Fix github issue #89.

The current source-to-source debugger transformation cannot handle the
predicates introduced by higher-order specialization. There are likely similar
issues with other HLDS->HLDS transformations.

The fix (for now) is to disable most HLDS->HLDS transformations in .ssdebug
grades.

compiler/handle_options.m:
    Disable most HLDS->HLDS optimizations when the ss-trace level is
    shallow or deep.

    Add an XXX comment about a separate issue.

compiler/ssdebug.m:
    Add an XXX comment about predicates produced by higher-order
    specialization.

tests/WS_FLAGS.ws:
    Add a missing include.

tests/valid/Mmakefile:
tests/valid/Mercury.options:
tests/valid/gh89.m:
    Add the test case from github issue 89.
This commit is contained in:
Julien Fischer
2020-05-17 01:12:39 +10:00
parent 556c3d4dc7
commit 06c9ecbb6b
6 changed files with 90 additions and 0 deletions

View File

@@ -1635,8 +1635,32 @@ convert_options_to_globals(OptionTable0, OpMode, Target, GC_Method,
add_error(phase_options, TraceHLSpec, !Specs)
),
% Source-to-source debugging requires disabling many HLDS->HLDS
% optimizations. This is so that the trace being generated relates to the
% source code and also because the SSDB transformation cannot (yet) handle
% the specialised predicates introduced by many optimizations.
(
( SSTraceLevel = shallow
; SSTraceLevel = deep
),
globals.set_option(allow_inlining, bool(no), !Globals),
globals.set_option(optimize_unused_args, bool(no), !Globals),
globals.set_option(optimize_higher_order, bool(no), !Globals),
globals.set_option(type_specialization, bool(no), !Globals),
globals.set_option(user_guided_type_specialization, bool(no),
!Globals),
globals.set_option(deforestation, bool(no), !Globals),
globals.set_option(constraint_propagation, bool(no), !Globals),
globals.set_option(local_constraint_propagation, bool(no), !Globals),
globals.set_option(optimize_duplicate_calls, bool(no), !Globals),
globals.set_option(optimize_constructor_last_call, bool(no), !Globals)
;
SSTraceLevel = none
),
% The pthreads headers on some architectures (Solaris, Linux)
% don't work with -ansi.
% XXX we don't pass -ansi to the C compiler anymore.
option_implies(parallel, ansi_c, bool(no), !Globals),
option_neg_implies(inline_builtins, constant_propagation, bool(no),

View File

@@ -1035,6 +1035,8 @@ ssdebug_process_proc_erroneous(SSTraceLevel, PredId, ProcId,
list(prog_var)::out, list(prog_var)::out, list(mer_mode)::out) is det.
get_stripped_headvars(PredInfo, ProcInfo, FullHeadVars, HeadVars, ArgModes) :-
% XXX using orig_arity here does not work when the predicate is one
% produced by higher-order specialization. See tests/valid/gh89.m.
PredArity = pred_info_orig_arity(PredInfo),
proc_info_get_headvars(ProcInfo, FullHeadVars),
proc_info_get_argmodes(ProcInfo, FullArgModes),

View File

@@ -2,6 +2,7 @@
-I@WORKSPACE@/library
-I@WORKSPACE@/browser
-I@WORKSPACE@/mdbcomp
-I@WORKSPACE@/ssdb
--c-include-directory @WORKSPACE@/boehm_gc
--c-include-directory @WORKSPACE@/boehm_gc/include
--c-include-directory @WORKSPACE@/runtime

View File

@@ -54,6 +54,9 @@ MCFLAGS-exists_cast_bug = --trace rep -O0 --optimize-saved-vars-const
MCFLAGS-explicit_quant = --halt-at-warn
MCFLAGS-foreign_underscore_var = --halt-at-warn
MCFLAGS-fzn_debug_abort = --trace rep
# XXX we should pass --ssdb-trace deep or --ss-debug to gh89
# but that currently doesn't work in non ssdb grades.
MCFLAGS-gh89 = --intermodule-optimization -O3 --no-warn-missing-opt-files
MCFLAGS-higher_order4 = -O3
MCFLAGS-higher_order_implied_mode = -O-1
MCFLAGS-ho_and_type_spec_bug = -O4

View File

@@ -135,6 +135,7 @@ OTHER_PROGS = \
func_default_modes \
func_in_head \
gh65 \
gh89 \
github_50 \
hawkins_switch_bug \
headvar_not_found \

59
tests/valid/gh89.m Normal file
View File

@@ -0,0 +1,59 @@
%---------------------------------------------------------------------------%
% vim: ts=4 sw=4 et ft=mercury
%---------------------------------------------------------------------------%
%
% This is a regression test for github issue #89.
%
% Any copyright is dedicated to the Public Domain.
% https://creativecommons.org/publicdomain/zero/1.0/
%
% Released by Transnat Games for testing purposes.
%
% Crashes with the following compiler flags:
% mmc --make crash -O 6 --intermodule-optimization --use-grade-subdirs \
% --grade=hlc.gc.spf.ssdebug
%
% Crashes as:
% Making Mercury/hlc.gc.spf.ssdebug/x86_64-unknown-openbsd6.5/Mercury/cs/crash.c
% Uncaught Mercury exception:
% Software Error: list.m: predicate `list.det_drop'/3: Unexpected: index out of range
% ** Error making `Mercury/hlc.gc.spf.ssdebug/x86_64-unknown-openbsd6.5/Mercury/cs/crash.c'.
%
% The crash requires an optimization level (I'm not sure which, 1 is fine but 6
% is not), intermodule optimizations, grade subdirs, and an ssdebug grade.
%
%---------------------------------------------------------------------------%
%
% The cause of the abort is the the source-to-source debugging transformation
% does not understand how to process the specializations produced by the
% higher-order optimization. Until it does, we have disabled the use of that
% optimization (and a bunch of others) by default with source-to-source
% debugging.
%
%---------------------------------------------------------------------------%
:- module gh89.
:- interface.
:- use_module array.
:- pred measure(array.array(int)::in, int::out, int::out) is det.
%---------------------------------------------------------------------------%
%---------------------------------------------------------------------------%
:- implementation.
%------------------------------------------------------------------------------%
:- pred measure(_, int, int, int, int).
:- mode measure(in, in, out, in, out) is det.
measure(_, W, W, H, H).
%------------------------------------------------------------------------------%
measure(Glyphs, W, H) :-
% There has to be something that actually uses the first argument to this
% predicate, otherwise the crash doesn't happen.
array.foldl2(measure, Glyphs, 0, W, 0, H).