diff --git a/src/stub/src/arch/i086/cc_test.c b/src/stub/src/arch/i086/cc_test.c
index a945a33a..b60d3fea 100644
--- a/src/stub/src/arch/i086/cc_test.c
+++ b/src/stub/src/arch/i086/cc_test.c
@@ -103,9 +103,6 @@ void __pascal p4fshlv_v(unsigned char v, uint32_t __far *a) { *a <<= v; }
 uint32_t __pascal p4nshlv(unsigned char v, uint32_t __near *a) { return *a <<= v; }
 uint32_t __pascal p4fshlv(unsigned char v, uint32_t __far *a) { return *a <<= v; }
 
-uint32_t __cdecl shlv_2(uint16_t h, uint16_t l, unsigned v)
-{ uint16_t x = l >> (16 - v); l <<= v; h <<= v; h |= x; return h * 65536ul + l; }
-
 hptrdiff_t __cdecl hptr2int(hptr a) { return (hptrdiff_t) a; }
 hptr __cdecl int2hptr(hptrdiff_t a) { return (hptr) a; }
 
diff --git a/src/stub/src/arch/i086/cleanasm.py b/src/stub/src/arch/i086/cleanasm.py
index e54b5fef..eeb4306d 100644
--- a/src/stub/src/arch/i086/cleanasm.py
+++ b/src/stub/src/arch/i086/cleanasm.py
@@ -195,29 +195,42 @@ def main(argv):
         if opts.call_rewrite and inst in ["call"]:
             k, v = parse_label(inst, args)
             if v[:2] == [1, 2]:     # external 2-byte
+                if k == "__aNahdiff":
+                    s = [
+                        ["push", "word ptr [bp+8]"],
+                        ["push", "word ptr [bp+6]"],
+                        ["push", r"word ptr \[bp([+-]\d+)\]$"],
+                        ["push", r"word ptr \[bp([+-]\d+)\]$"],
+                    ]
+                    dpos = omatch(i-1, -4, s)
+                    if dpos:
+                        orewrite_inst(i, "*DEL*", "", dpos)
+                        continue
                 if k in ["__LMUL", "__U4M",]:
+                    s1 = [
+                        ["mov",  "bx,0x300"],
+                        ["xor",  "cx,cx"],
+                    ]
+                    s2 = [
+                        ["shl",  "ax,1"],
+                        ["rcl",  "dx,1"],
+                    ]
+                    dpos1 = omatch(i-1, -2, s1)
+                    dpos2 = omatch(i+1,  2, s2)
+                    if dpos1 and dpos2:
+                        orewrite_inst(i, "M_U4M_dxax_0x0600", "", dpos1 + dpos2)
+                        continue
                     s = [
                         ["mov",  "bx,word ptr [bx]"],
                         ["xor",  "cx,cx"],
                     ]
                     dpos = omatch(i-1, -2, s, debug=0)
                     if 0 and dpos:
-                        orewrite_inst(i, "M_LMUL_dxax_00bx_ptr", "", dpos)
+                        orewrite_inst(i, "M_U4M_dxax_00bx_ptr", "", dpos)
                         continue
                     dpos = omatch(i-1, -1, s)
                     if dpos:
-                        orewrite_inst(i, "M_LMUL_dxax_00bx", "", dpos)
-                        continue
-                if k == "__aNahdiff":
-                    s = [
-                        ["push", "word ptr [bp+8]"],
-                        ["push", "word ptr [bp+6]"],
-                        ["push", "word ptr [bp-66]"],
-                        ["push", "word ptr [bp-68]"],
-                    ]
-                    dpos = omatch(i-1, -4, s)
-                    if dpos:
-                        orewrite_inst(i, "*DEL*", "", dpos)
+                        orewrite_inst(i, "M_U4M_dxax_00bx", "", dpos)
                         continue
                 if k == "__PIA":
                     s = [
@@ -257,6 +270,36 @@ def main(argv):
             if dpos:
                 orewrite_inst(i, "M_shld_8", "", dpos)
                 continue
+            s1 = [
+                ["mov",  r"^c[lx],0x8$"],
+                ["shl",  "si,1"],
+                ["rcl",  "di,1"],
+            ]
+            s2 = [
+                ["les",  r"^bx,dword ptr \[bp([+-]\d+)\]$"],
+            ]
+            dpos1 = omatch(i-1, -3, s1)
+            dpos2 = omatch(i+1,  1, s2)
+            if 1 and dpos1 and dpos2:
+                # bx and cx are free for use
+                orewrite_inst(i, "M_shld_disi_8_bxcx", "", dpos1)
+                continue
+            s1 = [
+                ["mov",  "ax,si"],
+                ["mov",  r"^c[lx],0x8$"],
+                ["shl",  "ax,1"],
+                ["rcl",  "di,1"],
+            ]
+            s2 = [
+                ["mov",  "si,ax"],
+                ["les",  r"^bx,dword ptr \[bp([+-]\d+)\]$"],
+            ]
+            dpos1 = omatch(i-1, -4, s1)
+            dpos2 = omatch(i+1,  2, s2)
+            if 1 and dpos1 and dpos2:
+                # bx and cx are free for use
+                orewrite_inst(i, "M_shld_diax_8_bxcx", "", dpos1[-3:])
+                continue
             s1 = [
                 ["mov",  r"^c[lx],0x8$"],
                 ["shl",  r"^word ptr \[bp([+-]\d+)\],1$"],
diff --git a/src/stub/src/arch/i086/lzma_d_cf.S b/src/stub/src/arch/i086/lzma_d_cf.S
index 20667d02..aa4699af 100644
--- a/src/stub/src/arch/i086/lzma_d_cf.S
+++ b/src/stub/src/arch/i086/lzma_d_cf.S
@@ -119,14 +119,14 @@
         M_PTC_JNE .Lf005
 .Lf007:
         mov     ax,0x1
-        jmp     .Lf139
+        jmp     .Lf125
 .Lf008:
         mov     ax,word ptr [bp-18]
         cmp     ax,word ptr [bp+26]
         jb      .Lf011
         je      .Lf010
 .Lf009:
-        jmp     .Lf136
+        jmp     .Lf122
 .Lf010:
         mov     ax,word ptr [bp-20]
         cmp     ax,word ptr [bp+24]
@@ -143,18 +143,14 @@
         add     dx,ax
         mov     word ptr [bp-14],dx
         cmp     di,0x100
-        jae     .Lf013
+        jae     .Lf012
         mov     ax,word ptr [bp-12]
         mov     dx,word ptr [bp-10]
         mov     bx,word ptr [bp-30]
         mov     cx,word ptr [bp-28]
         M_PTC
         je      .Lf007
-        mov     cx,0x8
-.Lf012:
-        shl     si,1
-        rcl     di,1
-        loop    .Lf012
+        M_shld_disi_8_bxcx
         les     bx,dword ptr [bp-12]
         mov     al,byte ptr es:[bx]
         mov     byte ptr [bp-90],al
@@ -171,24 +167,24 @@
         M_PIA1
         mov     word ptr [bp-12],ax
         mov     word ptr [bp-10],dx
-.Lf013:
+.Lf012:
         M_shrd_11_disi_bp -96 -98
         mov     bx,word ptr [bp-14]
         mov     bx,word ptr [bx]
-        M_LMUL_dxax_00bx
+        M_U4M_dxax_00bx
         mov     word ptr [bp-8],ax
         mov     word ptr [bp-94],dx
         mov     ax,word ptr [bp-4]
         cmp     ax,dx
-        jb      .Lf016
-        je      .Lf015
+        jb      .Lf015
+        je      .Lf014
+.Lf013:
+        jmp     .Lf034
 .Lf014:
-        jmp     .Lf037
-.Lf015:
         mov     ax,word ptr [bp-6]
         cmp     ax,word ptr [bp-8]
-        jae     .Lf014
-.Lf016:
+        jae     .Lf013
+.Lf015:
         mov     word ptr [bp-26],0x1
         mov     si,word ptr [bp-8]
         mov     di,dx
@@ -211,26 +207,23 @@
         mov     dx,word ptr [bp-18]
         and     dx,word ptr [bp-84]
         mov     cx,word ptr [bp-78]
-        jcxz    .Lf018
-.Lf017:
+        jcxz    .Lf017
+.Lf016:
         shl     ax,1
         rcl     dx,1
-        loop    .Lf017
-.Lf018:
+        loop    .Lf016
+.Lf017:
         add     ax,word ptr [bp-90]
         adc     dx,bx
-        mov     bx,0x300
-        M_LMUL_dxax_00bx
-        shl     ax,1
-        rcl     dx,1
+        M_U4M_dxax_0x0600
         mov     dx,word ptr [bp-34]
         add     dx,0xe6c
         add     dx,ax
         mov     word ptr [bp-14],dx
         cmp     word ptr [bp-16],0x7
-        jge     .Lf019
-        jmp     .Lf027
-.Lf019:
+        jge     .Lf018
+        jmp     .Lf025
+.Lf018:
         mov     bx,word ptr [bp-20]
         sub     bx,word ptr [bp-102]
         mov     cx,word ptr [bp-18]
@@ -243,7 +236,7 @@
         mov     al,byte ptr es:[bx]
         xor     ah,ah
         mov     word ptr [bp-76],ax
-.Lf020:
+.Lf019:
         shl     word ptr [bp-76],1
         mov     ax,word ptr [bp-76]
         xor     al,al
@@ -258,19 +251,15 @@
         add     dx,ax
         mov     word ptr [bp-46],dx
         cmp     di,0x100
-        jae     .Lf023
+        jae     .Lf021
         mov     ax,word ptr [bp-12]
         mov     dx,word ptr [bp-10]
         mov     bx,word ptr [bp-30]
         mov     cx,word ptr [bp-28]
-        M_PTC_JNE .Lf021
+        M_PTC_JNE .Lf020
         jmp     .Lf007
-.Lf021:
-        mov     cx,0x8
-.Lf022:
-        shl     si,1
-        rcl     di,1
-        loop    .Lf022
+.Lf020:
+        M_shld_disi_8_bxcx
         les     bx,dword ptr [bp-12]
         mov     al,byte ptr es:[bx]
         mov     byte ptr [bp-90],al
@@ -287,21 +276,21 @@
         M_PIA1
         mov     word ptr [bp-12],ax
         mov     word ptr [bp-10],dx
-.Lf023:
+.Lf021:
         M_shrd_11_disi_bp -96 -98
         mov     bx,word ptr [bp-46]
         mov     bx,word ptr [bx]
-        M_LMUL_dxax_00bx
+        M_U4M_dxax_00bx
         mov     word ptr [bp-8],ax
         mov     word ptr [bp-94],dx
         mov     ax,word ptr [bp-4]
         cmp     ax,dx
-        jb      .Lf024
-        jne     .Lf025
+        jb      .Lf022
+        jne     .Lf023
         mov     ax,word ptr [bp-6]
         cmp     ax,word ptr [bp-8]
-        jae     .Lf025
-.Lf024:
+        jae     .Lf023
+.Lf022:
         mov     si,word ptr [bp-8]
         mov     di,dx
         mov     ax,0x800
@@ -312,9 +301,9 @@
         add     word ptr [bx],ax
         shl     word ptr [bp-26],1
         cmp     word ptr [bp-66],0x0
-        jne     .Lf027
-        jmp     .Lf026
-.Lf025:
+        jne     .Lf025
+        jmp     .Lf024
