diff --git a/bochs/patches/patch.perf-regparm-cclark b/bochs/patches/patch.perf-regparm-cclark
new file mode 100644
index 000000000..7efe3c708
--- /dev/null
+++ b/bochs/patches/patch.perf-regparm-cclark
@@ -0,0 +1,2006 @@
+----------------------------------------------------------------------
+Patch name: patch.perf-regparm-cclark
+Author: Conn Clark
+Date: Wed Feb 12 15:21:37 CET 2003
+Status: applyable, to be updated
+
+Detailed description:
+
+	Here is the patch for my modifications to source in bochs-2.0.1.tar.gz. 
+All my changes were done to this code except I did update the vga.cc and 
+vga.h files from the CVS tree in a failed attempt to 8 bit color 
+working. The patch was generated by using the command "diff -u" if it 
+wasn't done right let me know and I will try again. (I have only created 
+a patch once before in my life and I did it wrong)
+
+	Basicly I have found that Bochs, like all other large properly written 
+well organized programs (No offense), spends a good portion of its time 
+just passing parameters from one function to another. To reduce the 
+effect of this I have used the GCC function atribute "regparm(n)" to 
+declare that certain functions use the register calling convention. The 
+  "n" in "regparm(n)" is the number of variables to pass in registers. 
+All other code and logic in bochs were left untouched. Note that as far 
+as I know this only works on GCC 3.2.X for X86 processors. Also note 
+that there is a trick to using it that is not well documented in the gcc 
+man pages. See examples below for implementation details.
+
+
+/**** Normal code ****/
+
+/* funct declaration */
+int	foo(int a, int b, int c, int d);
+
+/* function */
+int
+  foo( int a, int b, int c, int d)
+{
+int e;
+
+e = a + b + c + d;
+
+return e;
+}
+
+
+
+/**** Code using function attribute ****/
+
+/* funct declaration */
+int	foo(int a, int b, int c, int d) __attribute__((regparm(3)));
+
+/* function */
+int __attribute__((regparm(3)))
+  foo( int a, int b, int c, int d)
+{
+int e;
+
+e = a + b + c;
+e -= d;
+
+return e;
+}
+
+The trick is you put the function attribute after the function 
+declaration and before its terminating semi-colon. However in the 
+function you place it after the return type and before the function name.
+
+	The behavior of the above example is as follows varables a, b, and c 
+are passed in registers and d is passed on the stack (I think, if I'm 
+wrong a is passed on the stack). The best canidates to apply this 
+optimization to are functions that make immeadiate use of the registered 
+variables. functions that do not use the registered variables right away 
+may have to store the variables to the stack temporarily which may(but 
+not always) result in slower code for the function. This may still 
+create an overall benefit because using the register calling convention 
+may speed up the function calling this one offsetting the penalty.
+
+	The one place in Bochs where I think the register calling convention 
+may do some real good is the individual functions for each instruction 
+and the Resolve??Mod?Rm functions. I did not get this to work succesfuly 
+when I attempted it the first time so I gave up and applied the 
+optimization to everything else I could. It should be possible though I 
+just am not quite familiar with the code to pull it off yet.
+
+
+Here are the extra compiler flags I used to create my executable with.
+
+-O2 -march=pentium2 -mmmx -fthread-jumps -fexpensive-optimizations 
+-frename-registers -fomit-frame-pointer -finline-functions 
+-minline-all-stringops
+
+This may effect which functions it pays to use the register calling 
+convention with. The only real way to tell is to look at the assembly 
+code generated for the function and mabey the calling function or to 
+benchmark bochs for each function changed. It tends to help more than 
+hinder.
+
+Here are some details on the machine I used.
+All the work I did was on an overclocked  416 MHz dual processor ( I 
+know the second processor doesn't help) pentium2 with 512MB of 75Mhz 
+SDRAM and a PCI Diamond stealth 3D 3000 (Ancient I know, but in its day 
+it rocked). The Os is Red hat 8.0 patched up to date as of 1/10/03 . 
+Redhat 8.0 uses GCC 3.2.1 .
+
+My IPS was about 4000000 before I tampered with the source and now its 
+about 5000000.
+
+I don't know how you think this could be implemented in a way that looks 
+nice and works for every one not using Linux on an X86, but good luck. I 
+just hope this helps somebody.
+
+If you have any questions let me know,
+
+	Conn
+
+-----------------------------
+
+In reading through this one more time I have noticed that I forgot to 
+mention what options I specified  for ./configure . This patch may not 
+work with some of the other configurations. Here they are.
+
+--enable-new-pit --enable-ne2000 --enable-pci --enable-4meg-pages 
+--enable-pae --enable-guest2host-tlb --enable-repeat-speedups 
+--enable-icache --enable-global-pages --enable-host-specific-asms 
+--enable-vbe --enable-mmx --enable-fpu --enable-cdrom --enable-sse 
+--enable-sb16=linux --with-x11 --enable-all-optimizations --disable-plugins
+
+
+-----------------------------
+
+Notes from cbothamy :
+- add an --enable-regparm option to configure
+- enable-regparm if --enable-all-optimizations
+- change all __attribute__((regparm(X))) to GCC_ATTRIBUTE_REGPARM(X)
+- #define GCC_ATTRIBUTE_REGPARM(X) to  GCC_ATTRIBUTE((GCC_REGPARM(X)))
+- #define GCC_REGPARM(X) to regparm(X) only on x86 with gcc > 3.2
+or - #define GCC_FASTCALL to __attribute__((regparm(3))) and only use it
+or something like that
+
+Patch was created with:
+  cvs diff -u
+Apply patch to what version:
+  cvs checked out on Wed Feb 12 15:21:37 CET 2003
+Instructions:
+  To patch, go to main bochs directory.
+  Type "patch -p0 < THIS_PATCH_FILE".
+----------------------------------------------------------------------
+Index: pc_system.cc
+===================================================================
+RCS file: /cvsroot/bochs/bochs/pc_system.cc,v
+retrieving revision 1.31
+diff -u -r1.31 pc_system.cc
+--- pc_system.cc	25 Oct 2002 11:44:33 -0000	1.31
++++ pc_system.cc	12 Feb 2003 14:12:31 -0000
+@@ -126,7 +126,7 @@
+ // Read from the IO memory address space
+ //
+ 
+-  Bit32u
++  Bit32u __attribute__((regparm(2)))
+ bx_pc_system_c::inp(Bit16u addr, unsigned io_len)
+ {
+   Bit32u ret;
+@@ -141,13 +141,13 @@
+ // Write to the IO memory address space.
+ //
+ 
+-  void
++  void __attribute__((regparm(3)))
+ bx_pc_system_c::outp(Bit16u addr, Bit32u value, unsigned io_len)
+ {
+   bx_devices.outp(addr, value, io_len);
+ }
+ 
+-  void
++  void __attribute__((regparm(1)))
+ bx_pc_system_c::set_enable_a20(Bit8u value)
+ {
+ #if BX_CPU_LEVEL < 2
+Index: pc_system.h
+===================================================================
+RCS file: /cvsroot/bochs/bochs/pc_system.h,v
+retrieving revision 1.23
+diff -u -r1.23 pc_system.h
+--- pc_system.h	15 Dec 2002 16:00:41 -0000	1.23
++++ pc_system.h	12 Feb 2003 14:12:31 -0000
+@@ -187,9 +187,9 @@
+ 
+   bx_pc_system_c(void);
+ 
+-  Bit32u  inp(Bit16u addr, unsigned io_len);
+-  void    outp(Bit16u addr, Bit32u value, unsigned io_len);
+-  void    set_enable_a20(Bit8u value);
++  Bit32u  inp(Bit16u addr, unsigned io_len) __attribute__((regparm(2)));
++  void    outp(Bit16u addr, Bit32u value, unsigned io_len) __attribute__((regparm(3)));
++  void    set_enable_a20(Bit8u value) __attribute__((regparm(1)));
+   bx_bool get_enable_a20(void);
+   void    exit(void);
+ 
+Index: cpu/access.cc
+===================================================================
+RCS file: /cvsroot/bochs/bochs/cpu/access.cc,v
+retrieving revision 1.35
+diff -u -r1.35 access.cc
+--- cpu/access.cc	13 Nov 2002 21:00:03 -0000	1.35
++++ cpu/access.cc	12 Feb 2003 14:12:31 -0000
+@@ -46,7 +46,7 @@
+ 
+ 
+ 
+-  void
++  void __attribute__((regparm(3)))
+ BX_CPU_C::write_virtual_checks(bx_segment_reg_t *seg, bx_address offset,
+                                unsigned length)
+ {
+@@ -139,7 +139,7 @@
+     }
+ }
+ 
+-  void
++  void __attribute__((regparm(3)))
+ BX_CPU_C::read_virtual_checks(bx_segment_reg_t *seg, bx_address offset,
+                               unsigned length)
+ {
+@@ -261,7 +261,7 @@
+ 
+ 
+ 
+-  char *
++  char * __attribute__((regparm(1)))
+ BX_CPU_C::strseg(bx_segment_reg_t *seg)
+ {
+   if (seg == &BX_CPU_THIS_PTR sregs[0]) return("ES");
+@@ -277,7 +277,7 @@
+ }
+ 
+ 
+-  void
++  void __attribute__((regparm(3)))
+ BX_CPU_C::write_virtual_byte(unsigned s, bx_address offset, Bit8u *data)
+ {
+   bx_address laddr;
+@@ -342,7 +342,7 @@
+   goto accessOK;
+ }
+ 
+-  void
++  void __attribute__((regparm(3)))
+ BX_CPU_C::write_virtual_word(unsigned s, bx_address offset, Bit16u *data)
+ {
+   bx_address laddr;
+@@ -409,7 +409,7 @@
+   goto accessOK;
+ }
+ 
+-  void
++  void __attribute__((regparm(3)))
+ BX_CPU_C::write_virtual_dword(unsigned s, bx_address offset, Bit32u *data)
