diff --git a/proof/crypto_scalarmult/curve25519/amd64/mulx/extracted_ct_proof.ec b/proof/crypto_scalarmult/curve25519/amd64/mulx/extracted_ct_proof.ec
index 32f4af8e..1ce2b48a 100644
--- a/proof/crypto_scalarmult/curve25519/amd64/mulx/extracted_ct_proof.ec
+++ b/proof/crypto_scalarmult/curve25519/amd64/mulx/extracted_ct_proof.ec
@@ -2,14 +2,14 @@ require import Extracted_ct.
 
 equiv jade_scalarmult_curve25519_amd64_mulx : 
   M.jade_scalarmult_curve25519_amd64_mulx ~ M.jade_scalarmult_curve25519_amd64_mulx :
-    ={qp, np, pp, M.leakages} ==> ={M.leakages}.
+    ={q, n, p, M.leakages} ==> ={M.leakages}.
 proof.
 proc; inline *; sim => />.
 qed.
 
 equiv jade_scalarmult_curve25519_amd64_mulx_base : 
   M.jade_scalarmult_curve25519_amd64_mulx_base ~ M.jade_scalarmult_curve25519_amd64_mulx_base :
-    ={qp, np, M.leakages} ==> ={M.leakages}.
+    ={q, n, M.leakages} ==> ={M.leakages}.
 proof.
 proc; inline *; sim => />.
 qed.
diff --git a/proof/crypto_scalarmult/curve25519/amd64/ref4/extracted_ct_proof.ec b/proof/crypto_scalarmult/curve25519/amd64/ref4/extracted_ct_proof.ec
index 41c0ff8f..4190ab41 100644
--- a/proof/crypto_scalarmult/curve25519/amd64/ref4/extracted_ct_proof.ec
+++ b/proof/crypto_scalarmult/curve25519/amd64/ref4/extracted_ct_proof.ec
@@ -2,14 +2,14 @@ require import Extracted_ct.
 
 equiv jade_scalarmult_curve25519_amd64_ref4 : 
   M.jade_scalarmult_curve25519_amd64_ref4 ~ M.jade_scalarmult_curve25519_amd64_ref4 :
-    ={qp, np, pp, M.leakages} ==> ={M.leakages}.
+    ={q, n, p, M.leakages} ==> ={M.leakages}.
 proof.
 proc; inline *; sim => />.
 qed.
 
 equiv jade_scalarmult_curve25519_amd64_ref4_base : 
   M.jade_scalarmult_curve25519_amd64_ref4_base ~ M.jade_scalarmult_curve25519_amd64_ref4_base :
-    ={qp, np, M.leakages} ==> ={M.leakages}.
+    ={q, n, M.leakages} ==> ={M.leakages}.
 proof.
 proc; inline *; sim => />.
 qed.
diff --git a/proof/crypto_scalarmult/curve25519/amd64/ref5/extracted_ct_proof.ec b/proof/crypto_scalarmult/curve25519/amd64/ref5/extracted_ct_proof.ec
index 99a89f59..be5a7147 100644
--- a/proof/crypto_scalarmult/curve25519/amd64/ref5/extracted_ct_proof.ec
+++ b/proof/crypto_scalarmult/curve25519/amd64/ref5/extracted_ct_proof.ec
@@ -2,14 +2,14 @@ require import Extracted_ct.
 
 equiv jade_scalarmult_curve25519_amd64_ref5 : 
   M.jade_scalarmult_curve25519_amd64_ref5 ~ M.jade_scalarmult_curve25519_amd64_ref5 :
-    ={qp, np, pp, M.leakages} ==> ={M.leakages}.
+    ={q, n, p, M.leakages} ==> ={M.leakages}.
 proof.
 proc; inline *; sim => />.
 qed.
 