+.Lf023:
         sub     si,word ptr [bp-8]
         sbb     di,dx
         mov     ax,word ptr [bp-8]
@@ -330,35 +319,31 @@
         inc     ax
         mov     word ptr [bp-26],ax
         cmp     word ptr [bp-66],0x0
-        je      .Lf027
-.Lf026:
+        je      .Lf025
+.Lf024:
         cmp     word ptr [bp-26],0x100
-        jge     .Lf027
-        jmp     .Lf020
-.Lf027:
+        jge     .Lf025
+        jmp     .Lf019
+.Lf025:
         mov     ax,word ptr [bp-26]
         cmp     ax,0x100
-        jl      .Lf028
-        jmp     .Lf034
-.Lf028:
+        jl      .Lf026
+        jmp     .Lf031
+.Lf026:
         shl     ax,1
         mov     bx,word ptr [bp-14]
         add     bx,ax
         mov     word ptr [bp-40],bx
         cmp     di,0x100
-        jae     .Lf031
+        jae     .Lf028
         mov     ax,word ptr [bp-12]
         mov     dx,word ptr [bp-10]
         mov     bx,word ptr [bp-30]
         mov     cx,word ptr [bp-28]
-        M_PTC_JNE .Lf029
+        M_PTC_JNE .Lf027
         jmp     .Lf007
-.Lf029:
-        mov     cx,0x8
-.Lf030:
-        shl     si,1
-        rcl     di,1
-        loop    .Lf030
+.Lf027:
+        M_shld_disi_8_bxcx
         les     bx,dword ptr [bp-12]
         mov     al,byte ptr es:[bx]
         mov     byte ptr [bp-90],al
@@ -375,21 +360,21 @@
         M_PIA1
         mov     word ptr [bp-12],ax
         mov     word ptr [bp-10],dx
-.Lf031:
+.Lf028:
         M_shrd_11_disi_bp -96 -98
         mov     bx,word ptr [bp-40]
         mov     bx,word ptr [bx]
-        M_LMUL_dxax_00bx
+        M_U4M_dxax_00bx
         mov     word ptr [bp-8],ax
         mov     word ptr [bp-94],dx
         mov     ax,word ptr [bp-4]
         cmp     ax,dx
-        jb      .Lf032
-        jne     .Lf033
+        jb      .Lf029
+        jne     .Lf030
         mov     ax,word ptr [bp-6]
         cmp     ax,word ptr [bp-8]
-        jae     .Lf033
-.Lf032:
+        jae     .Lf030
+.Lf029:
         mov     si,word ptr [bp-8]
         mov     di,dx
         mov     ax,0x800
@@ -399,8 +384,8 @@
         shr     ax,cl
         add     word ptr [bx],ax
         shl     word ptr [bp-26],1
-        jmp     .Lf027
-.Lf033:
+        jmp     .Lf025
+.Lf030:
         sub     si,word ptr [bp-8]
         sbb     di,dx
         mov     ax,word ptr [bp-8]
@@ -415,8 +400,8 @@
         add     ax,ax
         inc     ax
         mov     word ptr [bp-26],ax
-        jmp     .Lf027
-.Lf034:
+        jmp     .Lf025
+.Lf031:
         mov     al,byte ptr [bp-26]
         mov     byte ptr [bp-2],al
         mov     bx,word ptr [bp-20]
@@ -432,19 +417,19 @@
         mov     byte ptr es:[bx],al
         mov     ax,word ptr [bp-16]
         cmp     ax,0x4
-        jge     .Lf035
+        jge     .Lf032
         xor     ax,ax
         mov     word ptr [bp-16],ax
         jmp     .Lf008
-.Lf035:
+.Lf032:
         cmp     ax,0xa
-        jge     .Lf036
+        jge     .Lf033
         sub     word ptr [bp-16],0x3
         jmp     .Lf008
-.Lf036:
+.Lf033:
         sub     word ptr [bp-16],0x6
         jmp     .Lf008
-.Lf037:
+.Lf034:
         sub     si,word ptr [bp-8]
         sbb     di,dx
         mov     ax,word ptr [bp-8]
@@ -462,19 +447,15 @@
         add     dx,ax
         mov     word ptr [bp-14],dx
         cmp     di,0x100
-        jae     .Lf040
+        jae     .Lf036
         mov     ax,word ptr [bp-12]
         mov     dx,word ptr [bp-10]
         mov     bx,word ptr [bp-30]
         mov     cx,word ptr [bp-28]
-        M_PTC_JNE .Lf038
+        M_PTC_JNE .Lf035
         jmp     .Lf007
-.Lf038:
-        mov     cx,0x8
-.Lf039:
-        shl     si,1
-        rcl     di,1
-        loop    .Lf039
+.Lf035:
+        M_shld_disi_8_bxcx
         les     bx,dword ptr [bp-12]
         mov     al,byte ptr es:[bx]
         mov     byte ptr [bp-90],al
@@ -491,21 +472,21 @@
         M_PIA1
         mov     word ptr [bp-12],ax
         mov     word ptr [bp-10],dx
-.Lf040:
+.Lf036:
         M_shrd_11_disi_bp -96 -98
         mov     bx,word ptr [bp-14]
         mov     bx,word ptr [bx]
-        M_LMUL_dxax_00bx
+        M_U4M_dxax_00bx
         mov     word ptr [bp-8],ax
         mov     word ptr [bp-94],dx
         mov     ax,word ptr [bp-4]
         cmp     ax,dx
-        jb      .Lf041
-        jne     .Lf044
+        jb      .Lf037
+        jne     .Lf040
         mov     ax,word ptr [bp-6]
         cmp     ax,word ptr [bp-8]
-        jae     .Lf044
-.Lf041:
+        jae     .Lf040
+.Lf037:
         mov     si,word ptr [bp-8]
         mov     di,dx
         mov     ax,0x800
@@ -527,17 +508,17 @@
         mov     ax,word ptr [bp-100]
         mov     word ptr [bp-50],ax
         cmp     word ptr [bp-16],0x7
-        jge     .Lf042
+        jge     .Lf038
         xor     ax,ax
-        jmp     .Lf043
-.Lf042:
+        jmp     .Lf039
+.Lf038:
         mov     ax,0x3
-.Lf043:
+.Lf039:
         mov     word ptr [bp-16],ax
         mov     bx,word ptr [bp-34]
         add     bx,0x664
-        jmp     .Lf076
-.Lf044:
+        jmp     .Lf068
+.Lf040:
         sub     si,word ptr [bp-8]
         sbb     di,dx
         mov     ax,word ptr [bp-8]
@@ -555,19 +536,15 @@
         add     ax,dx
         mov     word ptr [bp-14],ax
         cmp     di,0x100
-        jae     .Lf047
+        jae     .Lf042
         mov     ax,word ptr [bp-12]
         mov     dx,word ptr [bp-10]
         mov     bx,word ptr [bp-30]
         mov     cx,word ptr [bp-28]
-        M_PTC_JNE .Lf045
+        M_PTC_JNE .Lf041
         jmp     .Lf007
-.Lf045:
-        mov     cx,0x8
-.Lf046:
-        shl     si,1
-        rcl     di,1
-        loop    .Lf046
+.Lf041:
+        M_shld_disi_8_bxcx
         les     bx,dword ptr [bp-12]
         mov     al,byte ptr es:[bx]
         mov     byte ptr [bp-90],al
@@ -584,11 +561,11 @@
         M_PIA1
         mov     word ptr [bp-12],ax
         mov     word ptr [bp-10],dx
-.Lf047:
+.Lf042:
         M_shrd_11_disi_bp -96 -98
         mov     bx,word ptr [bp-14]
         mov     bx,word ptr [bx]
-        M_LMUL_dxax_00bx
+        M_U4M_dxax_00bx
         mov     word ptr [bp-8],ax
         mov     word ptr [bp-94],dx
         mov     ax,word ptr [bp-12]
@@ -598,15 +575,15 @@
         mov     word ptr [bp-80],dx
         mov     ax,word ptr [bp-4]
         cmp     ax,word ptr [bp-94]
-        jb      .Lf050
-        je      .Lf049
-.Lf048:
-        jmp     .Lf060
-.Lf049:
+        jb      .Lf045
+        je      .Lf044
+.Lf043:
+        jmp     .Lf054
+.Lf044:
         mov     ax,word ptr [bp-6]
         cmp     ax,word ptr [bp-8]
-        jae     .Lf048
-.Lf050:
+        jae     .Lf043
+.Lf045:
         mov     si,word ptr [bp-8]
         mov     di,word ptr [bp-94]
         mov     ax,0x800
@@ -625,20 +602,16 @@
         add     dx,ax
         mov     word ptr [bp-14],dx
         cmp     di,0x100
-        jae     .Lf053
+        jae     .Lf047
         mov     ax,word ptr [bp-12]
         mov     dx,word ptr [bp-10]
         mov     bx,word ptr [bp-30]
         mov     cx,word ptr [bp-28]
-        M_PTC_JNE .Lf051
+        M_PTC_JNE .Lf046
         jmp     .Lf007
-.Lf051:
+.Lf046:
         mov     ax,si
-        mov     cx,0x8
-.Lf052:
-        shl     ax,1
-        rcl     di,1
-        loop    .Lf052
+        M_shld_diax_8_bxcx
         mov     si,ax
         les     bx,dword ptr [bp-12]
         mov     al,byte ptr es:[bx]
@@ -655,21 +628,21 @@
         mov     word ptr [bp-12],bx
         mov     ax,word ptr [bp-80]
         mov     word ptr [bp-10],ax
-.Lf053:
+.Lf047:
         M_shrd_11_disi_bp -96 -98
         mov     bx,word ptr [bp-14]
         mov     bx,word ptr [bx]
-        M_LMUL_dxax_00bx
+        M_U4M_dxax_00bx
         mov     word ptr [bp-8],ax
         mov     word ptr [bp-94],dx
         mov     ax,word ptr [bp-4]
         cmp     ax,dx
-        jb      .Lf054
-        jne     .Lf056
+        jb      .Lf048
+        jne     .Lf050
         mov     ax,word ptr [bp-6]
         cmp     ax,word ptr [bp-8]
-        jae     .Lf059
-.Lf054:
+        jae     .Lf053
+.Lf048:
         mov     si,word ptr [bp-8]
         mov     di,dx
         mov     ax,0x800
@@ -680,18 +653,18 @@
         add     word ptr [bx],ax
         mov     ax,word ptr [bp-18]
         or      ax,word ptr [bp-20]
-        jne     .Lf055
+        jne     .Lf049
         jmp     .Lf007
-.Lf055:
+.Lf049:
         cmp     word ptr [bp-16],0x7
-        jge     .Lf057
+        jge     .Lf051
         mov     ax,0x9
-        jmp     .Lf058
-.Lf056:
-        jmp     .Lf059
-.Lf057:
+        jmp     .Lf052
+.Lf050:
+        jmp     .Lf053
+.Lf051:
         mov     ax,0xb
-.Lf058:
+.Lf052:
         mov     word ptr [bp-16],ax
         mov     bx,word ptr [bp-20]
         sub     bx,word ptr [bp-102]
@@ -716,7 +689,7 @@
         mov     al,byte ptr [bp-2]
         mov     byte ptr es:[bx],al
         jmp     .Lf008
-.Lf059:
+.Lf053:
         sub     si,word ptr [bp-8]
         sbb     di,dx
         mov     ax,word ptr [bp-8]
@@ -727,8 +700,8 @@
         mov     ax,word ptr [bx]
         shr     ax,cl
         sub     word ptr [bx],ax
-        jmp     .Lf073
-.Lf060:
+        jmp     .Lf065
+.Lf054:
         sub     si,word ptr [bp-8]
         sbb     di,word ptr [bp-94]
         mov     ax,word ptr [bp-8]
@@ -747,19 +720,15 @@
         add     dx,ax
         mov     word ptr [bp-14],dx
         cmp     di,0x100
-        jae     .Lf063
+        jae     .Lf056
         mov     ax,word ptr [bp-12]
         mov     dx,word ptr [bp-10]
