--- utils/clflags.ml.orig	Thu May 15 09:30:31 1997
+++ utils/clflags.ml	Thu Jun 25 00:07:14 1998
@@ -24,6 +24,7 @@
 and print_types = ref false             (* -i *)
 and make_archive = ref false            (* -a *)
 and debug = ref false                   (* -g *)
+and profile = ref false			(* -p *)
 and fast = ref false                    (* -unsafe *)
 and link_everything = ref false         (* -linkall *)
 and custom_runtime = ref false          (* -custom *)
--- driver/optmain.ml.orig	Tue Nov 18 12:14:54 1997
+++ driver/optmain.ml	Thu Jun 25 00:12:55 1998
@@ -81,6 +81,8 @@
              "<file>  Set output file name to <file> (default a.out)";
        "-output-obj", Arg.Unit(fun () -> output_c_object := true),
              "Output a C object file instead of an executable";
+       "-p", Arg.Set profile, 
+             "Make a profiled binary";
        "-pp", Arg.String(fun s -> preprocessor := Some s),
              "<command>  Pipe sources through preprocessor <command>";
        "-S", Arg.Set keep_asm_file, " Keep intermediate assembly file";
--- asmcomp/i386/emit.mlp.orig	Fri Nov 14 06:02:04 1997
+++ asmcomp/i386/emit.mlp	Thu Jun 25 01:21:04 1998
@@ -28,6 +28,8 @@
 
 let stack_offset = ref 0
 
+let profile_lbl = ref 0
+
 (* Layout of the stack frame *)
 
 let frame_size () =                     (* includes return address *)
@@ -64,6 +66,9 @@
 let emit_label lbl =
   emit_string label_prefix; emit_int lbl
 
+let emit_profile_label () =
+  emit_string label_prefix; emit_string "P"; emit_int !profile_lbl
+
 (* Some data directives have different names under Solaris *)
 
 let word_dir =
@@ -725,6 +730,22 @@
   emit_align 4;
   `	.globl	{emit_symbol fundecl.fun_name}\n`;
   `{emit_symbol fundecl.fun_name}:\n`;
+  if !Clflags.profile then
+  begin
+    `	pushl %ebp;\n`;
+    `	movl %esp, %ebp\n`;
+    `.data\n`;
+    emit_align 4;
+    `{emit_profile_label ()}:\n`;
+    `	.long 0\n`;
+    `.text\n`;
+    `	pushl %edx\n`;
+    `	movl ${emit_profile_label ()},%edx\n`;
+    `	call mcount\n`;
+    `	popl %edx\n`;
+    `	popl %ebp\n`;
+    incr profile_lbl;
+  end;
   let n = frame_size() - 4 in
   if n > 0 then
     `	subl	${emit_int n}, %esp\n`;
@@ -772,6 +793,7 @@
 (* Beginning / end of an assembly file *)
 
 let begin_assembly() =
+  profile_lbl:=0;
   let lbl_begin = Compilenv.current_unit_name() ^ "_data_begin" in
   `	.data\n`;
   `	.globl	{emit_symbol lbl_begin}\n`;
--- asmcomp/asmlink.ml.orig	Wed Jul  2 14:14:34 1997
+++ asmcomp/asmlink.ml	Thu Jun 25 15:16:19 1998
@@ -171,7 +171,9 @@
   close_out oc
 
 let call_linker file_list startup_file =
-  let libname = "libasmrun" ^ ext_lib in
+  let libname = 
+    (if !Clflags.profile then "libasmprun"
+      else "libasmrun") ^ ext_lib in
   let runtime_lib =
     try
       find_in_path !load_path libname
@@ -181,11 +183,12 @@
     match Config.system with
       "win32" ->
         if not !Clflags.output_c_object then
-          Printf.sprintf "%s /Fe%s -I%s %s %s %s %s %s %s"
+          Printf.sprintf "%s /Fe%s -I%s %s %s %s %s %s %s %s"
             Config.native_c_compiler
             !Clflags.exec_name
             Config.standard_library
             (String.concat " " (List.rev !Clflags.ccopts))
+            (if !Clflags.profile then "-pg" else "")
             startup_file
             (String.concat " " (List.rev file_list))
             (String.concat " " (List.rev !Clflags.ccobjs))
@@ -199,11 +202,12 @@
             (String.concat " " (List.rev file_list))
     | _ ->
         if not !Clflags.output_c_object then
-          Printf.sprintf "%s -o %s -I%s %s %s %s -L%s %s %s %s"
+          Printf.sprintf "%s -o %s -I%s %s %s %s %s -L%s %s %s %s"
             Config.native_c_compiler
             !Clflags.exec_name
             Config.standard_library
             (String.concat " " (List.rev !Clflags.ccopts))
+            (if !Clflags.profile then "-pg" else "")
             startup_file
             (String.concat " " (List.rev file_list))
             Config.standard_library
--- asmrun/Makefile.orig	Mon Nov 17 05:38:16 1997
+++ asmrun/Makefile	Fri Jul 17 18:50:03 1998
@@ -4,6 +4,9 @@
 FLAGS=-I../byterun -DNATIVE_CODE -DTARGET_$(ARCH) -DSYS_$(SYSTEM) 
 CFLAGS=$(FLAGS) -O $(NATIVECCCOMPOPTS)
 DFLAGS=$(FLAGS) -g -DDEBUG $(NATIVECCCOMPOPTS)
+PFLAGS=$(FLAGS) -pg $(NATIVECCCOMPOPTS)
+
+ASPPPFLAGS=$(ASPPFLAGS) -DPROFILE
 
