diff --git a/litedram/gen.py b/litedram/gen.py
index 06fb4da..d1fa927 100755
--- a/litedram/gen.py
+++ b/litedram/gen.py
@@ -582,10 +582,15 @@ class LiteDRAMCore(SoCCore):
         # ECP5DDRPHY.
         elif core_config["sdram_phy"] in  [litedram_phys.ECP5DDRPHY]:
             assert core_config["memtype"] in ["DDR3"]
+            kwargs = {}
+            if core_config.get("dm_swap", False):
+                kwargs['dm_remapping'] = {0:1, 1:0}
+
             self.submodules.ddrphy = sdram_phy = core_config["sdram_phy"](
                 pads         = platform.request("ddram"),
                 sys_clk_freq = sys_clk_freq,
-                cmd_delay    = core_config.get("cmd_delay", 0))
+                cmd_delay    = core_config.get("cmd_delay", 0),
+                **kwargs)
             self.ddrphy.settings.add_electrical_settings(**electrical_settings_kwargs)
             self.comb += crg.stop.eq(self.ddrphy.init.stop)
             self.comb += crg.reset.eq(self.ddrphy.init.reset)
diff --git a/litedram/phy/ecp5ddrphy.py b/litedram/phy/ecp5ddrphy.py
index 4f1010e..2aa49ca 100644
--- a/litedram/phy/ecp5ddrphy.py
+++ b/litedram/phy/ecp5ddrphy.py
@@ -117,7 +117,8 @@ class ECP5DDRPHY(Module, AutoCSR):
         cl           = None,
         cwl          = None,
         cmd_delay    = 0,
-        clk_polarity = 0):
+        clk_polarity = 0,
+        dm_remapping = None):
         assert isinstance(cmd_delay, int) and cmd_delay < 128
         pads        = PHYPadsCombiner(pads)
         memtype     = "DDR3"
@@ -127,6 +128,8 @@ class ECP5DDRPHY(Module, AutoCSR):
         nranks      = 1 if not hasattr(pads, "cs_n") else len(pads.cs_n)
         databits    = len(pads.dq)
         nphases     = 2
+        if not dm_remapping:
+            dm_remapping = {}
         assert databits%8 == 0
 
         # Init -------------------------------------------------------------------------------------
@@ -327,7 +330,7 @@ class ECP5DDRPHY(Module, AutoCSR):
             dm_o_data_d     = Signal(8)
             dm_o_data_muxed = Signal(4)
             for n in range(8):
-                self.comb += dm_o_data[n].eq(dfi.phases[n//4].wrdata_mask[n%4*databits//8 + i])
+                self.comb += dm_o_data[n].eq(dfi.phases[n//4].wrdata_mask[n%4*databits//8+dm_remapping.get(i, i)])
             self.sync += dm_o_data_d.eq(dm_o_data)
             dm_bl8_cases = {}
             dm_bl8_cases[0] = dm_o_data_muxed.eq(dm_o_data[:4])