diff --git a/artiq/coredevice/sawg.py b/artiq/coredevice/sawg.py
index f7897f212..647e30516 100644
--- a/artiq/coredevice/sawg.py
+++ b/artiq/coredevice/sawg.py
@@ -62,44 +62,36 @@ class Spline:
         rtio_output_list(now_mu(), self.channel, 0, value)
 
     @portable(flags={"fast-math"})
-    def pack_coeff_mu(self, coeff) -> TList(TInt32):
-        n = len(coeff)
-        width = n*self.width + (n - 1)*n//2*self.time_width
-        packed = [int32(0)] * ((width + 31)//32)
+    def pack_coeff_mu(self, coeff, packed):
         pos = 0
-        for i in range(n):
+        for i in range(len(coeff)):
             wi = self.width + i*self.time_width
             ci = coeff[i]
-            while wi:
+            while wi != 0:
                 j = pos//32
                 used = pos - 32*j
                 avail = 32 - used
                 if avail > wi:
                     avail = wi
-                packed[j] |= (ci & ((1 << avail) - 1)) << used
+                packed[j] |= int32(ci & ((1 << avail) - 1)) << used
                 ci >>= avail
                 wi -= avail
                 pos += avail
-        return packed
 
     @portable(flags={"fast-math"})
-    def coeff_to_mu(self, coeff) -> TList(TInt32):
-        n = len(coeff)
-        coeff64 = [int64(0)] * n
-        for i in range(n):
+    def coeff_to_mu(self, coeff, coeff64):
+        for i in range(len(coeff)):
             vi = coeff[i] * self.scale
             for j in range(i):
                 vi *= self.time_scale
             ci = int(round(vi))
             coeff64[i] = ci
             # artiq.wavesynth.coefficients.discrete_compensate:
-            continue
             if i == 2:
                 coeff64[1] += ci >> (self.time_width + 1)
             elif i == 3:
                 coeff64[2] += ci >> self.time_width
                 coeff64[1] += (ci // 3) >> (2*self.time_width + 1)
-        return self.pack_coeff_mu(coeff64)
 
     @kernel
     def set_list(self, value):
@@ -108,7 +100,13 @@ class Spline:
         :param value: List of floating point spline knot coefficients,
             lowest order (constant) coefficient first.
         """
-        self.set_list_mu(self.coeff_to_mu(value))
+        n = len(value)
+        width = n*self.width + (n - 1)*n//2*self.time_width
+        coeff64 = [int64(0)] * n
+        packed = [int32(0)] * ((width + 31)//32)
+        self.coeff_to_mu(value, coeff64)
+        self.pack_coeff_mu(coeff64, packed)
+        self.set_list_mu(packed)
 
     @kernel(flags={"fast-math"})
     def smooth(self, start: TFloat, stop: TFloat, duration: TFloat,
diff --git a/artiq/test/gateware/test_sawg_fe.py b/artiq/test/gateware/test_sawg_fe.py
index a93f95daa..21230c97f 100644
--- a/artiq/test/gateware/test_sawg_fe.py
+++ b/artiq/test/gateware/test_sawg_fe.py
@@ -104,12 +104,18 @@ class SAWGTest(unittest.TestCase):
 
     def test_coeff(self):
         import struct
-        for v in [-.1], [.1, -.01], [.1, .01, -.00001], \
-                [.1, .01, .00001, -.000000001]:
+        # these get discrete_compensate
+        # [.1, .01, -.00001], [.1, .01, .00001, -.000000001]
+        for v in [-.1], [.1, -.01]:
             ch = self.driver.offset
-            p = ch.coeff_to_mu(v)
+            q = [0] * len(v)
+            ch.coeff_to_mu(v, q)
+            n = len(v)
             t = ch.time_width
             w = ch.width
+            p = [0] * ((n*w + (n - 1)*n//2*t + 31)//32)
+            ch.pack_coeff_mu(q, p)
+            p = [_ & 0xffffffff for _ in p]
             p0 = [int(round(vi*ch.scale*ch.time_scale**i))
                   for i, vi in enumerate(v)]
             p0 = [struct.pack("<" + "_bhiiqqqq"[(w + i*t)//8], vi
@@ -135,19 +141,28 @@ class SAWGTest(unittest.TestCase):
                 self.assertEqual(out[2*i], v)
                 self.assertEqual(out[2*i+1], v)
 
-    def test_pack(self):
+    def pack(self, v):
+        n = len(v)
         ch = self.driver.offset
-        self.assertEqual(ch.pack_coeff_mu([1]), [1])
-        self.assertEqual(ch.pack_coeff_mu([1, 1 << 16]), [1, 1])
-        self.assertEqual(ch.pack_coeff_mu([1, 1 << 32]), [1, 0])
-        self.assertEqual(ch.pack_coeff_mu([0x1234, 0xa5a5a5a5]),
+        t = ch.time_width
+        w = ch.width
+        p = [0] * ((n*w + (n - 1)*n//2*t + 31)//32)
+        ch.pack_coeff_mu(v, p)
+        p = [_ & 0xffffffff for _ in p]
+        return p
+
+    def test_pack(self):
+        self.assertEqual(self.pack([1]), [1])
+        self.assertEqual(self.pack([1, 1 << 16]), [1, 1])
+        self.assertEqual(self.pack([1, 1 << 32]), [1, 0])
+        self.assertEqual(self.pack([0x1234, 0xa5a5a5a5]),
                          [0xa5a51234, 0xa5a5])
-        self.assertEqual(ch.pack_coeff_mu([1, 2, 3, 4]),
+        self.assertEqual(self.pack([1, 2, 3, 4]),
                          [0x20001, 0x30000, 0, 4, 0])
-        self.assertEqual(ch.pack_coeff_mu([-1, -2, -3, -4]),
+        self.assertEqual(self.pack([-1, -2, -3, -4]),
                          [0xfffeffff, 0xfffdffff, 0xffffffff,
                           0xfffffffc, 0xffffffff])
-        self.assertEqual(ch.pack_coeff_mu([0, -1, 0, -1]),
+        self.assertEqual(self.pack([0, -1, 0, -1]),
                          [0xffff0000, 0x0000ffff, 0,
                           0xffffffff, 0xffffffff])
 
@@ -171,3 +186,6 @@ class SAWGTest(unittest.TestCase):
         ch.set(.2)
         delay_mu(1*8)
         out = self.run_channel(self.rtio_manager.outputs)
+        # import matplotlib.pyplot as plt
+        # plt.plot(out)
+        # plt.show()