mirror of
https://github.com/Mercury-Language/mercury.git
synced 2026-04-15 17:33:38 +00:00
c8aabea968
compiler/call_gen.m:
When a builtin generates an output that the rest of the computation
ignores, don't update the state of the output variable(s), because
the presence of a state for a variable that shouldn't exist
can cause a compiler abort later.
compiler/code_gen.m:
Pass the needed info to call_gen.m.
tests/valid/dead_get_io_state.m:
A new test case for this bug, whose module comment explains
the chain of events leading to the compiler abort.
tests/valid/Mmakefile:
Enable the new test case.
73 lines
2.8 KiB
Mathematica
73 lines
2.8 KiB
Mathematica
%---------------------------------------------------------------------------%
|
|
% vim: ts=4 sw=4 et ft=mercury
|
|
%---------------------------------------------------------------------------%
|
|
%
|
|
% Before the fix committed on 2023 oct 6, this code resulted in
|
|
% this compiler abort in LLDS grades:
|
|
%
|
|
% Uncaught Mercury exception:
|
|
% Software Error: predicate
|
|
% `ll_backend.var_locn.clobber_lval_in_var_state_map'/6:
|
|
% Unexpected: empty state
|
|
%
|
|
% The chain of events that lead to this abort was the following.
|
|
%
|
|
% - The trace goal transformation in goal_expr_to_goal.m adds
|
|
%
|
|
% - a call to unsafe_get_io_state just before the call to string.format and
|
|
% - a call to unsafe_set_io_state just after the call to unexpected.
|
|
%
|
|
% Since the goal inside the trace goal's scope does not refer to !IO,
|
|
% the variable returned by unsafe_get_io_state as the initial I/O state
|
|
% is also the variable passed back to unsafe_set_io_state.
|
|
%
|
|
% At this point, this variable is a nonlocal variable in both of these calls.
|
|
%
|
|
% - Mode checking, noticing that the call to unexpected cannot succeed,
|
|
% deletes the call to unsafe_set_io_state. This makes the variable
|
|
% returned by unsafe_get_io_state *local* to that call.
|
|
%
|
|
% - Simplification declines to delete the call to unsafe_get_io_state,
|
|
% because goal_can_loop_or_throw_imaf returns can_loop_or_throw for this
|
|
% call, even though this builtin cannot do either of those things.
|
|
%
|
|
% - The LLDS code generator recognizes that unsafe_get_io_state is a builtin,
|
|
% and invokes generate_builtin to generate code for. generate_builtin
|
|
% applied the builtin's translation by magically binding the variable
|
|
% representing the I/O state WITHOUT CHECKING WHETHER THIS VARIABLE IS LIVE.
|
|
%
|
|
% - The call to the string.format (or rather, the calls that format_call.m
|
|
% expands the call to string.format into) require flushing live variables
|
|
% to the stack, since the call may clobber all the registers. The variable
|
|
% bound by unsafe_get_io_state had no stack slot (values of dummy types never
|
|
% do) and was not known to be a constant, so having the call clobbering
|
|
% all the registers would have destroyed its state. This is why the code
|
|
% generator raised the exception shown above.
|
|
%
|
|
% The fix was to make generate_builtin ignore bindings if they are made to
|
|
% variables that are ignored outside the builtin.
|
|
%
|
|
|
|
:- module dead_get_io_state.
|
|
:- interface.
|
|
|
|
:- pred test(int::in, int::in, int::out) is det.
|
|
|
|
:- implementation.
|
|
|
|
:- import_module int.
|
|
:- import_module list.
|
|
:- import_module require.
|
|
:- import_module string.
|
|
|
|
test(A, B, Z) :-
|
|
( if A < B then
|
|
trace [io(!IO)] (
|
|
string.format("A = %d, B = %d\n", [i(A), i(B)], Msg),
|
|
unexpected($pred, Msg)
|
|
),
|
|
Z = A
|
|
else
|
|
Z = B
|
|
).
|