Differences between revisions 17 and 18
Revision 17 as of 2020-07-11 13:31:52
Size: 9992
Editor: ThomasKoenig
Comment: Adjusted tree dump name, some advice on debugging the cmpiler.
Revision 18 as of 2020-07-22 15:20:18
Size: 10097
Editor: ThomasKoenig
Comment: Mentioned --enable-valgrind-annotations.
Deletions are marked like this. Additions are marked like this.
Line 113: Line 113:
You may get better valgrind results if you use {{{--enable-valgrind-annotations}}} for the bootstrap.

Quickstart Guide to Hacking Gfortran

Starting to help in developing an existing compiler can be a daunting task. This document aims to go give a new developer a foothold in the many lines of code. It is still very preliminary. Feel free to ask any of the gfortran regulars (or on the gfortran mailing list) for advice or help.

Gfortran is a front end to gcc. Its task is to parse Fortran code and to convert it to an intermediate form, which is then handed off to other parts of gcc (the so-called "middle end") which do further optimization and finally the translation to the assembly language ("back end").

This document is only about the gfortran front end, the other parts of gcc have their own documentation. It assumes you know how to build gcc, including gfortran.

What does the front end do?

The front end's action can be grouped into four phases:

Parsing
This converts source code into a stream of tokens which describe the language. Because Fortran does not have reserved keywords, the gfortran runs a series of matchers against code trying to find one that matches a statement. On failing a match an error message may be queued, and another matcher tried. If all attempts at matching fail, the error queue is dumped to the user.
(parse.c, scanner.c and primary.c)
Resolution
This resolves things left over from the parsing phase, such as types of expressions, and compile-time simplification of constants. Many errors are issued in this phase. At the end of this phase, the abstract syntax tree is finished.
(resolve.c and (for intrinsics) iresolve.c, expr.c, array.c, interface.c and simplify.c)
Front-end optimization
This does some optimization since there is some information in the Fortran language that can not easily be handled by the later stages.
(frontend-passes.c)
Translation
This translates the Fortran abstract syntax tree into a tree stucture suitable for the middle end.
(trans-*.c)

Examining gfortran data structures

There are a few useful options to look at gfortran internals. Compiling a file with gfortran -fdump-fortran-original foo.f90 dumps the internal representation of the Fortran abstract syntax tree to standard output. The code which generates output for this option can be found in dump-parse-tree.c, which can serve as a good starting point for examining gfortran's data structures.

Using gfortran -fdump-tree-original foo.f90 will generate a file named a-foo.f90.004t.original which contains a C-like representation of what the compiler handed off to the middle end. Most code errors can be found from examining this file.

Another interesting option is -fdump-ipa-cgraph, which will dump information about the middle end's symbol table to a-foo.f90.000i.cgraph.

Some documentation on the data files can be found in the GNU Fortran Compiler Internals document.

Using a debugger on the gfortran compiler

You need to run the debugger (usually gdb) on the f951 executable. This can be found in your gcc build directory. Assuming that this is ~/gcc-bin, the executable is in ~/gcc-bin/gcc/f951.

A good starting point is to run gdb with

$ gdb ~/gcc-bin/gcc/f951
(gdb) break show_expr
(gdb) run -fdump-fortran-original hello.f90

and then examine the expressions there. You can find some documentation on the gfc_code and gfc_expr expressions you will encounter in the gcc-internals.texi file in the gfortran source directory.

Another interesting variable to look at is gfc_current_ns. It contains the code found under gfc_current_ns->code and symbols (i.e. variable names, functions etc.) found under gfc_current_ns->sym_root. This is a gfc_symtree pointer. Looking at the first symbol will require you to look at *(gfc_current_ns->sym_root->n.sym).

If you are looking for the source of a particular error, you can set a breakpoint in gfc_error. Be prepared for a large number of false positives, because the parser calls gfc_error frequently for constructs that it may recognize later. It may be a better idea to grep for the error message in the gfortran source files, and then set a break point there.

If you want to inspect a particular internal data structure which is pointed to by a pointer *p , a good first try is to use

(gdb) call debug(p)

on it. There are also some special functions like gfc_debug_code and gfc_debug_expr which you can also call. These functions can also be found in dump-parse-tree.c.

If you want to extend the debugging facilities in dump-parse-tree.c, feel free. This code has no user impact, so it can be extended easily.