 equiv jade_scalarmult_curve25519_amd64_ref5_base : 
   M.jade_scalarmult_curve25519_amd64_ref5_base ~ M.jade_scalarmult_curve25519_amd64_ref5_base :
-    ={qp, np, M.leakages} ==> ={M.leakages}.
+    ={q, n, M.leakages} ==> ={M.leakages}.
 proof.
 proc; inline *; sim => />.
 qed.
diff --git a/src/crypto_scalarmult/curve25519/amd64/common/51/decode_u5.jinc b/src/crypto_scalarmult/curve25519/amd64/common/51/decode_u5.jinc
deleted file mode 100644
index bc467d49..00000000
--- a/src/crypto_scalarmult/curve25519/amd64/common/51/decode_u5.jinc
+++ /dev/null
@@ -1,53 +0,0 @@
-inline fn __decode_u_coordinate5(reg u64[4] t) -> reg u64[5]
-{
-  reg u64[5] u;
-  reg u64 mask;
-
-  mask = 0x7ffffffffffff;
-  
-  //u[0] = t[0] & mask; // 51; 13 left
-  u[0] = t[0];
-  u[0] &= mask;
-  
-  //u[1] = (t[1] << 13) || (t[0] >> 51) & mask; // 38; 26 left
-  u[1] = t[1];
-  u[1] <<= 13;
-  t[0] >>= 51;
-  u[1] |= t[0];
-  u[1] &= mask;
-  
-  //u[2] = (t[2] << 26) || (t[1] >> 38) & mask; // 25; 39 left
-  u[2] = t[2];
-  u[2] <<= 26;
-  t[1] >>= 38;
-  u[2] |= t[1];
-  u[2] &= mask;
-  
-  //u[3] = (t[3] << 39) || (t[2] >> 25) & mask; // 12; '52' left
-  u[3] = t[3];
-  u[3] <<= 39;
-  t[2] >>= 25;
-  u[3] |= t[2];
-  u[3] &= mask;
-  
-  //u[4] = (t[3] >> 12) & mask;
-  u[4] = t[3];
-  u[4] >>= 12;
-  u[4] &= mask;
-
-  return u;
-}
-
-inline fn __decode_u_coordinate_base5() -> reg u64[5]
-{
-  reg u64[5] u;
-
-  u[0] = 9;
-  u[1] = 0;
-  u[2] = 0;
-  u[3] = 0;
-  u[4] = 0;
-
-  return u;
-}
-
diff --git a/src/crypto_scalarmult/curve25519/amd64/common/51/init_points5.jinc b/src/crypto_scalarmult/curve25519/amd64/common/51/init_points5.jinc
deleted file mode 100644
index fe705c48..00000000
--- a/src/crypto_scalarmult/curve25519/amd64/common/51/init_points5.jinc
+++ /dev/null
@@ -1,56 +0,0 @@
-inline fn __init_points5(
-  reg u64[5] initr)
-  ->
-  stack u64[5],
-  reg   u64[5],
-  stack u64[5],
-  stack u64[5]
-{
-  inline int i;
-  stack u64[5] x2 x3 z3;
-  reg u64[5] z2r;
-  reg u64 z;
-
-  ?{}, z = #set0();
-
-  x2[0] = 1;
-  z2r[0] = 0;
-  x3 = #copy(initr);
-  z3[0] = 1;
-
-  for i=1 to 5
-  { x2[i] = z;
-    z2r[i] = z;
-    z3[i] = z;
-  }
-
-  //     (1,   0, init, 1)
-  return x2, z2r, x3,  z3;
-}
-
-inline fn __init_points5_x3()
-  ->
-  stack u64[5],
-  reg   u64[5],
-  stack u64[5]
-{
-  inline int i;
-  stack u64[5] f1s f3s;
-  reg   u64[5] f2;
-  reg   u64 z;
-
-  ?{}, z = #set0();
-
-  f1s[0] = 1;
-  f2[0]  = 1;
-  f3s[0] = 1;
-
-  for i=1 to 5
-  { f1s[i] = z;
-    f2[i]  = z;
-    f3s[i] = z;
-  }
-
-  return f1s, f2, f3s;
-}
-
diff --git a/src/crypto_scalarmult/curve25519/amd64/common/51/load5.jinc b/src/crypto_scalarmult/curve25519/amd64/common/51/load5.jinc
new file mode 100644
index 00000000..9511e3c7
--- /dev/null
+++ b/src/crypto_scalarmult/curve25519/amd64/common/51/load5.jinc
@@ -0,0 +1,113 @@
+inline fn __decode_u_coordinate5(reg u64 up) -> reg u64[5]
+{
+  inline int i;
+  reg u64[4] t;
+  reg u64[5] u;
+  reg u64 mask;
+
+  for i=0 to 4
+  { t[i] = [up + 8*i]; }
+  mask = 0x7ffffffffffff;
+  
+  //u[0] = t[0] & mask; // 51; 13 left
+  u[0] = t[0];
+  u[0] &= mask;
+  
+  //u[1] = (t[1] << 13) || (t[0] >> 51) & mask; // 38; 26 left
+  u[1] = t[1];
+  u[1] <<= 13;
+  t[0] >>= 51;
+  u[1] |= t[0];
+  u[1] &= mask;
+  
+  //u[2] = (t[2] << 26) || (t[1] >> 38) & mask; // 25; 39 left
+  u[2] = t[2];
+  u[2] <<= 26;
+  t[1] >>= 38;
+  u[2] |= t[1];
+  u[2] &= mask;
+  
+  //u[3] = (t[3] << 39) || (t[2] >> 25) & mask; // 12; '52' left
+  u[3] = t[3];
+  u[3] <<= 39;
+  t[2] >>= 25;
+  u[3] |= t[2];
+  u[3] &= mask;
+  
+  //u[4] = (t[3] >> 12) & mask;
+  u[4] = t[3];
+  u[4] >>= 12;
+  u[4] &= mask;
+
+  return u;
+}
+
+inline fn __decode_u_coordinate_base5() -> reg u64[5]
+{
+  reg u64[5] u;
+
+  u[0] = 9;
+  u[1] = 0;
+  u[2] = 0;
+  u[3] = 0;
+  u[4] = 0;
+
+  return u;
+}
+
+inline fn __init_points5(
+  reg u64[5] initr)
+  ->
+  stack u64[5],
+  reg   u64[5],
+  stack u64[5],
+  stack u64[5]
+{
+  inline int i;
+  stack u64[5] x2 x3 z3;
+  reg u64[5] z2r;
+  reg u64 z;
+
+  ?{}, z = #set0();
+
+  x2[0] = 1;
+  z2r[0] = 0;
+  x3 = #copy(initr);
+  z3[0] = 1;
+
+  for i=1 to 5
+  { x2[i] = z;
+    z2r[i] = z;
+    z3[i] = z;
+  }
+
+  //     (1,   0, init, 1)
+  return x2, z2r, x3,  z3;
+}
+
+inline fn __init_points5_x3()
+  ->
+  stack u64[5],
+  reg   u64[5],