+ {
+   bx_address laddr;
+@@ -477,7 +477,7 @@
+ }
+ 
+ 
+-  void
++  void __attribute__((regparm(3)))
+ BX_CPU_C::read_virtual_byte(unsigned s, bx_address offset, Bit8u *data)
+ {
+   bx_address laddr;
+@@ -529,7 +529,7 @@
+ }
+ 
+ 
+-  void
++  void __attribute__((regparm(3)))
+ BX_CPU_C::read_virtual_word(unsigned s, bx_address offset, Bit16u *data)
+ {
+   bx_address laddr;
+@@ -583,7 +583,7 @@
+ }
+ 
+ 
+-  void
++  void __attribute__((regparm(3)))
+ BX_CPU_C::read_virtual_dword(unsigned s, bx_address offset, Bit32u *data)
+ {
+   bx_address laddr;
+@@ -642,7 +642,7 @@
+ // address translation info is kept across read/write calls //
+ //////////////////////////////////////////////////////////////
+ 
+-  void
++  void __attribute__((regparm(3)))
+ BX_CPU_C::read_RMW_virtual_byte(unsigned s, bx_address offset, Bit8u *data)
+ {
+   bx_address laddr;
+@@ -711,7 +711,7 @@
+ }
+ 
+ 
+-  void
++  void __attribute__((regparm(3)))
+ BX_CPU_C::read_RMW_virtual_word(unsigned s, bx_address offset, Bit16u *data)
+ {
+   bx_address laddr;
+@@ -779,7 +779,7 @@
+   goto accessOK;
+ }
+ 
+-  void
++  void __attribute__((regparm(3)))
+ BX_CPU_C::read_RMW_virtual_dword(unsigned s, bx_address offset, Bit32u *data)
+ {
+   bx_address laddr;
+@@ -847,7 +847,7 @@
+   goto accessOK;
+ }
+ 
+-  void
++  void __attribute__((regparm(1)))
+ BX_CPU_C::write_RMW_virtual_byte(Bit8u val8)
+ {
+   BX_INSTR_MEM_DATA(CPU_ID, BX_CPU_THIS_PTR address_xlation.paddress1, 1, BX_WRITE);
+@@ -864,7 +864,7 @@
+     }
+ }
+ 
+-  void
++  void __attribute__((regparm(1)))
+ BX_CPU_C::write_RMW_virtual_word(Bit16u val16)
+ {
+   BX_INSTR_MEM_DATA(CPU_ID, BX_CPU_THIS_PTR address_xlation.paddress1, 2, BX_WRITE);
+@@ -931,7 +931,7 @@
+ }
+ 
+ 
+-  void
++  void __attribute__((regparm(3)))
+ BX_CPU_C::write_virtual_qword(unsigned s, bx_address offset, Bit64u *data)
+ {
+   bx_address laddr;
+@@ -999,7 +999,7 @@
+ }
+ 
+ 
+-  void
++  void __attribute__((regparm(3)))
+ BX_CPU_C::read_virtual_qword(unsigned s, bx_address offset, Bit64u *data)
+ {
+   bx_address laddr;
+@@ -1089,7 +1089,7 @@
+     }
+ }
+ 
+-  void
++  void __attribute__((regparm(3)))
+ BX_CPU_C::read_RMW_virtual_qword(unsigned s, bx_address offset, Bit64u *data)
+ {
+   bx_address laddr;
+Index: cpu/cpu.h
+===================================================================
+RCS file: /cvsroot/bochs/bochs/cpu/cpu.h,v
+retrieving revision 1.127
+diff -u -r1.127 cpu.h
+--- cpu/cpu.h	9 Feb 2003 13:30:39 -0000	1.127
++++ cpu/cpu.h	12 Feb 2003 14:12:34 -0000
+@@ -2689,23 +2689,23 @@
+     BX_CPU_THIS_PTR eipPageWindowSize = 0;
+     }
+ 
+-  BX_SMF void write_virtual_checks(bx_segment_reg_t *seg, bx_address offset, unsigned length);
+-  BX_SMF void read_virtual_checks(bx_segment_reg_t *seg, bx_address offset, unsigned length);
+-  BX_SMF void write_virtual_byte(unsigned seg, bx_address offset, Bit8u *data);
+-  BX_SMF void write_virtual_word(unsigned seg, bx_address offset, Bit16u *data);
+-  BX_SMF void write_virtual_dword(unsigned seg, bx_address offset, Bit32u *data);
+-  BX_SMF void write_virtual_qword(unsigned seg, bx_address offset, Bit64u *data);
+-  BX_SMF void read_virtual_byte(unsigned seg, bx_address offset, Bit8u *data);
+-  BX_SMF void read_virtual_word(unsigned seg, bx_address offset, Bit16u *data);
+-  BX_SMF void read_virtual_dword(unsigned seg, bx_address offset, Bit32u *data);
+-  BX_SMF void read_virtual_qword(unsigned seg, bx_address offset, Bit64u *data);
+-
+-  BX_SMF void read_RMW_virtual_byte(unsigned seg, bx_address offset, Bit8u *data);
+-  BX_SMF void read_RMW_virtual_word(unsigned seg, bx_address offset, Bit16u *data);
+-  BX_SMF void read_RMW_virtual_dword(unsigned seg, bx_address offset, Bit32u *data);
+-  BX_SMF void read_RMW_virtual_qword(unsigned seg, bx_address offset, Bit64u *data);
+-  BX_SMF void write_RMW_virtual_byte(Bit8u val8);
+-  BX_SMF void write_RMW_virtual_word(Bit16u val16);
++  BX_SMF void write_virtual_checks(bx_segment_reg_t *seg, bx_address offset, unsigned length) __attribute__((regparm(3)));
++  BX_SMF void read_virtual_checks(bx_segment_reg_t *seg, bx_address offset, unsigned length) __attribute__((regparm(3)));
++  BX_SMF void write_virtual_byte(unsigned seg, bx_address offset, Bit8u *data) __attribute__((regparm(3)));
++  BX_SMF void write_virtual_word(unsigned seg, bx_address offset, Bit16u *data) __attribute__((regparm(3)));
++  BX_SMF void write_virtual_dword(unsigned seg, bx_address offset, Bit32u *data) __attribute__((regparm(3)));
++  BX_SMF void write_virtual_qword(unsigned seg, bx_address offset, Bit64u *data) __attribute__((regparm(3)));
++  BX_SMF void read_virtual_byte(unsigned seg, bx_address offset, Bit8u *data) __attribute__((regparm(3)));
++  BX_SMF void read_virtual_word(unsigned seg, bx_address offset, Bit16u *data) __attribute__((regparm(3)));
++  BX_SMF void read_virtual_dword(unsigned seg, bx_address offset, Bit32u *data) __attribute__((regparm(3)));
++  BX_SMF void read_virtual_qword(unsigned seg, bx_address offset, Bit64u *data) __attribute__((regparm(3)));
++
++  BX_SMF void read_RMW_virtual_byte(unsigned seg, bx_address offset, Bit8u *data) __attribute__((regparm(3)));
++  BX_SMF void read_RMW_virtual_word(unsigned seg, bx_address offset, Bit16u *data) __attribute__((regparm(3)));
++  BX_SMF void read_RMW_virtual_dword(unsigned seg, bx_address offset, Bit32u *data) __attribute__((regparm(3)));
++  BX_SMF void read_RMW_virtual_qword(unsigned seg, bx_address offset, Bit64u *data) __attribute__((regparm(3)));
++  BX_SMF void write_RMW_virtual_byte(Bit8u val8) __attribute__((regparm(1)));
++  BX_SMF void write_RMW_virtual_word(Bit16u val16) __attribute__((regparm(1)));
+   BX_SMF void write_RMW_virtual_dword(Bit32u val32);
+   BX_SMF void write_RMW_virtual_qword(Bit64u val64);
+ 
+@@ -2723,13 +2723,13 @@
+ 
+ 
+   BX_SMF void access_linear(bx_address address, unsigned length, unsigned pl,
+-                     unsigned rw, void *data);
+-  BX_SMF Bit32u itranslate_linear(bx_address laddr, unsigned pl);
+-  BX_SMF Bit32u dtranslate_linear(bx_address laddr, unsigned pl, unsigned rw);
++                     unsigned rw, void *data) __attribute__((regparm(3)));
++  BX_SMF Bit32u itranslate_linear(bx_address laddr, unsigned pl) __attribute__((regparm(2)));
++  BX_SMF Bit32u dtranslate_linear(bx_address laddr, unsigned pl, unsigned rw) __attribute__((regparm(3)));
+   BX_SMF void TLB_flush(bx_bool invalidateGlobal);
+   BX_SMF void TLB_init(void);
+   BX_SMF void set_INTR(bx_bool value);
+-  BX_SMF char *strseg(bx_segment_reg_t *seg);
++  BX_SMF char *strseg(bx_segment_reg_t *seg) __attribute__((regparm(1)));
+   BX_SMF void interrupt(Bit8u vector, bx_bool is_INT, bx_bool is_error_code,
+                  Bit16u error_code);
+ #if BX_CPU_LEVEL >= 2
+@@ -2738,17 +2738,17 @@
+ #endif
+   BX_SMF int  int_number(bx_segment_reg_t *seg);
+   BX_SMF void shutdown_cpu(void);
+-  BX_SMF void CR3_change(bx_address value);
+-  BX_SMF void pagingCR0Changed(Bit32u oldCR0, Bit32u newCR0);
+-  BX_SMF void pagingCR4Changed(Bit32u oldCR4, Bit32u newCR4);
++  BX_SMF void CR3_change(bx_address value) __attribute__((regparm(1)));
++  BX_SMF void pagingCR0Changed(Bit32u oldCR0, Bit32u newCR0) __attribute__((regparm(2)));
++  BX_SMF void pagingCR4Changed(Bit32u oldCR4, Bit32u newCR4) __attribute__((regparm(2)));
+   BX_SMF void pagingA20Changed(void);
+ 
+   BX_SMF void reset(unsigned source);
+ 
+-  BX_SMF void jump_protected(bxInstruction_c *, Bit16u cs, bx_address disp);
+-  BX_SMF void call_protected(bxInstruction_c *, Bit16u cs, bx_address disp);
+-  BX_SMF void return_protected(bxInstruction_c *, Bit16u pop_bytes);
+-  BX_SMF void iret_protected(bxInstruction_c *);
++  BX_SMF void jump_protected(bxInstruction_c *, Bit16u cs, bx_address disp) __attribute__((regparm(3)));
++  BX_SMF void call_protected(bxInstruction_c *, Bit16u cs, bx_address disp) __attribute__((regparm(3)));
++  BX_SMF void return_protected(bxInstruction_c *, Bit16u pop_bytes) __attribute__((regparm(2)));
++  BX_SMF void iret_protected(bxInstruction_c *) __attribute__((regparm(1)));
+   BX_SMF void validate_seg_regs(void);
+   BX_SMF void stack_return_to_v86(Bit32u new_eip, Bit32u raw_cs_selector,
+                                      Bit32u flags32);
+@@ -2763,38 +2763,38 @@
+ #if BX_SUPPORT_X86_64
+   BX_SMF void get_RSP_from_TSS(unsigned pl, Bit64u *rsp);
+ #endif
+-  BX_SMF void write_flags(Bit16u flags, bx_bool change_IOPL, bx_bool change_IF);
++  BX_SMF void write_flags(Bit16u flags, bx_bool change_IOPL, bx_bool change_IF) __attribute__((regparm(3)));
+   BX_SMF void write_eflags(Bit32u eflags, bx_bool change_IOPL, bx_bool change_IF,
+                     bx_bool change_VM, bx_bool change_RF);
+-  BX_SMF void writeEFlags(Bit32u eflags, Bit32u changeMask); // Newer variant.
++  BX_SMF void writeEFlags(Bit32u eflags, Bit32u changeMask) __attribute__((regparm(2))); // Newer variant.