         mov     bx,word ptr [bp-30]
         mov     cx,word ptr [bp-28]
-        M_PTC_JNE .Lf061
+        M_PTC_JNE .Lf055
         jmp     .Lf007
-.Lf061:
-        mov     cx,0x8
-.Lf062:
-        shl     si,1
-        rcl     di,1
-        loop    .Lf062
+.Lf055:
+        M_shld_disi_8_bxcx
         les     bx,dword ptr [bp-12]
         mov     al,byte ptr es:[bx]
         mov     byte ptr [bp-90],al
@@ -775,21 +744,21 @@
         mov     word ptr [bp-12],bx
         mov     ax,word ptr [bp-80]
         mov     word ptr [bp-10],ax
-.Lf063:
+.Lf056:
         M_shrd_11_disi_bp -96 -98
         mov     bx,word ptr [bp-14]
         mov     bx,word ptr [bx]
-        M_LMUL_dxax_00bx
+        M_U4M_dxax_00bx
         mov     word ptr [bp-8],ax
         mov     word ptr [bp-94],dx
         mov     ax,word ptr [bp-4]
         cmp     ax,dx
-        jb      .Lf064
-        jne     .Lf065
+        jb      .Lf057
+        jne     .Lf058
         mov     ax,word ptr [bp-6]
         cmp     ax,word ptr [bp-8]
-        jae     .Lf065
-.Lf064:
+        jae     .Lf058
+.Lf057:
         mov     si,word ptr [bp-8]
         mov     di,dx
         mov     ax,0x800
@@ -800,8 +769,8 @@
         add     word ptr [bx],ax
         mov     ax,word ptr [bp-52]
         mov     cx,word ptr [bp-50]
-        jmp     .Lf072
-.Lf065:
+        jmp     .Lf064
+.Lf058:
         sub     si,word ptr [bp-8]
         sbb     di,dx
         mov     ax,word ptr [bp-8]
@@ -819,19 +788,15 @@
         add     ax,dx
         mov     word ptr [bp-14],ax
         cmp     di,0x100
-        jae     .Lf068
+        jae     .Lf060
         mov     ax,word ptr [bp-12]
         mov     dx,word ptr [bp-10]
         mov     bx,word ptr [bp-30]
         mov     cx,word ptr [bp-28]
-        M_PTC_JNE .Lf066
+        M_PTC_JNE .Lf059
         jmp     .Lf007
-.Lf066:
-        mov     cx,0x8
-.Lf067:
-        shl     si,1
-        rcl     di,1
-        loop    .Lf067
+.Lf059:
+        M_shld_disi_8_bxcx
         les     bx,dword ptr [bp-12]
         mov     al,byte ptr es:[bx]
         mov     byte ptr [bp-90],al
@@ -848,21 +813,21 @@
         M_PIA1
         mov     word ptr [bp-12],ax
         mov     word ptr [bp-10],dx
-.Lf068:
+.Lf060:
         M_shrd_11_disi_bp -96 -98
         mov     bx,word ptr [bp-14]
         mov     bx,word ptr [bx]
-        M_LMUL_dxax_00bx
+        M_U4M_dxax_00bx
         mov     word ptr [bp-8],ax
         mov     word ptr [bp-94],dx
         mov     ax,word ptr [bp-4]
         cmp     ax,dx
-        jb      .Lf069
-        jne     .Lf070
+        jb      .Lf061
+        jne     .Lf062
         mov     ax,word ptr [bp-6]
         cmp     ax,word ptr [bp-8]
-        jae     .Lf070
-.Lf069:
+        jae     .Lf062
+.Lf061:
         mov     si,word ptr [bp-8]
         mov     di,dx
         mov     ax,0x800
@@ -873,8 +838,8 @@
         add     word ptr [bx],ax
         mov     ax,word ptr [bp-56]
         mov     cx,word ptr [bp-54]
-        jmp     .Lf071
-.Lf070:
+        jmp     .Lf063
+.Lf062:
         sub     si,word ptr [bp-8]
         sbb     di,dx
         mov     ax,word ptr [bp-8]
@@ -891,47 +856,43 @@
         mov     word ptr [bp-68],dx
         mov     dx,word ptr [bp-54]
         mov     word ptr [bp-74],dx
-.Lf071:
+.Lf063:
         mov     dx,word ptr [bp-52]
         mov     word ptr [bp-56],dx
         mov     dx,word ptr [bp-50]
         mov     word ptr [bp-54],dx
-.Lf072:
+.Lf064:
         mov     dx,word ptr [bp-102]
         mov     word ptr [bp-52],dx
         mov     dx,word ptr [bp-100]
         mov     word ptr [bp-50],dx
         mov     word ptr [bp-102],ax
         mov     word ptr [bp-100],cx
-.Lf073:
+.Lf065:
         cmp     word ptr [bp-16],0x7
-        jge     .Lf074
+        jge     .Lf066
         mov     ax,0x8
-        jmp     .Lf075
-.Lf074:
+        jmp     .Lf067
+.Lf066:
         mov     ax,0xb
-.Lf075:
+.Lf067:
         mov     word ptr [bp-16],ax
         mov     bx,word ptr [bp-34]
         add     bx,0xa68
-.Lf076:
+.Lf068:
         mov     word ptr [bp-14],bx
         mov     bx,word ptr [bp-14]
         mov     word ptr [bp-22],bx
         cmp     di,0x100
-        jae     .Lf079
+        jae     .Lf070
         mov     ax,word ptr [bp-12]
         mov     dx,word ptr [bp-10]
         mov     bx,word ptr [bp-30]
         mov     cx,word ptr [bp-28]
-        M_PTC_JNE .Lf077
+        M_PTC_JNE .Lf069
         jmp     .Lf007
-.Lf077:
-        mov     cx,0x8
-.Lf078:
-        shl     si,1
-        rcl     di,1
-        loop    .Lf078
+.Lf069:
+        M_shld_disi_8_bxcx
         les     bx,dword ptr [bp-12]
         mov     al,byte ptr es:[bx]
         mov     byte ptr [bp-90],al
@@ -948,21 +909,21 @@
         M_PIA1
         mov     word ptr [bp-12],ax
         mov     word ptr [bp-10],dx
-.Lf079:
+.Lf070:
         M_shrd_11_disi_bp -96 -98
         mov     bx,word ptr [bp-22]
         mov     bx,word ptr [bx]
-        M_LMUL_dxax_00bx
+        M_U4M_dxax_00bx
         mov     word ptr [bp-8],ax
         mov     word ptr [bp-94],dx
         mov     ax,word ptr [bp-4]
         cmp     ax,dx
-        jb      .Lf080
-        jne     .Lf082
+        jb      .Lf071
+        jne     .Lf073
         mov     ax,word ptr [bp-6]
         cmp     ax,word ptr [bp-8]
-        jae     .Lf082
-.Lf080:
+        jae     .Lf073
+.Lf071:
         mov     si,word ptr [bp-8]
         mov     di,dx
         mov     ax,0x800
@@ -980,10 +941,10 @@
         mov     word ptr [bp-22],dx
         xor     ax,ax
         mov     word ptr [bp-70],ax
-.Lf081:
+.Lf072:
         mov     word ptr [bp-60],0x3
-        jmp     .Lf088
-.Lf082:
+        jmp     .Lf078
+.Lf073:
         sub     si,word ptr [bp-8]
         sbb     di,dx
         mov     ax,word ptr [bp-8]
@@ -999,19 +960,15 @@
         inc     bx
         mov     word ptr [bp-22],bx
         cmp     di,0x100
-        jae     .Lf085
+        jae     .Lf075
         mov     ax,word ptr [bp-12]
         mov     dx,word ptr [bp-10]
         mov     bx,word ptr [bp-30]
         mov     cx,word ptr [bp-28]
-        M_PTC_JNE .Lf083
+        M_PTC_JNE .Lf074
         jmp     .Lf007
-.Lf083:
-        mov     cx,0x8
-.Lf084:
-        shl     si,1
-        rcl     di,1
-        loop    .Lf084
+.Lf074:
+        M_shld_disi_8_bxcx
         les     bx,dword ptr [bp-12]
         mov     al,byte ptr es:[bx]
         mov     byte ptr [bp-90],al
@@ -1028,21 +985,21 @@
         M_PIA1
         mov     word ptr [bp-12],ax
         mov     word ptr [bp-10],dx
-.Lf085:
+.Lf075:
         M_shrd_11_disi_bp -96 -98
         mov     bx,word ptr [bp-22]
         mov     bx,word ptr [bx]
-        M_LMUL_dxax_00bx
+        M_U4M_dxax_00bx
         mov     word ptr [bp-8],ax
         mov     word ptr [bp-94],dx
         mov     ax,word ptr [bp-4]
         cmp     ax,dx
-        jb      .Lf086
-        jne     .Lf087
+        jb      .Lf076
+        jne     .Lf077
         mov     ax,word ptr [bp-6]
         cmp     ax,word ptr [bp-8]
-        jae     .Lf087
-.Lf086:
+        jae     .Lf077
+.Lf076:
         mov     si,word ptr [bp-8]
         mov     di,dx
         mov     ax,0x800
@@ -1059,8 +1016,8 @@
         add     dx,ax
         mov     word ptr [bp-22],dx
         mov     word ptr [bp-70],0x8
-        jmp     .Lf081
-.Lf087:
+        jmp     .Lf072
+.Lf077:
         sub     si,word ptr [bp-8]
         sbb     di,dx
         mov     ax,word ptr [bp-8]
@@ -1076,30 +1033,26 @@
         mov     word ptr [bp-22],bx
         mov     word ptr [bp-70],0x10
         mov     word ptr [bp-60],0x8
-.Lf088:
+.Lf078:
         mov     ax,word ptr [bp-60]
         mov     word ptr [bp-62],ax
         mov     word ptr [bp-24],0x1
-.Lf089:
+.Lf079:
         mov     ax,word ptr [bp-24]
         shl     ax,1
         mov     bx,word ptr [bp-22]
         add     bx,ax
         mov     word ptr [bp-38],bx
         cmp     di,0x100
-        jae     .Lf092
+        jae     .Lf081
         mov     ax,word ptr [bp-12]
         mov     dx,word ptr [bp-10]
         mov     bx,word ptr [bp-30]
         mov     cx,word ptr [bp-28]
-        M_PTC_JNE .Lf090
+        M_PTC_JNE .Lf080
         jmp     .Lf007
-.Lf090:
-        mov     cx,0x8
-.Lf091:
-        shl     si,1
-        rcl     di,1
-        loop    .Lf091
+.Lf080:
+        M_shld_disi_8_bxcx
         les     bx,dword ptr [bp-12]
         mov     al,byte ptr es:[bx]
         mov     byte ptr [bp-90],al
@@ -1116,21 +1069,21 @@
         M_PIA1
         mov     word ptr [bp-12],ax
         mov     word ptr [bp-10],dx
-.Lf092:
+.Lf081:
         M_shrd_11_disi_bp -96 -98
         mov     bx,word ptr [bp-38]
         mov     bx,word ptr [bx]
-        M_LMUL_dxax_00bx
+        M_U4M_dxax_00bx
         mov     word ptr [bp-8],ax
         mov     word ptr [bp-94],dx
         mov     ax,word ptr [bp-4]
         cmp     ax,dx
-        jb      .Lf093
-        jne     .Lf094
+        jb      .Lf082
+        jne     .Lf083
         mov     ax,word ptr [bp-6]
         cmp     ax,word ptr [bp-8]
-        jae     .Lf094
-.Lf093:
+        jae     .Lf083
+.Lf082:
         mov     si,word ptr [bp-8]
         mov     di,dx
         mov     ax,0x800
@@ -1140,8 +1093,8 @@
         shr     ax,cl
         add     word ptr [bx],ax
         shl     word ptr [bp-24],1
-        jmp     .Lf095
-.Lf094:
+        jmp     .Lf084
+.Lf083:
         sub     si,word ptr [bp-8]
         sbb     di,dx
         mov     ax,word ptr [bp-8]
@@ -1156,11 +1109,11 @@
         add     ax,ax