+  stack u64[5]
+{
+  inline int i;
+  stack u64[5] f1s f3s;
+  reg   u64[5] f2;
+  reg   u64 z;
+
+  ?{}, z = #set0();
+
+  f1s[0] = 1;
+  f2[0]  = 1;
+  f3s[0] = 1;
+
+  for i=1 to 5
+  { f1s[i] = z;
+    f2[i]  = z;
+    f3s[i] = z;
+  }
+
+  return f1s, f2, f3s;
+}
+
diff --git a/src/crypto_scalarmult/curve25519/amd64/common/64/decode_u4.jinc b/src/crypto_scalarmult/curve25519/amd64/common/64/decode_u4.jinc
deleted file mode 100644
index 04b1029c..00000000
--- a/src/crypto_scalarmult/curve25519/amd64/common/64/decode_u4.jinc
+++ /dev/null
@@ -1,18 +0,0 @@
-inline fn __decode_u_coordinate4(reg u64[4] u) -> reg u64[4]
-{
-  u[3] &= 0x7fffffffffffffff;
-  return u;
-}
-
-inline fn __decode_u_coordinate_base4() -> reg u64[4]
-{
-  reg u64[4] u;
-
-  u[0] = 9;
-  u[1] = 0;
-  u[2] = 0;
-  u[3] = 0;
-
-  return u;
-}
-
diff --git a/src/crypto_scalarmult/curve25519/amd64/common/64/init_points4.jinc b/src/crypto_scalarmult/curve25519/amd64/common/64/load4.jinc
similarity index 69%
rename from src/crypto_scalarmult/curve25519/amd64/common/64/init_points4.jinc
rename to src/crypto_scalarmult/curve25519/amd64/common/64/load4.jinc
index e6d72583..5f01ce07 100644
--- a/src/crypto_scalarmult/curve25519/amd64/common/64/init_points4.jinc
+++ b/src/crypto_scalarmult/curve25519/amd64/common/64/load4.jinc
@@ -1,3 +1,27 @@
+inline fn __decode_u_coordinate4(reg u64 up) -> reg u64[4]
+{
+  inline int i;
+  reg u64[4] u;
+
+  for i=0 to 4
+  { u[i] = [up + 8*i]; }
+  u[3] &= 0x7fffffffffffffff;
+
+  return u;
+}
+
+inline fn __decode_u_coordinate_base4() -> reg u64[4]
+{
+  reg u64[4] u;
+
+  u[0] = 9;
+  u[1] = 0;
+  u[2] = 0;
+  u[3] = 0;
+
+  return u;
+}
+
 inline fn __init_points4(
   reg u64[4] initr)
   ->
diff --git a/src/crypto_scalarmult/curve25519/amd64/common/decode.jinc b/src/crypto_scalarmult/curve25519/amd64/common/decode.jinc
new file mode 100644
index 00000000..19fbf18e
--- /dev/null
+++ b/src/crypto_scalarmult/curve25519/amd64/common/decode.jinc
@@ -0,0 +1,34 @@
+inline fn __decode_scalar(reg u64 kp) -> stack u8[32]
+{
+  inline int i;
+  stack u8[32] k;
+  reg u64 t;
+
+  for i=0 to 4
+  { t = [kp + 8*i];
+    k[u64 i] = t; }
+
+  k[0]  &= 0xf8;
+  k[31] &= 0x7f;
+  k[31] |= 0x40;
+
+  return k;
+}
+
+inline fn __decode_scalar_shl1(reg u64 kp) -> stack u64[4]
+{
+  inline int i;
+  reg u64[4] k;
+  stack u64[4] ks;
+
+  for i=0 to 4
+  { k[i] = [kp + 8*i]; }
+  k[3] <<= 1;
+  k[0] &= 0xfffffffffffffff8;
+  k[3] |= 0x8000000000000000;
+
+  ks = #copy(k);
+
+  return ks;
+}
+
diff --git a/src/crypto_scalarmult/curve25519/amd64/common/decode_scalar.jinc b/src/crypto_scalarmult/curve25519/amd64/common/decode_scalar.jinc
deleted file mode 100644
index ce2e08bb..00000000
--- a/src/crypto_scalarmult/curve25519/amd64/common/decode_scalar.jinc
+++ /dev/null
@@ -1,28 +0,0 @@
-inline fn __decode_scalar(reg u64[4] k) -> stack u8[32]
-{
-  inline int i;
-  stack u8[32] ks;
-
-  for i=0 to 4
-  { ks[u64 i] = k[i]; }
-
-  ks[0]  &= 0xf8;
-  ks[31] &= 0x7f;
-  ks[31] |= 0x40;
-
-  return ks;
-}
-
-inline fn __decode_scalar_shl1(reg u64[4] k) -> stack u64[4]
-{
-  stack u64[4] ks;
-
-  k[3] <<= 1;
-  k[0] &= 0xfffffffffffffff8;
-  k[3] |= 0x8000000000000000;
-
-  ks = #copy(k);
-
-  return ks;
-}
-
diff --git a/src/crypto_scalarmult/curve25519/amd64/common/load_store4.jinc b/src/crypto_scalarmult/curve25519/amd64/common/load_store4.jinc
deleted file mode 100644
index d417cfba..00000000
--- a/src/crypto_scalarmult/curve25519/amd64/common/load_store4.jinc
+++ /dev/null
@@ -1,19 +0,0 @@
-inline fn __load4(reg u64 p) -> reg u64[4]
-{
-  inline int i;
-  reg u64[4] a;
-
-  for i=0 to 4
-  { a[i] = [p + 8*i]; }
-
-  return a;
-}
-
-inline fn __store4(reg u64 p, reg u64[4] a)
-{
-  inline int i;
-
-  for i=0 to 4
-  { [p + 8*i] = a[i]; }
-}
-
diff --git a/src/crypto_scalarmult/curve25519/amd64/mulx/curve25519.jinc b/src/crypto_scalarmult/curve25519/amd64/mulx/curve25519.jinc
index 280c44f6..d6eae363 100644
--- a/src/crypto_scalarmult/curve25519/amd64/mulx/curve25519.jinc
+++ b/src/crypto_scalarmult/curve25519/amd64/mulx/curve25519.jinc
@@ -1,11 +1,10 @@
 from Jade require "crypto_scalarmult/curve25519/amd64/common/bit.jinc"
-from Jade require "crypto_scalarmult/curve25519/amd64/common/decode_scalar.jinc"
+from Jade require "crypto_scalarmult/curve25519/amd64/common/decode.jinc"
 