+   BX_SMF Bit16u read_flags(void);
+   BX_SMF Bit32u read_eflags(void);
+   BX_SMF Bit32u get_segment_base(unsigned seg);
+ 
+-  BX_SMF Bit8u   inp8(Bit16u addr);
+-  BX_SMF void    outp8(Bit16u addr, Bit8u value);
+-  BX_SMF Bit16u  inp16(Bit16u addr);
+-  BX_SMF void    outp16(Bit16u addr, Bit16u value);
+-  BX_SMF Bit32u  inp32(Bit16u addr);
+-  BX_SMF void    outp32(Bit16u addr, Bit32u value);
++  BX_SMF Bit8u   inp8(Bit16u addr) __attribute__((regparm(1)));
++  BX_SMF void    outp8(Bit16u addr, Bit8u value) __attribute__((regparm(2)));
++  BX_SMF Bit16u  inp16(Bit16u addr) __attribute__((regparm(1)));
++  BX_SMF void    outp16(Bit16u addr, Bit16u value) __attribute__((regparm(2)));
++  BX_SMF Bit32u  inp32(Bit16u addr) __attribute__((regparm(1)));
++  BX_SMF void    outp32(Bit16u addr, Bit32u value) __attribute__((regparm(2)));
+   BX_SMF bx_bool allow_io(Bit16u addr, unsigned len);
+   BX_SMF void    enter_protected_mode(void);
+   BX_SMF void    enter_real_mode(void);
+-  BX_SMF void    parse_selector(Bit16u raw_selector, bx_selector_t *selector);
+-  BX_SMF void    parse_descriptor(Bit32u dword1, Bit32u dword2, bx_descriptor_t *temp);
++  BX_SMF void    parse_selector(Bit16u raw_selector, bx_selector_t *selector) __attribute__((regparm(2)));
++  BX_SMF void    parse_descriptor(Bit32u dword1, Bit32u dword2, bx_descriptor_t *temp) __attribute__((regparm(3)));
+   BX_SMF void    load_ldtr(bx_selector_t *selector, bx_descriptor_t *descriptor);
+-  BX_SMF void    load_cs(bx_selector_t *selector, bx_descriptor_t *descriptor, Bit8u cpl);
+-  BX_SMF void    load_ss(bx_selector_t *selector, bx_descriptor_t *descriptor, Bit8u cpl);
++  BX_SMF void    load_cs(bx_selector_t *selector, bx_descriptor_t *descriptor, Bit8u cpl) __attribute__((regparm(3)));
++  BX_SMF void    load_ss(bx_selector_t *selector, bx_descriptor_t *descriptor, Bit8u cpl) __attribute__((regparm(3)));
+   BX_SMF void    fetch_raw_descriptor(bx_selector_t *selector,
+-                               Bit32u *dword1, Bit32u *dword2, Bit8u exception);
+-  BX_SMF void    load_seg_reg(bx_segment_reg_t *seg, Bit16u new_value);
++                               Bit32u *dword1, Bit32u *dword2, Bit8u exception) __attribute__((regparm(3)));
++  BX_SMF void    load_seg_reg(bx_segment_reg_t *seg, Bit16u new_value) __attribute__((regparm(2)));
+ #if BX_SUPPORT_X86_64
+   BX_SMF void    loadSRegLMNominal(unsigned seg, unsigned selector,
+                                    bx_address base, unsigned dpl);
+ #endif
+   BX_SMF bx_bool fetch_raw_descriptor2(bx_selector_t *selector,
+-                                Bit32u *dword1, Bit32u *dword2);
+-  BX_SMF void    push_16(Bit16u value16);
++                                Bit32u *dword1, Bit32u *dword2) __attribute__((regparm(3)));
++  BX_SMF void    push_16(Bit16u value16) __attribute__((regparm(1)));
+   BX_SMF void    push_32(Bit32u value32);
+ #if BX_SUPPORT_X86_64
+   BX_SMF void    push_64(Bit64u value64);
+@@ -2804,7 +2804,7 @@
+ #if BX_SUPPORT_X86_64
+   BX_SMF void    pop_64(Bit64u *value64_ptr);
+ #endif
+-  BX_SMF bx_bool can_push(bx_descriptor_t *descriptor, Bit32u esp, Bit32u bytes);
++  BX_SMF bx_bool can_push(bx_descriptor_t *descriptor, Bit32u esp, Bit32u bytes) __attribute__((regparm(3)));
+   BX_SMF bx_bool can_pop(Bit32u bytes);
+   BX_SMF void    sanity_checks(void);
+ 
+Index: cpu/ctrl_xfer_pro.cc
+===================================================================
+RCS file: /cvsroot/bochs/bochs/cpu/ctrl_xfer_pro.cc,v
+retrieving revision 1.21
+diff -u -r1.21 ctrl_xfer_pro.cc
+--- cpu/ctrl_xfer_pro.cc	3 Oct 2002 04:49:47 -0000	1.21
++++ cpu/ctrl_xfer_pro.cc	12 Feb 2003 14:12:35 -0000
+@@ -41,7 +41,7 @@
+ 
+ 
+ #if BX_CPU_LEVEL >= 2
+-  void
++  void __attribute__((regparm(3)))
+ BX_CPU_C::jump_protected(bxInstruction_c *i, Bit16u cs_raw, bx_address dispBig)
+ {
+   bx_descriptor_t  descriptor;
+@@ -456,7 +456,7 @@
+ 
+ 
+ #if BX_CPU_LEVEL >= 2
+-  void
++  void __attribute__((regparm(3)))
+ BX_CPU_C::call_protected(bxInstruction_c *i, Bit16u cs_raw, bx_address dispBig)
+ {
+   bx_selector_t cs_selector;
+@@ -1041,7 +1041,7 @@
+ 
+ 
+ #if BX_CPU_LEVEL >= 2
+-  void
++  void __attribute__((regparm(2)))
+ BX_CPU_C::return_protected(bxInstruction_c *i, Bit16u pop_bytes)
+ {
+   Bit16u raw_cs_selector, raw_ss_selector;
+@@ -1379,7 +1379,7 @@
+ 
+ 
+ #if BX_CPU_LEVEL >= 2
+-  void
++  void __attribute__((regparm(1)))
+ BX_CPU_C::iret_protected(bxInstruction_c *i)
+ {
+   Bit16u raw_cs_selector, raw_ss_selector;
+Index: cpu/flag_ctrl_pro.cc
+===================================================================
+RCS file: /cvsroot/bochs/bochs/cpu/flag_ctrl_pro.cc,v
+retrieving revision 1.13
+diff -u -r1.13 flag_ctrl_pro.cc
+--- cpu/flag_ctrl_pro.cc	25 Oct 2002 11:44:35 -0000	1.13
++++ cpu/flag_ctrl_pro.cc	12 Feb 2003 14:12:35 -0000
+@@ -32,7 +32,7 @@
+ 
+ 
+ 
+-  void
++  void __attribute__((regparm(2)))
+ BX_CPU_C::writeEFlags(Bit32u flags, Bit32u changeMask)
+ {
+   Bit32u supportMask, newEFlags;
+@@ -58,7 +58,7 @@
+ }
+ 
+ 
+-  void
++  void __attribute__((regparm(3)))
+ BX_CPU_C::write_flags(Bit16u flags, bx_bool change_IOPL, bx_bool change_IF)
+ {
+   Bit32u changeMask = 0x0dd5;
+Index: cpu/io_pro.cc
+===================================================================
+RCS file: /cvsroot/bochs/bochs/cpu/io_pro.cc,v
+retrieving revision 1.11
+diff -u -r1.11 io_pro.cc
+--- cpu/io_pro.cc	25 Oct 2002 11:44:35 -0000	1.11
++++ cpu/io_pro.cc	12 Feb 2003 14:12:35 -0000
+@@ -37,7 +37,7 @@
+ 
+ 
+ 
+-  Bit16u
++  Bit16u __attribute__((regparm(1)))
+ BX_CPU_C::inp16(Bit16u addr)
+ {
+   Bit16u ret16;
+@@ -54,7 +54,7 @@
+   return( ret16 );
+ }
+ 
+-  void
++  void __attribute__((regparm(2)))
+ BX_CPU_C::outp16(Bit16u addr, Bit16u value)
+ {
+   /* If CPL <= IOPL, then all IO addresses are accessible.
+@@ -72,7 +72,7 @@
+   BX_OUTP(addr, value, 2);
+ }
+ 
+-  Bit32u
++  Bit32u __attribute__((regparm(1)))
+ BX_CPU_C::inp32(Bit16u addr)
+ {
+   Bit32u ret32;
+@@ -89,7 +89,7 @@
+   return( ret32 );
+ }
+ 
+-  void
++  void __attribute__((regparm(2)))
+ BX_CPU_C::outp32(Bit16u addr, Bit32u value)
+ {
+   /* If CPL <= IOPL, then all IO addresses are accessible.
+@@ -107,7 +107,7 @@
+   BX_OUTP(addr, value, 4);
+ }
+ 
+-  Bit8u
++  Bit8u __attribute__((regparm(1)))
+ BX_CPU_C::inp8(Bit16u addr)
+ {
+   Bit8u ret8;
+@@ -125,7 +125,7 @@
+ }
+ 
+ 
+-  void
++  void __attribute__((regparm(2)))
+ BX_CPU_C::outp8(Bit16u addr, Bit8u value)
+ {
+   /* If CPL <= IOPL, then all IO addresses are accessible.
+Index: cpu/mmx.cc
+===================================================================
+RCS file: /cvsroot/bochs/bochs/cpu/mmx.cc,v
+retrieving revision 1.28
+diff -u -r1.28 mmx.cc
+--- cpu/mmx.cc	9 Jan 2003 05:21:22 -0000	1.28
++++ cpu/mmx.cc	12 Feb 2003 14:12:37 -0000
+@@ -27,7 +27,8 @@
+ 
+ #if BX_SUPPORT_MMX || BX_SUPPORT_SSE != 0
+ 
+-Bit8s SaturateWordSToByteS(Bit16s value)
++Bit8s  __attribute__((regparm(1)))
++SaturateWordSToByteS(Bit16s value)
+ {
+ /*
+   SaturateWordSToByteS   converts   a signed 16-bit value to a
+@@ -40,7 +41,8 @@
+   return value;
+ }
+ 
+-Bit16s SaturateDwordSToWordS(Bit32s value)
++Bit16s  __attribute__((regparm(1)))
++SaturateDwordSToWordS(Bit32s value)
+ {
+ /*
+   SaturateDwordSToWordS  converts  a  signed 32-bit value to a
+@@ -54,7 +56,8 @@
+   return value;
+ }
+ 
+-Bit8u SaturateWordSToByteU(Bit16s value)
++Bit8u  __attribute__((regparm(1)))
++SaturateWordSToByteU(Bit16s value)
+ {
+ /*
+   SaturateWordSToByteU  converts a signed 16-bit value to an
+@@ -67,7 +70,8 @@
+   return value;
+ }
+ 
+-Bit16u SaturateDwordSToWordU(Bit32s value)
++Bit16u  __attribute__((regparm(1)))
++SaturateDwordSToWordU(Bit32s value)
+ {
+ /*
+   SaturateDwordSToWordU  converts  a signed 32-bit value
+Index: cpu/paging.cc
+===================================================================
+RCS file: /cvsroot/bochs/bochs/cpu/paging.cc,v
+retrieving revision 1.38
+diff -u -r1.38 paging.cc
+--- cpu/paging.cc	4 Nov 2002 05:38:12 -0000	1.38
++++ cpu/paging.cc	12 Feb 2003 14:12:37 -0000
+@@ -387,7 +387,7 @@
+ // ==============================================================
+ 
+ 
+-  void
++  void __attribute__((regparm(2)))
+ BX_CPU_C::pagingCR0Changed(Bit32u oldCR0, Bit32u newCR0)
+ {
+   // Modification of PG,PE flushes TLB cache according to docs.
+@@ -400,7 +400,7 @@
+     BX_INFO(("pagingCR0Changed(0x%x -> 0x%x):", oldCR0, newCR0));
+ }
+ 
+-  void
++  void __attribute__((regparm(2)))
+ BX_CPU_C::pagingCR4Changed(Bit32u oldCR4, Bit32u newCR4)
+ {
+   // Modification of PGE,PAE,PSE flushes TLB cache according to docs.