 COBJS=startup.o main.o fail.o roots.o signals.o \
   misc.o freelist.o major_gc.o minor_gc.o memory.o alloc.o compare.o ints.o \
@@ -15,22 +18,28 @@
 
 OBJS=$(COBJS) $(ASMOBJS)
 DOBJS=$(COBJS:.o=.d.o) $(ASMOBJS)
+POBJS=$(COBJS:.o=.p.o) $(ARCH).p.o
 
-all: libasmrun.a
+all: libasmrun.a libasmprun.a
 
 libasmrun.a: $(OBJS)
 	rm -f libasmrun.a
 	ar rc libasmrun.a $(OBJS)
 	$(RANLIB) libasmrun.a
 
+libasmprun.a: $(POBJS)
+	rm -f libasmprun.a
+	ar rc libasmprun.a $(POBJS)
+	$(RANLIB) libasmprun.a
+
 libasmrund.a: $(DOBJS)
 	rm -f libasmrund.a
 	ar rc libasmrund.a $(DOBJS)
 	$(RANLIB) libasmrund.a
 
 install:
-	cp libasmrun.a $(LIBDIR)
-	cd $(LIBDIR); $(RANLIB) libasmrun.a
+	cp libasmrun.a libasmprun.a $(LIBDIR)
+	cd $(LIBDIR); $(RANLIB) libasmrun.a; $(RANLIB) libasmprun.a
 
 power.o: power-$(SYSTEM).o
 	cp power-$(SYSTEM).o power.o
@@ -98,16 +107,22 @@
 clean::
 	rm -f $(LINKEDFILES)
 
-.SUFFIXES: .S .d.o
+.SUFFIXES: .S .d.o .p.o
 
 .S.o:
 	$(ASPP) $(ASPPFLAGS) -o $*.o $*.S
 
+.S.p.o:
+	$(ASPP) $(ASPPPFLAGS) -o $*.p.o $*.S
+
 .c.d.o:
 	@ if test -f $*.o; then mv $*.o $*.f.o; else :; fi
 	$(CC) -c $(DFLAGS) $<
 	mv $*.o $*.d.o
 	@ if test -f $*.f.o; then mv $*.f.o $*.o; else :; fi
+
+.c.p.o:
+	$(CC) -c $(PFLAGS) -o $*.p.o $*.c
 
 clean::
 	rm -f *.o *.s *.a *~
--- asmrun/i386.S.orig	Fri Jul 17 18:35:12 1998
+++ asmrun/i386.S	Fri Jul 17 18:46:41 1998
@@ -41,7 +41,20 @@
         leal    4(%esp), %eax
         movl    %eax, G(caml_bottom_of_stack)
     /* Build array of registers, save it into caml_gc_regs */
-L105:   pushl   %ebp
+.L105:
+        pushl   %ebp
+#ifdef PROFILE
+        movl    %esp,%ebp
+.data
+        .align  4
+.LP0:
+        .long   0
+.text
+        pushl   %edx
+        movl    $.LP0,%edx
+        call    mcount
+        popl    %edx
+#endif /* PROFILE */
         pushl   %edi
         pushl   %esi
         pushl   %edx
@@ -64,58 +77,114 @@
 