Function inlining in the compiler can sometimes make the code hard to follow. One way to deal with that is to go into the gcc subdirectory and to edit the Makefile there to set all code to -O0 instead of -O2. Touch gfortran.h and recompile from there; the resulting compiler will be much easier to debug. If you use gdb, you can also replace the -g option by -ggdb3; this will make gdb recognize marcos, which can be useful. If you do this, don't forget to run your regression tests with normal optimization turned on, because the compiler will only flag some warnings for optimization levels above -O0.

Examining tree structures

Let's say you are looking at the middle end code generated in a tree variable named stmt somewhere in trans-*.c. The best way to look at this is to try

(gdb) call debug(stmt)

which will dump the code using the same internal representation as the -fdump-tree-original option. If the tree you are looking at contains a declaration, this will return an empty line. In this case, you can use

(gdb) call debug_tree(stmt)

which will return a complete, but somewhat hard to read, representation of the declaration.

Breaking on an internal error

If you want to look at an internal error, try setting a breakpoint in fancy_abort. Stepping up from this will lead you to the gcc_unreachable () call where something went wrong.

Using valgrind

Using valgrind with the compiler itself

You may get better valgrind results if you use --enable-valgrind-annotations for the bootstrap.

If you want to use valgrind on the compiler itself, start it (for example) with

$ valgrind --expensive-definedness-checks=yes ~/gcc-bin/gcc/f951 foo.f90

After the gfortran front end itself has finished, there will probably be a lot of messages about sparseset_p. This is a known false positive, you can then stop the run. --expensive-definedness-checks=yes is needed because there is at least one false positive without it - see PR 89747.

Using valgrind with test cases

Make sure you compile with -g, so that error messages will be more informative.

Debugging tree code generated by gfortran

Unfortunately, it is not possible to directly debug the tree code generated by gfortran, no debug info is emitted for the artificial variables.

However, debugging assembly code annotated via -fverbose-asm can help a lot. Assuming you would like to debug a file foo.f90 containting the simple program called foo.f90 with the contents

  real :: a(2,2)
  call random_number(a)
  print *,minloc(a)
end

The command

$ gfortran -S -fverbose-asm -fdump-tree-original-uid -fdump-tree-optimized-uid foo.f90

generates an assembly file foo.s which contains the statements from foo.f90.t*.optimized as comments. In this case, it is important to leave out the -g option. The file foo.f90.*t.optimized then contains GIMPLE statements like

  parm.0.dim[1].lbound = 1;
  parm.0.dim[1].ubound = 2;
  parm.0.dim[1].stride = 2;
  parm.0.data = &a[0];
  parm.0.offset = -3;
  _gfortran_arandom_r4 (&parm.0)

and the assembly file foo.s has the GIMPLE statements as comments, like this:

        movq    $1, -504(%rbp)  #, MEM[(struct array02_real(kind=4) *)_62].dim[1].lbound
        movq    $2, -496(%rbp)  #, MEM[(struct array02_real(kind=4) *)_62].dim[1].ubound
        movq    $2, -512(%rbp)  #, MEM[(struct array02_real(kind=4) *)_62].dim[1].stride
        leaq    -32(%rbp), %rax #, tmp86
        movq    %rax, -576(%rbp)        # tmp86, MEM[(struct array02_real(kind=4) *)_62].data
        movq    $-3, -568(%rbp) #, MEM[(struct array02_real(kind=4) *)_62].offset
        leaq    -576(%rbp), %rax        #, tmp87
        movq    %rax, %rdi      # tmp87,
        movl    $0, %eax        #,
        call    _gfortran_arandom_r4    #

The -uid part of the options makes sure that you can recognize the variables from different passes.

Translating foo.s with

$ gfortran -g foo.s

and using the debugger on ./a.out, breaking on MAIN__ and single-stepping through the assembler file as if it was the source file, with the optimized file for comparison, allows fairly good debugging.

What kind of PR to start with

All currently open bugs reports (called PRs) can be found in the gcc bugtracker called bugzilla if you set the product to fortran.

Traditionally, internal compiler errors on invalid code (gcc bugzilla keyword ice-on-invalid-code). have been considered relatively easy. But you may always find a hard one...

Happy hacking!

None: GFortranHacking (last edited 2020-07-22 15:20:18 by ThomasKoenig)