diff --git a/src/stub/src/arch/arm/v4t/nrv2b_d32.S b/src/stub/src/arch/arm/v4t/nrv2b_d32.S
index 0d1e2df7..e0dc8f25 100644
--- a/src/stub/src/arch/arm/v4t/nrv2b_d32.S
+++ b/src/stub/src/arch/arm/v4t/nrv2b_d32.S
@@ -46,14 +46,17 @@
 #define   jnextb0     GETBIT; bcc
 #define   jnextb1     GETBIT; bcs
 
+#ifndef PURE_THUMB
 ucl_nrv2b_decompress_32: .globl ucl_nrv2b_decompress_32  @ ARM mode
         .type ucl_nrv2b_decompress_32, %function
 /* error = (*)(char const *src, int len_src, char *dst, int *plen_dst) */
-        adr r12,1+go_thumb_n2b  @ load pc-relative address
+        adr r12,1+.thumb_nrv2b_d32  @ load pc-relative address
         bx  r12  @ enter THUMB mode
-
         .code 16  @ THUMB mode
-go_thumb_n2b:
+        .thumb_func
+#endif
+
+.thumb_nrv2b_d32:
         add r1,len,src  @ r1= eof_src;
         push {r1,r2,r3, r4,r5,r6,r7, lr}
         mov bits,#1; neg off,bits  @ off= -1 initial condition
@@ -121,8 +124,12 @@ get1b:  @ Out: Carry= next_bit
         ldrb tmp, [src],#1; orr bits,bits,tmp,lsl #3*8
         adcs bits,bits,bits  @ Carry= next_bit; set bit0 flag from CarryIn
         bx lr
+        .code 16  @ THUMB mode
 
+#ifndef PURE_THUMB
         .size ucl_nrv2b_decompress_32, .-ucl_nrv2b_decompress_32
+#endif
+
 /*
 vi:ts=8:et:nowrap
  */
diff --git a/src/stub/src/arch/arm/v4t/nrv2b_d8.S b/src/stub/src/arch/arm/v4t/nrv2b_d8.S
index c9f98a2d..f32415ee 100644
--- a/src/stub/src/arch/arm/v4t/nrv2b_d8.S
+++ b/src/stub/src/arch/arm/v4t/nrv2b_d8.S
@@ -50,12 +50,13 @@
 ucl_nrv2b_decompress_8: .globl ucl_nrv2b_decompress_8  @ ARM mode
         .type ucl_nrv2b_decompress_8, %function
 /* error = (*)(char const *src, int len_src, char *dst, int *plen_dst) */
-        adr r12,1+.go_thumb_n2b  @ load pc-relative address
+        adr r12,1+.thumb_nrv2b_d8  @ load pc-relative address
         bx  r12  @ enter THUMB mode
         .code 16  @ THUMB mode
         .thumb_func
 #endif
-.go_thumb_n2b:
+
+.thumb_nrv2b_d8:
         add r1,len,src  @ r1= eof_src;
         push {r1,r2,r3, r4,r5,r6,r7, lr}
         mov bits,#1; neg off,bits  @ off= -1 initial condition
@@ -118,7 +119,10 @@ copy_n2b:
         sub len,#1; bne copy_n2b
         b top_n2b
 
+#ifndef PURE_THUMB
         .size ucl_nrv2b_decompress_8, .-ucl_nrv2b_decompress_8
+#endif
+
 /*
 vi:ts=8:et:nowrap
  */
diff --git a/src/stub/src/arch/arm/v4t/nrv2e_d32.S b/src/stub/src/arch/arm/v4t/nrv2e_d32.S
index 79ca8b83..acafb6f1 100644
--- a/src/stub/src/arch/arm/v4t/nrv2e_d32.S
+++ b/src/stub/src/arch/arm/v4t/nrv2e_d32.S
@@ -50,12 +50,16 @@
 #define   jnextb0     GETBIT; bcc
 #define   jnextb1     GETBIT; bcs
 
