diff --git a/src/stub/Makefile b/src/stub/Makefile
index 01608172..3ae7032e 100644
--- a/src/stub/Makefile
+++ b/src/stub/Makefile
@@ -637,13 +637,19 @@ mipsel.r3000-ps1-%.h : tc_list = mipsel.r3000-ps1 default
 
 tc.mipsel.r3000-ps1.app-asm5900 = perl -w $(srcdir)/src/arch/mips/mipsel.r3000/app-asm5900.pl
 tc.mipsel.r3000-ps1.asm5900     = asm5900
+tc.mipsel.r3000-ps1.gas         = asm5900
 
 mipsel.r3000-ps1-%.h : $(srcdir)/src/mipsel.r3000-ps1.asm
 	# call gpp_inc to generate .d file
 	$(call tc,gpp_inc) --mode=c --MMD=$@ --MF=tmp/$T.tmp1.d $< -o /dev/null
 	$(call tc,pp-asm) $(PP_FLAGS) $< -o tmp/$T.tmp1
-	$(call tc,app-asm5900) tmp/$T.tmp1 tmp/$T.tmp2
-	$(call tc,asm5900) --nologo -q -ltmp/$T.bin.lst tmp/$T.tmp2 -otmp/$T.bin
+	#$(call tc,app-asm5900) tmp/$T.tmp1 tmp/$T.tmp2
+	#$(call tc,asm5900) --nologo -q -ltmp/$T.bin.lst tmp/$T.tmp2 -otmp/$T.bin
+
+	$(call tc,m-objcopy) --strip-unneeded tmp/$T.bin
+	$(call tc,m-objcopy) -R .text -R .data -R .bss tmp/$T.bin
+	$(call tc,m-objcopy) -R .note -R .comment tmp/$T.bin
+	$(call tc,m-objdump) -trwh tmp/$T.bin >> tmp/$T.bin
 	$(call tc,bin2h) --ident=$(IDENT_NAME) tmp/$T.bin $@
 
 mipsel.r3000-ps1-boot.h:    PP_FLAGS  += -DCDBOOT
diff --git a/src/stub/src/arch/mips/mipsel.r3000/bits.ash b/src/stub/src/arch/mips/mipsel.r3000/bits.ash
index 22a4b089..299eb2d7 100644
--- a/src/stub/src/arch/mips/mipsel.r3000/bits.ash
+++ b/src/stub/src/arch/mips/mipsel.r3000/bits.ash
@@ -1,3 +1,4 @@
+/*
 ;  bits.ash -- bit access for decompression
 ;
 ;  This file is part of the UCL data compression library.
@@ -24,14 +25,15 @@
 ;  <markus@oberhumer.com>               <jssg@users.sourceforge.net>
 ;  http://www.oberhumer.com/opensource/ucl/
 ;
+*/
 
 #ifndef _MR3K_STD_CONF_
 #define _MR3K_STD_CONF_
 
 
-;//////////////////////////////////////
-;// register defines
-;//////////////////////////////////////
+//////////////////////////////////////
+// register defines
+//////////////////////////////////////
 
 #define src         a0
 #define dst         a2
@@ -48,40 +50,41 @@
 #define m_pos       v1
 
 
-;//////////////////////////////////////
-;// optimized branch macros
-;//////////////////////////////////////
+//////////////////////////////////////
+// optimized branch macros
+//////////////////////////////////////
 
-beqz2gb     MACRO   reg,label,nrv_bb
-    IF (!small) && (nrv_bb == 8)
-            beqz    reg,label+4
+.macro     beqz2gb  reg,label,nrv_bb
+    IF (!small) && (\nrv_bb == 8)
+            beqz    \reg,\label+4
             andi    var,bb,0x007F
     ELSE
-            beqz    reg,label
+            beqz    \reg,\label
             nop
     ENDIF
-beqz2gb     ENDM
+.endm
 
-b2gb        MACRO   label,nrv_bb
+.macro      b2gb    label,nrv_bb
     IF (!small)
-            b       label+4
-        IF (nrv_bb == 8)
+            b       \label+4
+        IF (\nrv_bb == 8)
             andi    var,bb,0x007F
-        ELSE ;(nrv_bb == 32)
+        ELSE // ;(nrv_bb == 32)
             addiu   bc,-1
         ENDIF
     ELSE
-            b       label
+            b       \label
             nop
     ENDIF
-b2gb        ENDM
+.endm
 
