c", script])
 
     def get_ir(self, chain, config):
-        # On ECP5, force IR to 0x32.
         cfg_str = open(config).read()
-        ecp5 = "ecp5" in cfg_str
-        altera = "10m50" in cfg_str # TODO: or cyclone 10
-        if ecp5:
+        # Lattice ECP5.
+        if "ecp5" in cfg_str:
             chain = 0x32
-        elif altera:
-            chain = 0xC
-        # Else IR = 1 + CHAIN.
+        # Intel Max10.
+        elif "10m50" in cfg_str:
+            chain = 0xc
+        # Xilinx ZynqMP.
+        elif "zynqmp" in cfg_str:
+            chain = {
+                1: 0x902, # USER1.
+                2: 0x903, # USER2.
+                3: 0x922, # USER3.
+                4: 0x923, # USER4.
+            }[chain]
+        # Xilinx 7-Series.
         else:
-            chain = 0x1 + chain
+            chain = {
+                1: 0x02, # USER1.
+                2: 0x03, # USER2.
+                3: 0x22, # USER3.
+                4: 0x23, # USER4.
+            }[chain]
         return chain
 
+    def get_endstate(self, config):
+        cfg_str = open(config).read()
+        # Lattice ECP5.
+        if "ecp5" in cfg_str:
+            return "-endstate DRPAUSE" # CHECKME: Can we avoid it?
+        # Intel Max10.
+        elif "10m50" in cfg_str:
+            return "-endstate DRPAUSE" # CHECKME: Is it required on Intel?
+        else:
+            return ""
+
     def stream(self, port=20000, chain=1):
         """
         Create a TCP server to stream data to/from the internal JTAG TAP of the FPGA
@@ -68,7 +91,9 @@ class OpenOCD(GenericProgrammer):
           - TX data  : bit 1 to 8
           - TX valid : bit 9
         """
-        config = self.find_config()
+        config   = self.find_config()
+        ir       = self.get_ir(chain, config)
+        endstate = self.get_endstate(config)
         cfg = """
 proc jtagstream_poll {tap tx n} {
     set m [string length $tx]
@@ -81,7 +106,11 @@ proc jtagstream_poll {tap tx n} {
         #echo tx[scan $txj %c]
     }
     set txi [concat {*}$txi]
-    set rxi [split [drscan $tap {*}$txi -endstate DRPAUSE] " "]
+"""
+        cfg += f"""
+    set rxi [split [drscan $tap {{*}}$txi {endstate}] " "]
+"""
+        cfg += """
     #echo $txi:$rxi
     set rx ""
     set writable 1
@@ -153,7 +182,7 @@ proc jtagstream_serve {tap port} {
         write_to_file("stream.cfg", cfg)
         script = "; ".join([
             "init",
-            "irscan $_CHIPNAME.tap {:d}".format(self.get_ir(chain, config)),
+            "irscan $_CHIPNAME.tap {:d}".format(ir),
             "jtagstream_serve $_CHIPNAME.tap {:d}".format(port),
             "exit",
         ])
diff --git a/litex/soc/cores/jtag.py b/litex/soc/cores/jtag.py
index d3c3c553f..74e5928c0 100644
--- a/litex/soc/cores/jtag.py
+++ b/litex/soc/cores/jtag.py
@@ -304,11 +304,10 @@ class XilinxJTAG(Module):
 
     @staticmethod
     def get_primitive(device):
-        # TODO: Add support for all devices.
         prim_dict = {
             # Primitive Name   Ðevice (startswith)
             "BSCAN_SPARTAN6" : ["xc6"],
-            "BSCANE2"        : ["xc7", "xcku", "xcvu", "xczu"],
+            "BSCANE2"        : ["xc7a", "xc7k", "xc7v", "xc7z"] +  ["xcku", "xcvu", "xczu"],
         }
         for prim, prim_devs in prim_dict.items():
             for prim_dev in prim_devs:
@@ -316,6 +315,15 @@ class XilinxJTAG(Module):
                     return prim
         return None
 
+    @staticmethod
+    def get_tdi_delay(device):
+        # Input delay if 1 TCK on TDI on Zynq/ZynqMPSoC devices.
+        delay_dict = {"xc7z" : 1, "xczu" : 1}
+        for dev, delay in delay_dict.items():
+            if device.lower().startswith(dev):
+                return 1
+        return 0
+
 # ECP5 JTAG ----------------------------------------------------------------------------------------
 
 class ECP5JTAG(Module):
@@ -393,15 +401,14 @@ class JTAGPHY(Module):
 
         # # #
 
-        valid = Signal()
-        data  = Signal(data_width)
-        count = Signal(max=data_width)
 
         # JTAG TAP ---------------------------------------------------------------------------------
         if jtag is None:
+            jtag_tdi_delay = 0
             # Xilinx.
             if XilinxJTAG.get_primitive(device) is not None:
                 jtag = XilinxJTAG(primitive=XilinxJTAG.get_primitive(device), chain=chain)
+                jtag_tdi_delay = XilinxJTAG.get_tdi_delay(device)
             # Lattice.
             elif device[:5] == "LFE5U":
                 jtag = ECP5JTAG()
@@ -436,16 +443,29 @@ class JTAGPHY(Module):
             ]
             sink, source = tx_cdc.source, rx_cdc.sink
 
+        # JTAG TDI/TDO Delay -----------------------------------------------------------------------
+        jtag_tdi = jtag.tdi
+        jtag_tdo = jtag.tdo
+        if jtag_tdi_delay:
+            jtag_tdi_sr = Signal(data_width + 2 - jtag_tdi_delay)
+            self.sync.jtag += If(jtag.shift, jtag_tdi_sr.eq(Cat(jtag.tdi, jtag_tdi_sr)))
+            jtag_tdi = jtag_tdi_sr[-1]
+
         # JTAG Xfer FSM ----------------------------------------------------------------------------
+        valid = Signal()
+        ready = Signal()
+        data  = Signal(data_width)
+        count = Signal(max=data_width)
+
         fsm = FSM(reset_state="XFER-READY")
         fsm = ClockDomainsRenamer("jtag")(fsm)
         fsm = ResetInserter()(fsm)
         self.submodules += fsm
         self.comb += fsm.reset.eq(jtag.reset | jtag.capture)
         fsm.act("XFER-READY",
-            jtag.tdo.eq(source.ready),
+            jtag_tdo.eq(ready),
             If(jtag.shift,
-                sink.ready.eq(jtag.tdi),
+                sink.ready.eq(jtag_tdi),
                 NextValue(valid, sink.valid),
                 NextValue(data,  sink.data),
                 NextValue(count, 0),
@@ -453,20 +473,21 @@ class JTAGPHY(Module):
             )
         )
         fsm.act("XFER-DATA",
-            jtag.tdo.eq(data),
+            jtag_tdo.eq(data),
             If(jtag.shift,
                 NextValue(count, count + 1),
-                NextValue(data, Cat(data[1:], jtag.tdi)),
+                NextValue(data, Cat(data[1:], jtag_tdi)),
                 If(count == (data_width - 1),
                     NextState("XFER-VALID")
                 )
             )
         )
         fsm.act("XFER-VALID",
-            jtag.tdo.eq(valid),
+            jtag_tdo.eq(valid),
             If(jtag.shift,
-                source.valid.eq(jtag.tdi),
+                source.valid.eq(jtag_tdi),
+                source.data.eq(data),
+                NextValue(ready, source.ready),
                 NextState("XFER-READY")
             )
         )
-        self.comb += source.data.eq(data)