+@@ -411,7 +411,7 @@
+     BX_INFO(("pagingCR4Changed(0x%x -> 0x%x):", oldCR4, newCR4));
+ }
+ 
+-  void
++  void __attribute__((regparm(1)))
+ BX_CPU_C::CR3_change(bx_address value)
+ {
+   if (bx_dbg.paging) {
+@@ -570,7 +570,7 @@
+ // Translate a linear address to a physical address, for
+ // a data access (D)
+ 
+-  Bit32u
++  Bit32u __attribute__((regparm(3)))
+ BX_CPU_C::dtranslate_linear(bx_address laddr, unsigned pl, unsigned rw)
+ {
+   bx_address   lpf;
+@@ -991,7 +991,7 @@
+ // Translate a linear address to a physical address, for
+ // an instruction fetch access (I)
+ 
+-  Bit32u
++  Bit32u __attribute__((regparm(2)))
+ BX_CPU_C::itranslate_linear(bx_address laddr, unsigned pl)
+ {
+   return dtranslate_linear(laddr, pl, BX_READ);
+@@ -1066,7 +1066,7 @@
+ 
+ 
+ 
+-  void
++  void __attribute__((regparm(3)))
+ BX_CPU_C::access_linear(bx_address laddr, unsigned length, unsigned pl,
+     unsigned rw, void *data)
+ {
+Index: cpu/segment_ctrl_pro.cc
+===================================================================
+RCS file: /cvsroot/bochs/bochs/cpu/segment_ctrl_pro.cc,v
+retrieving revision 1.23
+diff -u -r1.23 segment_ctrl_pro.cc
+--- cpu/segment_ctrl_pro.cc	25 Oct 2002 11:44:35 -0000	1.23
++++ cpu/segment_ctrl_pro.cc	12 Feb 2003 14:12:37 -0000
+@@ -37,7 +37,7 @@
+ 
+ 
+ 
+-  void
++  void __attribute__((regparm(2)))
+ BX_CPU_C::load_seg_reg(bx_segment_reg_t *seg, Bit16u new_value)
+ {
+ #if BX_CPU_LEVEL >= 3
+@@ -376,7 +376,7 @@
+ 
+ 
+ #if BX_CPU_LEVEL >= 2
+-  void
++  void __attribute__((regparm(2)))
+ BX_CPU_C::parse_selector(Bit16u raw_selector, bx_selector_t *selector)
+ {
+   selector->value  = raw_selector;
+@@ -386,7 +386,7 @@
+ }
+ #endif
+ 
+-  void
++  void __attribute__((regparm(3)))
+ BX_CPU_C::parse_descriptor(Bit32u dword1, Bit32u dword2, bx_descriptor_t *temp)
+ {
+   Bit8u AR_byte;
+@@ -504,7 +504,7 @@
+     }
+ }
+ 
+-  void
++  void __attribute__((regparm(2)))
+ BX_CPU_C::load_ldtr(bx_selector_t *selector, bx_descriptor_t *descriptor)
+ {
+   /* check for null selector, if so invalidate LDTR */
+@@ -527,7 +527,7 @@
+   BX_CPU_THIS_PTR ldtr.cache.valid = 1;
+ }
+ 
+-  void
++  void __attribute__((regparm(3)))
+ BX_CPU_C::load_cs(bx_selector_t *selector, bx_descriptor_t *descriptor,
+            Bit8u cpl)
+ {
+@@ -559,7 +559,7 @@
+ #endif
+ }
+ 
+-  void
++  void __attribute__((regparm(3)))
+ BX_CPU_C::load_ss(bx_selector_t *selector, bx_descriptor_t *descriptor, Bit8u cpl)
+ {
+   BX_CPU_THIS_PTR sregs[BX_SEG_REG_SS].selector = *selector;
+@@ -581,7 +581,7 @@
+ }
+ 
+ #if BX_CPU_LEVEL >= 2
+-  void
++  void __attribute__((regparm(3)))
+ BX_CPU_C::fetch_raw_descriptor(bx_selector_t *selector,
+                         Bit32u *dword1, Bit32u *dword2, Bit8u exception_no)
+ {
+@@ -624,7 +624,7 @@
+ 
+ 
+ 
+-  bx_bool
++  bx_bool __attribute__((regparm(3)))
+ BX_CPU_C::fetch_raw_descriptor2(bx_selector_t *selector,
+                         Bit32u *dword1, Bit32u *dword2)
+ {
+Index: cpu/stack_pro.cc
+===================================================================
+RCS file: /cvsroot/bochs/bochs/cpu/stack_pro.cc,v
+retrieving revision 1.12
+diff -u -r1.12 stack_pro.cc
+--- cpu/stack_pro.cc	7 Nov 2002 15:42:14 -0000	1.12
++++ cpu/stack_pro.cc	12 Feb 2003 14:12:37 -0000
+@@ -36,7 +36,7 @@
+ 
+ 
+ 
+-  void
++  void __attribute__((regparm(1)))
+ BX_CPU_C::push_16(Bit16u value16)
+ {
+ BailBigRSP("push_16");
+@@ -238,7 +238,7 @@
+ 
+ 
+ #if BX_CPU_LEVEL >= 2
+-  bx_bool
++  bx_bool __attribute__((regparm(3)))
+ BX_CPU_C::can_push(bx_descriptor_t *descriptor, Bit32u esp, Bit32u bytes)
+ {
+ #if BX_SUPPORT_X86_64
+Index: cpu/xmm.h
+===================================================================
+RCS file: /cvsroot/bochs/bochs/cpu/xmm.h,v
+retrieving revision 1.8
+diff -u -r1.8 xmm.h
+--- cpu/xmm.h	23 Jan 2003 18:33:35 -0000	1.8
++++ cpu/xmm.h	12 Feb 2003 14:12:38 -0000
+@@ -89,7 +89,7 @@
+   is  represented  by  the saturated value -128 (0x80). If it is greater
+   than 127, it is represented by the saturated value 127 (0x7F).
+ */
+-Bit8s SaturateWordSToByteS(Bit16s value);
++Bit8s SaturateWordSToByteS(Bit16s value) __attribute__((regparm(1)));
+ 
+ /*
+   SaturateDwordSToWordS  converts  a  signed 32-bit value to a
+@@ -98,7 +98,7 @@
+   greater  than  32767,  it  is represented by the saturated value 32767
+   (0x7FFF).
+ */
+-Bit16s SaturateDwordSToWordS(Bit32s value);
++Bit16s SaturateDwordSToWordS(Bit32s value) __attribute__((regparm(1)));
+ 
+ /*
+   SaturateWordSToByteU  converts a signed 16-bit value to an
+@@ -106,7 +106,7 @@
+   is  represented  by  the  saturated value zero (0x00).If it is greater
+   than 255 it is represented by the saturated value 255 (0xFF).
+ */
+-Bit8u SaturateWordSToByteU(Bit16s value);
++Bit8u SaturateWordSToByteU(Bit16s value) __attribute__((regparm(1)));
+ 
+ /*
+   SaturateDwordSToWordU  converts  a signed 32-bit value
+@@ -115,6 +115,6 @@
+   (0x0000).  If  it  is greater  than  65535,  it  is represented by
+   the saturated value 65535 (0xFFFF).
+ */
+-Bit16u SaturateDwordSToWordU(Bit32s value);
++Bit16u SaturateDwordSToWordU(Bit32s value) __attribute__((regparm(1)));
+ 
+ #endif
+Index: fpu/fpu_proto.h
+===================================================================
+RCS file: /cvsroot/bochs/bochs/fpu/fpu_proto.h,v
+retrieving revision 1.4
+diff -u -r1.4 fpu_proto.h
+--- fpu/fpu_proto.h	6 Oct 2001 04:35:13 -0000	1.4
++++ fpu/fpu_proto.h	12 Feb 2003 14:12:38 -0000
+@@ -93,20 +93,20 @@
+ extern void FPU_etc(void);
+ /* fpu_tags.c */
+ extern int FPU_gettag0(void);
+-extern int FPU_gettagi(int stnr);
+-extern int FPU_gettag(int regnr);
+-extern void FPU_settag0(int tag);
+-extern void FPU_settagi(int stnr, int tag);
+-extern void FPU_settag(int regnr, int tag);
+-extern int FPU_Special(FPU_REG const *ptr);
+-extern int isNaN(FPU_REG const *ptr);
++extern int FPU_gettagi(int stnr) __attribute__((regparm(1)));
++extern int FPU_gettag(int regnr) __attribute__((regparm(1)));
++extern void FPU_settag0(int tag) __attribute__((regparm(1)));
++extern void FPU_settagi(int stnr, int tag) __attribute__((regparm(2)));
++extern void FPU_settag(int regnr, int tag) __attribute__((regparm(2)));
++extern int FPU_Special(FPU_REG const *ptr) __attribute__((regparm(1)));
++extern int isNaN(FPU_REG const *ptr) __attribute__((regparm(1)));
+ extern void FPU_pop(void);
+-extern int FPU_empty_i(int stnr);
++extern int FPU_empty_i(int stnr) __attribute__((regparm(1)));
+ extern int FPU_stackoverflow(FPU_REG **st_new_ptr);
+ extern void FPU_sync_tags(void);
+-extern void FPU_copy_to_regi(FPU_REG const *r, u_char tag, int stnr);
+-extern void FPU_copy_to_reg1(FPU_REG const *r, u_char tag);
+-extern void FPU_copy_to_reg0(FPU_REG const *r, u_char tag);
++extern void FPU_copy_to_regi(FPU_REG const *r, u_char tag, int stnr) __attribute__((regparm(3)));
++extern void FPU_copy_to_reg1(FPU_REG const *r, u_char tag) __attribute__((regparm(2)));
++extern void FPU_copy_to_reg0(FPU_REG const *r, u_char tag) __attribute__((regparm(2)));
+ /* fpu_trig.c */
+ extern void FPU_triga(void);
+ extern void FPU_trigb(void);
+@@ -146,32 +146,32 @@
+ /* reg_constant.c */
+ extern void fconst(void);
+ /* reg_ld_str.c */
+-extern int FPU_load_extended(long double *s, int stnr);
+-extern int FPU_load_double(double *dfloat, FPU_REG *loaded_data);
+-extern int FPU_load_single(float *single, FPU_REG *loaded_data);
+-extern int FPU_load_int64(s64 *_s);
+-extern int FPU_load_int32(s32 *_s, FPU_REG *loaded_data);
+-extern int FPU_load_int16(s16 *_s, FPU_REG *loaded_data);
+-extern int FPU_load_bcd(u_char *s);
++extern int FPU_load_extended(long double *s, int stnr) __attribute__((regparm(2)));
++extern int FPU_load_double(double *dfloat, FPU_REG *loaded_data) __attribute__((regparm(2)));
++extern int FPU_load_single(float *single, FPU_REG *loaded_data) __attribute__((regparm(2)));
++extern int FPU_load_int64(s64 *_s) __attribute__((regparm(1)));
++extern int FPU_load_int32(s32 *_s, FPU_REG *loaded_data) __attribute__((regparm(2)));
++extern int FPU_load_int16(s16 *_s, FPU_REG *loaded_data) __attribute__((regparm(2)));
++extern int FPU_load_bcd(u_char *s) __attribute__((regparm(1)));
+ extern int FPU_store_extended(FPU_REG *st0_ptr, u_char st0_tag,
+-			      long double *d);
+-extern int FPU_store_double(FPU_REG *st0_ptr, u_char st0_tag, double *dfloat);
+-extern int FPU_store_single(FPU_REG *st0_ptr, u_char st0_tag, float *single);
+-extern int FPU_store_int64(FPU_REG *st0_ptr, u_char st0_tag, s64 *d);
+-extern int FPU_store_int32(FPU_REG *st0_ptr, u_char st0_tag, s32 *d);
+-extern int FPU_store_int16(FPU_REG *st0_ptr, u_char st0_tag, s16 *d);
+-extern int FPU_store_bcd(FPU_REG *st0_ptr, u_char st0_tag, u_char *d);
+-extern int FPU_round_to_int(FPU_REG *r, u_char tag);
+-extern u_char *fldenv(fpu_addr_modes addr_modes, u_char *s);
+-extern void frstor(fpu_addr_modes addr_modes, u_char *data_address);
+-extern u_char *fstenv(fpu_addr_modes addr_modes, u_char *d);
+-extern void fsave(fpu_addr_modes addr_modes, u_char *data_address);
+-extern int FPU_tagof(FPU_REG *ptr);
++			      long double *d) __attribute__((regparm(3)));
++extern int FPU_store_double(FPU_REG *st0_ptr, u_char st0_tag, double *dfloat) __attribute__((regparm(3)));
++extern int FPU_store_single(FPU_REG *st0_ptr, u_char st0_tag, float *single) __attribute__((regparm(3)));
++extern int FPU_store_int64(FPU_REG *st0_ptr, u_char st0_tag, s64 *d) __attribute__((regparm(3)));