 
-;//////////////////////////////////////
-;// ucl memcpy
-;//////////////////////////////////////
+//////////////////////////////////////
+// ucl memcpy
+//////////////////////////////////////
 
-uclmcpy     MACRO   retoffset,nrv_bb
+.macro      uclmcpy retoffset,nrv_bb
+            local   wordchk, prepbytecpy, bytecopy
 #   ifdef FAST
             slti    var,m_off,4
             bnez    var,prepbytecpy
@@ -98,7 +101,7 @@ wordchk:
             addiu   m_pos,4
             bnez    m_len,wordchk
             addiu   dst,4
-            b2gb    retoffset,nrv_bb
+            b2gb    \retoffset,\nrv_bb
 prepbytecpy:
 #   else
             addiu   m_len,1
@@ -111,82 +114,90 @@ bytecopy:
             addiu   m_pos,1
             bnez    m_len,bytecopy
             addiu   dst,1
-            b2gb    retoffset,nrv_bb
-uclmcpy     ENDM
+            b2gb    \retoffset,\nrv_bb
+.endm
 
 
-;//////////////////////////////////////
-;// init decompressor
-;//////////////////////////////////////
+//////////////////////////////////////
+// init decompressor
+//////////////////////////////////////
 
-init        MACRO   nrv_bb
+.macro      init    nrv_bb
             move    bb,zero
-    IF (nrv_bb == 32)
+    IF (\nrv_bb == 32)
             move    bc,bb
     ENDIF
             li      last_m_off,1
             move    src_ilen,src
-init        ENDM
+.endm
+
+#ifndef LOCLABELHACK
+# define LOCLABELHACK 0
+#else
+# undef __LL
+#
+#endif
 
 
-;//////////////////////////////////////
-;// 32bit getbit macro
-;//////////////////////////////////////
+//////////////////////////////////////
+// 32bit getbit macro
+//////////////////////////////////////
 
-gbit_le32   MACRO
+.macro      gbit_le32
+local .L1
     IF (!small)
             addiu   bc,-1
     ENDIF
-            bgez    bc,\@z
+            bgez    bc,.L1
             srlv    var,bb,bc
             li      bc,31
             lwr     bb,0(src_ilen)
             lwl     bb,3(src_ilen)
             addiu   src_ilen,4
             srlv    var,bb,bc
-\@z:
+.L1:
     IF (small)
             jr      ra
     ENDIF
             andi    var,0x0001
-gbit_le32   ENDM
+.endm
 
 
-;//////////////////////////////////////
-;// 8bit getbit macro
-;//////////////////////////////////////
+//////////////////////////////////////
+// 8bit getbit macro
+//////////////////////////////////////
 
-gbit_8      MACRO
+.macro      gbit_8
+local .L2
     IF (!small)
             andi    var,bb,0x007F
     ENDIF
-            bnez    var,\@x
+            bnez    var,.L2
             sll     bb,1
             lbu     var,0(src_ilen)
             addiu   src_ilen,1
             sll     var,1
             addiu   bb,var,1
-\@x:
+.L2:
             srl     var,bb,8
     IF (small)
             jr      ra
     ENDIF
             andi    var,0x0001
-gbit_8      ENDM
+.endm
 
+//////////////////////////////////////
+// getbit call macro for small version
+//////////////////////////////////////
 
-;//////////////////////////////////////
-;// getbit call macro for small version
-;//////////////////////////////////////
-
-gbit_call   MACRO   subroutine,nrv_bb
-            bal     subroutine
-    IF (nrv_bb == 8)
+.macro      gbit_call subroutine,nrv_bb
+            bal     \subroutine
+    IF (\nrv_bb == 8)
             andi    var,bb,0x007F
     ELSE
             addiu   bc,-1
     ENDIF
-gbit_call   ENDM
+.endm
 
