This document explains a couple of things that are specific to VMS. There are currently two "chapters", the first deals with cross-assembly issues, and the second deals with the VMS debugger and GNU-CC. *********************************************************************** ****************** Notes for Cross Assembly with VMS ****************** *********************************************************************** If you wish to build gas on a non-VMS system to cross-assemble, you should use: configure ${hosttype} -target=vms and then follow the usual procedure. The object files generated on Unix will be correct from a binary point of view, but the real trick is getting them to the VMS machine. The format of the object file is a variable-length record, but each record contains binary data. gas writes the records in the same format that VMS would expect, namely a two-byte count followed by that number of bytes. If you try to copy the file to a VMS system using ftp, the ftp protocol will screw up the file by looking for nulls (record terminator for unix) and it will insert it's own record terminators at that point. This will obviously corrupt the file. If you try to transfer the file with ftp in binary mode, the file itself will not be corrupt, but VMS will think that the file contains fixed-length records of 512 bytes. You can use the public-domain FILE utility to change this with a command like: $FILE foo.o/type=variable If you do not have this utility available, the following program can be used to perform this task: #include <fab.h> #define RME$C_SETRFM 1 struct FAB * fab; main(int argc, char * argv[]){ int i, status; fab = (struct FAB*) malloc(sizeof(struct FAB)); *fab = cc$rms_fab; /* initialize FAB*/ fab->fab$b_fac = FAB$M_PUT; fab->fab$l_fop |= FAB$M_ESC; fab->fab$l_ctx = RME$C_SETRFM; fab->fab$w_ifi = 0; for(i=1;i<argc;i++){ printf("Setting %s to variable length records.\n",argv[i]); fab->fab$l_fna = argv[i]; fab->fab$b_fns = strlen(argv[i]); status = sys$open(fab,0,0); if((status & 7) != 1) lib$signal(status); fab->fab$b_rfm = FAB$C_VAR; status = sys$modify(fab,0,0); if((status & 7) != 1) lib$signal(status); status = sys$close(fab,0,0); if((status & 7) != 1) lib$signal(status); }; } If you have NFS running on the VMS system, what you need to do depends upon which NFS software you are running on the VMS system. There are a number of different TCP/IP packages for VMS available, and only very limited testing has been performed. In the tests that has been done so far, the contents of the file will always be correct when transferring the file via NFS, but the record attributes may or may not be correct. One proprietary TCP/IP/NFS package for VMS is known to automatically fix the record attributes of the object file if you NFS mount a unix disk from the VMS system, and if the file has a ".obj" extension on the unix system. Other TCP/IP packages might do this for you as well, but they have not been checked. No matter what method you use to get the file to the VMS system, it is always a good idea to check to make sure that it is the correct type by doing a "$dir/full" on the object file. The desired record attributes will be "None". Undesirable record attributes will be "Stream-LF" or anything else. Once you get the files on the VMS system, you can check their integrity with the "$anal/obj" command. (Naturally at some point you should rename the .o files to .obj). As far as the debugger is concerned, the records will be correct, but the debugger will not be able to find the source files, since it only has the file name, and not the full directory specification. You must give the debugger some help by telling it which directories to search for the individual files - once you have done this you should be able to proceed normally. It is a good idea to use names for your files which will be valid under VMS, since otherwise you will have no way of getting the debugger to find the source file when deugging. The reason for this is that the object file normally contins specific information that the debugger can use to positively identify a file, and if you are assembling on a unix system this information simply does not exist in a meaningful way. You must help the debugger by using the "SET FILE=" command to tell the debugger where to look for source files. The debugger records will be correct, except that the debugger will not be initially able to find the source files. You can use the "SET FILE" command to tell the debugger where to look for the source files. I have only tested this with a SVr4 i486 machine, and everything seems to work OK, with the limited testing that I have done. Other machines may or may not work. You should read the chapters on cross-compilers in the gcc manual before fooling with this. Since gas does not need to do any floating point arithmetic, the floating point constants that are generated here should be correct - the only concern is with constant folding in the main compiler. The range and precision of floats and doubles are similar on the 486 (with a builtin 80387) and the VAX, although there is a factor of 2 to 4 difference in the range. The double, as implemented on the 486, is quite similar to the G_FLOAT on the VAX. *********************************************************************** ****************** Notes for using GNU CC with the VMS debugger******** *********************************************************************** 1) You should be aware that GNU-C, as with any other decent compiler, will do things when optimization is turned on that you may not expect. Sometimes intermediate results are not written to variables, if they are only used in one place, and sometimes variables that are not used at all will not be written to the symbol table. Also, parameters to inline functions are often inaccessible. You can see the assembly code equivalent by using KP7 in the debugger, and from this you can tell if in fact a variable should have the value that you expect. You can find out if a variable lives withing a register by doing a 'show symbol/addr'. 2) Overly complex data types, such as: int (*(*(*(*(*(* sarr6)[1])[1])[2])[3])[4])[5]; will not be debugged properly, since the debugging record overflows an internal debugger buffer. gcc-as will convert these to *void as far as the debugger symbol table is concerned, which will avoid any problems, and the assembler will give you a message informing you that this has happened. 