Diagnostics should be worded in terms of the user’s source code, and the source language, rather than GCC’s own implementation details.
A good diagnostic is actionable: it should assist the user in taking action.
Consider what an end user will want to do when encountering a diagnostic.
Given an error, an end user will think: “How do I fix this?”
Given a warning, an end user will think:
A good diagnostic provides pertinent information to allow the user to easily answer the above questions.
A perfect compiler would issue a warning on every aspect of the user’s source code that ought to be fixed, and issue no other warnings. Naturally, this ideal is impossible to achieve.
Warnings should have a good signal-to-noise ratio: we should have few false positives (falsely issuing a warning when no warning is warranted) and few false negatives (failing to issue a warning when one is justified).
Note that a false positive can mean, in practice, a warning that the user doesn’t agree with. Ideally a diagnostic should contain enough information to allow the user to make an informed choice about whether they should care (and how to fix it), but a balance must be drawn against overloading the user with irrelevant data.
GCC is typically used in two different ways:
Keep both of these styles of usage in mind when implementing diagnostics.
Provide the user with details that allow them to identify what the problem is. For example, the vaguely-worded message:
demo.c:1:1: warning: 'noinline' attribute ignored [-Wattributes] 1 | int foo __attribute__((noinline)); | ^~~
doesn’t tell the user why the attribute was ignored, or what kind of
entity the compiler thought the attribute was being applied to (the
source location for the diagnostic is also poor;
see discussion of input_location
).
A better message would be:
demo.c:1:24: warning: attribute 'noinline' on variable 'foo' was ignored [-Wattributes] 1 | int foo __attribute__((noinline)); | ~~~ ~~~~~~~~~~~~~~~^~~~~~~~~ demo.c:1:24: note: attribute 'noinline' is only applicable to functions
which spells out the missing information (and fixes the location information, as discussed below).
The above example uses a note to avoid a combinatorial explosion of possible messages.
It’s worth testing a new warning on many instances of real-world code, written by different people, and seeing what it complains about, and what it doesn’t complain about.
This may suggest heuristics that silence common false positives.
It may also suggest ways to improve the precision of the message.
Many diagnostics relate to a mismatch between two different places in the user’s source code. Examples include:
In each case, the diagnostic should indicate both pertinent locations (so that the user can easily see the problem and how to fix it).
The standard way to do this is with a note (via inform
). For
example:
auto_diagnostic_group d; if (warning_at (loc, OPT_Wduplicated_cond, "duplicated %<if%> condition")) inform (EXPR_LOCATION (t), "previously used here");
which leads to:
demo.c: In function 'test': demo.c:5:17: warning: duplicated 'if' condition [-Wduplicated-cond] 5 | else if (flag > 3) | ~~~~~^~~ demo.c:3:12: note: previously used here 3 | if (flag > 3) | ~~~~~^~~
The inform
call should be guarded by the return value from the
warning_at
call so that the note isn’t emitted when the warning
is suppressed.
For cases involving punctuation where the locations might be near
each other, they can be conditionally consolidated via
gcc_rich_location::add_location_if_nearby
:
auto_diagnostic_group d; gcc_rich_location richloc (primary_loc); bool added secondary = richloc.add_location_if_nearby (secondary_loc); error_at (&richloc, "main message"); if (!added secondary) inform (secondary_loc, "message for secondary");
This will emit either one diagnostic with two locations:
demo.c:42:10: error: main message (foo) ~ ^
or two diagnostics:
demo.c:42:4: error: main message foo) ^ demo.c:40:2: note: message for secondary ( ^
GCC’s location_t
type can support both ordinary locations,
and locations relating to a macro expansion.
As of GCC 6, ordinary locations changed from supporting just a point in the user’s source code to supporting three points: the caret location, plus a start and a finish:
a = foo && bar; ~~~~^~~~~~ | | | | | finish | caret start
Tokens coming out of libcpp have locations of the form caret == start
,
such as for foo
here:
a = foo && bar; ^~~ | | | finish caret == start
Compound expressions should be reported using the location of the expression as a whole, rather than just of one token within it.
For example, in -Wformat
, rather than underlining just the first
token of a bad argument:
printf("hello %i %s", (long)0, "world"); ~^ ~ %li
the whole of the expression should be underlined, so that the user can easily identify what is being referred to:
printf("hello %i %s", (long)0, "world"); ~^ ~~~~~~~ %li
Avoid using the input_location
global, and the diagnostic functions
that implicitly use it—use error_at
and warning_at
rather
than error
and warning
, and provide the most appropriate
location_t
value available at that phase of the compilation. It’s
possible to supply secondary location_t
values via
rich_location
.