         inc     ax
         mov     word ptr [bp-24],ax
-.Lf095:
+.Lf084:
         dec     word ptr [bp-62]
-        je      .Lf096
-        jmp     .Lf089
-.Lf096:
+        je      .Lf085
+        jmp     .Lf079
+.Lf085:
         mov     cl,byte ptr [bp-60]
         mov     ax,0x1
         shl     ax,cl
@@ -1168,15 +1121,15 @@
         mov     ax,word ptr [bp-70]
         add     word ptr [bp-24],ax
         cmp     word ptr [bp-16],0x4
-        jl      .Lf097
-        jmp     .Lf130
-.Lf097:
+        jl      .Lf086
+        jmp     .Lf116
+.Lf086:
         add     word ptr [bp-16],0x7
         mov     ax,word ptr [bp-24]
         cmp     ax,0x4
-        jl      .Lf098
+        jl      .Lf087
         mov     ax,0x3
-.Lf098:
+.Lf087:
         mov     cl,0x7
         mov     dx,ax
         shl     dx,cl
@@ -1186,26 +1139,22 @@
         mov     word ptr [bp-14],ax
         mov     word ptr [bp-92],0x6
         mov     word ptr [bp-32],0x1
-.Lf099:
+.Lf088:
         mov     ax,word ptr [bp-32]
         shl     ax,1
         mov     bx,word ptr [bp-14]
         add     bx,ax
         mov     word ptr [bp-42],bx
         cmp     di,0x100
-        jae     .Lf102
+        jae     .Lf090
         mov     ax,word ptr [bp-12]
         mov     dx,word ptr [bp-10]
         mov     bx,word ptr [bp-30]
         mov     cx,word ptr [bp-28]
-        M_PTC_JNE .Lf100
+        M_PTC_JNE .Lf089
         jmp     .Lf007
-.Lf100:
-        mov     cx,0x8
-.Lf101:
-        shl     si,1
-        rcl     di,1
-        loop    .Lf101
+.Lf089:
+        M_shld_disi_8_bxcx
         les     bx,dword ptr [bp-12]
         mov     al,byte ptr es:[bx]
         mov     byte ptr [bp-90],al
@@ -1222,21 +1171,21 @@
         M_PIA1
         mov     word ptr [bp-12],ax
         mov     word ptr [bp-10],dx
-.Lf102:
+.Lf090:
         M_shrd_11_disi_bp -96 -98
         mov     bx,word ptr [bp-42]
         mov     bx,word ptr [bx]
-        M_LMUL_dxax_00bx
+        M_U4M_dxax_00bx
         mov     word ptr [bp-8],ax
         mov     word ptr [bp-94],dx
         mov     ax,word ptr [bp-4]
         cmp     ax,dx
-        jb      .Lf103
-        jne     .Lf104
+        jb      .Lf091
+        jne     .Lf092
         mov     ax,word ptr [bp-6]
         cmp     ax,word ptr [bp-8]
-        jae     .Lf104
-.Lf103:
+        jae     .Lf092
+.Lf091:
         mov     si,word ptr [bp-8]
         mov     di,dx
         mov     ax,0x800
@@ -1246,8 +1195,8 @@
         shr     ax,cl
         add     word ptr [bx],ax
         shl     word ptr [bp-32],1
-        jmp     .Lf105
-.Lf104:
+        jmp     .Lf093
+.Lf092:
         sub     si,word ptr [bp-8]
         sbb     di,dx
         mov     ax,word ptr [bp-8]
@@ -1262,15 +1211,15 @@
         add     ax,ax
         inc     ax
         mov     word ptr [bp-32],ax
-.Lf105:
+.Lf093:
         dec     word ptr [bp-92]
-        je      .Lf106
-        jmp     .Lf099
-.Lf106:
+        je      .Lf094
+        jmp     .Lf088
+.Lf094:
         sub     word ptr [bp-32],0x40
         mov     ax,word ptr [bp-32]
         cmp     ax,0x4
-        jl      .Lf109
+        jl      .Lf097
         sar     ax,1
         dec     ax
         mov     word ptr [bp-36],ax
@@ -1281,14 +1230,14 @@
         xor     ax,ax
         mov     word ptr [bp-100],ax
         cmp     word ptr [bp-32],0xe
-        jge     .Lf110
+        jge     .Lf098
         mov     cx,word ptr [bp-36]
-        jcxz    .Lf108
-.Lf107:
+        jcxz    .Lf096
+.Lf095:
         shl     word ptr [bp-102],1
         rcl     word ptr [bp-100],1
-        loop    .Lf107
-.Lf108:
+        loop    .Lf095
+.Lf096:
         mov     dx,word ptr [bp-102]
         shl     dx,1
         mov     ax,word ptr [bp-34]
@@ -1300,26 +1249,22 @@
         dec     dx
         dec     dx
         mov     word ptr [bp-14],dx
-        jmp     .Lf119
-.Lf109:
-        jmp     .Lf128
-.Lf110:
+        jmp     .Lf106
+.Lf097:
+        jmp     .Lf114
+.Lf098:
         sub     word ptr [bp-36],0x4
-.Lf111:
+.Lf099:
         cmp     di,0x100
-        jae     .Lf114
+        jae     .Lf101
         mov     ax,word ptr [bp-12]
         mov     dx,word ptr [bp-10]
         mov     bx,word ptr [bp-30]
         mov     cx,word ptr [bp-28]
-        M_PTC_JNE .Lf112
+        M_PTC_JNE .Lf100
         jmp     .Lf007
-.Lf112:
-        mov     cx,0x8
-.Lf113:
-        shl     si,1
-        rcl     di,1
-        loop    .Lf113
+.Lf100:
+        M_shld_disi_8_bxcx
         les     bx,dword ptr [bp-12]
         mov     al,byte ptr es:[bx]
         mov     byte ptr [bp-90],al
@@ -1336,59 +1281,55 @@
         M_PIA1
         mov     word ptr [bp-12],ax
         mov     word ptr [bp-10],dx
-.Lf114:
+.Lf101:
         shr     di,1
         rcr     si,1
         shl     word ptr [bp-102],1
         rcl     word ptr [bp-100],1
         mov     ax,word ptr [bp-4]
         cmp     di,ax
-        jb      .Lf115
-        jne     .Lf116
+        jb      .Lf102
+        jne     .Lf103
         cmp     si,word ptr [bp-6]
-        ja      .Lf116
-.Lf115:
+        ja      .Lf103
+.Lf102:
         sub     word ptr [bp-6],si
         sbb     word ptr [bp-4],di
         or      byte ptr [bp-102],0x1
-.Lf116:
+.Lf103:
         dec     word ptr [bp-36]
-        je      .Lf117
-        jmp     .Lf111
-.Lf117:
+        je      .Lf104
+        jmp     .Lf099
+.Lf104:
         mov     bx,word ptr [bp-34]
         add     bx,0x644
         mov     word ptr [bp-14],bx
         mov     cx,0x4
-.Lf118:
+.Lf105:
         shl     word ptr [bp-102],1
         rcl     word ptr [bp-100],1
-        loop    .Lf118
+        loop    .Lf105
         mov     word ptr [bp-36],0x4
-.Lf119:
+.Lf106:
         mov     ax,0x1
         mov     word ptr [bp-64],ax
         mov     word ptr [bp-48],ax
-.Lf120:
+.Lf107:
         mov     ax,word ptr [bp-48]
         shl     ax,1
         mov     bx,word ptr [bp-14]
         add     bx,ax
         mov     word ptr [bp-44],bx
         cmp     di,0x100
-        jae     .Lf123
+        jae     .Lf109
         mov     ax,word ptr [bp-12]
         mov     dx,word ptr [bp-10]
         mov     bx,word ptr [bp-30]
         mov     cx,word ptr [bp-28]
-        M_PTC_JNE .Lf121
+        M_PTC_JNE .Lf108
         jmp     .Lf007
-.Lf121:
-        mov     cx,0x8
-.Lf122:
-        shl     si,1
-        rcl     di,1
-        loop    .Lf122
+.Lf108:
+        M_shld_disi_8_bxcx
         les     bx,dword ptr [bp-12]
         mov     al,byte ptr es:[bx]
         mov     byte ptr [bp-90],al
@@ -1405,21 +1346,21 @@
         M_PIA1
         mov     word ptr [bp-12],ax
         mov     word ptr [bp-10],dx
-.Lf123:
+.Lf109:
         M_shrd_11_disi_bp -96 -98
         mov     bx,word ptr [bp-44]
         mov     bx,word ptr [bx]
-        M_LMUL_dxax_00bx
+        M_U4M_dxax_00bx
         mov     word ptr [bp-8],ax
         mov     word ptr [bp-94],dx
         mov     ax,word ptr [bp-4]
         cmp     ax,dx
-        jb      .Lf124
-        jne     .Lf125
+        jb      .Lf110
+        jne     .Lf111
         mov     ax,word ptr [bp-6]
         cmp     ax,word ptr [bp-8]
-        jae     .Lf125
-.Lf124:
+        jae     .Lf111
+.Lf110:
         mov     si,word ptr [bp-8]
         mov     di,dx
         mov     ax,0x800
@@ -1429,8 +1370,8 @@
         shr     ax,cl
         add     word ptr [bx],ax
         shl     word ptr [bp-48],1
-        jmp     .Lf126
-.Lf125:
+        jmp     .Lf112
+.Lf111:
         sub     si,word ptr [bp-8]
         sbb     di,dx
         mov     ax,word ptr [bp-8]
@@ -1449,37 +1390,37 @@
         cwd
         or      word ptr [bp-102],ax
         or      word ptr [bp-100],dx
-.Lf126:
+.Lf112:
         shl     word ptr [bp-64],1
         dec     word ptr [bp-36]
-        je      .Lf127
-        jmp     .Lf120
-.Lf127:
-        jmp     .Lf129
-.Lf128:
+        je      .Lf113
+        jmp     .Lf107
+.Lf113:
+        jmp     .Lf115
+.Lf114:
         cwd
         mov     word ptr [bp-102],ax
         mov     word ptr [bp-100],dx
-.Lf129:
+.Lf115:
         add     word ptr [bp-102],0x1
         adc     word ptr [bp-100],0x0
         mov     ax,word ptr [bp-100]
         or      ax,word ptr [bp-102]
-        jne     .Lf130
-        jmp     .Lf136
-.Lf130:
+        jne     .Lf116
+        jmp     .Lf122
+.Lf116:
         add     word ptr [bp-24],0x2
         mov     ax,word ptr [bp-100]
         cmp     ax,word ptr [bp-18]
-        jbe     .Lf132
-.Lf131:
+        jbe     .Lf118
+.Lf117:
         jmp     .Lf007
-.Lf132:
-        jne     .Lf133
+.Lf118:
+        jne     .Lf119
         mov     ax,word ptr [bp-102]
         cmp     ax,word ptr [bp-20]
-        ja      .Lf131
-.Lf133:
+        ja      .Lf117
+.Lf119:
         mov     bx,word ptr [bp-20]
         sub     bx,word ptr [bp-102]
         mov     cx,word ptr [bp-18]
@@ -1504,34 +1445,34 @@
         mov     al,byte ptr [bp-2]
         mov     byte ptr es:[bx],al
         cmp     word ptr [bp-24],0x0
-        jne     .Lf135
-.Lf134:
+        jne     .Lf121
+.Lf120:
         jmp     .Lf008
-.Lf135:
+.Lf121:
         mov     ax,word ptr [bp-18]
         cmp     ax,word ptr [bp+26]
-        jb      .Lf133
-        jne     .Lf134
+        jb      .Lf119
+        jne     .Lf120
         mov     ax,word ptr [bp-20]
         cmp     ax,word ptr [bp+24]
-        jb      .Lf133
-        jmp     .Lf134
-.Lf136:
+        jb      .Lf119
+        jmp     .Lf120
+.Lf122:
         cmp     di,0x100
-        jae     .Lf138
+        jae     .Lf124
         mov     ax,word ptr [bp-12]