-from Jade require "crypto_scalarmult/curve25519/amd64/common/64/decode_u4.jinc"
-from Jade require "crypto_scalarmult/curve25519/amd64/common/64/init_points4.jinc"
 from Jade require "crypto_scalarmult/curve25519/amd64/common/64/add4.jinc"
 from Jade require "crypto_scalarmult/curve25519/amd64/common/64/sub4.jinc"
 from Jade require "crypto_scalarmult/curve25519/amd64/common/64/cswap4.jinc"
+from Jade require "crypto_scalarmult/curve25519/amd64/common/64/load4.jinc"
 from Jade require "crypto_scalarmult/curve25519/amd64/common/64/tobytes4.jinc"
 
 require "mul4.jinc"
@@ -142,27 +141,38 @@ inline fn __curve25519_internal_mulx(stack u8[32] k, reg u64[4] u) -> reg u64[4]
   return r;
 }
 
-inline fn __curve25519_mulx(reg u64[4] _k _u) -> reg u64[4]
+inline fn __curve25519_mulx(reg u64 rp kp up)
 {
+  inline int i;
   stack u8[32] k;
   reg u64[4] u r;
+  stack u64 rps;
 
-  k = __decode_scalar(_k);
-  u = __decode_u_coordinate4(_u);
+  rps = rp;
+
+  k = __decode_scalar(kp);
+  u = __decode_u_coordinate4(up);
   r = __curve25519_internal_mulx(k, u);
 
-  return r;
+  rp = rps;
+  for i=0 to 4
+  { [rp + 8*i] = r[i]; }
 }
 
-inline fn __curve25519_mulx_base(reg u64[4] _k) -> reg u64[4]
+inline fn __curve25519_mulx_base(reg u64 rp kp)
 {
+  inline int i;
   stack u8[32] k;
   reg u64[4] u r;
+  stack u64 rps;
+
+  rps = rp;
 
-  k = __decode_scalar(_k);
+  k = __decode_scalar(kp);
   u = __decode_u_coordinate_base4();
   r = __curve25519_internal_mulx(k, u);
 
-  return r;
+  rp = rps;
+  for i=0 to 4
+  { [rp + 8*i] = r[i]; }
 }
-
diff --git a/src/crypto_scalarmult/curve25519/amd64/mulx/invert4.jinc b/src/crypto_scalarmult/curve25519/amd64/mulx/invert4.jinc
index 22cce58b..727e3bca 100644
--- a/src/crypto_scalarmult/curve25519/amd64/mulx/invert4.jinc
+++ b/src/crypto_scalarmult/curve25519/amd64/mulx/invert4.jinc
@@ -1,6 +1,9 @@
 require "mul4.jinc"
 require "sqr4.jinc"
 
+// TODO:
+// - rewrite invert4 (& aux. functions) once reg arrays are supported in non-inline functions
+// - this should reduce the size of scalarmult.s to half
 inline fn __invert4(reg u64[4] f) -> reg u64[4]
 {
   reg u32 i;
@@ -10,90 +13,90 @@ inline fn __invert4(reg u64[4] f) -> reg u64[4]
   fs = #copy(f);
 
   // z2 = z1^2^1
-  t0  = _sqr4_rr_(f);
+  t0  = __sqr4_rr(f);
   t0s = #copy(t0);
 
   // z8 = z2^2^2
-  t1  = _sqr4_rr_(t0);
-  t1  = _sqr4_rr_(t1);
+  t1  = __sqr4_rr(t0);
+  t1  = __sqr4_rr(t1);
 