For example, in the example of imprecise wording above, generating the
diagnostic using warning
:
// BAD: implicitly uses input_location
warning (OPT_Wattributes, "%qE attribute ignored", name);
leads to:
// BAD: uses input_location
demo.c:1:1: warning: 'noinline' attribute ignored [-Wattributes]
1 | int foo __attribute__((noinline));
| ^~~
which thus happened to use the location of the int
token, rather
than that of the attribute. Using warning_at
with the location of
the attribute, providing the location of the declaration in question
as a secondary location, and adding a note:
auto_diagnostic_group d; gcc_rich_location richloc (attrib_loc); richloc.add_range (decl_loc); if (warning_at (OPT_Wattributes, &richloc, "attribute %qE on variable %qE was ignored", name)) inform (attrib_loc, "attribute %qE is only applicable to functions");
would lead to:
// OK: use location of attribute, with a secondary location demo.c:1:24: warning: attribute 'noinline' on variable 'foo' was ignored [-Wattributes] 1 | int foo __attribute__((noinline)); | ~~~ ~~~~~~~~~~~~~~~^~~~~~~~~ demo.c:1:24: note: attribute 'noinline' is only applicable to functions
See the diagnostics section of the GCC coding conventions.
In the C++ front end, when comparing two types in a message, use ‘%H’ and ‘%I’ rather than ‘%T’, as this allows the diagnostics subsystem to highlight differences between template-based types. For example, rather than using ‘%qT’:
// BAD: a pair of %qT used in C++ front end for type comparison error_at (loc, "could not convert %qE from %qT to %qT", expr, TREE_TYPE (expr), type);
which could lead to:
error: could not convert 'map<int, double>()' from 'map<int,double>' to 'map<int,int>'
using ‘%H’ and ‘%I’ (via ‘%qH’ and ‘%qI’):
// OK: compare types in C++ front end via %qH and %qI error_at (loc, "could not convert %qE from %qH to %qI", expr, TREE_TYPE (expr), type);
allows the above output to be simplified to:
error: could not convert 'map<int, double>()' from 'map<[...],double>' to 'map<[...],int>'
where the double
and int
are colorized to highlight them.
Text should be quoted by either using the ‘q’ modifier in a directive such as ‘%qE’, or by enclosing the quoted text in a pair of ‘%<’ and ‘%>’ directives, and never by using explicit quote characters. The directives handle the appropriate quote characters for each language and apply the correct color or highlighting.
The following elements should be quoted in GCC diagnostics:
Other elements such as numbers that do not refer to numeric constants that appear in the source code should not be quoted. For example, in the message:
argument %d of %qE must be a pointer type
since the argument number does not refer to a numerical constant in the source code it should not be quoted.
As of GCC 15, the diagnostics subsystem has a concept of “highlight colors”. These should be used to consistently colorize both the text within diagnostic messages and underlined ranges of quoted source when highlighting mismatches, for all messages with a logically-related group of diagnostics.
See diagnostic-highlight-colors.h for symbolic names for color codes, covering e.g.
highlight_colors::expected
versus highlight_colors::actual
highlight_colors::lhs
versus highlight_colors::rhs
For example, given:
error: invalid operands to binary + (have 'S' {aka 'struct s'} and 'T' {aka 'struct t'}) return callee_4a () + callee_4b (); ~~~~~~~~~~~~ ^ ~~~~~~~~~~~~ | | | T {aka struct t} S {aka struct s}
highlight_colors::lhs
(which equates to the color name highlight-a
)
highlight_colors::rhs
(which equates to the color name highlight-b
)
Doing so ought to make it easier for the user to understand what the diagnostic is telling them.
When issuing followup note
diagnostics, all diagnostics within the
group should use a consistent scheme to highlight the mismatching elements,
so that color contrasts the differences. For example, given:
warning: format ‘%i’ expects argument of type ‘int’, but argument 2 has type ‘const char *’ [-Wformat=] 279 | printf("hello " INT_FMT " world", msg); | ^~~~~~~~ ~~~ | | | const char * note: format string is defined here 278 | #define INT_FMT "%i" | ~^ | | | int | %s
%i
and int
referring to the format string and the
expected type due to it should be colorized as highlight-a
both
in the diagnostics message and in the range quoted in the range
.
const char *
in the diagnostic message and in the quoted
range should be colorized as highlight-b
.
This can be implemented by using e.g.
highlight_colors::actual
and
highlight_colors::expected
when adding ranges to
rich_location
instances, and e.g. by using the %e
format
code for pretty_printer
to use a pp_element *
, and using
appropriate member functions of pp_element to add colorization.