         mov     dx,word ptr [bp-10]
         mov     bx,word ptr [bp-30]
         mov     cx,word ptr [bp-28]
-        M_PTC_JNE .Lf137
+        M_PTC_JNE .Lf123
         jmp     .Lf007
-.Lf137:
+.Lf123:
         mov     ax,word ptr [bp-12]
         mov     dx,word ptr [bp-10]
         M_PIA1
         mov     word ptr [bp-12],ax
         mov     word ptr [bp-10],dx
-.Lf138:
+.Lf124:
         mov     ax,word ptr [bp-12]
         mov     dx,word ptr [bp-10]
         mov     bx,word ptr [bp+10]
@@ -1546,7 +1487,7 @@
         mov     ax,word ptr [bp-18]
         mov     word ptr [bx+2],ax
         xor     ax,ax
-.Lf139:
+.Lf125:
         mov     sp,bp
         pop     bp
         pop     di
diff --git a/src/stub/src/arch/i086/lzma_d_cs.S b/src/stub/src/arch/i086/lzma_d_cs.S
index 9f4f0dba..e756f10d 100644
--- a/src/stub/src/arch/i086/lzma_d_cs.S
+++ b/src/stub/src/arch/i086/lzma_d_cs.S
@@ -119,14 +119,14 @@
         M_PTC_JNE .Ls005
 .Ls007:
         mov     ax,0x1
-        jmp     .Ls139
+        jmp     .Ls125
 .Ls008:
         mov     ax,word ptr [bp-18]
         cmp     ax,word ptr [bp+26]
         jb      .Ls011
         je      .Ls010
 .Ls009:
-        jmp     .Ls136
+        jmp     .Ls122
 .Ls010:
         mov     ax,word ptr [bp-20]
         cmp     ax,word ptr [bp+24]
@@ -143,18 +143,14 @@
         add     dx,ax
         mov     word ptr [bp-14],dx
         cmp     di,0x100
-        jae     .Ls013
+        jae     .Ls012
         mov     ax,word ptr [bp-12]
         mov     dx,word ptr [bp-10]
         mov     bx,word ptr [bp-30]
         mov     cx,word ptr [bp-28]
         M_PTC
         je      .Ls007
-        mov     cx,0x8
-.Ls012:
-        shl     si,1
-        rcl     di,1
-        loop    .Ls012
+        M_shld_disi_8_bxcx
         les     bx,dword ptr [bp-12]
         mov     al,byte ptr es:[bx]
         mov     byte ptr [bp-90],al
@@ -171,24 +167,24 @@
         M_PIA1
         mov     word ptr [bp-12],ax
         mov     word ptr [bp-10],dx
-.Ls013:
+.Ls012:
         M_shrd_11_disi_bp -96 -98
         mov     bx,word ptr [bp-14]
         mov     bx,word ptr [bx]
-        M_LMUL_dxax_00bx
+        M_U4M_dxax_00bx
         mov     word ptr [bp-8],ax
         mov     word ptr [bp-94],dx
         mov     ax,word ptr [bp-4]
         cmp     ax,dx
-        jb      .Ls016
-        je      .Ls015
+        jb      .Ls015
+        je      .Ls014
+.Ls013:
+        jmp     .Ls034
 .Ls014:
-        jmp     .Ls037
-.Ls015:
         mov     ax,word ptr [bp-6]
         cmp     ax,word ptr [bp-8]
-        jae     .Ls014
-.Ls016:
+        jae     .Ls013
+.Ls015:
         mov     word ptr [bp-26],0x1
         mov     si,word ptr [bp-8]
         mov     di,dx
@@ -211,26 +207,23 @@
         mov     dx,word ptr [bp-18]
         and     dx,word ptr [bp-84]
         mov     cx,word ptr [bp-78]
-        jcxz    .Ls018
-.Ls017:
+        jcxz    .Ls017
+.Ls016:
         shl     ax,1
         rcl     dx,1
-        loop    .Ls017
-.Ls018:
+        loop    .Ls016
+.Ls017:
         add     ax,word ptr [bp-90]
         adc     dx,bx
-        mov     bx,0x300
-        M_LMUL_dxax_00bx
-        shl     ax,1
-        rcl     dx,1
+        M_U4M_dxax_0x0600
         mov     dx,word ptr [bp-34]
         add     dx,0xe6c
         add     dx,ax
         mov     word ptr [bp-14],dx
         cmp     word ptr [bp-16],0x7
-        jge     .Ls019
-        jmp     .Ls027
-.Ls019:
+        jge     .Ls018
+        jmp     .Ls025
+.Ls018:
         mov     bx,word ptr [bp-20]
         sub     bx,word ptr [bp-102]
         mov     cx,word ptr [bp-18]
@@ -243,7 +236,7 @@
         mov     al,byte ptr es:[bx]
         xor     ah,ah
         mov     word ptr [bp-76],ax
-.Ls020:
+.Ls019:
         shl     word ptr [bp-76],1
         mov     ax,word ptr [bp-76]
         xor     al,al
@@ -258,19 +251,15 @@
         add     dx,ax
         mov     word ptr [bp-46],dx
         cmp     di,0x100
-        jae     .Ls023
+        jae     .Ls021
         mov     ax,word ptr [bp-12]
         mov     dx,word ptr [bp-10]
         mov     bx,word ptr [bp-30]
         mov     cx,word ptr [bp-28]
-        M_PTC_JNE .Ls021
+        M_PTC_JNE .Ls020
         jmp     .Ls007
-.Ls021:
-        mov     cx,0x8
-.Ls022:
-        shl     si,1
-        rcl     di,1
-        loop    .Ls022
+.Ls020:
+        M_shld_disi_8_bxcx
         les     bx,dword ptr [bp-12]
         mov     al,byte ptr es:[bx]
         mov     byte ptr [bp-90],al
@@ -287,21 +276,21 @@
         M_PIA1
         mov     word ptr [bp-12],ax
         mov     word ptr [bp-10],dx
-.Ls023:
+.Ls021:
         M_shrd_11_disi_bp -96 -98
         mov     bx,word ptr [bp-46]
         mov     bx,word ptr [bx]
-        M_LMUL_dxax_00bx
+        M_U4M_dxax_00bx
         mov     word ptr [bp-8],ax
         mov     word ptr [bp-94],dx
         mov     ax,word ptr [bp-4]
         cmp     ax,dx
-        jb      .Ls024
-        jne     .Ls025
+        jb      .Ls022
+        jne     .Ls023
         mov     ax,word ptr [bp-6]
         cmp     ax,word ptr [bp-8]
-        jae     .Ls025
-.Ls024:
+        jae     .Ls023
+.Ls022:
         mov     si,word ptr [bp-8]
         mov     di,dx
         mov     ax,0x800
@@ -312,9 +301,9 @@
         add     word ptr [bx],ax
         shl     word ptr [bp-26],1
         cmp     word ptr [bp-66],0x0
-        jne     .Ls027
-        jmp     .Ls026
-.Ls025:
+        jne     .Ls025
+        jmp     .Ls024
+.Ls023:
         sub     si,word ptr [bp-8]
         sbb     di,dx
         mov     ax,word ptr [bp-8]
@@ -330,35 +319,31 @@
         inc     ax
         mov     word ptr [bp-26],ax
         cmp     word ptr [bp-66],0x0
-        je      .Ls027
-.Ls026:
+        je      .Ls025
+.Ls024:
         cmp     word ptr [bp-26],0x100
-        jge     .Ls027
-        jmp     .Ls020
-.Ls027:
+        jge     .Ls025
+        jmp     .Ls019
+.Ls025:
         mov     ax,word ptr [bp-26]
         cmp     ax,0x100
-        jl      .Ls028
-        jmp     .Ls034
-.Ls028:
+        jl      .Ls026
+        jmp     .Ls031
+.Ls026:
         shl     ax,1
         mov     bx,word ptr [bp-14]
         add     bx,ax
         mov     word ptr [bp-40],bx
         cmp     di,0x100
-        jae     .Ls031
+        jae     .Ls028
         mov     ax,word ptr [bp-12]
         mov     dx,word ptr [bp-10]
         mov     bx,word ptr [bp-30]
         mov     cx,word ptr [bp-28]
-        M_PTC_JNE .Ls029
+        M_PTC_JNE .Ls027
         jmp     .Ls007
-.Ls029:
-        mov     cx,0x8
-.Ls030:
-        shl     si,1
-        rcl     di,1
-        loop    .Ls030
+.Ls027:
+        M_shld_disi_8_bxcx
         les     bx,dword ptr [bp-12]
         mov     al,byte ptr es:[bx]
         mov     byte ptr [bp-90],al
@@ -375,21 +360,21 @@
         M_PIA1
         mov     word ptr [bp-12],ax
         mov     word ptr [bp-10],dx
-.Ls031:
+.Ls028:
         M_shrd_11_disi_bp -96 -98
         mov     bx,word ptr [bp-40]
         mov     bx,word ptr [bx]
-        M_LMUL_dxax_00bx
+        M_U4M_dxax_00bx
         mov     word ptr [bp-8],ax
         mov     word ptr [bp-94],dx
         mov     ax,word ptr [bp-4]
         cmp     ax,dx
-        jb      .Ls032
-        jne     .Ls033
+        jb      .Ls029
+        jne     .Ls030
         mov     ax,word ptr [bp-6]
         cmp     ax,word ptr [bp-8]
-        jae     .Ls033
-.Ls032:
+        jae     .Ls030
+.Ls029:
         mov     si,word ptr [bp-8]
         mov     di,dx
         mov     ax,0x800
@@ -399,8 +384,8 @@
         shr     ax,cl
         add     word ptr [bx],ax
         shl     word ptr [bp-26],1
-        jmp     .Ls027
-.Ls033:
+        jmp     .Ls025
+.Ls030:
         sub     si,word ptr [bp-8]
         sbb     di,dx
         mov     ax,word ptr [bp-8]
@@ -415,8 +400,8 @@
         add     ax,ax
         inc     ax
         mov     word ptr [bp-26],ax
-        jmp     .Ls027
-.Ls034:
+        jmp     .Ls025
+.Ls031:
         mov     al,byte ptr [bp-26]
         mov     byte ptr [bp-2],al
         mov     bx,word ptr [bp-20]
@@ -432,19 +417,19 @@
         mov     byte ptr es:[bx],al
         mov     ax,word ptr [bp-16]
         cmp     ax,0x4
-        jge     .Ls035
+        jge     .Ls032
         xor     ax,ax
         mov     word ptr [bp-16],ax
         jmp     .Ls008
-.Ls035:
+.Ls032:
         cmp     ax,0xa
-        jge     .Ls036
+        jge     .Ls033
         sub     word ptr [bp-16],0x3
         jmp     .Ls008
-.Ls036:
+.Ls033:
         sub     word ptr [bp-16],0x6
         jmp     .Ls008
-.Ls037:
+.Ls034:
         sub     si,word ptr [bp-8]
         sbb     di,dx
         mov     ax,word ptr [bp-8]
@@ -462,19 +447,15 @@
         add     dx,ax
         mov     word ptr [bp-14],dx
         cmp     di,0x100
-        jae     .Ls040
+        jae     .Ls036
         mov     ax,word ptr [bp-12]
         mov     dx,word ptr [bp-10]
         mov     bx,word ptr [bp-30]
         mov     cx,word ptr [bp-28]
-        M_PTC_JNE .Ls038
+        M_PTC_JNE .Ls035
         jmp     .Ls007
-.Ls038:
-        mov     cx,0x8
-.Ls039:
-        shl     si,1
-        rcl     di,1
-        loop    .Ls039
+.Ls035:
+        M_shld_disi_8_bxcx
         les     bx,dword ptr [bp-12]
         mov     al,byte ptr es:[bx]
         mov     byte ptr [bp-90],al
@@ -491,21 +472,21 @@
         M_PIA1
         mov     word ptr [bp-12],ax
         mov     word ptr [bp-10],dx
-.Ls040:
+.Ls036:
         M_shrd_11_disi_bp -96 -98
         mov     bx,word ptr [bp-14]
         mov     bx,word ptr [bx]
