The crash that this diff fixes occurred when giving a command such as
"print Var^1" to mdb, where the first argument of Var is a direct arg.
runtime/mercury_ml_expand_body.h:
When deconstructing a term with a direct arg, return NULL
as the value of expand_info->chosen_arg_word_sized_ptr.
The crash occurred when we returned a non-null pointer,
which violated the expectations of trace/mercury_trace_vars.c
and its callers. (Not surprising, since those that function and
its callers were written long before the direct_arg optimization
was added to the system.)
runtime/mercury_deconstruct.h:
Document the rationale behind the above changes. (The contents of
mercury_ml_expand_body.h are #included in mercury_deconstruct.c.)
trace/mercury_trace_vars.c:
Add the debugging code I used to track down this issue, in disabled form.
Fix missing copyright year.
trace/mercury_trace_browse.c:
Delete obsolete comment.
Fix missing copyright years.
tests/debugger/direct_arg_test.{m,inp,exp}:
A test case for this bug.
tests/debugger/Mmakefile:
Enable the new test case.
compiler/hlds_out_type_table.m:
When dumping out the data constructors in the type table,
if a constructor has names for some of its fields,
put the name and the type of each field on different lines.
In the original test case for this bug, of which direct_arg_test.m
is an extreme simplification, pretty much every line overflows
without this.
Also, factor out some duplicated code, and replace bools with values
of a bespoke type.
runtime/Mmakefile:
Some time ago, autoconf seems to have changed the text it puts
into the files it generates, including mercury_conf.h.
This broke the old code we had in this file that looked for
the old text. Fix this by updating the pattern we look for.
Document updates to autoconf as possible reasons for any future
reoccurrence of this bug.
Document the reason why we "standardize" the autoconfigured contents
of mercury_conf.h.
Fix programming style.
configure.ac:
The changes in this file are only stylistic; they made it easier
to track down the above bug.
Put a comment about "order matters" *before* the lists whose order
matters. Define the first of two lists being concatenated first
(since order matters not just for code correctness, but also for
reading comprehension :-().
Fix misleading indentation.
Add an XXX about some (seemingly) unneeded code.
configure.ac:
runtime/mercury_conf.h.in:
runtime/mercury_goto.h:
Delete workaround for systems where the assembler did not
support the .type directive.
runtime/mercury_deep_rec_depth_body.h:
Instead assuming that the code #including this file has a variable
named CSD in scope, get the caller to pass that variable via a macro.
library/profiling_builtin.m:
Pass that variable using that macro. Since the macro definition is
not a comment, it won't get a singleton warning. Mark the CSD argument
of the relevant predicates as not-a-singleton.
... when targeting C.
library/string.m:
On 27 Oct 2021, Julien sped up one of these operations
by avoiding the use of sprintf. Apply the same technique,
suitably generalized, to all other similar operations.
Use macros to reduce the amount of code duplication needed.
runtime/mercury_conf.h.in:
Fix an old bug: spell MR_MERCURY_IS_{32,64}_BITS correctly,
to match the name that is (conditionally) defined by configure.
configure.ac:
Delete references to MERCURY_IS_{32,64}_BITS, which turn out
to be totally unused.
runtime/mercury_grade.h:
Rename the grade modifier, and the C macro that represents it.
compiler/options.m:
Rename the --c-debug-grade option to --target-debug-grade.
compiler/compute_grade.m:
Rename the grade modifier, and the option that represents it.
Restrict the .target_debut grade modifier to MLDS grades.
compiler/handle_options.m:
Implement --target-debug-grade by having it imply --target-debug.
compiler/compile_target_code.m:
compiler/link_target_code.m:
Pay attention to either --target-debug-grade (for purposes related
to the grade itself) and to --target-debug (for all other purposes).
scripts/canonical_grade.in:
scripts/canonical_grade.sh-subr:
scripts/final_grade_options.sh-subr:
scripts/init_grade_options.sh-subr:
scripts/parse_grade_options.sh-subr:
Parse target_debug grade modifiers and --target-debug-grade options
instead of c_debug grade modifiers and --c-debug-grade options.
Add (normally commented-out) infrastructure to make it easier
to debug changes.
Restrict the .target_debut grade modifier to MLDS grades.
scripts/mgnuc.in:
scripts/mgnuc_file_opts.sh-subr:
Rename some variables to clarify the distinction between the
--target-debug option (which, like -g, enabled debugging of only one file)
and the --target-debug-grade option (which enables it for the whole
program).
configure.ac:
Make it easier to debug grade-related changes by recording
both autoconfigured and user-supplied grades that the rejected by
the canonical_grade script.