 
 #endif  //_MR3K_STD_CONF_
diff --git a/src/stub/src/arch/mips/mipsel.r3000/macros.ash b/src/stub/src/arch/mips/mipsel.r3000/macros.ash
index 65528389..fa9b06f5 100644
--- a/src/stub/src/arch/mips/mipsel.r3000/macros.ash
+++ b/src/stub/src/arch/mips/mipsel.r3000/macros.ash
@@ -1,3 +1,4 @@
+/*
 ;  macros.ash -- macros
 ;
 ;  This file is part of the UCL data compression library.
@@ -24,29 +25,74 @@
 ;  <markus@oberhumer.com>               <jssg@users.sourceforge.net>
 ;  http://www.oberhumer.com/opensource/ucl/
 ;
+*/
 
-; Load Address macro
+.macro  section name
+        .section \name
+.endm
 
-la MACRO reg,addr
-        lui     reg,(addr >> 16)+((addr >> 15)&1)
-        addiu   reg,reg,((addr & 0FFFFh))
-la ENDM
+.set    noat
+.altmacro
 
-; Load Immidiate macro
+#define zero    $0
+#define at      $1
+#define v0      $2
+#define v1      $3
+#define a0      $4
+#define a1      $5
+#define a2      $6
+#define a3      $7
+#define t0      $8
+#define t1      $9
+#define t2      $10
+#define t3      $11
+#define t4      $12
+#define t5      $13
+#define t6      $14
+#define t7      $15
+#define sp      $29
+#define ra      $31
 
-li MACRO reg,imm
-IF (imm == -1)
-        addiu    reg, zero, -1
+#define IF      .if
+#define ELSE    .else
+#define ENDIF   .endif
+
+#define DW      .long
+#define DB      .byte
+
+// Load Address macro
+
+.macro  la      reg,addr
+        lui     \reg,(\addr >> 16)+((\addr >> 15)&1)
+        addiu   \reg,\reg,((\addr & 0xFFFF))
+.endm
+
+// Load Immidiate macro
+
+#if 0
+.macro  li      reg,imm
+IF (\imm == -1)
+        addiu    \reg, zero, -1
 ELSE
-        IF ((imm>>16)!=0)
-                lui     reg,imm >> 16
-                ori     reg,reg,(imm & 0FFFFh)
+        IF ((\imm>>16)!=0)
+                lui     \reg,\imm >> 16
+                ori     \reg,\reg,(\imm & 0xFFFF)
         ELSE
-                IF ((imm&0xffff)>0x8000)
-                        ori     reg,zero,(imm & 0FFFFh)
+                IF ((\imm&0xffff)>0x8000)
+                        ori     \reg,zero,(\imm & 0xFFFF)
                 ELSE
-                        addiu   reg,zero,(imm & 0FFFFh)
+                        addiu   \reg,zero,(\imm & 0xFFFF)
                 ENDIF
         ENDIF
 ENDIF
-li ENDM
+.endm
+#endif
+
+.macro  subiu   reg, p1, p2
+        .if p2
+        addiu   \reg, p1, -p2
+        .else
+        addiu   \reg, -p1
+        .endif
+.endm
+
diff --git a/src/stub/src/arch/mips/mipsel.r3000/nrv2e_d.ash b/src/stub/src/arch/mips/mipsel.r3000/nrv2e_d.ash
index 2f62cad7..db6cc617 100644
--- a/src/stub/src/arch/mips/mipsel.r3000/nrv2e_d.ash
+++ b/src/stub/src/arch/mips/mipsel.r3000/nrv2e_d.ash
@@ -1,3 +1,4 @@
+/*
 ;  n2e_d.ash -- NRV2E decompressor in Mips R3000 assembly
 ;
 ;  This file is part of the UCL data compression library.
@@ -24,11 +25,13 @@
 ;  <markus@oberhumer.com>               <jssg@users.sourceforge.net>
 ;  http://www.oberhumer.com/opensource/ucl/
 ;
+*/
 