3) You must, of course, compile and link with /debug. If you link without debug, you still get traceback table in the executable, but there is no symbol table for variables. 4) Included in the patches to VMS.C are fixes to two bugs that are unrelated to the changes that I have made. One of these made it impossible to debug small programs sometimes, and the other caused the debugger to become confused about which routine it was in, and give this incorrect info in tracebacks. 5) If you are using the GNU-C++ compiler, you should modify the compiler driver file GNU_CC:[000000]GCC.COM (or GXX.COM). If you have a seperate GXX.COM, then you need to change one line in GXX.COM to: $ if f$locate("D",p2) .ne. P2_Length then Debug = " ""-G0""" Notice zero---> ^ If you are using a GCC.COM that does both C and C++, add the following lines to GCC.COM: $! $! Use old style debugging records for VMS $! $ if (Debug.nes."" ).and. Plus then Debug = " ""-G0""" after the variables Plus and Debug are set. The reason for this, is that C++ compiler by default generates debugging records that are more complex, with many new syntactical elements that allow for the new features of the language. The -G0 switch tells the C++ compiler to use the old style debugging records. Until the debugger understands C++ there is not any point to try and use the expanded syntax. 6) When you have nested scopes, i.e.: main(){ int i; {int i; {int i; };};} and you say "EXAM i" the debugger needs to figure out which variable you actually want to reference. I have arranged things to define a block to the debugger when you use brackets to enter a new scope, so in the example above, the variables would be described as: TEST\main\i TEST\main\$0\i TEST\main\$0\$0\i At each level, the block name is a number with a dollar sign prefix, the numbers start with 0 and count upward. When you say EXAM i, the debugger looks at the current PC, and decides which block it is currently in. It works from the innermost level outward until it finds a block that has the variable "i" defined. You can always specify the scope explicitly. 7) With C++, there can be a lot of inline functions, and it would be rather restrictive to force the user to debug the program by converting all of the inline functions to normal functions. What I have done is to essentially "add" (with the debugger) source lines from the include files that contain the inline functions. Thus when you step into an inline function it appears as if you have called the function, and you can examine variables and so forth. There are several *very* important differences, however. First of all, since there is no function call involved, you cannot step over the inline function call - you always step into it. Secondly, since the same source lines are used in many locations, there is a seperate copy of the source for *each* usage. Without this, breakpoints do not work, since we must have a 1-to-1 mapping between source lines and PC. Since you cannot step over inline function calls, it can be a real pain if you are not really interested in what is going on for that function call. What I have done is to use the "-D" switch for the assembler to toggle the following behavior. With the "-D" switch, all inline functions are included in the object file, and you can debug everything. Without the "-D" switch (default case with VMS implementation), inline functions are included *only* if they did not come from system header files (i.e. from GNU_CC_INCLUDE: or GNU_GXX_INCLUDE:). Thus, without the switch the user only debugs his/her own inline functions, and not the system ones. (This is especially useful if you do a lot of stream I/O in C++). This probably will not provide enough granularity for many users, but for now this is still somewhat experimental, and I would like to reflect upon it and get some feedback before I go any further. Possible solutions include an interactive prompting, a logical name, or a new command line option in gcc.c (which is then passed through somehow to the guts of the assembler). The inline functions from header files appear after the source code for the source file. This has the advantage that the source file itself is numbered with the same line numbers that you get with an editor. In addition, the entire header file is not included, since the assembler makes a list of the min and max source lines that are used, and only includes those lines from the first to the last actually used. (It is easy to change it to include the whole file). 8) When you are debugging C++ objects, the object "this" is refered to as "$this". Actually, the compiler writes it as ".this", but the period is not good for the debugger, so I have a routine to convert it to a $. (It actually converts all periods to $, but only for variables, since this was intended to allow us to access "this". 9) If you use the asm("...") keyword for global symbols, you will not be able to see that symbol with the debugger. The reason is that there are two records for the symbol stored in the data structures of the assembler. One contains the info such as psect number and offset, and the other one contains the information having to do with the data type of the variable. In order to debug as symbol, you need to be able to coorelate these records, and the only way to do this is by name. The record with the storage attributes will take the name used in the asm directive, and the record that specifies the data type has the actual variable name, and thus when you use the asm directive to change a variable name, the symbol becomes invisible. 10) Older versions of the compiler ( GNU-C 1.37.92 and earlier) place global constants in the text psect. This is unfortunate, since to the linker this appears to be an entry point. I sent a patch to the compiler to RMS, which will generate a .const section for these variables, and patched the assembler to put these variables into a psect just like that for normal variables, except that they are marked NOWRT. static constants are still placed in the text psect, since there is no need for any external access.