Conform to the changes above.
README.sanitizers:
doc/user_guide.texi:
grade_lib/grade_spec.m:
grade_lib/grade_string.m:
scripts/ml.in:
tests/warnings/help_text.err_exp:
tools/lmc.in:
tools/test_mercury:
Conform to the changes above.
scripts/Mmake.vars.in:
Add some XXXs about style.
compiler/print_help.m:
Switch to a coding scheme that allows us to add a prefix to the
current optdb struct's help text *after* we generate that help text.
Use this capability to mark private options as such.
Do not wrap argument "names" that are not actually names.
compiler/options.m:
Fix mismatches between option's default values and their help structures
(such as value being maybe_string, implying an argument, but the
help structure lacking an argument.
Add some XXXs.
compiler/globals.m:
Add some XXXs.
compiler/use_local_vars.m:
Fix indentation.
compiler/handle_options.m:
runtime/mercury_region.h:
Fix comments.
runtime/mercury_memory.c:
Document why the implementation of the MR_GC_malloc family of functions
does not really affect performance.
runtime/mercury_conf_param.h:
Fix a comment.
runtime/mercury_memory.c:
In the functions that call GC_MALLOC, GC_MALLOC_ATOMIC,
GC_MALLOC_UNCOLLECTABLE or GC_REALLOC, do not check whether
the pointers they return are NULL, since the oom (out-of-memory)
handler we specify for Boehm gc will have already checked for NULL,
and it will abort the program instead of returning NULL.
runtime/mercury_memory.h:
Make some comments more readable.
runtime/mercury_wrapper.c:
Document mercury_memory.c's reliance on the oom handler.
Check for MR_BOEHM_GC explicitly, not via an inference.
tools/generate_errno_name:
Use a secondary switch to handle errno names that may be defined to
the same value as other names.
runtime/mercury_errno_name.c:
Regenerate this file.
On AIX, both ENOTEMPTY and EEXIST are defined to the same value,
which caused the code generated for MR_errno_name() to fail to compile.
tools/generate_errno_name:
Handle the case that ENOTEMPTY == EEXIST in the generated function.
Likewise for other errno name aliases.
runtime/mercury_errno_name.c:
Regenerate this file.
The size argument to strncat() gives the number of characters to copy
from the input string, *not* the size of the destination buffer.
runtime/mercury_deep_profiling.c:
Add a safe string concatenation function, MR_safe_strcat.
Use MR_safe_strcat instead of strncat.
runtime/mercury_memory.h:
State explicitly that the C functions in this module all abort
when they cannot allocate the memory they were asked to allocate.
Put the documentation of some functions in the same order as their
definitions.
compiler/trace_gen.m:
Given that the standard trace slots are (by design) always the same,
generate code that does not redundantly specify them.
runtime/mercury_trace_base.h:
Add the three macros that trace_gen.m can now generate references to.
Fix a problem that arises in the deep profiler if the program being profiled
was using both intermodule optimization and inlining. The issue was that
even though runtime/mercury_deep_profiling.c had access, for every call site
to the full context of that call site, containing both the file name and
the line number, it wrote out *only* the line number. The deep profiler
then got the file name from the file name stored in the proc_static structure
of the procedure containing the call site.
This works very close to 100% of the time, because
- user-written programs just about never use ":- pragma source_file", and
- in the absence of such pragmas, all the goals in a procedure will be
from the same file as the procedure's context.
However, if
- a call site calls a procedure in another module,
- the compiler has access to the code of that procedure from a .opt file, and
- the compiler decides to inline that call,
then the call, whose context is in the original source file, will be replaced
by the code of the procedure from the .opt file, whose context will NOT have
the same file name. Any description of this call site will list
- the code from the .opt file (such as the callee's callee),
- the file name from the original source file, and
- the line number from the .opt file.
This mismatch is very confusing, which is why this diff fixes it.
runtime/mercury_deep_profiling.c:
Fix this by writing out the file name part, as well as the line number
part, of each call site. The space impact is not as big as one might
expect, because compiler/deep_profiling.m already had an optimization
that set the filename part of each call site context to the empty string
if it was identical to the context of the procedure. Therefore for all
call sites that do NOT exhibit the bug that this diff fixes, the space
cost is only a single NULL character in the profiling data file.