+small = 0
 #if (NRV_BB==8)
 #   ifdef SMALL
         IF (!small)
-            DEFINE  small
+            small = 1
         ENDIF
 #       define  UCL_DECOMPRESSOR    ucl_nrv2e_decompress_8_small
 #       define  GETBIT              gbit_call gbit_sub,NRV_BB
@@ -42,7 +45,7 @@
 #       define  GETBIT              gbit_call gbit_sub,NRV_BB
 #   else
         IF (small)
-            UNDEF  small
+            small = 0
         ENDIF
 #       define  UCL_DECOMPRESSOR    ucl_nrv2e_decompress_32
 #       define  GETBIT              gbit_le32
@@ -51,20 +54,23 @@
 #   error "define NRV_BB first!"
 #endif
 
-
 #include "bits.ash"
 
+/*
 ; ------------- DECOMPRESSION -------------
 ; On entry:
 ;   a0  src pointer
 ;   a2  dest pointer
+*/
 
 
-UCL_DECOMPRESSOR    PROC
+.macro  UCL_DECOMPRESSOR
+        local   n2e_18, n2e_4C, n2e_E4, n2e_124, n2e_150, n2e_184
+        local   n2e_1E0, n2e_244, gbit_sub, n2e_decomp_done
         init    NRV_BB
 n2e_18:
         GETBIT
-        beqz    var,n2e_4C
+        beqz    var, n2e_4C
         li      m_off,1
         lbu     var,0(src_ilen)
         addiu   src_ilen,1
@@ -139,7 +145,8 @@ gbit_sub:
 #endif
 
 n2e_decomp_done:
-UCL_DECOMPRESSOR    ENDP
+.endm
+        UCL_DECOMPRESSOR
 
 #undef  UCL_DECOMPRESSOR
 #undef  GETBIT
diff --git a/src/stub/src/mipsel.r3000-ps1.asm b/src/stub/src/mipsel.r3000-ps1.asm
index 0d60c9c9..29acd6fc 100644
--- a/src/stub/src/mipsel.r3000-ps1.asm
+++ b/src/stub/src/mipsel.r3000-ps1.asm
@@ -1,3 +1,4 @@
+/*
 ;  l_ps1.asm -- ps1/exe program entry & decompressor
 ;
 ;  This file is part of the UPX executable compressor.
@@ -28,6 +29,7 @@
 ;  Jens Medoch
 ;  <jssg@users.sourceforge.net>
 ;
+*/
 
 
 #include "arch/mips/mipsel.r3000/macros.ash"
@@ -38,105 +40,103 @@
 #define  SZ_REG  4
 
 #if CDBOOT
-regs            MACRO   _w, marker
-                _w      a0,SZ_REG*0(sp)
-                _w      a1,SZ_REG*1(sp)
-                _w      a2,SZ_REG*2(sp)
-                _w      a3,SZ_REG*3(sp)
-                _w      v0,SZ_REG*4(sp)
-                _w      v1,SZ_REG*5(sp)
-IF (marker != 0)
-                DW      marker
+.macro          regs    _w, marker
+                \_w     a0,SZ_REG*0(sp)
+                \_w     a1,SZ_REG*1(sp)
+                \_w     a2,SZ_REG*2(sp)
+                \_w     a3,SZ_REG*3(sp)
+                \_w     v0,SZ_REG*4(sp)
+                \_w     v1,SZ_REG*5(sp)
+IF (\marker != 0)
+                DW      \marker
 ENDIF
-                _w      ra,SZ_REG*6(sp)
-regs            ENDM
+                \_w     ra,SZ_REG*6(sp)
+.endm
 
 #define  REG_SZ (7*SZ_REG)
 
 #else //console
 
-regs            MACRO   _w, marker
-                _w      at,SZ_REG*0(sp)
-                _w      a0,SZ_REG*1(sp)
-                _w      a1,SZ_REG*2(sp)
-                _w      a2,SZ_REG*3(sp)
-                _w      a3,SZ_REG*4(sp)
-                _w      v0,SZ_REG*5(sp)
-                _w      v1,SZ_REG*6(sp)
-                _w      ra,SZ_REG*7(sp)
-IF (marker != 0)
-                DW      marker
+.macro          regs    _w, marker
+                \_w     at,SZ_REG*0(sp)
+                \_w     a0,SZ_REG*1(sp)
+                \_w     a1,SZ_REG*2(sp)
+                \_w     a2,SZ_REG*3(sp)
+                \_w     a3,SZ_REG*4(sp)
+                \_w     v0,SZ_REG*5(sp)
+                \_w     v1,SZ_REG*6(sp)
+                \_w     ra,SZ_REG*7(sp)
+IF (\marker != 0)
+                DW      \marker
                 addu    sp,at
 ENDIF
-regs            ENDM
+.endm
 