         .align  FUNCTION_ALIGN
 G(caml_alloc1):
+#ifdef PROFILE
+        pushl   %ebp
+        movl    %esp,%ebp
+.data
+        .align  4
+.LP1:
+        .long   0
+.text
+        pushl   %edx
+        movl    $.LP1,%edx
+        call    mcount
+        popl    %edx
+        popl    %ebp
+#endif /* PROFILE */
         movl    G(young_ptr), %eax
         subl    $8, %eax
         movl    %eax, G(young_ptr)
         cmpl    G(young_limit), %eax
-        jb      L100
+        jb      .L100
         ret
-L100:   movl    0(%esp), %eax
+.L100:   movl    0(%esp), %eax
         movl    %eax, G(caml_last_return_address)
         leal    4(%esp), %eax
         movl    %eax, G(caml_bottom_of_stack)
-        call    L105
+        call    .L105
         jmp     G(caml_alloc1)
 
         .align  FUNCTION_ALIGN
 G(caml_alloc2):
+#ifdef PROFILE
+        pushl   %ebp
+        movl    %esp,%ebp
+.data
+        .align  4
+.LP2:
+        .long   0
+.text
+        pushl   %edx
+        movl    $.LP2,%edx
+        call    mcount
+        popl    %edx
+        popl    %ebp
+#endif /* PROFILE */
         movl    G(young_ptr), %eax
         subl    $12, %eax
         movl    %eax, G(young_ptr)
         cmpl    G(young_limit), %eax
-        jb      L101
+        jb      .L101
         ret
-L101:   movl    0(%esp), %eax
+.L101:   movl    0(%esp), %eax
         movl    %eax, G(caml_last_return_address)
         leal    4(%esp), %eax
         movl    %eax, G(caml_bottom_of_stack)
-        call    L105
+        call    .L105
         jmp     G(caml_alloc2)
 
         .align  FUNCTION_ALIGN
 G(caml_alloc3):
+#ifdef PROFILE
+        pushl   %ebp
+        movl    %esp,%ebp
+.data
+        .align  4
+.LP3:
+        .long   0
+.text
+        pushl   %edx
+        movl    $.LP3,%edx
+        call    mcount
+        popl    %edx
+        popl    %ebp
+#endif /* PROFILE */
         movl    G(young_ptr), %eax
         subl    $16, %eax
         movl    %eax, G(young_ptr)
         cmpl    G(young_limit), %eax
-        jb      L102
+        jb      .L102
         ret
-L102:   movl    0(%esp), %eax
+.L102:   movl    0(%esp), %eax
         movl    %eax, G(caml_last_return_address)
         leal    4(%esp), %eax
         movl    %eax, G(caml_bottom_of_stack)
-        call    L105
+        call    .L105
         jmp     G(caml_alloc3)
 
         .align  FUNCTION_ALIGN
 G(caml_alloc):
+#ifdef PROFILE
+        pushl   %ebp
+        movl    %esp,%ebp
+.data
+        .align  4
+.LP4:
+        .long   0
+.text
+        pushl   %edx
+        movl    $.LP4,%edx
+        call    mcount
+        popl    %edx
+        popl    %ebp
+#endif /* PROFILE */
         subl    G(young_ptr), %eax      /* eax = size - young_ptr */
         negl    %eax                    /* eax = young_ptr - size */
         cmpl    G(young_limit), %eax
-        jb      L103
+        jb      .L103
         movl    %eax, G(young_ptr)
         ret
-L103:   subl    G(young_ptr), %eax      /* eax = - size */
+.L103:   subl    G(young_ptr), %eax      /* eax = - size */
         negl    %eax                    /* eax = size */
         pushl   %eax                    /* save desired size */
         subl    %eax, G(young_ptr)      /* must update young_ptr */
@@ -123,7 +192,7 @@
         movl    %eax, G(caml_last_return_address)
         leal    8(%esp), %eax
         movl    %eax, G(caml_bottom_of_stack)
-        call    L105
+        call    .L105
         popl    %eax                    /* recover desired size */
         jmp     G(caml_alloc)
 