See the terminology and markup section of the GCC coding conventions.
GCC’s diagnostic subsystem can emit fix-it hints: small suggested edits to the user’s source code.
They are printed by default underneath the code in question. They can also be viewed via -fdiagnostics-generate-patch and -fdiagnostics-parseable-fixits. With the latter, an IDE ought to be able to offer to automatically apply the suggested fix.
Fix-it hints contain code fragments, and thus they should not be marked for translation.
Fix-it hints can be added to a diagnostic by using a rich_location
rather than a location_t
- the fix-it hints are added to the
rich_location
using one of the various add_fixit
member
functions of rich_location
. They are documented with
rich_location
in libcpp/line-map.h.
It’s easiest to use the gcc_rich_location
subclass of
rich_location
found in gcc-rich-location.h, as this
implicitly supplies the line_table
variable.
For example:
if (const char *suggestion = hint.suggestion ()) { gcc_rich_location richloc (location); richloc.add_fixit_replace (suggestion); error_at (&richloc, "%qE does not name a type; did you mean %qs?", id, suggestion); }
which can lead to:
spellcheck-typenames.C:73:1: error: 'singed' does not name a type; did you mean 'signed'? 73 | singed char ch; | ^~~~~~ | signed
Non-trivial edits can be built up by adding multiple fix-it hints to one
rich_location
. It’s best to express the edits in terms of the
locations of individual tokens. Various handy functions for adding
fix-it hints for idiomatic C and C++ can be seen in
gcc-rich-location.h.
When implementing a fix-it hint, please verify that the suggested edit leads to fixed, compilable code. (Unfortunately, this currently must be done by hand using -fdiagnostics-generate-patch. It would be good to have an automated way of verifying that fix-it hints actually fix the code).
For example, a “gotcha” here is to forget to add a space when adding a
missing reserved word. Consider a C++ fix-it hint that adds
typename
in front of a template declaration. A naive way to
implement this might be:
gcc_rich_location richloc (loc); // BAD: insertion is missing a trailing space richloc.add_fixit_insert_before ("typename"); error_at (&richloc, "need %<typename%> before %<%T::%E%> because " "%qT is a dependent scope", parser->scope, id, parser->scope);
When applied to the code, this might lead to:
T::type x;
being “corrected” to:
typenameT::type x;
In this case, the correct thing to do is to add a trailing space after
typename
:
gcc_rich_location richloc (loc); // OK: note that here we have a trailing space richloc.add_fixit_insert_before ("typename "); error_at (&richloc, "need %<typename%> before %<%T::%E%> because " "%qT is a dependent scope", parser->scope, id, parser->scope);
leading to this corrected code:
typename T::type x;
It’s best to express deletion suggestions in terms of deletion fix-it hints, rather than replacement fix-it hints. For example, consider this:
auto_diagnostic_group d; gcc_rich_location richloc (location_of (retval)); tree name = DECL_NAME (arg); richloc.add_fixit_replace (IDENTIFIER_POINTER (name)); warning_at (&richloc, OPT_Wredundant_move, "redundant move in return statement");
which is intended to e.g. replace a std::move
with the underlying
value:
return std::move (retval); ~~~~~~~~~~^~~~~~~~ retval
where the change has been expressed as replacement, replacing with the name of the declaration. This works for simple cases, but consider this case:
#ifdef SOME_CONFIG_FLAG # define CONFIGURY_GLOBAL global_a #else # define CONFIGURY_GLOBAL global_b #endif int fn () { return std::move (CONFIGURY_GLOBAL /* some comment */); }
The above implementation erroneously strips out the macro and the comment in the fix-it hint:
return std::move (CONFIGURY_GLOBAL /* some comment */); ~~~~~~~~~~^~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~ global_a
and thus this resulting code:
return global_a;
It’s better to do deletions in terms of deletions; deleting the
std::move (
and the trailing close-paren, leading to
this:
return std::move (CONFIGURY_GLOBAL /* some comment */); ~~~~~~~~~~^~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~ CONFIGURY_GLOBAL /* some comment */
and thus this result:
return CONFIGURY_GLOBAL /* some comment */;
Unfortunately, the pertinent location_t
values are not always
available.
In the rare cases where you need to suggest more than one mutually
exclusive solution to a problem, this can be done by emitting
multiple notes and calling
rich_location::fixits_cannot_be_auto_applied
on each note’s
rich_location
. If this is called, then the fix-it hints in
the rich_location
will be printed, but will not be added to
generated patches.