-        M_LMUL_dxax_00bx
+        M_U4M_dxax_00bx
         mov     word ptr [bp-8],ax
         mov     word ptr [bp-94],dx
         mov     ax,word ptr [bp-4]
         cmp     ax,dx
-        jb      .Ls041
-        jne     .Ls044
+        jb      .Ls037
+        jne     .Ls040
         mov     ax,word ptr [bp-6]
         cmp     ax,word ptr [bp-8]
-        jae     .Ls044
-.Ls041:
+        jae     .Ls040
+.Ls037:
         mov     si,word ptr [bp-8]
         mov     di,dx
         mov     ax,0x800
@@ -527,17 +508,17 @@
         mov     ax,word ptr [bp-100]
         mov     word ptr [bp-50],ax
         cmp     word ptr [bp-16],0x7
-        jge     .Ls042
+        jge     .Ls038
         xor     ax,ax
-        jmp     .Ls043
-.Ls042:
+        jmp     .Ls039
+.Ls038:
         mov     ax,0x3
-.Ls043:
+.Ls039:
         mov     word ptr [bp-16],ax
         mov     bx,word ptr [bp-34]
         add     bx,0x664
-        jmp     .Ls076
-.Ls044:
+        jmp     .Ls068
+.Ls040:
         sub     si,word ptr [bp-8]
         sbb     di,dx
         mov     ax,word ptr [bp-8]
@@ -555,19 +536,15 @@
         add     ax,dx
         mov     word ptr [bp-14],ax
         cmp     di,0x100
-        jae     .Ls047
+        jae     .Ls042
         mov     ax,word ptr [bp-12]
         mov     dx,word ptr [bp-10]
         mov     bx,word ptr [bp-30]
         mov     cx,word ptr [bp-28]
-        M_PTC_JNE .Ls045
+        M_PTC_JNE .Ls041
         jmp     .Ls007
-.Ls045:
-        mov     cx,0x8
-.Ls046:
-        shl     si,1
-        rcl     di,1
-        loop    .Ls046
+.Ls041:
+        M_shld_disi_8_bxcx
         les     bx,dword ptr [bp-12]
         mov     al,byte ptr es:[bx]
         mov     byte ptr [bp-90],al
@@ -584,11 +561,11 @@
         M_PIA1
         mov     word ptr [bp-12],ax
         mov     word ptr [bp-10],dx
-.Ls047:
+.Ls042:
         M_shrd_11_disi_bp -96 -98
         mov     bx,word ptr [bp-14]
         mov     bx,word ptr [bx]
-        M_LMUL_dxax_00bx
+        M_U4M_dxax_00bx
         mov     word ptr [bp-8],ax
         mov     word ptr [bp-94],dx
         mov     ax,word ptr [bp-12]
@@ -598,15 +575,15 @@
         mov     word ptr [bp-80],dx
         mov     ax,word ptr [bp-4]
         cmp     ax,word ptr [bp-94]
-        jb      .Ls050
-        je      .Ls049
-.Ls048:
-        jmp     .Ls060
-.Ls049:
+        jb      .Ls045
+        je      .Ls044
+.Ls043:
+        jmp     .Ls054
+.Ls044:
         mov     ax,word ptr [bp-6]
         cmp     ax,word ptr [bp-8]
-        jae     .Ls048
-.Ls050:
+        jae     .Ls043
+.Ls045:
         mov     si,word ptr [bp-8]
         mov     di,word ptr [bp-94]
         mov     ax,0x800
@@ -625,20 +602,16 @@
         add     dx,ax
         mov     word ptr [bp-14],dx
         cmp     di,0x100
-        jae     .Ls053
+        jae     .Ls047
         mov     ax,word ptr [bp-12]
         mov     dx,word ptr [bp-10]
         mov     bx,word ptr [bp-30]
         mov     cx,word ptr [bp-28]
-        M_PTC_JNE .Ls051
+        M_PTC_JNE .Ls046
         jmp     .Ls007
-.Ls051:
+.Ls046:
         mov     ax,si
-        mov     cx,0x8
-.Ls052:
-        shl     ax,1
-        rcl     di,1
-        loop    .Ls052
+        M_shld_diax_8_bxcx
         mov     si,ax
         les     bx,dword ptr [bp-12]
         mov     al,byte ptr es:[bx]
@@ -655,21 +628,21 @@
         mov     word ptr [bp-12],bx
         mov     ax,word ptr [bp-80]
         mov     word ptr [bp-10],ax
-.Ls053:
+.Ls047:
         M_shrd_11_disi_bp -96 -98
         mov     bx,word ptr [bp-14]
         mov     bx,word ptr [bx]
-        M_LMUL_dxax_00bx
+        M_U4M_dxax_00bx
         mov     word ptr [bp-8],ax
         mov     word ptr [bp-94],dx
         mov     ax,word ptr [bp-4]
         cmp     ax,dx
-        jb      .Ls054
-        jne     .Ls056
+        jb      .Ls048
+        jne     .Ls050
         mov     ax,word ptr [bp-6]
         cmp     ax,word ptr [bp-8]
-        jae     .Ls059
-.Ls054:
+        jae     .Ls053
+.Ls048:
         mov     si,word ptr [bp-8]
         mov     di,dx
         mov     ax,0x800
@@ -680,18 +653,18 @@
         add     word ptr [bx],ax
         mov     ax,word ptr [bp-18]
         or      ax,word ptr [bp-20]
-        jne     .Ls055
+        jne     .Ls049
         jmp     .Ls007
-.Ls055:
+.Ls049:
         cmp     word ptr [bp-16],0x7
-        jge     .Ls057
+        jge     .Ls051
         mov     ax,0x9
-        jmp     .Ls058
-.Ls056:
-        jmp     .Ls059
-.Ls057:
+        jmp     .Ls052
+.Ls050:
+        jmp     .Ls053
+.Ls051:
         mov     ax,0xb
-.Ls058:
+.Ls052:
         mov     word ptr [bp-16],ax
         mov     bx,word ptr [bp-20]
         sub     bx,word ptr [bp-102]
@@ -716,7 +689,7 @@
         mov     al,byte ptr [bp-2]
         mov     byte ptr es:[bx],al
         jmp     .Ls008
-.Ls059:
+.Ls053:
         sub     si,word ptr [bp-8]
         sbb     di,dx
         mov     ax,word ptr [bp-8]
@@ -727,8 +700,8 @@
         mov     ax,word ptr [bx]
         shr     ax,cl
         sub     word ptr [bx],ax
-        jmp     .Ls073
-.Ls060:
+        jmp     .Ls065
+.Ls054:
         sub     si,word ptr [bp-8]
         sbb     di,word ptr [bp-94]
         mov     ax,word ptr [bp-8]
@@ -747,19 +720,15 @@
         add     dx,ax
         mov     word ptr [bp-14],dx
         cmp     di,0x100
-        jae     .Ls063
+        jae     .Ls056
         mov     ax,word ptr [bp-12]
         mov     dx,word ptr [bp-10]
         mov     bx,word ptr [bp-30]
         mov     cx,word ptr [bp-28]
-        M_PTC_JNE .Ls061
+        M_PTC_JNE .Ls055
         jmp     .Ls007
-.Ls061:
-        mov     cx,0x8
-.Ls062:
-        shl     si,1
-        rcl     di,1
-        loop    .Ls062
+.Ls055:
+        M_shld_disi_8_bxcx
         les     bx,dword ptr [bp-12]
         mov     al,byte ptr es:[bx]
         mov     byte ptr [bp-90],al
@@ -775,21 +744,21 @@
         mov     word ptr [bp-12],bx
         mov     ax,word ptr [bp-80]
         mov     word ptr [bp-10],ax
-.Ls063:
+.Ls056:
         M_shrd_11_disi_bp -96 -98
         mov     bx,word ptr [bp-14]
         mov     bx,word ptr [bx]
-        M_LMUL_dxax_00bx
+        M_U4M_dxax_00bx
         mov     word ptr [bp-8],ax
         mov     word ptr [bp-94],dx
         mov     ax,word ptr [bp-4]
         cmp     ax,dx
-        jb      .Ls064
-        jne     .Ls065
+        jb      .Ls057
+        jne     .Ls058
         mov     ax,word ptr [bp-6]
         cmp     ax,word ptr [bp-8]
-        jae     .Ls065
-.Ls064:
+        jae     .Ls058
+.Ls057:
         mov     si,word ptr [bp-8]
         mov     di,dx
         mov     ax,0x800
@@ -800,8 +769,8 @@
         add     word ptr [bx],ax
         mov     ax,word ptr [bp-52]
         mov     cx,word ptr [bp-50]
-        jmp     .Ls072
-.Ls065:
+        jmp     .Ls064
+.Ls058:
         sub     si,word ptr [bp-8]
         sbb     di,dx
         mov     ax,word ptr [bp-8]
@@ -819,19 +788,15 @@
         add     ax,dx
         mov     word ptr [bp-14],ax
         cmp     di,0x100
-        jae     .Ls068
+        jae     .Ls060
         mov     ax,word ptr [bp-12]
         mov     dx,word ptr [bp-10]
         mov     bx,word ptr [bp-30]
         mov     cx,word ptr [bp-28]
-        M_PTC_JNE .Ls066
+        M_PTC_JNE .Ls059
         jmp     .Ls007
-.Ls066:
-        mov     cx,0x8
-.Ls067:
-        shl     si,1
-        rcl     di,1
-        loop    .Ls067
+.Ls059:
+        M_shld_disi_8_bxcx
         les     bx,dword ptr [bp-12]
         mov     al,byte ptr es:[bx]
         mov     byte ptr [bp-90],al
@@ -848,21 +813,21 @@
         M_PIA1
         mov     word ptr [bp-12],ax
         mov     word ptr [bp-10],dx
-.Ls068:
+.Ls060:
         M_shrd_11_disi_bp -96 -98
         mov     bx,word ptr [bp-14]
         mov     bx,word ptr [bx]
-        M_LMUL_dxax_00bx
+        M_U4M_dxax_00bx
         mov     word ptr [bp-8],ax
         mov     word ptr [bp-94],dx
         mov     ax,word ptr [bp-4]
         cmp     ax,dx
-        jb      .Ls069
-        jne     .Ls070
+        jb      .Ls061
+        jne     .Ls062
         mov     ax,word ptr [bp-6]
         cmp     ax,word ptr [bp-8]
-        jae     .Ls070
-.Ls069:
+        jae     .Ls062
+.Ls061:
         mov     si,word ptr [bp-8]
         mov     di,dx
         mov     ax,0x800
@@ -873,8 +838,8 @@
         add     word ptr [bx],ax
         mov     ax,word ptr [bp-56]
         mov     cx,word ptr [bp-54]
-        jmp     .Ls071
-.Ls070:
+        jmp     .Ls063
+.Ls062:
         sub     si,word ptr [bp-8]
         sbb     di,dx
         mov     ax,word ptr [bp-8]
@@ -891,47 +856,43 @@
         mov     word ptr [bp-68],dx
         mov     dx,word ptr [bp-54]
         mov     word ptr [bp-74],dx
-.Ls071:
+.Ls063:
         mov     dx,word ptr [bp-52]
         mov     word ptr [bp-56],dx
         mov     dx,word ptr [bp-50]
         mov     word ptr [bp-54],dx
-.Ls072:
+.Ls064:
         mov     dx,word ptr [bp-102]
         mov     word ptr [bp-52],dx
         mov     dx,word ptr [bp-100]
         mov     word ptr [bp-50],dx
         mov     word ptr [bp-102],ax
         mov     word ptr [bp-100],cx
-.Ls073:
+.Ls065:
         cmp     word ptr [bp-16],0x7
-        jge     .Ls074
+        jge     .Ls066
         mov     ax,0x8
-        jmp     .Ls075
-.Ls074:
+        jmp     .Ls067
+.Ls066:
         mov     ax,0xb
-.Ls075:
+.Ls067:
         mov     word ptr [bp-16],ax
         mov     bx,word ptr [bp-34]
         add     bx,0xa68
-.Ls076:
+.Ls068:
         mov     word ptr [bp-14],bx