 #define  REG_SZ (8*SZ_REG)
 
 #endif //CDBOOT
 
-                ORG      0
-start:
-
 #if CDBOOT
-; =============
-; ============= ENTRY POINT
-; =============
-; for cd-boot only
+// =============
+// ============= ENTRY POINT
+// =============
+// for cd-boot only
 
-;       __PS1START__
-                li      t0,'PSVR'       ; prepare to compute value
-                subu    t0,s0,t0        ; get stored header offset in mem
+section         PS1START
+                li      t0, PSVR  // prepare to compute value
+                subu    t0,s0,t0        // get stored header offset in mem
                 jr      t0
-                subiu   sp,REG_SZ       ; adjust stack
+                subiu   sp, REG_SZ         // adjust stack
 cutpoint:
-;       __PS1ENTRY__
-                regs    sw,0            ; push used regs
-                li      a0,'CPDO'       ; load COMPDATA offset
-;               li      a1,'CDSZ'       ; compressed data size - disabled!
-;       __PS1CONHL__
-                li      a2,'DECO'
-;       __PS1CONHI__
-                lui     a2,HI('DECO')
+section         PS1ENTRY
+                regs    sw,0            // push used regs
+                li      a0, CPDO  // load COMPDATA offset
+//               li      a1,'CDSZ'       // compressed data size - disabled!
+section         PS1CONHL
+                li      a2, DECO
+section         PS1CONHI
+                lui     a2,HI( DECO)
 
 
 #else //CONSOLE
-; =============
-; ============= ENTRY POINT
-; =============
-; for console- / cd-boot
+// =============
+// ============= ENTRY POINT
+// =============
+// for console- / cd-boot
 
-;       __PS1START__
-                addiu   at,zero,'LS'    ; size of decomp. routine
-                subu    sp,at           ; adjust the stack with this size
-                regs    sw,0            ; push used regs
-                subiu   a2,at,REG_SZ    ; a2 = counter copyloop
-                addiu   a3,sp,REG_SZ    ; get offset for decomp. routine
+section         PS1START
+                addiu   at,zero, LS // size of decomp. routine
+                subu    sp,at           // adjust the stack with this size
+                regs    sw,0            // push used regs
+                subiu   a2,at,REG_SZ    // a2 = counter copyloop
+                addiu   a3,sp,REG_SZ    // get offset for decomp. routine
                 move    a1,a3
-                li      a0,'DCRT'       ; load decompression routine's offset
+                lui     a0, DCRT    // load decompression routine's offset
+//                ori     a0, LO(DCRT)    // load decompression routine's offset
 copyloop:
-                lw      at,0(a0)        ; memcpy *a0 -> at -> *a1
+                lw      at,0(a0)        // memcpy *a0 -> at -> *a1
                 addiu   a2,-4
                 sw      at,0(a1)
                 addiu   a0,4
                 bnez    a2,copyloop
                 addiu   a1,4
 
-;       __PS1PADCD__
-                addiu   a0,'PC'         ; a0 = pointer compressed data
-;       __PS1CONHL__
-                lui     a2,HI('DECO')   ; load DECOMPDATA HI offset
+section         PS1PADCD
+                addiu   a0, PC    // a0 = pointer compressed data
+section         PS1CONHL
+                lui     a2,HI( DECO)  // load DECOMPDATA HI offset
                 jr      a3
-                ori     a2,LO('DECO')   ; load DECOMPDATA LO offset
-;       __PS1CONHI__
+                ori     a2,LO( DECO)  // load DECOMPDATA LO offset
+section         PS1CONHI
                 jr      a3
-                lui     a2,HI('DECO')   ; same for HI only !(offset&0xffff)
+                lui     a2,HI( DECO)  // same for HI only !(offset&0xffff)
 cutpoint:
-;       __PS1ENTRY__
+section         PS1ENTRY
 
 #endif //CDBOOT
-; =============
-; ============= DECOMPRESSION
-; =============
+// =============
+// ============= DECOMPRESSION
+// =============
 