+#ifndef PURE_THUMB
 ucl_nrv2e_decompress_32: .globl ucl_nrv2e_decompress_32  @ ARM mode
         .type ucl_nrv2e_decompress_32, %function
 /* error = (*)(char const *src, int len_src, char *dst, int *plen_dst) */
-        adr r12,1+thumb_nrv2e_d32; bx r12  @ enter THUMB mode
+        adr r12,1+.thumb_nrv2e_d32; bx r12  @ enter THUMB mode
         .code 16  @ THUMB mode
-thumb_nrv2e_d32:
+        .thumb_func
+#endif
+
+.thumb_nrv2e_d32:
         add r1,len,src  @ r1= eof_src;
         push {r1,r2,r3, r4,r5,r6,r7, lr}
         mov bits,#1; neg off,bits  @ off= -1 initial condition
@@ -129,8 +133,11 @@ get32:  @ In: Carry set (unchanged until final adcs)
         ldrb tmp, [src],#1; orr bits,bits,tmp,lsl #3*8
         adcs bits,bits,bits  @ Set Carry out
         bx lr
+        .code 16  @ THUMB mode
 
+#ifndef PURE_THUMB
         .size ucl_nrv2e_decompress_32, .-ucl_nrv2e_decompress_32
+#endif
 
 /*
 vi:ts=8:et:nowrap
diff --git a/src/stub/src/arch/arm/v4t/nrv2e_d8.S b/src/stub/src/arch/arm/v4t/nrv2e_d8.S
index 6206ee43..05a597e3 100644
--- a/src/stub/src/arch/arm/v4t/nrv2e_d8.S
+++ b/src/stub/src/arch/arm/v4t/nrv2e_d8.S
@@ -80,6 +80,7 @@ ucl_nrv2e_decompress_8: .globl ucl_nrv2e_decompress_8  @ ARM mode
         .code 16  @ THUMB mode
         .thumb_func
 #endif
+
 .thumb_nrv2e_d8:
         push {r2,r3, r4,r5,r6,r7, lr}
 #define sp_DST0 0  /* stack offset of original dst */
@@ -90,7 +91,7 @@ ucl_nrv2e_decompress_8: .globl ucl_nrv2e_decompress_8  @ ARM mode
         mov dstlim,tmp
 #endif  /*}*/
         mov bits,#1; neg off,bits  @ off= -1 initial condition
-        lsl bits,#31  @ 1<<31; refill next time
+        lsl bits,#31  @ 1<<31: refill next time
         mov wrnk,#5
         lsl wrnk,#8  @ 0x500  @ nrv2e M2_MAX_OFFSET
         b top_n2e
@@ -175,7 +176,10 @@ copy_n2e:
         strb tmp,[dst]; add dst,#1
         sub len,#1; bne copy_n2e
         b top_n2e
+
+#ifndef PURE_THUMB
         .size ucl_nrv2e_decompress_8, .-ucl_nrv2e_decompress_8
+#endif
 
 /*
 vi:ts=8:et:nowrap
diff --git a/src/stub/src/arm.v4t-wince.pe.S b/src/stub/src/arm.v4t-wince.pe.S
index cdd6b947..97782d14 100644
--- a/src/stub/src/arm.v4t-wince.pe.S
+++ b/src/stub/src/arm.v4t-wince.pe.S
@@ -296,19 +296,37 @@ section ProcessEnd
 
 section .ucl_nrv2b_decompress_8
 #include "arch/arm/v4t/nrv2b_d8.S"
+#undef GETBIT
+#undef wrnk
 
 section Call2B
-        bl      .go_thumb_n2b
+        bl      .thumb_nrv2b_d8
 
-#undef wrnk
-#undef GETBIT
 
 section .ucl_nrv2e_decompress_8
 #include "arch/arm/v4t/nrv2e_d8.S"
+#undef GETBIT
+#undef wrnk
 
 section Call2E
         bl      .thumb_nrv2e_d8
 
+
+#if 0
+section .ucl_nrv2b_decompress_32
+#include "arch/arm/v4t/nrv2b_d32.S"
+#undef GETBIT
+#undef wrnk
+#endif
+
+
+#if 0
+section .ucl_nrv2e_decompress_32
+#include "arch/arm/v4t/nrv2e_d32.S"
+#undef GETBIT
+#undef wrnk
+#endif
+
 ////////////////////////////////////////
 
 section CallLZMA