         mov     bx,word ptr [bp-14]
         mov     word ptr [bp-22],bx
         cmp     di,0x100
-        jae     .Ls079
+        jae     .Ls070
         mov     ax,word ptr [bp-12]
         mov     dx,word ptr [bp-10]
         mov     bx,word ptr [bp-30]
         mov     cx,word ptr [bp-28]
-        M_PTC_JNE .Ls077
+        M_PTC_JNE .Ls069
         jmp     .Ls007
-.Ls077:
-        mov     cx,0x8
-.Ls078:
-        shl     si,1
-        rcl     di,1
-        loop    .Ls078
+.Ls069:
+        M_shld_disi_8_bxcx
         les     bx,dword ptr [bp-12]
         mov     al,byte ptr es:[bx]
         mov     byte ptr [bp-90],al
@@ -948,21 +909,21 @@
         M_PIA1
         mov     word ptr [bp-12],ax
         mov     word ptr [bp-10],dx
-.Ls079:
+.Ls070:
         M_shrd_11_disi_bp -96 -98
         mov     bx,word ptr [bp-22]
         mov     bx,word ptr [bx]
-        M_LMUL_dxax_00bx
+        M_U4M_dxax_00bx
         mov     word ptr [bp-8],ax
         mov     word ptr [bp-94],dx
         mov     ax,word ptr [bp-4]
         cmp     ax,dx
-        jb      .Ls080
-        jne     .Ls082
+        jb      .Ls071
+        jne     .Ls073
         mov     ax,word ptr [bp-6]
         cmp     ax,word ptr [bp-8]
-        jae     .Ls082
-.Ls080:
+        jae     .Ls073
+.Ls071:
         mov     si,word ptr [bp-8]
         mov     di,dx
         mov     ax,0x800
@@ -980,10 +941,10 @@
         mov     word ptr [bp-22],dx
         xor     ax,ax
         mov     word ptr [bp-70],ax
-.Ls081:
+.Ls072:
         mov     word ptr [bp-60],0x3
-        jmp     .Ls088
-.Ls082:
+        jmp     .Ls078
+.Ls073:
         sub     si,word ptr [bp-8]
         sbb     di,dx
         mov     ax,word ptr [bp-8]
@@ -999,19 +960,15 @@
         inc     bx
         mov     word ptr [bp-22],bx
         cmp     di,0x100
-        jae     .Ls085
+        jae     .Ls075
         mov     ax,word ptr [bp-12]
         mov     dx,word ptr [bp-10]
         mov     bx,word ptr [bp-30]
         mov     cx,word ptr [bp-28]
-        M_PTC_JNE .Ls083
+        M_PTC_JNE .Ls074
         jmp     .Ls007
-.Ls083:
-        mov     cx,0x8
-.Ls084:
-        shl     si,1
-        rcl     di,1
-        loop    .Ls084
+.Ls074:
+        M_shld_disi_8_bxcx
         les     bx,dword ptr [bp-12]
         mov     al,byte ptr es:[bx]
         mov     byte ptr [bp-90],al
@@ -1028,21 +985,21 @@
         M_PIA1
         mov     word ptr [bp-12],ax
         mov     word ptr [bp-10],dx
-.Ls085:
+.Ls075:
         M_shrd_11_disi_bp -96 -98
         mov     bx,word ptr [bp-22]
         mov     bx,word ptr [bx]
-        M_LMUL_dxax_00bx
+        M_U4M_dxax_00bx
         mov     word ptr [bp-8],ax
         mov     word ptr [bp-94],dx
         mov     ax,word ptr [bp-4]
         cmp     ax,dx
-        jb      .Ls086
-        jne     .Ls087
+        jb      .Ls076
+        jne     .Ls077
         mov     ax,word ptr [bp-6]
         cmp     ax,word ptr [bp-8]
-        jae     .Ls087
-.Ls086:
+        jae     .Ls077
+.Ls076:
         mov     si,word ptr [bp-8]
         mov     di,dx
         mov     ax,0x800
@@ -1059,8 +1016,8 @@
         add     dx,ax
         mov     word ptr [bp-22],dx
         mov     word ptr [bp-70],0x8
-        jmp     .Ls081
-.Ls087:
+        jmp     .Ls072
+.Ls077:
         sub     si,word ptr [bp-8]
         sbb     di,dx
         mov     ax,word ptr [bp-8]
@@ -1076,30 +1033,26 @@
         mov     word ptr [bp-22],bx
         mov     word ptr [bp-70],0x10
         mov     word ptr [bp-60],0x8
-.Ls088:
+.Ls078:
         mov     ax,word ptr [bp-60]
         mov     word ptr [bp-62],ax
         mov     word ptr [bp-24],0x1
-.Ls089:
+.Ls079:
         mov     ax,word ptr [bp-24]
         shl     ax,1
         mov     bx,word ptr [bp-22]
         add     bx,ax
         mov     word ptr [bp-38],bx
         cmp     di,0x100
-        jae     .Ls092
+        jae     .Ls081
         mov     ax,word ptr [bp-12]
         mov     dx,word ptr [bp-10]
         mov     bx,word ptr [bp-30]
         mov     cx,word ptr [bp-28]
-        M_PTC_JNE .Ls090
+        M_PTC_JNE .Ls080
         jmp     .Ls007
-.Ls090:
-        mov     cx,0x8
-.Ls091:
-        shl     si,1
-        rcl     di,1
-        loop    .Ls091
+.Ls080:
+        M_shld_disi_8_bxcx
         les     bx,dword ptr [bp-12]
         mov     al,byte ptr es:[bx]
         mov     byte ptr [bp-90],al
@@ -1116,21 +1069,21 @@
         M_PIA1
         mov     word ptr [bp-12],ax
         mov     word ptr [bp-10],dx
-.Ls092:
+.Ls081:
         M_shrd_11_disi_bp -96 -98
         mov     bx,word ptr [bp-38]
         mov     bx,word ptr [bx]
-        M_LMUL_dxax_00bx
+        M_U4M_dxax_00bx
         mov     word ptr [bp-8],ax
         mov     word ptr [bp-94],dx
         mov     ax,word ptr [bp-4]
         cmp     ax,dx
-        jb      .Ls093
-        jne     .Ls094
+        jb      .Ls082
+        jne     .Ls083
         mov     ax,word ptr [bp-6]
         cmp     ax,word ptr [bp-8]
-        jae     .Ls094
-.Ls093:
+        jae     .Ls083
+.Ls082:
         mov     si,word ptr [bp-8]
         mov     di,dx
         mov     ax,0x800
@@ -1140,8 +1093,8 @@
         shr     ax,cl
         add     word ptr [bx],ax
         shl     word ptr [bp-24],1
-        jmp     .Ls095
-.Ls094:
+        jmp     .Ls084
+.Ls083:
         sub     si,word ptr [bp-8]
         sbb     di,dx
         mov     ax,word ptr [bp-8]
@@ -1156,11 +1109,11 @@
         add     ax,ax
         inc     ax
         mov     word ptr [bp-24],ax
-.Ls095:
+.Ls084:
         dec     word ptr [bp-62]
-        je      .Ls096
-        jmp     .Ls089
-.Ls096:
+        je      .Ls085
+        jmp     .Ls079
+.Ls085:
         mov     cl,byte ptr [bp-60]
         mov     ax,0x1
         shl     ax,cl
@@ -1168,15 +1121,15 @@
         mov     ax,word ptr [bp-70]
         add     word ptr [bp-24],ax
         cmp     word ptr [bp-16],0x4
-        jl      .Ls097
-        jmp     .Ls130
-.Ls097:
+        jl      .Ls086
+        jmp     .Ls116
+.Ls086:
         add     word ptr [bp-16],0x7
         mov     ax,word ptr [bp-24]
         cmp     ax,0x4
-        jl      .Ls098
+        jl      .Ls087
         mov     ax,0x3
-.Ls098:
+.Ls087:
         mov     cl,0x7
         mov     dx,ax
         shl     dx,cl
@@ -1186,26 +1139,22 @@
         mov     word ptr [bp-14],ax
         mov     word ptr [bp-92],0x6
         mov     word ptr [bp-32],0x1
-.Ls099:
+.Ls088:
         mov     ax,word ptr [bp-32]
         shl     ax,1
         mov     bx,word ptr [bp-14]
         add     bx,ax
         mov     word ptr [bp-42],bx
         cmp     di,0x100
-        jae     .Ls102
+        jae     .Ls090
         mov     ax,word ptr [bp-12]
         mov     dx,word ptr [bp-10]
         mov     bx,word ptr [bp-30]
         mov     cx,word ptr [bp-28]
-        M_PTC_JNE .Ls100
+        M_PTC_JNE .Ls089
         jmp     .Ls007
-.Ls100:
-        mov     cx,0x8
-.Ls101:
-        shl     si,1
-        rcl     di,1
-        loop    .Ls101
+.Ls089:
+        M_shld_disi_8_bxcx
         les     bx,dword ptr [bp-12]
         mov     al,byte ptr es:[bx]
         mov     byte ptr [bp-90],al
@@ -1222,21 +1171,21 @@
         M_PIA1
         mov     word ptr [bp-12],ax
         mov     word ptr [bp-10],dx
-.Ls102:
+.Ls090:
         M_shrd_11_disi_bp -96 -98
         mov     bx,word ptr [bp-42]
         mov     bx,word ptr [bx]
-        M_LMUL_dxax_00bx
+        M_U4M_dxax_00bx
         mov     word ptr [bp-8],ax
         mov     word ptr [bp-94],dx
         mov     ax,word ptr [bp-4]
         cmp     ax,dx
-        jb      .Ls103
-        jne     .Ls104
+        jb      .Ls091
+        jne     .Ls092
         mov     ax,word ptr [bp-6]
         cmp     ax,word ptr [bp-8]
-        jae     .Ls104
-.Ls103:
+        jae     .Ls092
+.Ls091:
         mov     si,word ptr [bp-8]
         mov     di,dx
         mov     ax,0x800
@@ -1246,8 +1195,8 @@
         shr     ax,cl
         add     word ptr [bx],ax
         shl     word ptr [bp-32],1
-        jmp     .Ls105
-.Ls104:
+        jmp     .Ls093
+.Ls092:
         sub     si,word ptr [bp-8]
         sbb     di,dx
         mov     ax,word ptr [bp-8]
@@ -1262,15 +1211,15 @@
         add     ax,ax
         inc     ax
         mov     word ptr [bp-32],ax
-.Ls105:
+.Ls093:
         dec     word ptr [bp-92]
-        je      .Ls106
-        jmp     .Ls099
-.Ls106:
+        je      .Ls094
+        jmp     .Ls088
+.Ls094:
         sub     word ptr [bp-32],0x40
         mov     ax,word ptr [bp-32]
         cmp     ax,0x4
-        jl      .Ls109
+        jl      .Ls097
         sar     ax,1
         dec     ax
         mov     word ptr [bp-36],ax
@@ -1281,14 +1230,14 @@
         xor     ax,ax
         mov     word ptr [bp-100],ax
         cmp     word ptr [bp-32],0xe
-        jge     .Ls110
+        jge     .Ls098
         mov     cx,word ptr [bp-36]
-        jcxz    .Ls108
-.Ls107:
+        jcxz    .Ls096
+.Ls095:
         shl     word ptr [bp-102],1
         rcl     word ptr [bp-100],1
-        loop    .Ls107
-.Ls108:
+        loop    .Ls095
+.Ls096:
         mov     dx,word ptr [bp-102]
         shl     dx,1
         mov     ax,word ptr [bp-34]
@@ -1300,26 +1249,22 @@
         dec     dx
         dec     dx
         mov     word ptr [bp-14],dx
-        jmp     .Ls119
-.Ls109:
-        jmp     .Ls128
-.Ls110:
+        jmp     .Ls106
+.Ls097:
+        jmp     .Ls114
+.Ls098:
         sub     word ptr [bp-36],0x4
-.Ls111:
+.Ls099:
         cmp     di,0x100
-        jae     .Ls114
+        jae     .Ls101