   // z9 = z1*z8
-  t1  = _mul4_rsr_(fs,t1);
+  t1  = __mul4_rsr(fs,t1);
   t1s = #copy(t1);
 
   // z11 = z2*z9
-  t0  = _mul4_rsr_(t0s,t1);
+  t0  = __mul4_rsr(t0s,t1);
   t0s = #copy(t0);
 
   // z22 = z11^2^1
-  t2 = _sqr4_rr_(t0);
+  t2 = __sqr4_rr(t0);
 
   // z_5_0 = z9*z22
-  t1  = _mul4_rsr_(t1s,t2);
+  t1  = __mul4_rsr(t1s,t2);
   t1s = #copy(t1);
 
   // z_10_5 = z_5_0^2^5
-  t2 = _sqr4_rr_(t1);
+  t2 = __sqr4_rr(t1);
   i = 4/2;
-  t2 = _it_sqr4_x2_(t2, i);
+  t2 = __it_sqr4_x2(t2, i);
   t2s = #copy(t2);
 
   // z_10_0 = z_10_5*z_5_0
-  t1  = _mul4_rsr_(t1s,t2);
+  t1  = __mul4_rsr(t1s,t2);
   t1s = #copy(t1);
 
   // z_20_10 = z_10_0^2^10
   i = 10/2;
-  t2 = _it_sqr4_x2_(t1, i);
+  t2 = __it_sqr4_x2(t1, i);
 
   // z_20_0 = z_20_10*z_10_0
-  t2  = _mul4_rsr_(t1s,t2);
+  t2  = __mul4_rsr(t1s,t2);
   t2s = #copy(t2);
 
   // z_40_20 = z_20_0^2^20
   i = 20/2;
-  t3 = _it_sqr4_x2_(t2, i);
+  t3 = __it_sqr4_x2(t2, i);
 
   // z_40_0 = z_40_20*z_20_0
-  t2  = _mul4_rsr_(t2s,t3);
+  t2  = __mul4_rsr(t2s,t3);
 
   // z_50_10 = z_40_0^2^10
   i = 10/2;
-  t2 = _it_sqr4_x2_(t2, i);
+  t2 = __it_sqr4_x2(t2, i);
 
   // z_50_0 = z_50_10*z_10_0
-  t1  = _mul4_rsr_(t1s,t2);
+  t1  = __mul4_rsr(t1s,t2);
   t1s = #copy(t1);
 
   // z_100_50 = z_50_0^2^50
   i = 50/2;
-  t2 = _it_sqr4_x2_(t1, i);
+  t2 = __it_sqr4_x2(t1, i);
 
   // z_100_0 = z_100_50*z_50_0
-  t2  = _mul4_rsr_(t1s,t2);
+  t2  = __mul4_rsr(t1s,t2);
   t2s = #copy(t2);
 
   // z_200_100 = z_100_0^2^100
   i = 100/2;
-  t3 = _it_sqr4_x2_(t2, i);
+  t3 = __it_sqr4_x2(t2, i);
 
   // z_200_0 = z_200_100*z_100_0
-  t2  = _mul4_rsr_(t2s,t3);
+  t2  = __mul4_rsr(t2s,t3);
 
   // z_250_50 = z_200_0^2^50
   i = 50/2;
-  t2 = _it_sqr4_x2_(t2, i);
+  t2 = __it_sqr4_x2(t2, i);
 
   // z_250_0 = z_250_50*z_50_0
-  t1  = _mul4_rsr_(t1s,t2);
+  t1  = __mul4_rsr(t1s,t2);
 
   // z_255_5 = z_250_0^2^5
   i = 4/2;
-  t1 = _it_sqr4_x2_(t1, i);
-  t1 = _sqr4_rr_(t1);
+  t1 = __it_sqr4_x2(t1, i);
+  t1 = __sqr4_rr(t1);
 
   // z_255_21 = z_255_5*z11
-  t1 = _mul4_rsr_(t0s,t1);
+  t1 = __mul4_rsr(t0s,t1);
 
   return t1;
 }
diff --git a/src/crypto_scalarmult/curve25519/amd64/mulx/mul4.jinc b/src/crypto_scalarmult/curve25519/amd64/mulx/mul4.jinc
index 54132927..0cbb33e8 100644
--- a/src/crypto_scalarmult/curve25519/amd64/mulx/mul4.jinc
+++ b/src/crypto_scalarmult/curve25519/amd64/mulx/mul4.jinc
@@ -168,54 +168,6 @@ inline fn __mul4_rsr(stack u64[4] fs, reg u64[4] g) -> reg u64[4]
   return h;
 }
 