Since this IS a change in the file format, bump the format version number
from 8 to 9.
deep_profiler/profile.m:
deep_profiler/read_profile.m:
Handle reading in both version 8 and version 9 profiling data files.
deep_profiler/create_report.m:
When creating descriptions of call sites, use the call site's filename
if it is not the empty string; if it is the empty string, then use
the containing procedure's file name, as we have done all along.
deep_profiler/display_report.m:
deep_profiler/dump.m:
deep_profiler/report.m:
Dump the new field in commands intended only for implementors.
deep_profiler/startup.m:
Conform to the changes above.
When this new runtime option is specified, the runtime system will use
Deep.{data,procrep} as the names of the files it writes out.
runtime/mercury_engine.h:
Add a flag to the engine that records whether this option has been
specified or not.
runtime/mercury_wrapper.c:
Set the flag if/when we see the --deep-std-name option.
runtime/mercury_deep_profiling.c:
If the new flag is set, use Deep.{data,procrep} as filenames.
doc/user_guide.texi:
Document the new option.
tools/bootcheck:
Specify the new option for bootchecks.
tests/debugger/Mmakefile:
tests/declarative_debugger/Mmakefile:
tests/hard_coded/Mmakefile:
tests/par_conj/Mmakefile:
tests/stm/Mmakefile:
When specifying a value of MERCURY_OPTIONS that overrides the value
set by tools/bootcheck, include --deep-std-name in that value.
runtime/mercury_deep_profiling.c:
Replace the "Deep" in "Deep.data" and "Deep.procrep" with
- the name of the executable program, and
- the date and time of the profiling run,
yielding names such as
mercury_compile_on_2024-10-05_at_08-43-34.data
mercury_compile_on_2024-10-05_at_08-43-34.procrep
doc/user_guide.texi:
Document the new filenames.
NEWS.md:
Announce the change.
like this:
mdb> stack
0 pred mutrec.q1/3-0 (det) (mutrec.m:133)
1 ┌ 2* pred mutrec.p2/3-0 (det) (mutrec.m:106 and others)
3 │ pred mutrec.p3/3-0 (det) (mutrec.m:122)
4 │ 2* pred mutrec.p2/3-0 (det) (mutrec.m:102 and others)
6 └ pred mutrec.p3/3-0 (det) (mutrec.m:122)
7 2* pred mutrec.p1/3-0 (det) (mutrec.m:82 and others)
9 pred mutrec.test/2-0 (det) (mutrec.m:42)
10 pred mutrec.main/2-0 (det) (mutrec.m:33)
This implements a suggestion by Peter Wang from *April 2012*.
runtime/mercury_stack_trace.h:
Add a new field to the data structure that represents each line
in a stack trace. This field represents the string, if any,
that implements the boxes around the calls in the clique.
This will be two characters, a box character and a space,
e.g. for the call at level 6, it will be "└ ".
runtime/mercury_stack_trace.c:
Use the new capability to add boxes to calls in cliques,
and equal-length padding to the calls that are not in cliques
but which share stack traces with calls that are.
Initialize some variables closer to their first use.
Group some related statements together. Add some comments.
Delete some already-acted-upon TODOs.
tests/debugger/mutrec.exp:
tests/debugger/mutrec_higher_order.exp:
Expect the updated clique markers.
NEWS.md:
doc/user_guide.texi:
Document the change.
Don't call GC_disable() before GC_INIT(), which we did to avoid GC
during early runtime initialisation. This once had a small benefit,
but will not work after we upgrade to Boehm GC with commit 52a538ff
"Fix handling of GC_gc_no counter wrap in GC_clear_stack".
runtime/mercury_wrapper.c:
As above.
runtime/mercury_atomic_ops.h:
Use compiler intrinsic for MR_ATOMIC_ADD_AND_FETCH_WORD_BODY
when compiling with clang, instead of falling back to the
compare and swap implementation.
Use compiler intrinsic for MR_ATOMIC_SUB_INT_BODY
when compiling with clang but not targeting x86/x86-64.
In those cases, MR_atomic_sub_int() was not being defined,
and causing a link error.
Use compiler intrinsic for MR_ATOMIC_DEC_AND_IS_ZERO_WORD_BODY
when compiling with clang but not targeting x86/x86-64.
runtime/mercury_deep_profiling.c:
runtime/mercury_memory_zones.c:
runtime/mercury_trace_base.c:
As above. clang -Wstrict-prototypes warns about functions with a
"()" argument list, even if there was a prototype declared with
"(void)".
runtime/mercury_trace_base.c:
Add code to print escape characters as \e.
Reorder the existing escape sequences.
