This is the mail archive of the gcc-patches@gcc.gnu.org mailing list for the GCC project.


Index Nav: [Date Index] [Subject Index] [Author Index] [Thread Index]
Message Nav: [Date Prev] [Date Next] [Thread Prev] [Thread Next]
Other format: [Raw text]

Re: Designs for better debug info in GCC


On Dec 18, 2007, Alexandre Oliva <aoliva@redhat.com> wrote:

> On Dec 17, 2007, Diego Novillo <dnovillo@google.com> wrote:
>> On 12/17/07 19:50, Alexandre Oliva wrote:
>>> Now, since you're so interested in it and you've already read the
>>> various perspectives on the issue that I listed in my yesterday's
>>> e-mail to you, would you help me improve this document, by letting me
>>> know what you believe to be missing from the selected postings on
>>> design strategies, rationales and goals:

>> No.  I am not interested in organizing your thoughts for you.

> Wow, nice shot!

Rats, this below-the-waistline attack really got me annoyed.

So annoyed that I spent the night writing up this consolidated design
document.

So, what do you say now?


	A plan to fix local variable debug information in GCC

		by Alexandre Oliva <aoliva@redhat.com>

			   2007-12-18 draft

== Introduction

The DWARF Debugging Information Format, version 3, determines the ways
a compiler can communicate the location of user variables at run time
to debug information consumers such as debuggers, program analysis
tools, run-time monitors, etc.

One possibility is that the location of a variable is fixed throughout
the execution of a function.  This is generally good enough for
unoptimized programs.

However, for optimized programs, the location of a variable can vary.
The variable may be live for some parts of a function, even in
multiple locations simultaneously.  At other parts, it may be
completely unavailable, or it may still be computable even if no
location actually holds its value.  The encoding, in these cases, can
be a location list: tuples with possibly-overlapping ranges of
instructions, and location expressions that determine a location or a
value for the variable.

Historically, GCC started with the simpler, fixed-location model.  In
fact, back then, there weren't debug information formats that could
represent anything better than this.

More recently, GCC gained code to keep track of varying locations, and
to emit debug information accordingly.  Unfortunately, very many
optimization passes discard information that would be necessary to
emit correct and complete variable location lists.

Coalescing, scalarizing, substituting, propagating, and many other
transformations prevent the late-running variable tracker from doing
an accurate job.  By the time it runs, many variables no longer show
up in the retained annotations, although they're still conceptually
available.

The variable tracker can't tell when a user variable overlaps with
another, and it can't tell when a variable is overwritten, if the
assignment is optimized away.  These limitations are inherent to a
model based on inspecting actual code and trying to make inferences
from that.  In order to be able to represent not only what remained in
the code, but also what was optimized, combined or otherwise
apparently-removed, additional information needs to be kept around.

This paper describes an approach to maintain this information.


== Goals

* Ensure that, for every user variable for which we emit debug
information, the information is correct, i.e., if it says the value of
a variable at a certain instruction is at certain locations, or is a
known constant, then the variable must not be at any other location at
that point, and the locations or values must match reasonable
expectations based on source code inspection.

* Defining "reasonable expectations" is tricky, for code reordering
typical of optimization can make room for numerous surprises.  I don't
have a precise definition for this yet, but very clearly to me saying
that a variable holds a value that it couldn't possibly hold (e.g.,
because it is only assigned that value in a code path that is
knowingly not taken) is a very clear indication that something is
amiss.  The general guiding rule is, if we aren't sure the information
is correct (or we're sure it isn't), we shouldn't pretend that it is.

* Try to ensure that, if the value of a variable is a known constant
at a certain point in the program, this information is present in
debug information.

* Try to ensure that, if the value of a variable is available or
computable at any location at a certain point in the program, this
information is present in debug information.

* Stop missing optimizations for the sake of preserving debug
information.