++extern int FPU_store_int32(FPU_REG *st0_ptr, u_char st0_tag, s32 *d) __attribute__((regparm(3)));
++extern int FPU_store_int16(FPU_REG *st0_ptr, u_char st0_tag, s16 *d) __attribute__((regparm(3)));
++extern int FPU_store_bcd(FPU_REG *st0_ptr, u_char st0_tag, u_char *d) __attribute__((regparm(3)));
++extern int FPU_round_to_int(FPU_REG *r, u_char tag) __attribute__((regparm(2)));
++extern u_char *fldenv(fpu_addr_modes addr_modes, u_char *s) __attribute__((regparm(2)));
++extern void frstor(fpu_addr_modes addr_modes, u_char *data_address) __attribute__((regparm(2)));
++extern u_char *fstenv(fpu_addr_modes addr_modes, u_char *d) __attribute__((regparm(2)));
++extern void fsave(fpu_addr_modes addr_modes, u_char *data_address) __attribute__((regparm(2)));
++extern int FPU_tagof(FPU_REG *ptr) __attribute__((regparm(1)));
+ /* reg_mul.c */
+ extern int FPU_mul(FPU_REG const *b, u_char tagb, int deststnr, int control_w);
+ 
+ extern int FPU_div(int flags, FPU_REG *regrm, int control_w); // bbd: changed arg2 from int to FPU_REG*
+ /* reg_convert.c */
+-extern int FPU_to_exp16(FPU_REG const *a, FPU_REG *x);
++extern int FPU_to_exp16(FPU_REG const *a, FPU_REG *x) __attribute__((regparm(2)));
+ #endif /* _FPU_PROTO_H */
+ 
+Index: fpu/fpu_system.h
+===================================================================
+RCS file: /cvsroot/bochs/bochs/fpu/fpu_system.h,v
+retrieving revision 1.5
+diff -u -r1.5 fpu_system.h
+--- fpu/fpu_system.h	23 Jan 2003 18:33:35 -0000	1.5
++++ fpu/fpu_system.h	12 Feb 2003 14:12:38 -0000
+@@ -115,10 +115,10 @@
+ #endif
+ 
+ 
+-extern unsigned fpu_get_user(void *ptr, unsigned len);
+-extern void fpu_put_user(unsigned val, void *ptr, unsigned len);
++extern unsigned fpu_get_user(void *ptr, unsigned len) __attribute__((regparm(2)));
++extern void fpu_put_user(unsigned val, void *ptr, unsigned len) __attribute__((regparm(2)));
+ 
+-extern void fpu_verify_area(unsigned what, void *ptr, unsigned n);
++extern void fpu_verify_area(unsigned what, void *ptr, unsigned n) __attribute__((regparm(3)));
+ extern void math_emulate_init(void);
+ extern unsigned fpu_get_ds(void);
+ extern void fpu_set_ax(u16);
+Index: fpu/fpu_tags.c
+===================================================================
+RCS file: /cvsroot/bochs/bochs/fpu/fpu_tags.c,v
+retrieving revision 1.2
+diff -u -r1.2 fpu_tags.c
+--- fpu/fpu_tags.c	6 Oct 2001 03:53:46 -0000	1.2
++++ fpu/fpu_tags.c	12 Feb 2003 14:12:38 -0000
+@@ -29,19 +29,22 @@
+ }
+ 
+ 
+-int FPU_gettagi(int stnr)
++int  __attribute__((regparm(1)))
++FPU_gettagi(int stnr)
+ {
+   return (fpu_tag_word >> (((top+stnr) & 7)*2)) & 3;
+ }
+ 
+ 
+-int FPU_gettag(int regnr)
++int  __attribute__((regparm(1)))
++FPU_gettag(int regnr)
+ {
+   return (fpu_tag_word >> ((regnr & 7)*2)) & 3;
+ }
+ 
+ 
+-void FPU_settag0(int tag)
++void  __attribute__((regparm(1)))
++FPU_settag0(int tag)
+ {
+   int regnr = top;
+   regnr &= 7;
+@@ -50,7 +53,8 @@
+ }
+ 
+ 
+-void FPU_settagi(int stnr, int tag)
++void  __attribute__((regparm(2)))
++FPU_settagi(int stnr, int tag)
+ {
+   int regnr = stnr+top;
+   regnr &= 7;
+@@ -59,7 +63,8 @@
+ }
+ 
+ 
+-void FPU_settag(int regnr, int tag)
++void  __attribute__((regparm(2)))
++FPU_settag(int regnr, int tag)
+ {
+   regnr &= 7;
+   fpu_tag_word &= ~(3 << (regnr*2));
+@@ -67,7 +72,8 @@
+ }
+ 
+ 
+-int FPU_Special(FPU_REG const *ptr)
++int  __attribute__((regparm(1)))
++FPU_Special(FPU_REG const *ptr)
+ {
+   int exp = exponent(ptr);
+ 
+@@ -81,14 +87,16 @@
+ }
+ 
+ 
+-int isNaN(FPU_REG const *ptr)
++int  __attribute__((regparm(1)))
++isNaN(FPU_REG const *ptr)
+ {
+   return ( (exponent(ptr) == EXP_BIAS+EXP_OVER)
+ 	   && !((ptr->sigh == 0x80000000) && (ptr->sigl == 0)) );
+ }
+ 
+ 
+-int FPU_empty_i(int stnr)
++int  __attribute__((regparm(1)))
++FPU_empty_i(int stnr)
+ {
+   int regnr = (top+stnr) & 7;
+ 
+@@ -104,19 +112,22 @@
+ }
+ 
+ 
+-void FPU_copy_to_regi(FPU_REG const *r, u_char tag, int stnr)
++void  __attribute__((regparm(3)))
++FPU_copy_to_regi(FPU_REG const *r, u_char tag, int stnr)
+ {
+   reg_copy(r, &st(stnr));
+   FPU_settagi(stnr, tag);
+ }
+ 
+-void FPU_copy_to_reg1(FPU_REG const *r, u_char tag)
++void  __attribute__((regparm(2)))
++FPU_copy_to_reg1(FPU_REG const *r, u_char tag)
+ {
+   reg_copy(r, &st(1));
+   FPU_settagi(1, tag);
+ }
+ 
+-void FPU_copy_to_reg0(FPU_REG const *r, u_char tag)
++void  __attribute__((regparm(2)))
++FPU_copy_to_reg0(FPU_REG const *r, u_char tag)
+ {
+   int regnr = top;
+   regnr &= 7;
+Index: fpu/reg_convert.c
+===================================================================
+RCS file: /cvsroot/bochs/bochs/fpu/reg_convert.c,v
+retrieving revision 1.2
+diff -u -r1.2 reg_convert.c
+--- fpu/reg_convert.c	6 Oct 2001 03:53:46 -0000	1.2
++++ fpu/reg_convert.c	12 Feb 2003 14:12:38 -0000
+@@ -15,7 +15,8 @@
+ #include "fpu_emu.h"
+ 
+ 
+-int FPU_to_exp16(FPU_REG const *a, FPU_REG *x)
++int  __attribute__((regparm(2)))
++FPU_to_exp16(FPU_REG const *a, FPU_REG *x)
+ {
+   int sign = getsign(a);
+ 
+Index: fpu/reg_ld_str.c
+===================================================================
+RCS file: /cvsroot/bochs/bochs/fpu/reg_ld_str.c,v
+retrieving revision 1.5
+diff -u -r1.5 reg_ld_str.c
+--- fpu/reg_ld_str.c	26 Dec 2002 14:47:02 -0000	1.5
++++ fpu/reg_ld_str.c	12 Feb 2003 14:12:38 -0000
+@@ -38,7 +38,8 @@
+ #define SINGLE_Emin (-126)       /* smallest valid exponent */
+ 
+ 
+-static u_char normalize_no_excep(FPU_REG *r, int exp, int sign)
++static u_char  __attribute__((regparm(3)))
++normalize_no_excep(FPU_REG *r, int exp, int sign)
+ {
+   u_char tag;
+ 
+@@ -53,7 +54,8 @@
+ }
+ 
+ 
+-int FPU_tagof(FPU_REG *ptr)
++int  __attribute__((regparm(1)))
++FPU_tagof(FPU_REG *ptr)
+ {
+   int exp;
+ 
+@@ -87,7 +89,8 @@
+ 
+ 
+ /* Get a long double from user memory */
+-int FPU_load_extended(long double *s, int stnr)
++int  __attribute__((regparm(2)))
++FPU_load_extended(long double *s, int stnr)
+ {
+   FPU_REG *sti_ptr = &st(stnr);
+ 
+@@ -107,7 +110,8 @@
+ 
+ 
+ /* Get a double from user memory */
+-int FPU_load_double(double *dfloat, FPU_REG *loaded_data)
++int  __attribute__((regparm(2)))
++FPU_load_double(double *dfloat, FPU_REG *loaded_data)
+ {
+   int exp, tag, negative;
+   u32 m64, l64;
+@@ -179,7 +183,8 @@
+ 
+ 
+ /* Get a float from user memory */
+-int FPU_load_single(float *single, FPU_REG *loaded_data)
++int  __attribute__((regparm(2)))
++FPU_load_single(float *single, FPU_REG *loaded_data)
+ {
+   u32 m32;
+   int exp, tag, negative;
+@@ -243,7 +248,8 @@
+ 
+ 
+ /* Get a 64bit quantity from user memory */
+-int FPU_load_int64(s64 *_s)
++int  __attribute__((regparm(1)))
++FPU_load_int64(s64 *_s)
+ {
+   s64 s;
+   int sign;
+@@ -285,7 +291,8 @@
+ 
+ 
+ /* Get a long from user memory */
+-int FPU_load_int32(s32 *_s, FPU_REG *loaded_data)
++int  __attribute__((regparm(2)))
++FPU_load_int32(s32 *_s, FPU_REG *loaded_data)
+ {
+   s32 s;
+   int negative;
+@@ -314,7 +321,8 @@
+ 
+ 
+ /* Get a short from user memory */
+-int FPU_load_int16(s16 *_s, FPU_REG *loaded_data)
++int  __attribute__((regparm(1)))
++FPU_load_int16(s16 *_s, FPU_REG *loaded_data)
+ {
+   s16 s, negative;
+ 
+@@ -343,7 +351,8 @@
+ 
+ 
+ /* Get a packed bcd array from user memory */
+-int FPU_load_bcd(u_char *s)
++int  __attribute__((regparm(1)))
++FPU_load_bcd(u_char *s)
+ {
+   FPU_REG *st0_ptr = &st(0);
+   int pos;
+@@ -386,7 +395,8 @@
+ /*===========================================================================*/
+ 
+ /* Put a long double into user memory */
+-int FPU_store_extended(FPU_REG *st0_ptr, u_char st0_tag, long double *d)
++int  __attribute__((regparm(3)))
++FPU_store_extended(FPU_REG *st0_ptr, u_char st0_tag, long double *d)
+ {
+   /*
+     The only exception raised by an attempt to store to an
+@@ -428,7 +438,8 @@
+ 
+ 
+ /* Put a double into user memory */
+-int FPU_store_double(FPU_REG *st0_ptr, u_char st0_tag, double *dfloat)
++int  __attribute__((regparm(3)))
++FPU_store_double(FPU_REG *st0_ptr, u_char st0_tag, double *dfloat)
+ {
+   u32 l[2];
+   u32 increment = 0;	/* avoid gcc warnings */
+@@ -647,7 +658,8 @@
+ 
+ 
+ /* Put a float into user memory */
+-int FPU_store_single(FPU_REG *st0_ptr, u_char st0_tag, float *single)
++int  __attribute__((regparm(3)))
++FPU_store_single(FPU_REG *st0_ptr, u_char st0_tag, float *single)
+ {
+   s32 templ;
+   u32 increment = 0;     	/* avoid gcc warnings */
+@@ -871,7 +883,8 @@
+ 
+ 
+ /* Put a 64bit quantity into user memory */
+-int FPU_store_int64(FPU_REG *st0_ptr, u_char st0_tag, s64 *d)
++int  __attribute__((regparm(3)))
++FPU_store_int64(FPU_REG *st0_ptr, u_char st0_tag, s64 *d)
+ {
+   FPU_REG t;
+   s64 tll;
+@@ -942,7 +955,8 @@
+ 
+ 
+ /* Put a long into user memory */
+-int FPU_store_int32(FPU_REG *st0_ptr, u_char st0_tag, s32 *d)
++int  __attribute__((regparm(3)))
++FPU_store_int32(FPU_REG *st0_ptr, u_char st0_tag, s32 *d)
+ {
+   FPU_REG t;
+   int precision_loss;
+@@ -999,7 +1013,8 @@
+ 
+ 
+ /* Put a short into user memory */
+-int FPU_store_int16(FPU_REG *st0_ptr, u_char st0_tag, s16 *d)
++int  __attribute__((regparm(3)))
++FPU_store_int16(FPU_REG *st0_ptr, u_char st0_tag, s16 *d)
+ {
+   FPU_REG t;
+   int precision_loss;
+@@ -1056,7 +1071,8 @@
+ 
+ 
+ /* Put a packed bcd array into user memory */
+-int FPU_store_bcd(FPU_REG *st0_ptr, u_char st0_tag, u_char *d)
++int  __attribute__((regparm(3)))
++FPU_store_bcd(FPU_REG *st0_ptr, u_char st0_tag, u_char *d)
+ {
+   FPU_REG t;
+   u64 ll;
+@@ -1142,7 +1158,8 @@
+ /* Overflow is signalled by a non-zero return value (in eax).