Fix a bug: a CR was printed as \b, not as \r.
library/mercury_term_lexer.m:
Add the two affected functions in mercury_trace_base.c
to the list of places that handle escape sequences.
library/mercury_term_lexer.m:
Convert any occurrences of the \e escape sequence to the escape character.
List all the places in the library, compiler and runtime that also handle
escape sequences, some of which handle all these sequences,
and some of which handle only subsets.
Sort the letters in recognized escape sequences.
compiler/parse_tree_out_pragma.m:
library/rtti_implementation.m:
library/term_io.m:
Add comments to all the other places that handle escape sequences
that direct readers to mercury_term_lexer.m as containing the master list
of such sequences.
Add commented-out code that, after stage 1 has been installed,
stage 2 should enable.
runtime/mercury_ml_expand_body.h:
runtime/mercury_string.c:
Turn escape characters back into their escape sequence form
for characters and strings.
tests/valid_seq/char_escape_opt_helper_1.m:
Test whether the compiler accepts \e as an escape sequence.
compiler/options.m:
Add a mechanism for detecting the presence of this diff in the
installed compiler.
... and turn them back on.
compiler/c_util.m:
Provide a version of output_quoted_string_c which accepts strings
that may not be well formed, and prints out any code_units that
do not belong to a well-formed UTF-8 code point using an octal
escape sequence.
Ask for some predicates with trivial bodies to be inlined.
compiler/llds_out_data.m:
Use the new predicate in c_util to output string constants.
This fixes the bug that could cause trie string switches to fail
in the presence of strings containing multi-byte code points.
In such cases, given the string constants that trie switches can generate,
which contain only part of such a multi-byte code unit sequence,
llds_out_data.m used to write out a C string constant that contained
the unicode replacement character instead of those code units.
compiler/string_switch.m:
If a trie node has four or more code units that can lead to matches, then
use binary search instead of linear search to find out what to do next.
To make this possible, separate the action of generating such search code
from the action of finding out what to do for each possible code unit
value.
Generate comments that can be helpful in tracking down bugs such as
the one described above.
compiler/switch_gen.m:
Allow trie string switches once again.
runtime/mercury_ml_expand_body.h:
Add an XXX for new behavior that I found the need for while
debugging this change.
tests/hard_coded/string_switch*.{m,exp}:
Add some new alternatives to each switch to create nodes
at which the LLDS code generator will now use binary search.
Add a query string to test one of these alternatives.
(All four test cases are identical, apart from their module names.)
Expect the changes for the existing keys caused by the new switch arms,
as well as the extra outputs for the new query string.
runtime/mercury_string.h:
Expand the comments on the functions that work with UTF-8 strings
by describing their behavior in the presence of non-well-formed
code unit sequences in more detail.
runtime/mercury_string.c:
Improve the style of the code, and the text of some comments.
For now, the implementation covers only non-lookup switches.
compiler/builtin_ops.m:
Generalize the existing offset_str_eq binary op by adding an optional
size parameter, which, if present, restricts the equality test to look at
the given number of code units at most.
compiler/llds_out_data.m:
compiler/mlds_to_c_data.m:
Generalize the output of binop rvals whose operation is offset_str_eq.
In llds_out_data.m, fix a bug in the original code. (This bug did not
lead to problems because before this diff, we never generated this op.)
compiler/string_switch_util.m:
Add a predicate that recognizes when a trie node that is NOT a leaf
nevertheless represents the top of a stick, which means that it has
only one possible next code unit, which itself may have only one
possible next code unit, and so on, until we reach a node that *does*
have two or more next code units. (One of those may be the code unit
of the string-ending NULL character.)
compiler/ml_string_switch.m:
Use the new predicate in string_switch_util.m to generate better code
for sticks. Instead of comparing each character in the stick individually
against the relevant code unit of the string being switched on, compare
them all at once using the new binary op.
compiler/ml_switch_gen.m:
Insist on both the host machine and the target machine
using the C backend.
compiler/string_switch.m:
Implement non-lookup trie switches. The code follows the approach used
in ml_string_switch.m as much as possible, but there are plenty of
differences caused by targeting the LLDS.
Rename some predicates to specify which switch implementation method
they belong to.
Write a comment just once, and refer to it from elsewhere instead of
duplicating it at each reference site.
compiler/switch_gen.m:
Enable the use of trie switches when the option values call for it,
and when the switch is not a lookup switch.
compiler/cse_detection.m:
Do not flood the output of mmc -V with messages that have nothing to do
with the module being compiled.
compiler/options.m:
Add a way to specify --no-allow-inlining on the command line.