 #ifndef FAST
 #   define FAST
@@ -157,11 +157,11 @@ cutpoint:
 #endif
 #define NRV_BB 8
 
-;       __PS1N2B08__
-#include "arch/mips/mipsel.r3000/nrv2b_d.ash"
-;       __PS1N2D08__
-#include "arch/mips/mipsel.r3000/nrv2d_d.ash"
-;       __PS1N2E08__
+section         PS1N2B08
+//#include "arch/mips/mipsel.r3000/nrv2b_d.ash"
+section         PS1N2D08
+//#include "arch/mips/mipsel.r3000/nrv2d_d.ash"
+section         PS1N2E08
 #include "arch/mips/mipsel.r3000/nrv2e_d.ash"
 
 #ifdef NRV_BB
@@ -169,28 +169,28 @@ cutpoint:
 #endif
 #define NRV_BB 32
 
-;       __PS1N2B32__
-#include "arch/mips/mipsel.r3000/nrv2b_d.ash"
-;       __PS1N2D32__
-#include "arch/mips/mipsel.r3000/nrv2d_d.ash"
-;       __PS1N2E32__
+section         PS1N2B32
+//#include "arch/mips/mipsel.r3000/nrv2b_d.ash"
+section         PS1N2D32
+//#include "arch/mips/mipsel.r3000/nrv2d_d.ash"
+section         PS1N2E32
 #include "arch/mips/mipsel.r3000/nrv2e_d.ash"
 
 
-; =============
+// =============
 
-;       __PS1MSETS__
-                ori     a0,zero,'SC'    ; amount of removed zeros at eof
-;       __PS1MSETB__
-                ori     a0,zero,'SC'    ; amount of removed zeros at eof
-                sll     a0,3            ; (cd mode 2 data sector alignment)
-;       __PS1MSETA__
+section         PS1MSETS
+                ori     a0,zero, SC // amount of removed zeros at eof
+section         PS1MSETB
+                ori     a0,zero, SC // amount of removed zeros at eof
+                sll     a0,3              // (cd mode 2 data sector alignment)
+section         PS1MSETA
 memset_aligned:
                 sw      zero,0(a2)
                 addiu   a0,-1
                 bnez    a0,memset_aligned
                 addiu   a2,4
-;       __PS1MSETU__
+section         PS1MSETU
 memset_unaligned:
                 swl     zero,3(a2)
                 swr     zero,0(a2)
@@ -198,41 +198,23 @@ memset_unaligned:
                 bnez    a0,memset_unaligned
                 addiu   a2,4
 
-; =============
+// =============
 
-;       __PS1EXITC__
-                li      t2,160          ; flushes
-                jalr    ra,t2           ; instruction
-                li      t1,68           ; cache
-                regs    lw, 'JPEP'      ; marker for the entry jump
+section         PS1EXITC
+                li      t2,160          // flushes
+                jalr    ra,t2           // instruction
+                li      t1,68           // cache
+                regs    lw, JPEP // marker for the entry jump
 
-; =============
+// =============
 
-;       __PS1PAHDR__
-                DB      85,80,88,33     ;  0  UPX_MAGIC_LE32
-        ; another magic for PackHeader::putPackHeader
-                DB      161,216,208,213 ;     UPX_MAGIC2_LE32
-                DW      0               ;  8  uncompressed adler32
-                DW      0               ; 12  compressed adler32
-                DW      0               ; 16  uncompressed len
-                DW      0               ; 20  compressed len
-                DW      0               ; 24  original file size
-                DB      0               ; 28  filter id
-                DB      0               ; 29  filter cto
-                DB      0               ;  unsused
-                DB      45              ; 31  header checksum
+#include        "include/header2.ash"
 
-; =============
+// =============
 
 #if !CDBOOT
-;       __PS1SREGS__
+section         PS1SREGS
                 DW      REG_SZ
 #endif //CDBOOT
 
-;       __PS1EOASM__
-eof:
-;               section .data
-                DW_UNALIGNED    -1
-                DH_UNALIGNED    eof
-
-; vi:ts=8:et:nowrap
+// vi:ts=8:et:nowrap