* Avoid using additional memory and CPU cycles that would be needed
only for debug information when compiling without generating debug
information


== Internal Representation

For historical reasons, GCC has two completely different, even if
nearly isomorphic, internal representations: trees and RTL.  This
decision has required a lot of code to be duplicated for low-level
manipulation and simplification of each of these representations.

Since tracking variables and their values must start early to ensure
correctness, and be carried throughout the complete optimization
process, it might seem tempting to introduce yet another
representation for debug information, decaying both isomorphic
representations into a single debug information representation.  The
drawbacks would be additional duplication of internal representation
manipulation code, and the possibility of increasing memory use out of
the need for representing information in yet another format.

Another concern is that even the simplest compiler transformations may
need to be reflected in debug information.  This might indicate a need
for modifying every point of transformation in every optimization pass
so as to propagate information into the debug information
representation.  This is undesirable, because it would be very
intrusive.

But then, keeping references to the correct values, expressions or
variables, as transformations are made, is precisely what optimization
passes have to do to perform their jobs correctly.  Finding a way to
take advantage of this is a very non-intrusive way of keeping debug
information accurate.  In fact, most transformations wouldn't need any
changes whatsoever: uses of variables in debug information can, in
most optimization passes, be handled just like any other uses.

Once this is established, a possible representation becomes almost
obvious: statements (in trees) or instructions (in rtl) that assert,
to the variable tracker, that a user variable or member is represented
by a given expression:

  # DEBUG var expr

By var, we mean a tree expression that denotes a user variable, for
now.  We envision trivially extending it to support components of
variables in the future.

By expr, we mean a tree or rtl expression that computes the value of
the variable at the point in which the statement or instruction
appears in the program.  A special value needs to be specified for
each representation that denotes a location or value that cannot be
determined or represented in debug information, for example, the
location of a variable that was completely optimized away.  It might
be useful to represent the expression as a list of expressions, and to
distinguish lvalues from rvalues, but for now let's keep this simple.


== Generating debug information

Generating initial annotations when entering SSA is early enough in
the translation that the program will still reflect very reliably the
original source code.  Annotations are only generated for user
variables that are GIMPLE registers, i.e., variables that represent
scalar values and that never have their address taken.  Other kinds of
variables don't have varying locations, so we don't need to worry
about them.

After every assignment to such a variable, we emit a DEBUG statement
that will preserve, throughout compilation, the information that, at
that point, the assigned variable was represented by that expression.
So, after turning an assignment such as the following into SSA form,
we emit the debug statement below right after it:

  x_1 = whatever;
  # DEBUG x x_1

Likewise, at control flow merge points, for each PHI node we introduce
in the SSA representation, we emit an annotation:

  # x_4 = PHI <x_1(3), x_2(4), x_3(7)>;
  # DEBUG x x_4

Then, we let tree optimizers do their jobs.  Whenever they rename,
renumber, coalesce, combine or otherwise optimize a variable, they
will automatically update debug statements that mention them as well.

In the rare cases in which the presence of such a statement might
prevent an optimization, we need to adjust the optimizer code such
that the optimization is not prevented.  This most often amounts to
skipping or otherwise ignoring debug statements.  In a few very rare
cases, special code might be needed to adjust debug statements
manually.

After transformation to RTL, the representation needs translation, but
conceptually it's still the same: a mapping from variable to
expression.  Again, optimizers will most often adjust debug
instructions automatically.

The exceptions can be handled at no cost: the test for whether an
element of the instruction stream is an instruction or some kind of
note, that never needs updating, is a range test, in its optimized
form.  By placing the identifier for a debug instruction at one of the
limits of this range, testing for both ranges requires identical code,
except for the constants.

Since most code that tests for INSN_P and handles instructions can and
should match debug instructions as well, in order to keep them up to
date, we extend INSN_P so as to match debug instructions, and modify
the exceptions, that need to skip debug instructions, by using an
alternate test, with the same meaning as the original definition of
INSN_P.  These simple and non-intrusive changes are relatively common,
but still, by far, the exception rather than the rule.