+    In the case of overflow, the returned significand always has the
+    largest possible value */
+-int FPU_round_to_int(FPU_REG *r, u_char tag)
++int  __attribute__((regparm(2)))
++FPU_round_to_int(FPU_REG *r, u_char tag)
+ {
+   u_char     very_big;
+   unsigned eax;
+@@ -1206,7 +1223,8 @@
+ 
+ /*===========================================================================*/
+ 
+-u_char *fldenv(fpu_addr_modes addr_modes, u_char *s)
++u_char  __attribute__((regparm(2)))
++*fldenv(fpu_addr_modes addr_modes, u_char *s)
+ {
+   u16 tag_word = 0;
+   u_char tag;
+@@ -1297,7 +1315,8 @@
+ }
+ 
+ 
+-void frstor(fpu_addr_modes addr_modes, u_char *data_address)
++void  __attribute__((regparm(2)))
++frstor(fpu_addr_modes addr_modes, u_char *data_address)
+ {
+   int i, regnr;
+   u_char *s = fldenv(addr_modes, data_address);
+@@ -1347,7 +1366,8 @@
+ }
+ 
+ 
+-u_char *fstenv(fpu_addr_modes addr_modes, u_char *d)
++u_char  __attribute__((regparm(2)))
++*fstenv(fpu_addr_modes addr_modes, u_char *d)
+ {
+   if ( (addr_modes.default_mode == VM86) ||
+       ((addr_modes.default_mode == PM16)
+@@ -1414,7 +1434,8 @@
+ }
+ 
+ 
+-void fsave(fpu_addr_modes addr_modes, u_char *data_address)
++void  __attribute__((regparm(2)))
++fsave(fpu_addr_modes addr_modes, u_char *data_address)
+ {
+   u_char *d;
+   int offset = (top & 7) * sizeof(FPU_REG), other = 8*sizeof(FPU_REG) - offset;
+Index: fpu/wmFPUemu_glue.cc
+===================================================================
+RCS file: /cvsroot/bochs/bochs/fpu/wmFPUemu_glue.cc,v
+retrieving revision 1.17
+diff -u -r1.17 wmFPUemu_glue.cc
+--- fpu/wmFPUemu_glue.cc	6 Jan 2003 02:20:46 -0000	1.17
++++ fpu/wmFPUemu_glue.cc	12 Feb 2003 14:12:39 -0000
+@@ -189,7 +189,7 @@
+ //BX_DEBUG(( "fpu_set_ax(0x%04x)", (unsigned) val16));
+ }
+ 
+-  void
++  void __attribute__((regparm(3)))
+ fpu_verify_area(unsigned what, void *ptr, unsigned n)
+ {
+   bx_segment_reg_t *seg;
+@@ -213,7 +213,7 @@
+ }
+ 
+ 
+-  unsigned
++  unsigned __attribute__((regparm(2)))
+ fpu_get_user(void *ptr, unsigned len)
+ {
+   Bit32u val32;
+@@ -238,7 +238,7 @@
+   return(val32);
+ }
+ 
+-  void
++  void __attribute__((regparm(2)))
+ fpu_put_user(unsigned val, void *ptr, unsigned len)
+ {
+   Bit32u val32;
+Index: iodev/cdrom.cc
+===================================================================
+RCS file: /cvsroot/bochs/bochs/iodev/cdrom.cc,v
+retrieving revision 1.58
+diff -u -r1.58 cdrom.cc
+--- iodev/cdrom.cc	30 Jan 2003 20:44:32 -0000	1.58
++++ iodev/cdrom.cc	12 Feb 2003 14:12:39 -0000
+@@ -1250,7 +1250,7 @@
+ #endif
+ }
+ 
+-  void
++  void __attribute__((regparm(2)))
+ cdrom_interface::read_block(uint8* buf, int lba)
+ {
+   // Read a single block from the CD
+Index: iodev/cdrom.h
+===================================================================
+RCS file: /cvsroot/bochs/bochs/iodev/cdrom.h,v
+retrieving revision 1.11
+diff -u -r1.11 cdrom.h
+--- iodev/cdrom.h	25 Oct 2002 11:44:38 -0000	1.11
++++ iodev/cdrom.h	12 Feb 2003 14:12:39 -0000
+@@ -47,7 +47,7 @@
+   uint32 capacity();
+ 
+   // Read a single block from the CD
+-  void read_block(uint8* buf, int lba);
++  void read_block(uint8* buf, int lba) __attribute__((regparm(1)));
+ 
+ private:
+   int fd;
+Index: iodev/devices.cc
+===================================================================
+RCS file: /cvsroot/bochs/bochs/iodev/devices.cc,v
+retrieving revision 1.50
+diff -u -r1.50 devices.cc
+--- iodev/devices.cc	28 Jan 2003 16:56:59 -0000	1.50
++++ iodev/devices.cc	12 Feb 2003 14:12:40 -0000
+@@ -599,7 +599,7 @@
+  * Read a byte of data from the IO memory address space
+  */
+ 
+-  Bit32u
++  Bit32u __attribute__((regparm(2)))
+ bx_devices_c::inp(Bit16u addr, unsigned io_len)
+ {
+   Bit8u handle;
+@@ -621,7 +621,7 @@
+  * Write a byte of data to the IO memory address space.
+  */
+ 
+-  void
++  void __attribute__((regparm(3)))
+ bx_devices_c::outp(Bit16u addr, Bit32u value, unsigned io_len)
+ {
+   Bit8u handle;
+Index: iodev/dma.cc
+===================================================================
+RCS file: /cvsroot/bochs/bochs/iodev/dma.cc,v
+retrieving revision 1.26
+diff -u -r1.26 dma.cc
+--- iodev/dma.cc	4 Nov 2002 17:29:12 -0000	1.26
++++ iodev/dma.cc	12 Feb 2003 14:12:40 -0000
+@@ -212,7 +212,7 @@
+   // static IO port read callback handler
+   // redirects to non-static class handler to avoid virtual functions
+ 
+-  Bit32u
++  Bit32u __attribute__((regparm(3)))
+ bx_dma_c::read_handler(void *this_ptr, Bit32u address, unsigned io_len)
+ {
+ #if !BX_USE_DMA_SMF
+@@ -222,7 +222,7 @@
+ }
+ 
+   /* 8237 DMA controller */
+-  Bit32u
++  Bit32u __attribute__((regparm(2)))
+ bx_dma_c::read( Bit32u   address, unsigned io_len)
+ {
+ #else
+@@ -343,7 +343,7 @@
+   // static IO port write callback handler
+   // redirects to non-static class handler to avoid virtual functions
+ 
+-  void
++  void __attribute__((regparm(3)))
+ bx_dma_c::write_handler(void *this_ptr, Bit32u address, Bit32u value, unsigned io_len)
+ {
+ #if !BX_USE_DMA_SMF
+@@ -354,7 +354,7 @@
+ 
+ 
+   /* 8237 DMA controller */
+-  void
++  void __attribute__((regparm(3)))
+ bx_dma_c::write(Bit32u   address, Bit32u   value, unsigned io_len)
+ {
+ #else
+Index: iodev/dma.h
+===================================================================
+RCS file: /cvsroot/bochs/bochs/iodev/dma.h,v
+retrieving revision 1.12
+diff -u -r1.12 dma.h
+--- iodev/dma.h	25 Oct 2002 11:44:39 -0000	1.12
++++ iodev/dma.h	12 Feb 2003 14:12:40 -0000
+@@ -63,11 +63,11 @@
+ 
+ private:
+ 
+-  static Bit32u read_handler(void *this_ptr, Bit32u address, unsigned io_len);
+-  static void   write_handler(void *this_ptr, Bit32u address, Bit32u value, unsigned io_len);
++  static Bit32u read_handler(void *this_ptr, Bit32u address, unsigned io_len) __attribute__((regparm(3)));
++  static void   write_handler(void *this_ptr, Bit32u address, Bit32u value, unsigned io_len) __attribute__((regparm(3)));
+ #if !BX_USE_DMA_SMF
+-  Bit32u   read( Bit32u   address, unsigned io_len);
+-  void     write(Bit32u   address, Bit32u   value, unsigned io_len);
++  Bit32u   read( Bit32u   address, unsigned io_len) __attribute__((regparm(2)));
++  void     write(Bit32u   address, Bit32u   value, unsigned io_len) __attribute__((regparm(1)));
+ #endif
+   BX_DMA_SMF void control_HRQ(bx_bool ma_sl);
+   BX_DMA_SMF void reset_controller(unsigned num);
+Index: iodev/harddrv.cc
+===================================================================
+RCS file: /cvsroot/bochs/bochs/iodev/harddrv.cc,v
+retrieving revision 1.95
+diff -u -r1.95 harddrv.cc
+--- iodev/harddrv.cc	1 Feb 2003 12:21:43 -0000	1.95
++++ iodev/harddrv.cc	12 Feb 2003 14:12:42 -0000
+@@ -2388,7 +2388,7 @@
+ }
+ 
+ 
+-  bx_bool
++  bx_bool __attribute__((regparm(2)))
+ bx_hard_drive_c::calculate_logical_address(Bit8u channel, off_t *sector)
+ {
+       off_t logical_sector;
+@@ -2418,7 +2418,7 @@
+       return true;
+ }
+ 
+-  void
++  void __attribute__((regparm(1)))
+ bx_hard_drive_c::increment_address(Bit8u channel)
+ {
+       BX_SELECTED_CONTROLLER(channel).sector_count--;
+@@ -2887,7 +2887,7 @@
+     }
+ }
+ 
+-  void
++  void __attribute__((regparm(3)))