-inline fn __mul4_rpr(reg ptr u64[4] fp, reg u64[4] g) -> reg u64[4]
-{
-  reg bool cf of;
-  reg u64[4] h r;
-  reg u64 _38 f z;
-
-  of, cf, _, _, _, z = #set0();
-
-  f = fp[0];
-  h, r, cf, of = __mul4_c0(      f, g, z, cf, of);
-
-  f = fp[1];
-  h, r, cf, of = __mul4_c1(h, r, f, g, z, cf, of);
-
-  f = fp[2];
-  h, r, cf, of = __mul4_c2(h, r, f, g, z, cf, of);
-
-  f = fp[3];
-  h, r, cf, of = __mul4_c3(h, r, f, g, z, cf, of);
-
-  _38 = 38;
-  h = __reduce4(h, r, _38, z, cf, of);
-
-  return h;
-}
-
-fn _mul4_rpr(reg ptr u64[4] fp, reg u64[4] g) -> reg u64[4]
-{
-  reg u64[4] h;
-
-  h = __mul4_rpr(fp, g);
-
-  return h;
-}
-
-inline fn _mul4_rsr_(stack u64[4] _fs, reg u64[4] _g) -> reg u64[4]
-{
-  reg ptr u64[4] fp;
-  reg u64[4] _h h g;
-
-  fp = _fs;
-  g = #copy(_g);
-  h = _mul4_rpr(fp, g);
-  _h = #copy(h);
-
-  return _h;
-}
-
 inline fn __mul4_ssr(stack u64[4] fs, reg u64[4] g) -> stack u64[4]
 {
   stack u64[4] hs;
diff --git a/src/crypto_scalarmult/curve25519/amd64/mulx/scalarmult.jazz b/src/crypto_scalarmult/curve25519/amd64/mulx/scalarmult.jazz
index d6146eca..04672100 100644
--- a/src/crypto_scalarmult/curve25519/amd64/mulx/scalarmult.jazz
+++ b/src/crypto_scalarmult/curve25519/amd64/mulx/scalarmult.jazz
@@ -1,39 +1,17 @@
-from Jade require "crypto_scalarmult/curve25519/amd64/common/load_store4.jinc"
 require "curve25519.jinc"
 
-export fn jade_scalarmult_curve25519_amd64_mulx(reg u64 qp np pp) -> reg u64
+export fn jade_scalarmult_curve25519_amd64_mulx(reg u64 q n p) -> reg u64
 {
   reg u64 r;
-  stack u64 qps;
-  reg u64[4] q n p;
-
-  qps = qp;
-  n   = __load4(np);
-  p   = __load4(pp);
-
-  q = __curve25519_mulx(n, p);
-
-  qp = qps;
-  __store4(qp, q);
-
+  __curve25519_mulx(q, n, p);
   ?{}, r = #set0();
   return r;
 }
 
-export fn jade_scalarmult_curve25519_amd64_mulx_base(reg u64 qp np) -> reg u64
+export fn jade_scalarmult_curve25519_amd64_mulx_base(reg u64 q n) -> reg u64
 {
   reg u64 r;
-  stack u64 qps;
-  reg u64[4] q n;
-
-  qps = qp;
-  n   = __load4(np);
-
-  q = __curve25519_mulx_base(n);
-
-  qp = qps;
-  __store4(qp, q);
-
+  __curve25519_mulx_base(q, n);
   ?{}, r = #set0();
   return r;
 }
diff --git a/src/crypto_scalarmult/curve25519/amd64/mulx/sqr4.jinc b/src/crypto_scalarmult/curve25519/amd64/mulx/sqr4.jinc
index f659aa9d..46fd5c62 100644
--- a/src/crypto_scalarmult/curve25519/amd64/mulx/sqr4.jinc
+++ b/src/crypto_scalarmult/curve25519/amd64/mulx/sqr4.jinc
@@ -29,7 +29,7 @@ inline fn __sqr4_rr(reg u64[4] f) -> reg u64[4]
 
   of, h[3]     = #ADOX ( h[3], t[3], of );
   cf, r[0]     = #ADCX ( r[0], t[4], cf );
-
+  
   (r[1], t[4]) = #MULX ( fx,   f[3]     ); // f1*f3
      of, r[0]  = #ADOX ( r[0], t[4], of );
 
@@ -81,24 +81,6 @@ inline fn __sqr4_rr(reg u64[4] f) -> reg u64[4]
   return h;
 }
 
-fn _sqr4_rr(reg u64[4] f) -> reg u64[4]
-{
-  reg u64[4] h;
-  h = __sqr4_rr(f);
-  return h;
-}
-
-inline fn _sqr4_rr_(reg u64[4] _f) -> reg u64[4]
-{
-  reg u64[4] _h h f;
-
-  f = #copy(_f);
-  h = _sqr4_rr(f);
-  _h = #copy(h);
-
-  return _h;
-}
-
 inline fn __it_sqr4_x2(reg u64[4] f, reg u32 i) -> reg u64[4]
 {
   reg bool zf;
@@ -118,21 +100,6 @@ inline fn __it_sqr4_x2(reg u64[4] f, reg u32 i) -> reg u64[4]
   return f;
 }
 
-fn _it_sqr4_x2(reg u64[4] f, reg u32 i) -> reg u64[4]
-{
-  f = __it_sqr4_x2(f, i);
-  return f;
-}
-
-inline fn _it_sqr4_x2_(reg u64[4] _f, reg u32 i) -> reg u64[4]
-{
-  reg u64[4] f;
-  f = #copy(_f);
-  f = _it_sqr4_x2(f, i);
-  return f;
-}
-
-
 inline fn __sqr4_ss(stack u64[4] fs) -> stack u64[4]
 {
   stack u64[4] hs;
diff --git a/src/crypto_scalarmult/curve25519/amd64/ref4/curve25519.jinc b/src/crypto_scalarmult/curve25519/amd64/ref4/curve25519.jinc
index 47c40757..4663cfcd 100644
--- a/src/crypto_scalarmult/curve25519/amd64/ref4/curve25519.jinc
+++ b/src/crypto_scalarmult/curve25519/amd64/ref4/curve25519.jinc
@@ -1,11 +1,10 @@
 from Jade require "crypto_scalarmult/curve25519/amd64/common/bit.jinc"
-from Jade require "crypto_scalarmult/curve25519/amd64/common/decode_scalar.jinc"
+from Jade require "crypto_scalarmult/curve25519/amd64/common/decode.jinc"
 
