This 5 character fix comes with a solver test as well as a functional
installer test essentially verifying the same thing. The solver test is
more useful when working on the solver. But the functional test is less
likely to be accidentally modified incorrectly during refactoring, as
every single package, version and link in the rather complex test
scenario is essential, and a modified version of the test may very well
still result in a successful installation but no longer verify the bug
described below.
Background:
In commit 451bab1c2c from May 19, 2012 I
refactored literals from complex objects into pure integers to reduce
memory consumption. The absolute value of an integer literal is the id
of the package it refers to in the package pool. The sign indicates
whether the package should be installed (positive) or removed (negative),
So a major part of the refactoring was swapping this call:
$literal->getPackageId()
For this:
abs($literal)
Unintentionally in line 554/523 I incorrectly applied this change to the
line:
$this->literalFromId(-$literal->getPackageId());
It was converted to:
-abs($literal);
The function literalFromId used to create a new literal object. By using
the abs() function this change essentially forces the resulting literal
to be negative, while the minus sign previously inverted the literal, so
positive into negative and vice versa.
This particular line is in a function meant to analyze a conflicting
decision during dependency resolution and to draw a conclusion from it,
then revert the state of the solver to an earlier position, and attempt
to solve the rest of the rules again with this new "learned" conclusion.
Because of this bug these conclusions could only ever occur in the
negative, e.g. "don't install package X". This is by far the most likely
scenario when the solver reaches this particular line, but there are
exceptions.
If you experienced a solver problem description that contained a
statement like "Conclusion: don't install vendor/package 1.2.3" which
directly contradicted other statements listed as part of the problem,
this could likely have been the cause.
The previous rule was not previously updated to point to the next rule when
removing a middle rule from the watch tree for a literal. This resulted in
jumping from one literal's watch tree to another's, which could then jump
back to the original and cause infinite loop in a case like #265.
Fixes#265
Nils Adermann
Gabriel Caruso
Jordi Boggiano
Niels Keurentjes
Jefferson Carpenter
Rob Bast
Possum
Robert Gruendler
Alexey Prilipko
Grégoire Pineau
Martin Hasoň
Volker Dusch