[12/22] Add source-ranges for trees

This patch adds a way to associate source range information with tree
expressions and decls, for later use by diagnostics.

It's a poor implementation which is unacceptable on multiple grounds:
for starters, it adds a source_range (8 bytes) to struct tree_exp and
to struct tree_decl_minimal.  (It also doesn't bootstrap).

Adding the source_range fields above covers most expressions, but
doesn't help with references to decls and to constants.  Consider:

  int test (int foo)
  {
    return foo * 100;
           ^^^   ^^^
  }

I want gcc to retain information so that diagnostics can underline
the "foo" and "100" above, but all we have is a VAR_DECL and an
INTEGER_CST.  The former's location is in at the top of the
function, and the latter has no location.

Hence the patch adds a new SOURCE_RANGE tree code.  This
is a kind of unary operator or wrapper, which wraps things that don't
have a range field themselves (further bloating the IR).

They get thrown away during gimplification.

This works for simple cases, but isn't yet complete: there are plenty
of places where the frontends will fail if they see a SOURCE_RANGE.

So, as I said, it's a poor implementation, but the followup patches
needed some way to record source range information for trees.

Some alternate ideas for how this could be implementated:

(a) somehow compress the location and range into the 4 bytes taken up
    by a source_location: are we really using all 32-bits?  I suspect
    that in real-world code ranges, there's enough closeness between the
    location, range start and range finish that we can pack them into
    32 bits for most cases, with some kind of lookaside for those that
    don't fit.

(b) introduce some kind of "DECL_USAGE" or "DECL_REF" tree node, an
    expression that references a decl, segregating decls from expression
    trees, putting the location information into the DECL_USAGE node.

(c) only bother tracking the information if -fdiagnostics-show-caret
    is enabled (note, though that it's on by default; if we go down
    this path, maybe it's another thing for torture testing?).

(d) only track information temporarily (e.g. in c_expr, rather than in
    tree), discarding it as the tree is built, or perhaps special-casing
    some places where it's particular worth preserving e.g. the ranges
    of the arguments at a callsite, so the user can easily identify
    whatever "argument 3" means.

etc.  Ideas?

gcc/c-family/ChangeLog:
	* c-common.c (c_fully_fold_internal): Capture existing souce_range,
	and store it on the result.
	* c-pretty-print.c (c_pretty_printer::expression): Handle
	SOURCE_RANGE.

gcc/c/ChangeLog:
	* c-typeck.c (array_to_pointer_conversion): Handle SOURCE_RANGE,
	and preserve any source range information.
	(build_function_call_vec): Handle SOURCE_RANGE.
	(lvalue_p): Likewise.
	(c_finish_return): Likewise.

gcc/ChangeLog:
	* gimplify.c (gimplify_expr): Throw away SOURCE_RANGEs.
	* print-tree.c (print_node): Print any source range information.
	* tree-core.h (struct tree_exp): Add a "range" field.
	(struct tree_decl_minimal): Likewise.
	* tree.c (build1_stat): Initialize EXPR_LOCATION_RANGE (t).
	(build_decl_stat): Add overload taking a source_range.
	(set_source_range): New functions.
	* tree.def (SOURCE_RANGE): New tree code.
	* tree.h (CAN_HAVE_RANGE_P): New.
	(EXPR_LOCATION_RANGE): New.
	(EXPR_RANGE_OR_LOC): New.
	(EXPR_HAS_RANGE): New.
	(DECL_LOCATION_RANGE): New.
	(build_decl_stat): New overload.
	(set_source_range): New decls.
---
 gcc/c-family/c-common.c       | 10 +++++++++-
 gcc/c-family/c-pretty-print.c |  4 ++++
 gcc/c/c-typeck.c              | 23 ++++++++++++++++++++++-
 gcc/gimplify.c                |  4 ++++
 gcc/print-tree.c              | 21 +++++++++++++++++++++
 gcc/tree-core.h               |  2 ++
 gcc/tree.c                    | 42 ++++++++++++++++++++++++++++++++++++++----
 gcc/tree.def                  |  2 ++
 gcc/tree.h                    | 21 +++++++++++++++++++++
 9 files changed, 123 insertions(+), 6 deletions(-)

[12/22] Add source-ranges for trees

Commit Message

Patch