         mov     ax,word ptr [bp-12]
         mov     dx,word ptr [bp-10]
         mov     bx,word ptr [bp-30]
         mov     cx,word ptr [bp-28]
-        M_PTC_JNE .Ls112
+        M_PTC_JNE .Ls100
         jmp     .Ls007
-.Ls112:
-        mov     cx,0x8
-.Ls113:
-        shl     si,1
-        rcl     di,1
-        loop    .Ls113
+.Ls100:
+        M_shld_disi_8_bxcx
         les     bx,dword ptr [bp-12]
         mov     al,byte ptr es:[bx]
         mov     byte ptr [bp-90],al
@@ -1336,59 +1281,55 @@
         M_PIA1
         mov     word ptr [bp-12],ax
         mov     word ptr [bp-10],dx
-.Ls114:
+.Ls101:
         shr     di,1
         rcr     si,1
         shl     word ptr [bp-102],1
         rcl     word ptr [bp-100],1
         mov     ax,word ptr [bp-4]
         cmp     di,ax
-        jb      .Ls115
-        jne     .Ls116
+        jb      .Ls102
+        jne     .Ls103
         cmp     si,word ptr [bp-6]
-        ja      .Ls116
-.Ls115:
+        ja      .Ls103
+.Ls102:
         sub     word ptr [bp-6],si
         sbb     word ptr [bp-4],di
         or      byte ptr [bp-102],0x1
-.Ls116:
+.Ls103:
         dec     word ptr [bp-36]
-        je      .Ls117
-        jmp     .Ls111
-.Ls117:
+        je      .Ls104
+        jmp     .Ls099
+.Ls104:
         mov     bx,word ptr [bp-34]
         add     bx,0x644
         mov     word ptr [bp-14],bx
         mov     cx,0x4
-.Ls118:
+.Ls105:
         shl     word ptr [bp-102],1
         rcl     word ptr [bp-100],1
-        loop    .Ls118
+        loop    .Ls105
         mov     word ptr [bp-36],0x4
-.Ls119:
+.Ls106:
         mov     ax,0x1
         mov     word ptr [bp-64],ax
         mov     word ptr [bp-48],ax
-.Ls120:
+.Ls107:
         mov     ax,word ptr [bp-48]
         shl     ax,1
         mov     bx,word ptr [bp-14]
         add     bx,ax
         mov     word ptr [bp-44],bx
         cmp     di,0x100
-        jae     .Ls123
+        jae     .Ls109
         mov     ax,word ptr [bp-12]
         mov     dx,word ptr [bp-10]
         mov     bx,word ptr [bp-30]
         mov     cx,word ptr [bp-28]
-        M_PTC_JNE .Ls121
+        M_PTC_JNE .Ls108
         jmp     .Ls007
-.Ls121:
-        mov     cx,0x8
-.Ls122:
-        shl     si,1
-        rcl     di,1
-        loop    .Ls122
+.Ls108:
+        M_shld_disi_8_bxcx
         les     bx,dword ptr [bp-12]
         mov     al,byte ptr es:[bx]
         mov     byte ptr [bp-90],al
@@ -1405,21 +1346,21 @@
         M_PIA1
         mov     word ptr [bp-12],ax
         mov     word ptr [bp-10],dx
-.Ls123:
+.Ls109:
         M_shrd_11_disi_bp -96 -98
         mov     bx,word ptr [bp-44]
         mov     bx,word ptr [bx]
-        M_LMUL_dxax_00bx
+        M_U4M_dxax_00bx
         mov     word ptr [bp-8],ax
         mov     word ptr [bp-94],dx
         mov     ax,word ptr [bp-4]
         cmp     ax,dx
-        jb      .Ls124
-        jne     .Ls125
+        jb      .Ls110
+        jne     .Ls111
         mov     ax,word ptr [bp-6]
         cmp     ax,word ptr [bp-8]
-        jae     .Ls125
-.Ls124:
+        jae     .Ls111
+.Ls110:
         mov     si,word ptr [bp-8]
         mov     di,dx
         mov     ax,0x800
@@ -1429,8 +1370,8 @@
         shr     ax,cl
         add     word ptr [bx],ax
         shl     word ptr [bp-48],1
-        jmp     .Ls126
-.Ls125:
+        jmp     .Ls112
+.Ls111:
         sub     si,word ptr [bp-8]
         sbb     di,dx
         mov     ax,word ptr [bp-8]
@@ -1449,37 +1390,37 @@
         cwd
         or      word ptr [bp-102],ax
         or      word ptr [bp-100],dx
-.Ls126:
+.Ls112:
         shl     word ptr [bp-64],1
         dec     word ptr [bp-36]
-        je      .Ls127
-        jmp     .Ls120
-.Ls127:
-        jmp     .Ls129
-.Ls128:
+        je      .Ls113
+        jmp     .Ls107
+.Ls113:
+        jmp     .Ls115
+.Ls114:
         cwd
         mov     word ptr [bp-102],ax
         mov     word ptr [bp-100],dx
-.Ls129:
+.Ls115:
         add     word ptr [bp-102],0x1
         adc     word ptr [bp-100],0x0
         mov     ax,word ptr [bp-100]
         or      ax,word ptr [bp-102]
-        jne     .Ls130
-        jmp     .Ls136
-.Ls130:
+        jne     .Ls116
+        jmp     .Ls122
+.Ls116:
         add     word ptr [bp-24],0x2
         mov     ax,word ptr [bp-100]
         cmp     ax,word ptr [bp-18]
-        jbe     .Ls132
-.Ls131:
+        jbe     .Ls118
+.Ls117:
         jmp     .Ls007
-.Ls132:
-        jne     .Ls133
+.Ls118:
+        jne     .Ls119
         mov     ax,word ptr [bp-102]
         cmp     ax,word ptr [bp-20]
-        ja      .Ls131
-.Ls133:
+        ja      .Ls117
+.Ls119:
         mov     bx,word ptr [bp-20]
         sub     bx,word ptr [bp-102]
         mov     cx,word ptr [bp-18]
@@ -1504,34 +1445,34 @@
         mov     al,byte ptr [bp-2]
         mov     byte ptr es:[bx],al
         cmp     word ptr [bp-24],0x0
-        jne     .Ls135
-.Ls134:
+        jne     .Ls121
+.Ls120:
         jmp     .Ls008
-.Ls135:
+.Ls121:
         mov     ax,word ptr [bp-18]
         cmp     ax,word ptr [bp+26]
-        jb      .Ls133
-        jne     .Ls134
+        jb      .Ls119
+        jne     .Ls120
         mov     ax,word ptr [bp-20]
         cmp     ax,word ptr [bp+24]
-        jb      .Ls133
-        jmp     .Ls134
-.Ls136:
+        jb      .Ls119
+        jmp     .Ls120
+.Ls122:
         cmp     di,0x100
-        jae     .Ls138
+        jae     .Ls124
         mov     ax,word ptr [bp-12]
         mov     dx,word ptr [bp-10]
         mov     bx,word ptr [bp-30]
         mov     cx,word ptr [bp-28]
-        M_PTC_JNE .Ls137
+        M_PTC_JNE .Ls123
         jmp     .Ls007
-.Ls137:
+.Ls123:
         mov     ax,word ptr [bp-12]
         mov     dx,word ptr [bp-10]
         M_PIA1
         mov     word ptr [bp-12],ax
         mov     word ptr [bp-10],dx
-.Ls138:
+.Ls124:
         mov     ax,word ptr [bp-12]
         mov     dx,word ptr [bp-10]
         mov     bx,word ptr [bp+10]
@@ -1546,7 +1487,7 @@
         mov     ax,word ptr [bp-18]
         mov     word ptr [bx+2],ax
         xor     ax,ax
-.Ls139:
+.Ls125:
         mov     sp,bp
         pop     bp
         pop     di
diff --git a/src/stub/src/arch/i086/lzma_m.h b/src/stub/src/arch/i086/lzma_m.h
index fef1cdbd..a8608cf9 100644
--- a/src/stub/src/arch/i086/lzma_m.h
+++ b/src/stub/src/arch/i086/lzma_m.h
@@ -144,11 +144,11 @@ L1:
 
 
 /*************************************************************************
-// support macros: LMUL, shld, shrd
+// support macros: U4M, shld, shrd
 **************************************************************************/
 
 // umul32: dx:ax = dx:ax * 00:bx
-.macro  M_LMUL_dxax_00bx
+.macro  M_U4M_dxax_00bx
         // mult high-word
         mov     cx, ax      // save ax
         mov     ax, dx
@@ -162,7 +162,7 @@ L1:
 
 
 // umul32: dx:ax = dx:ax * word ptr [bx]
-.macro  M_LMUL_dxax_00bx_ptr
+.macro  M_U4M_dxax_00bx_ptr
         // mult high-word
         mov     cx, ax      // save ax
         mov     ax, dx
@@ -176,7 +176,7 @@ L1:
 
 
 // umul32: dx:ax = ax:cx * 00:bx
-.macro  M_LMUL_axcx_00bx
+.macro  M_U4M_axcx_00bx
         // mult high-word
         mul     bx
         xchg    ax, cx      // save high-word result, get low
@@ -187,14 +187,95 @@ L1:
 .endm
 
 
+// umul32: dx:ax = dx:ax * 0x0600
+.macro  M_U4M_dxax_0x0600
+    // FIXME: compute clocks and optimize this
+#if 1
+        // code size: 18 bytes
+        // i086: > 140 clocks (mul needs 70 clocks)
+        // i286: >  26 clocks (mul needs 13 clocks)
+        mov     bx, 0x300
+        M_U4M_dxax_00bx
+        shl     ax
+        rcl     dx
+#elif 1
+        // code size: 14 bytes
+        // i086: > 140 clocks (mul needs 70 clocks)
+        // i286: >  26 clocks (mul needs 13 clocks)
+        mov     bx, 0x600
+        M_U4M_dxax_00bx
+#else
+        // code size: 16+8 == 24 bytes
+        // i086: 18+9 == 27 clocks
+        // i286: 16+8 == 24 clocks
+    // FIXME: can we further optimize this ?
+        shl     ax
+        rcl     dx          // dx:ax <<= 1      v * 2
+        mov     cx, dx
+        mov     bx, ax      // cx:bx = dx:ax    v * 2
+        shl     ax
+        rcl     dx          // dx:ax <<= 1      v * 4
+        add     ax, bx
+        adc     dx, cx      // dx:ax += cx:bx   v * 6
+        M_shld_8            // dx:ax <<= 8      v * 0x600
+#endif
+.endm
+
+
 // shld: dx:ax <<= 8
 .macro  M_shld_8
+        // code size: 8 bytes
+        // i086: 9 clocks
+        // i286: 8 clocks
         mov     dh, dl
         mov     dl, ah
         mov     ah, al
         xor     al, al
 .endm
 
+
+// shld: di:si <<= 8; bx and cx are free
+.macro  M_shld_disi_8_bxcx
+    // FIXME: compute clocks and optimize this
+#if 1
+        local   L1
+        mov     cx, 8
+L1:     shl     si
+        rcl     di
+        loop    L1
+#else
+        mov     bx, di
+        mov     cx, si
+        mov     bh, bl
+        mov     bl, ch
+        mov     ch, cl
+        xor     cl, cl
+        mov     di, bx
+        mov     si, cx
+#endif
+.endm
+
+
+// shld: di:ax <<= 8; bx and cx are free
+.macro  M_shld_diax_8_bxcx
+    // FIXME: compute clocks and optimize this
+#if 1
+        local   L1
+        mov     cx, 8
+L1:     shl     ax
+        rcl     di
+        loop    L1
+#else
+        mov     bx, di
+        mov     bh, bl
+        mov     bl, ah
+        mov     ah, al
+        xor     al, al
+        mov     di, bx
+#endif
+.endm
+
+
 .macro  M_shld_8_bp h l
         mov     dx, word ptr[bp+h]
         mov     ax, word ptr[bp+l]