+ bx_hard_drive_c::init_send_atapi_command(Bit8u channel, Bit8u command, int req_length, int alloc_length, bool lazy)
+ {
+       // BX_SELECTED_CONTROLLER(channel).byte_count is a union of BX_SELECTED_CONTROLLER(channel).cylinder_no;
+@@ -2961,7 +2961,7 @@
+       BX_SELECTED_DRIVE(channel).sense.ascq = 0;
+ }
+ 
+-void
++void __attribute__((regparm(1)))
+ bx_hard_drive_c::atapi_cmd_nop(Bit8u channel)
+ {
+       BX_SELECTED_CONTROLLER(channel).interrupt_reason.i_o = 1;
+@@ -2990,13 +2990,13 @@
+       memcpy(BX_SELECTED_CONTROLLER(channel).buffer + 8, src, size);
+ }
+ 
+-  void
++  void __attribute__((regparm(1)))
+ bx_hard_drive_c::ready_to_send_atapi(Bit8u channel)
+ {
+       raise_interrupt(channel);
+ }
+ 
+-void
++void __attribute__((regparm(1)))
+ bx_hard_drive_c::raise_interrupt(Bit8u channel)
+ {
+ 	BX_DEBUG(("raise_interrupt called, disable_irq = %02x", BX_SELECTED_CONTROLLER(channel).control.disable_irq));
+@@ -3387,12 +3387,14 @@
+       data[7] = 0x00;
+ }
+ 
+-uint16 read_16bit(const uint8* buf)
++uint16  __attribute__((regparm(1))) 
++read_16bit(const uint8* buf)
+ {
+       return (buf[0] << 8) | buf[1];
+ }
+ 
+-uint32 read_32bit(const uint8* buf)
++uint32  __attribute__((regparm(1)))
++read_32bit(const uint8* buf)
+ {
+       return (buf[0] << 24) | (buf[1] << 16) | (buf[2] << 8) | buf[3];
+ }
+Index: iodev/harddrv.h
+===================================================================
+RCS file: /cvsroot/bochs/bochs/iodev/harddrv.h,v
+retrieving revision 1.18
+diff -u -r1.18 harddrv.h
+--- iodev/harddrv.h	25 Oct 2002 11:44:40 -0000	1.18
++++ iodev/harddrv.h	12 Feb 2003 14:12:42 -0000
+@@ -241,8 +241,8 @@
+   error_recovery_t ();
+ };
+ 
+-uint16 read_16bit(const uint8* buf);
+-uint32 read_32bit(const uint8* buf);
++uint16 read_16bit(const uint8* buf) __attribute__((regparm(1)));
++uint32 read_32bit(const uint8* buf) __attribute__((regparm(1)));
+ 
+ 
+ #ifdef LOWLEVEL_CDROM
+@@ -315,18 +315,18 @@
+ 
+ private:
+ 
+-  BX_HD_SMF bx_bool calculate_logical_address(Bit8u channel, off_t *sector);
+-  BX_HD_SMF void increment_address(Bit8u channel);
++  BX_HD_SMF bx_bool calculate_logical_address(Bit8u channel, off_t *sector) __attribute__((regparm(2)));
++  BX_HD_SMF void increment_address(Bit8u channel) __attribute__((regparm(1)));
+   BX_HD_SMF void identify_drive(Bit8u channel);
+   BX_HD_SMF void identify_ATAPI_drive(Bit8u channel);
+   BX_HD_SMF void command_aborted(Bit8u channel, unsigned command);
+ 
+-  BX_HD_SMF void init_send_atapi_command(Bit8u channel, Bit8u command, int req_length, int alloc_length, bool lazy = false);
+-  BX_HD_SMF void ready_to_send_atapi(Bit8u channel);
+-  BX_HD_SMF void raise_interrupt(Bit8u channel);
++  BX_HD_SMF void init_send_atapi_command(Bit8u channel, Bit8u command, int req_length, int alloc_length, bool lazy = false) __attribute__((regparm(3)));
++  BX_HD_SMF void ready_to_send_atapi(Bit8u channel) __attribute__((regparm(1)));
++  BX_HD_SMF void raise_interrupt(Bit8u channel) __attribute__((regparm(1)));
+   BX_HD_SMF void atapi_cmd_error(Bit8u channel, sense_t sense_key, asc_t asc);
+   BX_HD_SMF void init_mode_sense_single(Bit8u channel, const void* src, int size);
+-  BX_HD_SMF void atapi_cmd_nop(Bit8u channel);
++  BX_HD_SMF void atapi_cmd_nop(Bit8u channel) __attribute__((regparm(1)));
+ 
+   // FIXME:
+   // For each ATA channel we should have one controller struct
+Index: iodev/iodev.h
+===================================================================
+RCS file: /cvsroot/bochs/bochs/iodev/iodev.h,v
+retrieving revision 1.29
+diff -u -r1.29 iodev.h
+--- iodev/iodev.h	28 Jan 2003 16:57:00 -0000	1.29
++++ iodev/iodev.h	12 Feb 2003 14:12:42 -0000
+@@ -287,8 +287,8 @@
+   bx_bool register_irq(unsigned irq, const char *name);
+   bx_bool unregister_irq(unsigned irq, const char *name);
+   void iodev_init(void);
+-  Bit32u inp(Bit16u addr, unsigned io_len);
+-  void   outp(Bit16u addr, Bit32u value, unsigned io_len);
++  Bit32u inp(Bit16u addr, unsigned io_len) __attribute__((regparm(2)));
++  void   outp(Bit16u addr, Bit32u value, unsigned io_len) __attribute__((regparm(3)));
+ 
+   static void timer_handler(void *);
+   void timer(void);
+Index: iodev/keyboard.cc
+===================================================================
+RCS file: /cvsroot/bochs/bochs/iodev/keyboard.cc,v
+retrieving revision 1.74
+diff -u -r1.74 keyboard.cc
+--- iodev/keyboard.cc	5 Jan 2003 01:37:21 -0000	1.74
++++ iodev/keyboard.cc	12 Feb 2003 14:12:43 -0000
+@@ -785,7 +785,7 @@
+ 
+ 
+ 
+-  void
++  void __attribute__((regparm(1)))
+ bx_keyb_c::set_kbd_clock_enable(Bit8u   value)
+ {
+   bx_bool prev_kbd_clock_enabled;
+@@ -930,7 +930,7 @@
+ //BX_DEBUG(( "#   out_buffer = %u", (unsigned) BX_KEY_THIS s.kbd_controller.kbd_output_buffer);
+ }
+ 
+-  bx_bool
++  bx_bool __attribute__((regparm(3)))
+ bx_keyb_c::mouse_enQ_packet(Bit8u   b1, Bit8u   b2, Bit8u   b3)
+ {
+   if ((BX_KEY_THIS s.mouse_internal_buffer.num_elements + 3) >= BX_MOUSE_BUFF_SIZE) {
+Index: iodev/keyboard.h
+===================================================================
+RCS file: /cvsroot/bochs/bochs/iodev/keyboard.h,v
+retrieving revision 1.20
+diff -u -r1.20 keyboard.h
+--- iodev/keyboard.h	26 Dec 2002 22:19:44 -0000	1.20
++++ iodev/keyboard.h	12 Feb 2003 14:12:43 -0000
+@@ -215,7 +215,7 @@
+   bx_bool stop_paste;  // stop the current paste operation on hardware reset
+ 
+   BX_KEY_SMF void     resetinternals(bx_bool powerup);
+-  BX_KEY_SMF void     set_kbd_clock_enable(Bit8u   value);
++  BX_KEY_SMF void     set_kbd_clock_enable(Bit8u   value) __attribute__((regparm(1)));
+   BX_KEY_SMF void     set_aux_clock_enable(Bit8u   value);
+   BX_KEY_SMF void     kbd_ctrl_to_kbd(Bit8u   value);
+   BX_KEY_SMF void     kbd_ctrl_to_mouse(Bit8u   value);
+@@ -223,7 +223,7 @@
+   BX_KEY_SMF void     kbd_enQ_imm(Bit8u   val);
+   BX_KEY_SMF void     activate_timer(void);
+   BX_KEY_SMF void     controller_enQ(Bit8u   data, unsigned source);
+-  BX_KEY_SMF bx_bool  mouse_enQ_packet(Bit8u   b1, Bit8u   b2, Bit8u   b3);
++  BX_KEY_SMF bx_bool  mouse_enQ_packet(Bit8u   b1, Bit8u   b2, Bit8u   b3) __attribute__((regparm(3)));
+   BX_KEY_SMF void     mouse_enQ(Bit8u   mouse_data);
+ 
+   static void   timer_handler(void *);
+Index: iodev/ne2k.cc
+===================================================================
+RCS file: /cvsroot/bochs/bochs/iodev/ne2k.cc,v
+retrieving revision 1.47
+diff -u -r1.47 ne2k.cc
+--- iodev/ne2k.cc	13 Dec 2002 18:27:07 -0000	1.47
++++ iodev/ne2k.cc	12 Feb 2003 14:12:43 -0000
+@@ -236,7 +236,7 @@
+ // The first 16 bytes contains the MAC address at even locations,
+ // and there is 16K of buffer memory starting at 16K
+ //
+-Bit32u
++Bit32u __attribute__((regparm(2)))
+ bx_ne2k_c::chipmem_read(Bit32u address, unsigned int io_len)
+ {
+   Bit32u retval = 0;
+@@ -266,7 +266,7 @@
+   return (0xff);
+ }
+ 
+-void
++void __attribute__((regparm(3)))
+ bx_ne2k_c::chipmem_write(Bit32u address, Bit32u value, unsigned io_len)
+ {
+   if ((io_len == 2) && (address & 0x1)) 
+@@ -290,7 +290,7 @@
+ // after that, insw/outsw instructions can be used to move
+ // the appropriate number of bytes to/from the device.