When optimizations are completed, including register allocation and
scheduling, it is time to pick up the debug instructions and emit
debug information out of them.  Conceptually, the debug instructions
represent points of assignment, at which a user variable ought to
evaluate to the annotated expression, maintained throughout
compilation.  However, when the value of a variable is live at more
than one location, it is important to note it, such that, if a
debugging session attempts to modify the variable, all copies are
modified.

The idea is to use some mechanism to determine equivalent expressions
throughout a function (say some variant of Global Value Numbering).
At debug instructions, we assert that the value of the named variable
is in the equivalence class represented by the expression.  As we scan
basic blocks forward and find that expressions in an equivalence class
are modified, we remove them from the equivalence class, and thus from
the list of available locations for the variable.  When such
expressions are further copied, we add them to equivalence classes.
At function calls and volatile asm statements, we remove
non-function-private memory slots from equivalence classes.  At
function calls, we also remove call-clobbered registers from
equivalence classes.  When no live expression remains in the
equivalence class that represents a variable, it is understood that
its value is no longer available.  At basic block confluences, we
combine information from the end states of the incoming blocks and the
debug statements added as a side effect of PHI nodes.

The end result is accurate debug information.  Also, except for
transformations that require special handling to update debug
annotations properly, debug information should come out as complete as
possible.


== Testability

Since debug annotations are added early, and, in most cases,
maintained up-to-date by the same code that optimizers use to maintain
executable code up-to-date, debug annotations are likely to remain
accurate throughout compilation.

The risk of this approach is that the annotations get in the way of
optimizations, thus causing executable code to vary depending on
whether or not debug information is to be generated.  The risk of
varying code could be removed at the expense of generating and
maintaining debug annotations throughout compilation and just throwing
them away at the end.  This is undesirable, for it would slow down
compilation without debug information and waste memory while at that.

Therefore, we've built testing mechanisms into the compiler to detect
cases in which the presence of debug annotations would cause code
changes.

The bootstrap-debug Makefile target, by default, compiles the second
bootstrap stage without debug information, and the third bootstrap
stage with it, and then compares all object files after stripping
them, a process that discards all debug information.

Furthermore, bootstrap4-debug, after bootstrap-debug and
prepare-bootstrap4-debug-lib-g0, rebuilds all target libraries without
debug information, and compares them with the stage3 target libraries,
built with debug information.

At the time of this writing, both tests pass on platforms
x86_64-linux-gnu and i686-linux-gnu, and ppc64-linux-gnu and
ia64-linux-gnu are getting close.

Additional testing mechanisms should be built in, to exercise a wider
range of internal GCC behaviors and extensions, for example, by
comparing the compiler output with and without debug information while
compiling all of its testsuite.

Even if testing mechanisms fail to catch an error, the generation of
debug annotations is controlled by a command-line option, such that
any code changes caused by it can be easily avoided, at the expense of
the quality of the debug information.

Testing for accuracy and completeness of debug information can be best
accomplished using a debugging environment.  For example, writing
programs of increasing complexity, adding functional-call or asm probe
points to stabilize the internal execution state, and then examining
the state of the program at these probe points in a debugger, shall
let us know how accurate and how complete variable location
information is.

Measuring accuracy is easy: if you ask for the value of a variable,
and get a value other than the expected, there's a bug in the
compiler.  If you get "unavailable", this can still be regarded as
accurate, for locations are always optional.  However, it might be
incomplete.  Telling whether the variable was indeed optimized away,
or whether the value is available or computable but the information is
missing, is a harder problem, but it's not part of the accuracy test,
but rather of the completeness test.

The completeness score for an unoptimized program might very often be
unachievable for optimized programs, not because the compiler is doing
a poor job at maintaining debug information, but rather because the
compiler is doing a good job at optimizing it, to the point that it is
no longer possible to determine the value of the inspected variable.