@@ -132,6 +201,20 @@
         .globl  G(caml_c_call)
         .align  FUNCTION_ALIGN
 G(caml_c_call):
+#ifdef PROFILE
+        pushl   %ebp
+        movl    %esp,%ebp
+.data
+        .align  4
+.LP5:
+        .long   0
+.text
+        pushl   %edx
+        movl    $.LP5,%edx
+        call    mcount
+        popl    %edx
+        popl    %ebp
+#endif /* PROFILE */
     /* Record lowest stack address and return address */
         movl    (%esp), %edx
         movl    %edx, G(caml_last_return_address)
@@ -145,6 +228,20 @@
         .globl  G(caml_start_program)
         .align  FUNCTION_ALIGN
 G(caml_start_program):
+#ifdef PROFILE
+        pushl   %ebp
+        movl    %esp,%ebp
+.data
+        .align  4
+.LP6:
+        .long   0
+.text
+        pushl   %edx
+        movl    $.LP6,%edx
+        call    mcount
+        popl    %edx
+        popl    %ebp
+#endif /* PROFILE */
     /* Save callee-save registers */
         pushl   %ebx
         pushl   %esi
@@ -153,18 +250,18 @@
     /* Initial entry point is caml_program */
         movl    $ G(caml_program), %esi
     /* Common code for caml_start_program and callback* */
-L106:
+.L106:
     /* Build a callback link */
         pushl   G(caml_gc_regs)
         pushl   G(caml_last_return_address)
         pushl   G(caml_bottom_of_stack)
     /* Build an exception handler */
-        pushl   $L108
+        pushl   $.L108
         pushl   G(caml_exception_pointer)
         movl    %esp, G(caml_exception_pointer)
     /* Call the Caml code */
         call    *%esi
-L107:
+.L107:
     /* Pop the exception handler */
         popl    G(caml_exception_pointer)
         popl    %esi    /* dummy register */
@@ -179,7 +276,7 @@
         popl    %ebx
     /* Return to caller. */
         ret
-L108:
+.L108:
     /* Exception handler*/
     /* Pop the callback link, restoring the global variables */
         popl    G(caml_bottom_of_stack)
@@ -195,6 +292,20 @@
         .globl  G(raise_caml_exception)
         .align  FUNCTION_ALIGN
 G(raise_caml_exception):
+#ifdef PROFILE
+        pushl        %ebp
+        movl        %esp,%ebp
+.data
+        .align        4
+.LP7:
+        .long        0
+.text
+        pushl        %edx
+        movl        $.LP7,%edx
+        call        mcount
+        popl        %edx
+        popl        %ebp
+#endif /* PROFILE */
         movl    4(%esp), %eax
         movl    G(caml_exception_pointer), %esp
         popl    G(caml_exception_pointer)
@@ -205,6 +316,20 @@
         .globl  G(callback)
         .align  FUNCTION_ALIGN
 G(callback):
+#ifdef PROFILE
+        pushl        %ebp
+        movl        %esp,%ebp
+.data
+        .align        4
+.LP8:
+        .long        0
+.text
+        pushl        %edx
+        movl        $.LP8,%edx
+        call        mcount
+        popl        %edx
+        popl        %ebp
+#endif /* PROFILE */
     /* Save callee-save registers */
         pushl   %ebx
         pushl   %esi
@@ -214,11 +339,25 @@
         movl    20(%esp), %ebx   /* closure */
         movl    24(%esp), %eax   /* argument */
         movl    0(%ebx), %esi    /* code pointer */
-        jmp     L106
+        jmp     .L106
 
         .globl  G(callback2)
         .align  FUNCTION_ALIGN
 G(callback2):
+#ifdef PROFILE
+        pushl        %ebp
+        movl        %esp,%ebp
+.data
+        .align        4
+.LP9:
+        .long        0
+.text
+        pushl        %edx
+        movl        $.LP9,%edx
+        call        mcount
+        popl        %edx
+        popl        %ebp
+#endif /* PROFILE */
     /* Save callee-save registers */
         pushl   %ebx
         pushl   %esi
@@ -229,11 +368,25 @@
         movl    24(%esp), %eax   /* first argument */
         movl    28(%esp), %ebx   /* second argument */
         movl    $ G(caml_apply2), %esi   /* code pointer */
-        jmp     L106
+        jmp     .L106
 
         .globl  G(callback3)
         .align  FUNCTION_ALIGN
 G(callback3):
+#ifdef PROFILE
+        pushl        %ebp
+        movl        %esp,%ebp
+.data
+        .align        4
+.LP10:
+        .long        0
+.text
+        pushl        %edx
+        movl        $.LP10,%edx
+        call        mcount
+        popl        %edx
+        popl        %ebp
+#endif /* PROFILE */
     /* Save callee-save registers */
         pushl   %ebx
         pushl   %esi
@@ -245,13 +398,13 @@
         movl    28(%esp), %ebx   /* second argument */
         movl    32(%esp), %ecx   /* third argument */
         movl    $ G(caml_apply3), %esi   /* code pointer */
-        jmp     L106
+        jmp     .L106
 
         .data
         .globl  G(system_frametable)
 G(system_frametable):
         .long   1               /* one descriptor */
-        .long   L107            /* return address into callback */
+        .long   .L107            /* return address into callback */
 #ifndef SYS_solaris
         .word   -1              /* negative frame size => use callback link */
         .word   0               /* no roots here */