+ //
+-Bit32u
++Bit32u __attribute__((regparm(2)))
+ bx_ne2k_c::asic_read(Bit32u offset, unsigned int io_len)
+ {
+   Bit32u retval = 0;
+Index: iodev/ne2k.h
+===================================================================
+RCS file: /cvsroot/bochs/bochs/iodev/ne2k.h,v
+retrieving revision 1.10
+diff -u -r1.10 ne2k.h
+--- iodev/ne2k.h	19 Nov 2002 18:56:39 -0000	1.10
++++ iodev/ne2k.h	12 Feb 2003 14:12:43 -0000
+@@ -209,14 +209,14 @@
+   BX_NE2K_SMF Bit32u read_cr(void);
+   BX_NE2K_SMF void   write_cr(Bit32u value);
+ 
+-  BX_NE2K_SMF Bit32u chipmem_read(Bit32u address, unsigned io_len);
+-  BX_NE2K_SMF Bit32u asic_read(Bit32u offset, unsigned io_len);
++  BX_NE2K_SMF Bit32u chipmem_read(Bit32u address, unsigned io_len) __attribute__((regparm(2)));
++  BX_NE2K_SMF Bit32u asic_read(Bit32u offset, unsigned io_len) __attribute__((regparm(2)));
+   BX_NE2K_SMF Bit32u page0_read(Bit32u offset, unsigned io_len);
+   BX_NE2K_SMF Bit32u page1_read(Bit32u offset, unsigned io_len);
+   BX_NE2K_SMF Bit32u page2_read(Bit32u offset, unsigned io_len);
+   BX_NE2K_SMF Bit32u page3_read(Bit32u offset, unsigned io_len);
+ 
+-  BX_NE2K_SMF void chipmem_write(Bit32u address, Bit32u value, unsigned io_len);
++  BX_NE2K_SMF void chipmem_write(Bit32u address, Bit32u value, unsigned io_len) __attribute__((regparm(3)));
+   BX_NE2K_SMF void asic_write(Bit32u address, Bit32u value, unsigned io_len);
+   BX_NE2K_SMF void page0_write(Bit32u address, Bit32u value, unsigned io_len);
+   BX_NE2K_SMF void page1_write(Bit32u address, Bit32u value, unsigned io_len);
+Index: iodev/pit82c54.cc
+===================================================================
+RCS file: /cvsroot/bochs/bochs/iodev/pit82c54.cc,v
+retrieving revision 1.20
+diff -u -r1.20 pit82c54.cc
+--- iodev/pit82c54.cc	27 Aug 2002 19:54:46 -0000	1.20
++++ iodev/pit82c54.cc	12 Feb 2003 14:12:43 -0000
+@@ -97,12 +97,14 @@
+     thisctr.OUTpin=data;
+   }
+ 
+-  void pit_82C54::set_count (counter_type & thisctr, Bit32u data) {
++  void  __attribute__((regparm(2)))
++pit_82C54::set_count (counter_type & thisctr, Bit32u data) {
+     thisctr.count=data & 0xFFFF;
+     set_binary_to_count(thisctr);
+   }
+ 
+-  void pit_82C54::set_count_to_binary(counter_type & thisctr) {
++  void  __attribute__((regparm(1)))
++pit_82C54::set_count_to_binary(counter_type & thisctr) {
+     if(thisctr.bcd_mode) {
+       thisctr.count=
+ 	(((thisctr.count_binary/1)%10)<<0) |
+@@ -115,7 +117,8 @@
+     }
+   }
+ 
+-  void pit_82C54::set_binary_to_count(counter_type & thisctr) {
++  void  __attribute__((regparm(1)))
++pit_82C54::set_binary_to_count(counter_type & thisctr) {
+     if(thisctr.bcd_mode) {
+       thisctr.count_binary=
+ 	(1*((thisctr.count>>0)&0xF)) +
+@@ -128,7 +131,8 @@
+     }
+   }
+ 
+-  void pit_82C54::decrement (counter_type & thisctr) {
++  void  __attribute__((regparm(1)))
++pit_82C54::decrement (counter_type & thisctr) {
+     if(!thisctr.count) {
+       if(thisctr.bcd_mode) {
+ 	thisctr.count=0x9999;
+@@ -181,7 +185,8 @@
+   void pit_82C54::reset (unsigned type) {
+   }
+ 
+-void pit_82C54::decrement_multiple(counter_type & thisctr, Bit32u cycles) {
++void  __attribute__((regparm(2)))
++pit_82C54::decrement_multiple(counter_type & thisctr, Bit32u cycles) {
+   while(cycles>0) {
+     if(cycles<=thisctr.count_binary) {
+       thisctr.count_binary-=cycles;
+@@ -276,7 +281,8 @@
+   }
+ }
+ 
+-  void pit_82C54::clock(Bit8u cnum) {
++  void  __attribute__((regparm(1)))
++pit_82C54::clock(Bit8u cnum) {
+     if(cnum>MAX_COUNTER) {
+       BX_ERROR(("Counter number too high in clock"));
+     } else {
+Index: iodev/pit82c54.h
+===================================================================
+RCS file: /cvsroot/bochs/bochs/iodev/pit82c54.h,v
+retrieving revision 1.11
+diff -u -r1.11 pit82c54.h
+--- iodev/pit82c54.h	27 Aug 2002 19:54:46 -0000	1.11
++++ iodev/pit82c54.h	12 Feb 2003 14:12:43 -0000
+@@ -94,17 +94,17 @@
+ 
+   void set_OUT (counter_type & thisctr, bool data);
+ 
+-  void set_count (counter_type & thisctr, Bit32u data);
++  void set_count (counter_type & thisctr, Bit32u data) __attribute__((regparm(2)));
+ 
+-  void set_count_to_binary (counter_type & thisctr);
++  void set_count_to_binary (counter_type & thisctr) __attribute__((regparm(1)));
+ 
+-  void set_binary_to_count (counter_type & thisctr);
++  void set_binary_to_count (counter_type & thisctr) __attribute__((regparm(1)));
+ 
+-  void decrement (counter_type & thisctr);
++  void decrement (counter_type & thisctr) __attribute__((regparm(1)));
+ 
+-  void decrement_multiple(counter_type & thisctr, Bit32u cycles);
++  void decrement_multiple(counter_type & thisctr, Bit32u cycles) __attribute__((regparm(2)));
+ 
+-  void clock(Bit8u cnum);
++  void clock(Bit8u cnum) __attribute__((regparm(1)));
+ 
+   void print_counter(counter_type & thisctr);
+ 
+Index: iodev/vga.cc
+===================================================================
+RCS file: /cvsroot/bochs/bochs/iodev/vga.cc,v
+retrieving revision 1.61
+diff -u -r1.61 vga.cc
+--- iodev/vga.cc	9 Feb 2003 08:25:22 -0000	1.61
++++ iodev/vga.cc	12 Feb 2003 14:12:44 -0000
+@@ -2226,11 +2226,11 @@
+ 
+ 
+ #if BX_SUPPORT_VBE
+-  Bit8u
++  Bit8u  __attribute__((regparm(1)))
+ bx_vga_c::vbe_mem_read(Bit32u addr)
+ {
+-  Bit32u offset;        
+-  
++  Bit32u offset;
++
+   if (addr >= VBE_DISPI_LFB_PHYSICAL_ADDRESS)
+   {
+     // LFB read
+@@ -2249,12 +2249,12 @@
+   return (BX_VGA_THIS s.vbe_memory[offset]);
+ }
+ 
+-  void
++  void __attribute__((regparm(2)))
+ bx_vga_c::vbe_mem_write(Bit32u addr, Bit8u value)
+ {
+-  Bit32u offset;        
++  Bit32u offset;
+   unsigned x_tileno, y_tileno;
+-  
++
+   if (addr >= VBE_DISPI_LFB_PHYSICAL_ADDRESS)
+   {
+     // LFB write
+Index: iodev/vga.h
+===================================================================
+RCS file: /cvsroot/bochs/bochs/iodev/vga.h,v
+retrieving revision 1.23
+diff -u -r1.23 vga.h
+--- iodev/vga.h	9 Feb 2003 08:25:22 -0000	1.23
++++ iodev/vga.h	12 Feb 2003 14:12:45 -0000
+@@ -106,11 +106,11 @@
+   virtual void   mem_write(Bit32u addr, Bit8u value);
+   virtual void   trigger_timer(void *this_ptr);
+ 
+-#if BX_SUPPORT_VBE 
+-  BX_VGA_SMF Bit8u  vbe_mem_read(Bit32u addr);
+-  BX_VGA_SMF void   vbe_mem_write(Bit32u addr, Bit8u value);  
++#if BX_SUPPORT_VBE
++  BX_VGA_SMF Bit8u  vbe_mem_read(Bit32u addr) __attribute__((regparm(1)));
++  BX_VGA_SMF void   vbe_mem_write(Bit32u addr, Bit8u value) __attribute__((regparm(2)));
+ #endif
+-  
++
+   virtual void   redraw_area(unsigned x0, unsigned y0,
+                              unsigned width, unsigned height);
+ 
+Index: memory/memory.cc
+===================================================================
+RCS file: /cvsroot/bochs/bochs/memory/memory.cc,v
+retrieving revision 1.25
+diff -u -r1.25 memory.cc
+--- memory/memory.cc	3 Nov 2002 17:17:11 -0000	1.25
++++ memory/memory.cc	12 Feb 2003 14:12:45 -0000
+@@ -36,7 +36,7 @@
+ 
+ #if BX_PROVIDE_CPU_MEMORY
+ 
+-  void
++  void __attribute__((regparm(3)))
+ BX_MEM_C::writePhysicalPage(BX_CPU_C *cpu, Bit32u addr, unsigned len, void *data)
+ {
+   Bit8u *data_ptr;
+@@ -233,7 +233,7 @@
+ }
+ 
+ 
+-  void
++  void __attribute__((regparm(3)))
+ BX_MEM_C::readPhysicalPage(BX_CPU_C *cpu, Bit32u addr, unsigned len, void *data)
+ {
+   Bit8u *data_ptr;
+Index: memory/memory.h
+===================================================================
+RCS file: /cvsroot/bochs/bochs/memory/memory.h,v
+retrieving revision 1.13
+diff -u -r1.13 memory.h
+--- memory/memory.h	25 Oct 2002 11:44:41 -0000	1.13
++++ memory/memory.h	12 Feb 2003 14:12:45 -0000
+@@ -60,23 +60,23 @@
+   BX_MEM_C(void);
+   BX_MEM_C(size_t memsize);
+   ~BX_MEM_C(void);
+-  BX_MEM_SMF void    alloc_vector_aligned (size_t bytes, size_t alignment);
++  BX_MEM_SMF void    alloc_vector_aligned (size_t bytes, size_t alignment) __attribute__((regparm(2)));
+   BX_MEM_SMF void    init_memory(int memsize);
+   BX_MEM_SMF void    readPhysicalPage(BX_CPU_C *cpu, Bit32u addr,
+-                                      unsigned len, void *data);
++                                      unsigned len, void *data) __attribute__((regparm(3)));
+   BX_MEM_SMF void    writePhysicalPage(BX_CPU_C *cpu, Bit32u addr,
+-                                       unsigned len, void *data);
++                                       unsigned len, void *data) __attribute__((regparm(3)));
+   BX_MEM_SMF void    load_ROM(const char *path, Bit32u romaddress);
+   BX_MEM_SMF Bit32u  get_memory_in_k(void);
+ #if BX_PCI_SUPPORT
+-  BX_MEM_SMF Bit8u*  pci_fetch_ptr(Bit32u addr);
++  BX_MEM_SMF Bit8u*  pci_fetch_ptr(Bit32u addr) __attribute__((regparm(1)));
+ #endif
+   BX_MEM_SMF bx_bool dbg_fetch_mem(Bit32u addr, unsigned len, Bit8u *buf);
+   BX_MEM_SMF bx_bool dbg_set_mem(Bit32u addr, unsigned len, Bit8u *buf);
+   BX_MEM_SMF bx_bool dbg_crc32(
+     unsigned long (*f)(unsigned char *buf, int len),
+     Bit32u addr1, Bit32u addr2, Bit32u *crc);
+-  BX_MEM_SMF Bit8u * getHostMemAddr(BX_CPU_C *cpu, Bit32u a20Addr, unsigned op);
++  BX_MEM_SMF Bit8u * getHostMemAddr(BX_CPU_C *cpu, Bit32u a20Addr, unsigned op) __attribute__((regparm(3)));
+   };
+ 
+ #if BX_PROVIDE_CPU_MEMORY==1
+Index: memory/misc_mem.cc
+===================================================================
+RCS file: /cvsroot/bochs/bochs/memory/misc_mem.cc,v
+retrieving revision 1.35
+diff -u -r1.35 misc_mem.cc
+--- memory/misc_mem.cc	29 Jan 2003 15:01:16 -0000	1.35
++++ memory/misc_mem.cc	12 Feb 2003 14:12:45 -0000
+@@ -63,7 +63,7 @@
+ 
+ 
+ #if BX_PROVIDE_CPU_MEMORY
+-void
++void __attribute__((regparm(2)))
+ BX_MEM_C::alloc_vector_aligned (size_t bytes, size_t alignment)
+ {
+   if (actual_vector != NULL) {
+@@ -196,7 +196,7 @@
+ #endif // #if BX_PROVIDE_CPU_MEMORY
+ 
+ #if BX_PCI_SUPPORT
+-  Bit8u*
++  Bit8u* __attribute__((regparm(1)))
+ BX_MEM_C::pci_fetch_ptr(Bit32u addr)
+ {
+   if (bx_options.Oi440FXSupport->get ()) {
+@@ -306,7 +306,7 @@
+ }
+ 
+ 
+-  Bit8u *
++  Bit8u * __attribute__((regparm(3)))
+ BX_MEM_C::getHostMemAddr(BX_CPU_C *cpu, Bit32u a20Addr, unsigned op)
+   // Return a host address corresponding to the guest physical memory
+   // address (with A20 already applied), given that the calling