This can help debug code generator changes like this, by disallowing
a transform that can modify the Mercury code whose compilation process
you are trying to debug. (The documentation of the --inlining option
implies that --no-inlining should do the same job, but it does not.)
The option is not documented for users.
compiler/string_encoding.m:
Provide a version of from_code_unit_list_in_encoding that allows
non-well-formed code unit sequences as input, and provide det versions
of both versions. This is for use by both string_switch.m and
ml_string_switch.m.
compiler/hlds_goal.m:
Document the properties of case_ids.
compiler/llds.m:
Document the possibility that string constants are not well formed.
compiler/bytecode.m:
compiler/code_util.m:
compiler/mlds_dump.m:
compiler/ml_global_data.m:
compiler/mlds_to_cs_data.m:
compiler/mlds_to_java_data.m:
compiler/opt_debug.m:
Conform to the changes above.
library/string.m:
Replace the non-exported test predicate internal_encoding_is_utf8 with
an exported function that returns an enum specifying the string encoding.
NEWS.md:
Announce the new function.
runtime/mercury_string.h:
Add the C macro that implements the new form of the offset_str_eq
binary op.
tests/hard_coded/string_switch4.{m,exp}:
We have long had three copies of the exact same code, in string_switch.m,
string_switch2.m and string_switch3.m, which were compiled with
- no smart switch implementation
- smart switch implementation forced to use the hash table method
- smart switch implementation forced to use binary search method
Add this new copy, which is compiled with
- smart switch implementation forced to use the new trie method
tests/hard_coded/Mmakefile:
Add the new test case.
tests/hard_coded/Mercury.options:
Update the options of the test cases, and specify them for the new.
tests/hard_coded/string_switch.m:
tests/hard_coded/string_switch2.m:
tests/hard_coded/string_switch3.m:
Update the top-of-module comment block to be identical in all four copies
of this module.
library/char.m:
Export a new predicate, char_int_is_surrogate, the duplicates
the job of the MR_is_surrogate macro in the runtime, for use by
the new check in string.m. Add a comment about the code duplication.
The new predicate is not documented for users.
runtime/mercury_string.h:
Add a comment about the code duplication.
library/string.m:
Use the new predicate in char.m to check for surrogates when converting
a code unit list to an utf8 string.
library/benchmarking.m:
Add this new predicate.
runtime/mercury_report_stats.[ch]:
Implement this predicate for C. The implementation is next to
the implementation of report_full_memory_stats.
NEWS.md:
Announce the new predicate.
compiler/mercury_compile_main.m:
Use the new predicate to avoid nuisance output from
report_full_memory_stats in grades where those stats are not available.
Make the Mercury implementation of copying attempt to preserve file permissions
if we are on system that supports stat() and chmod().
compiler/copy_util.m:
As above.
configure.ac:
runtime/mercury_conf.h.in:
Check for the presence of the chmod() function.
We had configure choose the format string length modifier for
MR_int_least64_t based on whether the type is an alias for "int",
"long", "long long" or "__int64". That does not work for some versions
of MinGW-w64, which warn about the "ll" length modifier even though
MR_int_least64_t is an alias for "long long". The reason is that
we would be calling the MSVC runtime *printf functions,
which require the "I64" length modifier instead of "ll".
The only place MR_INT_LEAST64_LENGTH_MODIFIER is used in the Mercury
system is in the hidden function float64_bits_string (which is also
no longer used by the Mercury compiler after the removal of the hl
grades). We can replace that use with PRIdLEAST64 from inttypes.h.
Users are unlikely to be using MR_INT_LEAST64_LENGTH_MODIFIER,
so it should be safe to remove it.
configure.ac:
runtime/mercury_conf.h.in:
Don't define MR_INT_LEAST64_LENGTH_MODIFIER.
library/float.m:
Replace use of MR_INT_LEAST64_LENGTH_MODIFIER with PRIdLEAST64
in float64_bits_string.
Mark all float32_bits_string and float64_bits_string foreign procs
as 'may_not_export_body'. There is no need to opt-export procedures
that will rarely be used, if ever.
runtime/mercury_types.h:
Update comment.
The configure script currently contains a check for the presence of the C99
header stdint.h. This is redundant since (1) *all* C compilers that are
currently capable of compiling Mercury support C99 and (2) most #includes of
stdint.h in the Mercury system are not, and have never been, protected by the
MR_HAVE_STDINT_H anyway.
configure.ac:
runtime/mercury_conf.h.in:
Do not check for stdint.h.
runtime/mercury_std.h:
Delete the one remaining use of the MR_HAVE_STDINT_H macro.
