mirror of
https://github.com/Mercury-Language/mercury.git
synced 2026-04-16 09:53:36 +00:00
c18cd1ff51
This fixes Mantis bug #512, whose symptom is a compiler abort in some very specific circumstances. The regression test for the bug below documents the long causal chain needed to tickle this abort. compiler/simplify_goal_conj.m: When deleting unreachable code, set the requanify flag, which also causes instmap_deltas to be recomputed. The immediate cause of the bug was an instmap_delta that should have been thus recomputed, but wasn't. Do this in both places where simplification can delete dead code, though the Mantis 512 test case involves only one. tests/valid/bug512.m: A regression test for this bug. tests/valid/Mercury.options: tests/valid/Mmakefile: Enable the new test case, and run it with the flags that tickle the bug it is guarding against.
92 lines
3.5 KiB
Mathematica
92 lines
3.5 KiB
Mathematica
%---------------------------------------------------------------------------%
|
|
% vim: ft=mercury ts=4 sw=4 et
|
|
%---------------------------------------------------------------------------%
|
|
%
|
|
% This is a regression test for Mantus bug #512. The command
|
|
%
|
|
% mmc -C -s asm_fast.gc -O5 --intermod-opt bug512.m
|
|
%
|
|
% used to generate this output:
|
|
%
|
|
% Uncaught Mercury exception:
|
|
% Software Error: predicate `ll_backend.liveness.require_equal'/4:
|
|
% Unexpected: branches of if-then-else disagree on liveness
|
|
% First:
|
|
% Rest: Cs_8
|
|
%
|
|
% The causal chain leading to the abort was quite complex. Here are
|
|
% its main steps.
|
|
%
|
|
% - The call char.to_upper is inlined. Its body is an if-then-else,
|
|
% with the else case copying the input arg to the output arg.
|
|
% In this case, that is C = C0.
|
|
%
|
|
% - Deforestation pushes the test C \= C0 into both arms of the if-then-else.
|
|
% This leaves the else case consisting of C = C0, not ( C = C0 ).
|
|
%
|
|
% - The follow-code optimization pushes Cs = [C, ':'] into both arms
|
|
% of the if-then-else. This puts Cs (variable number 8) into the nonlocals
|
|
% set of both arms.
|
|
%
|
|
% - The final pre-codegen simplify, the first invocation of simplification
|
|
% after deforestation, notices that after C = C0, the goal "C = C0" inside
|
|
% the negation can be replaced by "true", that therefore "not C = C0"
|
|
% can be replaced by "fail", and thus the whole else arm of the if-then-else
|
|
% can be replaced by "fail", whose instmap_delta is unreachable.
|
|
%
|
|
% So far, so good. But then ...
|
|
%
|
|
% - Due to the change in determinism, simplify_goal_plain_conj wraps
|
|
% a commit(dont_force_pruning) scope around the "fail" goal.
|
|
%
|
|
% - The replacement process does *not* set the flag that requests
|
|
% a requantification of the procedure body, followed by the recomputation
|
|
% of its instmap deltas. In the absence of the flag, simplification does
|
|
% neither.
|
|
%
|
|
% - Without the recomputation of instmap_deltas, the instmap_delta of the
|
|
% commit scope has the instmap_delta of the original else arm goal,
|
|
% which says that the else arm computes Cs_8, and its end is reachable.
|
|
%
|
|
% - The code of liveness.m starts by invoking quantification, which removes
|
|
% Cs_8 from both the nonlocals set and the instmap_delta of the else arm,
|
|
% but, crucially, leaves the end of the else arm reachable. If the ends
|
|
% of both arms are reachable, then the two arms should define the exact
|
|
% same set of variables, but in this case, they don't. This is therefore
|
|
% effectively a compiler-induced mode error. Compilation is not supposed
|
|
% to go beyond semantic analysis if the program contains any mode errors,
|
|
% so later passes rely on their absence.
|
|
%
|
|
% - The mode error caused by the incorrect reachability screws up the
|
|
% calculations of the second (deadness) pass of liveness.m.
|
|
%
|
|
% - That screwup then causes the fourth (resume_points) pass in livenes.m
|
|
% to throw the above exception.
|
|
%
|
|
% Note that the data in the abort message is backwards. Even though the
|
|
% then-arm computes Cs_8 and the else-arm does not, the incorrect post-death
|
|
% and post-birth sets that the deadness pass attaches to the then- and
|
|
% else-arms respectively makes the resume_point pass believe the exact
|
|
% opposite.
|
|
%
|
|
|
|
:- module bug512.
|
|
:- interface.
|
|
|
|
:- import_module string.
|
|
|
|
:- pred normalise_drive_colon(string::in, string::out) is semidet.
|
|
|
|
:- implementation.
|
|
|
|
:- import_module char.
|
|
:- import_module list.
|
|
|
|
normalise_drive_colon(S0, S) :-
|
|
string.to_char_list(S0, Cs0),
|
|
Cs0 = [C0, ':'],
|
|
char.to_upper(C0, C),
|
|
C \= C0,
|
|
Cs = [C, ':'],
|
|
string.from_char_list(Cs, S).
|