== Concerns

=== Memory consumption

Keeping more information around requires more memory; information
theory tells us that there's only so much information you can fit in a
bit.

In order to generate correct debug information, more information needs
to be retained throughout compilation.  The only way to arrange for
debug information to not require any additional memory is to waste
memory when not generating debug information.  But this is
undesirable.

Therefore, the better debug information we want, the more memory
overhead we're going to have to tolerate.

Of course at times we can trade memory for efficiency, using more
computationally expensive representations that are more compact.

At other times, we may trade memory for maintainability.  For example,
instead of emitting annotations as soon as we enter SSA mode, we could
emit them on demand, i.e., whenever we deleted, moved or significantly
modified an SSA assignment for which we would have emitted a debug
annotation.  Additional memory would be needed to mark assignments
that should have gained annotations but haven't, and care must be
taken to make sure that transformations aren't made without leaving a
correct debug statement in place.  It is not clear that this would
save significant memory, for a large fraction of relevant assignments
are modified or moved anyway, so it might very well be a
maintainability loss and a performance penalty for no measurable
memory gains.

Worst case, we may trade memory for debug information quality: if
memory use of this scheme is too high for some scenario, one can
disable debug information annotations through a command line option,
or disable debug information altogether.


=== Intrusiveness

Given that nearly all compiler transformations would require
reflection in debug information, any solution that doesn't take
advantage of this fact is bound to require changes all over the place.

Perhaps not so much for Tree-SSA passes, that are relatively
well-behaved and use a narrow API to make transformations, but very
clearly so for RTL passes, that very often modify instructions in
place, and at times even reuse locations assigned to user variables as
temporaries.

Even when we do use the strength of optimizers to maintain debug
information up to date, there are exceptions in which detailed
knowledge about the transformation taking place enables us to adjust
the annotations properly, if possible, or to discard location
information for the variable otherwise.

It is just not possible to hope that information can be maintained
accurate throughout compilation without any effort from optimizers, or
even through a trivial API for a debug information generator.  A
number of the exceptions that require detailed knowledge about the
ongoing transformation would be indistinguishable from other common
transformations that would have very different effects on debug
information.  At this point, any expectations of lower intrusiveness
by use of such an API vanish.

By letting optimizers do their jobs on debug annotations, and handling
exceptions only at the few locations where they are needed, trivially
in most such cases, we keep intrusiveness at a minimum.

Of course we could get even lower intrusiveness by accepting errors in
debug information, or accepting to generate different code depending
on debug information command-line options.  But these options
shouldn't be considered seriously.


=== Complexity

The annotations are conceptually trivial and they can be immediately
handled by optimizers.  It is hard to imagine a simpler design that
would still enable us to get right cases such as those in the examples
below.

Worrying about the representation of debug annotations as statements
or instructions, rather than notes, is missing the fact that, most of
the time, we do want them to be updated just like statements and
instructions.

Worrying about the representation of debug annotations in-line, rather
than an on-the-side representation, is a valid concern, but it's
addressed by the testability of the design, and the in-line
representation is highly advantageous, not only for using optimizers
to keep debug information accurate, but also for doing away with the
need for yet another internal representation and all the efforts into
maintaining it accurate.


=== Optimizations

Correct and more complete debugging information isn't supposed to
disable optimizations.  Keep in mind that enabling debug information
isn't supposed to modify the executable code in any way whatsoever.

The goal is to ensure that whatever debug information the compiler
generates actually matches the executable code, and that it is as
complete as viable.

The goal is not to disable optimizations so as to preserve variables
or code, such that it can be represented in debug information and
provide for a debugging experience more like that of code that is not
optimized.

If debug information disables any optimization, that's a bug that
needs fixing.

Now, while testing this design, a number of opportunities for
optimization that GCC missed were detected and fixed, others were
merely detected, and at least one optimization shortcoming kept in
place in order to get better debug information could be removed, for
the new debug information infrastructure enables the optimization to
be applied in its fullest extent.