-from Jade require "crypto_scalarmult/curve25519/amd64/common/64/decode_u4.jinc"
-from Jade require "crypto_scalarmult/curve25519/amd64/common/64/init_points4.jinc"
 from Jade require "crypto_scalarmult/curve25519/amd64/common/64/add4.jinc"
 from Jade require "crypto_scalarmult/curve25519/amd64/common/64/sub4.jinc"
 from Jade require "crypto_scalarmult/curve25519/amd64/common/64/cswap4.jinc"
+from Jade require "crypto_scalarmult/curve25519/amd64/common/64/load4.jinc"
 from Jade require "crypto_scalarmult/curve25519/amd64/common/64/tobytes4.jinc"
 
 require "mul4.jinc"
@@ -143,27 +142,38 @@ inline fn __curve25519_internal_ref4(stack u8[32] k, reg u64[4] u) -> reg u64[4]
   return r;
 }
 
-inline fn __curve25519_ref4(reg u64[4] _k _u) -> reg u64[4]
+inline fn __curve25519_ref4(reg u64 rp kp up)
 {
+  inline int i;
   stack u8[32] k;
   reg u64[4] u r;
+  stack u64 rps;
 
-  k = __decode_scalar(_k);
-  u = __decode_u_coordinate4(_u);
+  rps = rp;
+
+  k = __decode_scalar(kp);
+  u = __decode_u_coordinate4(up);
   r = __curve25519_internal_ref4(k, u);
 
-  return r;
+  rp = rps;
+  for i=0 to 4
+  { [rp + 8*i] = r[i]; }
 }
 
-inline fn __curve25519_ref4_base(reg u64[4] _k) -> reg u64[4]
+inline fn __curve25519_ref4_base(reg u64 rp kp)
 {
+  inline int i;
   stack u8[32] k;
   reg u64[4] u r;
+  stack u64 rps;
+
+  rps = rp;
 
-  k = __decode_scalar(_k);
+  k = __decode_scalar(kp);
   u = __decode_u_coordinate_base4();
   r = __curve25519_internal_ref4(k, u);
 
-  return r;
+  rp = rps;
+  for i=0 to 4
+  { [rp + 8*i] = r[i]; }
 }
-
diff --git a/src/crypto_scalarmult/curve25519/amd64/ref4/scalarmult.jazz b/src/crypto_scalarmult/curve25519/amd64/ref4/scalarmult.jazz
index 7750b8c6..9fb9ac00 100644
--- a/src/crypto_scalarmult/curve25519/amd64/ref4/scalarmult.jazz
+++ b/src/crypto_scalarmult/curve25519/amd64/ref4/scalarmult.jazz
@@ -1,39 +1,17 @@
-from Jade require "crypto_scalarmult/curve25519/amd64/common/load_store4.jinc"
 require "curve25519.jinc"
 
-export fn jade_scalarmult_curve25519_amd64_ref4(reg u64 qp np pp) -> reg u64
+export fn jade_scalarmult_curve25519_amd64_ref4(reg u64 q n p) -> reg u64
 {
   reg u64 r;
-  stack u64 qps;
-  reg u64[4] q n p;
-
-  qps = qp;
-  n   = __load4(np);
-  p   = __load4(pp);
-
-  q = __curve25519_ref4(n, p);
-
-  qp = qps;
-  __store4(qp, q);
-
+  __curve25519_ref4(q, n, p);
   ?{}, r = #set0();
   return r;
 }
 
-export fn jade_scalarmult_curve25519_amd64_ref4_base(reg u64 qp np) -> reg u64
+export fn jade_scalarmult_curve25519_amd64_ref4_base(reg u64 q n) -> reg u64
 {
   reg u64 r;
-  stack u64 qps;
-  reg u64[4] q n;
-
-  qps = qp;
-  n   = __load4(np);
-
-  q = __curve25519_ref4_base(n);
-
-  qp = qps;
-  __store4(qp, q);
-
+  __curve25519_ref4_base(q, n);
   ?{}, r = #set0();
   return r;
 }
diff --git a/src/crypto_scalarmult/curve25519/amd64/ref5/curve25519.jinc b/src/crypto_scalarmult/curve25519/amd64/ref5/curve25519.jinc
index 6963fcde..681969bd 100644
--- a/src/crypto_scalarmult/curve25519/amd64/ref5/curve25519.jinc
+++ b/src/crypto_scalarmult/curve25519/amd64/ref5/curve25519.jinc
@@ -1,11 +1,10 @@
 from Jade require "crypto_scalarmult/curve25519/amd64/common/bit.jinc"
-from Jade require "crypto_scalarmult/curve25519/amd64/common/decode_scalar.jinc"
+from Jade require "crypto_scalarmult/curve25519/amd64/common/decode.jinc"
 