== Examples

It is desirable to be able to represent constants and other
optimized-away values, rather than stating variables have values they
can no longer have:

int
x1 (int x)
{
  int i;

  i = 2;
  f(i);
  i = x;
  h();
  i = 7;
  g(i);
}

Even if variable i is completely optimized away, a debugger can still
print the correct values for i if we keep annotations such as:

  (debug (var_location i (const_int 2)))
  (set (reg arg0) (const_int 2))
  (call (mem (symbol_ref f)))
  (debug (var_location i unknown))
  (call (mem (symbol_ref h)))
  (debug (var_location i (const_int 7)))
  (set (reg arg0) (const_int 7))
  (call (mem (symbol_ref g)))

In this case, before the call to h, not only the assignment to i was
dead, but also the value of the incoming argument x had already been
clobbered.  If i had been assigned to another constant instead, debug
information could easily represent this.

Another example that covers PHI nodes and conditionals:

int
x2 (int x, int y, int z)
{
  int c = z;
  whatever0(c);
  c = x;
  whatever1();
  if (some_condition)
    {
      whatever2();
      c = y;
      whatever3();
    }
  whatever4(c);
}

With SSA infrastructure, this program can be optimized to:

int
x2 (int x, int y, int z)
{
  int c;
  # bb 1
  whatever0(z_0(D));
  whatever1();
  if (some_condition)
    {
      # bb 2
      whatever2();
      whatever3();
    }
  # bb 3
  # c_1 = PHI <x_2(D)(1), y_3(D)(2)>;
  whatever4(c_1);
}

Note how, without debug annotations, c is only initialized just before
the call to whatever4.  At all other points, the value of c would be
unavailable to the debugger, possibly even wrong.

If we were to annotate the SSA definitions forward-propagated into c
versions as applying to c, we'd end up with all of x_2, y_3 and z_0
applied to c throughout the entire function, in the absence of
additional markers.

Now, with the annotations proposed in this paper, what is initially:

int
x2 (int x, int y, int z)
{
  int c;
  # bb 1
  c_4 = z_0(D);
  # DEBUG c c_4
  whatever0(c_4);
  c_5 = x_2(D);
  # DEBUG c c_5
  whatever1();
  if (some_condition)
    {
      # bb 2
      whatever2();
      c_6 = y_3(D);
      # DEBUG c c_6
      whatever3();
    }
  
  # bb 3
  # c_1 = PHI <c_5(D)(1), c_6(D)(2)>
  # DEBUG c c_1
  whatever4(c_1);
}

is optimized into:

int
x2 (int x, int y, int z)
{
  int c;
  # bb 1
  # DEBUG c z_0(D)
  whatever0(z_0(D));
  # DEBUG c x_2(D)
  whatever1();
  if (some_condition)
    {
      # bb 2
      whatever2();
      # DEBUG y_3(D)
      whatever3();
    }
  # bb 3
  # c_1 = PHI <x_2(D)(1), y_3(D)(2)>;
  # DEBUG c c_1
  whatever4(c_1);
}

and then, at every one of the inspection points, we get the correct
value for variable c.


== Conclusion

This design enables a compiler to emit variable location debug
information that complies with the DWARF version 3 standard, and that
is likely to be as complete as theoretically possible, with an
implementation that is conceptually simple, relatively easy to
introduce, trivial to test and easy to maintain in the long run.  Not
wasting memory or CPU cycles during compilation without debug
information are welcome bonuses.

-- 
Alexandre Oliva         http://www.lsd.ic.unicamp.br/~oliva/
FSF Latin America Board Member         http://www.fsfla.org/
Red Hat Compiler Engineer   aoliva@{redhat.com, gcc.gnu.org}
Free Software Evangelist  oliva@{lsd.ic.unicamp.br, gnu.org}

Index Nav: [Date Index] [Subject Index] [Author Index] [Thread Index]
Message Nav: [Date Prev] [Date Next] [Thread Prev] [Thread Next]