-from Jade require "crypto_scalarmult/curve25519/amd64/common/51/decode_u5.jinc"
-from Jade require "crypto_scalarmult/curve25519/amd64/common/51/init_points5.jinc"
 from Jade require "crypto_scalarmult/curve25519/amd64/common/51/add5.jinc"
 from Jade require "crypto_scalarmult/curve25519/amd64/common/51/sub5.jinc"
 from Jade require "crypto_scalarmult/curve25519/amd64/common/51/cswap5.jinc"
+from Jade require "crypto_scalarmult/curve25519/amd64/common/51/load5.jinc"
 from Jade require "crypto_scalarmult/curve25519/amd64/common/51/tobytes5.jinc"
 
 require "mul5.jinc"
@@ -143,29 +142,40 @@ inline fn __curve25519_internal_ref5(stack u8[32] k, reg u64[5] u) -> reg u64[4]
   return r;
 }
 
-inline fn __curve25519_ref5(reg u64[4] _k _u) -> reg u64[4]
+inline fn __curve25519_ref5(reg u64 rp kp up)
 {
+  inline int i;
   stack u8[32] k;
   reg u64[5] u;
   reg u64[4] r;
+  stack u64 rps;
 
-  k = __decode_scalar(_k);
-  u = __decode_u_coordinate5(_u);
+  rps = rp;
+  k = __decode_scalar(kp);
+  u = __decode_u_coordinate5(up);
   r = __curve25519_internal_ref5(k, u);
 
-  return r;
+  rp = rps;
+  for i=0 to 4
+  { [rp + 8*i] = r[i]; }
 }
 
-inline fn __curve25519_ref5_base(reg u64[4] _k) -> reg u64[4]
+inline fn __curve25519_ref5_base(reg u64 rp kp)
 {
+  inline int i;
   stack u8[32] k;
   reg u64[5] u;
   reg u64[4] r;
+  stack u64 rps;
 
-  k = __decode_scalar(_k);
+  rps = rp;
+  k = __decode_scalar(kp);
   u = __decode_u_coordinate_base5();
   r = __curve25519_internal_ref5(k, u);
 
-  return r;
+  rp = rps;
+  for i=0 to 4
+  { [rp + 8*i] = r[i]; }
 }
 
+
diff --git a/src/crypto_scalarmult/curve25519/amd64/ref5/scalarmult.jazz b/src/crypto_scalarmult/curve25519/amd64/ref5/scalarmult.jazz
index 88556ba3..f171a571 100644
--- a/src/crypto_scalarmult/curve25519/amd64/ref5/scalarmult.jazz
+++ b/src/crypto_scalarmult/curve25519/amd64/ref5/scalarmult.jazz
@@ -1,40 +1,19 @@
-from Jade require "crypto_scalarmult/curve25519/amd64/common/load_store4.jinc"
 require "curve25519.jinc"
 
-export fn jade_scalarmult_curve25519_amd64_ref5(reg u64 qp np pp) -> reg u64
+export fn jade_scalarmult_curve25519_amd64_ref5(reg u64 q n p) -> reg u64
 {
   reg u64 r;
-  stack u64 qps;
-  reg u64[4] q n p;
-
-  qps = qp;
-  n   = __load4(np);
-  p   = __load4(pp);
-
-  q = __curve25519_ref5(n, p);
-
-  qp = qps;
-  __store4(qp, q);
-
+  __curve25519_ref5(q, n, p);
   ?{}, r = #set0();
   return r;
 }
 
-export fn jade_scalarmult_curve25519_amd64_ref5_base(reg u64 qp np) -> reg u64
+export fn jade_scalarmult_curve25519_amd64_ref5_base(reg u64 q n) -> reg u64
 {
   reg u64 r;
-  stack u64 qps;
-  reg u64[4] q n;
-
-  qps = qp;
-  n   = __load4(np);
-
-  q = __curve25519_ref5_base(n);
-
-  qp = qps;
-  __store4(qp, q);
-
+  __curve25519_ref5_base(q, n);
   ?{}, r = #set0();
   return r;
 }
 
+
diff --git a/test/Makefile b/test/Makefile
index 33545b61..e5ec002c 100644
--- a/test/Makefile
+++ b/test/Makefile
@@ -111,15 +111,13 @@ COMPILE     = $(CC) $(CFLAGS) -o $@ $(DEFINE) $(DNAMESPACES) $(INCLUDES) crypto_
 COMPILE_P   = $(CC) $(CFLAGS) -o $@ $(DEFINE) $(DNAMESPACES) $(INCLUDES) crypto_$(OPERATION)/$(@F).c $(PRINT)        $(RANDSRC) $(CIL)
 
 # --------------------------------------------------------------------
-.PHONY: __phony default compile-and-run reporter-and-err
+.PHONY: __phony default main
 
-default: compile-and-run
+default: main
 
-compile-and-run:
+main:
 	$(MAKE) CI=1 -C $(SRC) all
 	$(MAKE) CI=1 all
-
-reporter-and-err:
 	$(MAKE) CI=1 reporter
 	$(MAKE) CI=1 err